Evaluation of Multiplayer Games in Robot-Assisted Neurorehabilitation

DISS. ETH NO. 25607

EVALUATION OF MULTIPLAYER GAMES IN ROBOT-ASSISTED NEUROREHABILITATION

A thesis submitted to attain the degree of DOCTOR OF SCIENCES OF ETH ZURICH (Dr. sc. ETH Zurich)

presented by

KILIAN LEVIN BAUR MSc ETH ME, ETH Zurich born on 05.06.1984 citizen of Sarmenstorf (AG)

accepted on the recommendation of Robert Riener Verena Klamroth-Marganska Domen Novak

2018

“Whatever you can do or dream you can, begin it.” - Johann Wolfgang von Goethe

Abstract

Stroke is one of the main causes of chronic disability. For the reintegration of patients post-stroke into society, rehabilitation of motor functions is required. Intensity of rehabilitation training is crucial to regain the body functions of patients. Intensive training is only possible when the patient’s motivation to participate can be kept at a high level. To maintain a high level of motivation, patient-tailored training or multiplayer games can be stabilizing factors. However, the combination of patient-tailored training and multiplayer games introduces new challenges to the field of robot-assisted training. The goal of this work was to evaluate patient-tailored multiplayer games in device- and robot-assisted therapy. Therefore, elements of multiplayer games have been structured conceptually and the impli- cations of each element has been discussed using widely established models in the field: a model of game experience, i.e., the model of flow, and a model of motor performance, i.e., the challenge point framework. The term game experiences comprises motivational aspects such as perceived enjoyment or perceived effort of the subjects. Motor performance or game performance can be increased by more physical and mental effort by the subject. Studies were designed and executed with patient-tailored single and multiplayer applications to verify structure and models. The first research question was whether there is a general consent that multiplayer games in health-related disciplines increase the motivation and the effort of the players compared to single player games. I systematically searched for multiplayer games in health-related disciplines and found thirteen articles that met the inclusion criteria. ii Abstract

The studies reported in these articles were all using competitive, collaborative, and co-active multiplayer modes. However, these modes do not represent the full diversity of multiplayer modes. When played in multiplayer mode, I ascertained that nine studies showed positive effects on game experience and six studies showed improved game performance. None of the thirteen studies showed positive effects on the game experience and only two studies showed increased game performance when played in single player mode. I concluded that multiplayer modes should be considered for the design of future therapy settings. As indicated by some of the selected studies, I also concluded that difficulty adaptation for multiplayer games will play a role in future therapy designs. These conclusions are supported by the flow theory and the challenge point framework. The second research question was whether multiplayer games in stroke therapy increase motivation and effort of patients in the (sub-)acute phase post-stroke. I carried out a test using an air hockey computer game in single-player and two-player modes played by forty patients in the (sub-)acute phase post-stroke. I identified that two- player modes were preferred by sixteen out of forty patients. How- ever, I did not see any difference between single-player and two-player modes regarding motivation and effort. Furthermore, I found a significant difference between competitive and cooperative forms of games. Competitive forms of the game were more motivating and increased the exercise intensity. The third research question was whether robotic therapy devices can account for differences in skill level of patients in mulitplayer games by using the haptic modality. For competitive multiplayer games, I developed a haptic difficulty adaptation strategy to balance differences in skill level. Tests with simulated and healthy participants showed that the effect of difficulty adaptation seems more important than the provision of multiplayer games alone. The fourth research question was whether patients and their spouses or partners enjoy playing together in a therapy environment while using individual devices. Targeting a balanced patient-spouse game play, I extended the difficulty adaption strategy in competitive multiplayer games for applications with non-haptic devices. I further developed a virtual coupling of a non-haptic device with a haptic Abstract iii device. I successfully tested both the difficulty adaptation and the virtual coupling of haptic and non-haptic devices for feasibility in a study with patients and their partners or partners. The results of these single cases study show a tendency that patients are more motivated in such a multiplayer setting compared to playing the same games in single-player mode. The study further contributes to an investigation of whether the patient’s at-home-behavior changes when they train together with their spouses or partners compared to single player training sessions. Overall, reviewing the current literature in health-related multiplayer games and conducting my own multiplayer studies, I can confirm the potential of multiplayer games for robot assisted therapy. This potential can only be fully exploited when the multiplayer games also account for differences in skill level of the included players. These differences in skill level can be overcome by patient-tailoring the therapy conditions. Established models, such as the flow theory and the challenge point framework support the comparison of different patient- tailoring strategies regarding patient motivation and performance. This work contributes to these models by extending the existing models and facilitating the application of these models for studies using patient-tailored robot-assisted multiplayer games. We consider that this work complements developments in the field of robot-assisted training, in the sense that motivating game play is enabled between differently skilled players where every player can play with a free choice of device. iv Abstract v

Zusammenfassung

Schlaganfall ist eine der Hauptursachen fur¨ chronische Invalidität. Fur¨ die Reintegration von Schlaganfall Patienten in die Gesellschaft ist die Rehabilitation motorischer Funktionen erforderlich. Die Inten- sität des Rehabilitationstrainings ist entscheidend fur¨ die Wiederher- stellung der motorischer Funktionen der Patienten. Intensives Trai- ning ist nur möglich, wenn die Motivation der Patienten zur Teilnah- me auf hohem Niveau gehalten werden kann. Um ein hohes Maß an Motivation zu gewährleisten, können patientenspezifisches Training oder Mehrspieler-Modi unterstutzende¨ Faktoren sein. Die Kombina- tion aus patientenspezifischem Training und Mehrspieler-Modi bringt jedoch neue Herausforderungen auf dem Gebiet des robotergestutzten¨ Trainings mit sich. Ziel dieser Arbeit war es, patientenspezifische Mehrspieler-Modi in der geräte- und robotergestutzten¨ Therapie zu identifizieren und zu evaluieren. Dafur¨ wurden Elemente von Mehrspieler-Modi konzeptio- nell strukturiert und der Einfluss der einzelnen Elemente wurde in etablierten Modellen diskutiert: einem Modell zum Trainingserleb- nis, dem sogenannten Flow-Modell, und ein Modell zur Trainingslei- stung, dem sogenannten Challenge-Point-Framework. Der Begriff Er- lebnis beinhaltet motivationale Aspekte wie das wahrgenommenene Vergnugen¨ oder die wahrgenommene Anstrengung des Patienten. Die Trainingsleistung kann durch Einsatz vom Patienten erhöht werden. Ich habe Studien mit patientenspezfischen Einzel- und Mehrspieler- Modi geplant und durchgefuhrt,¨ um Struktur und Modelle zu verifi- zieren. Die erste Forschungsfrage war, ob es einen allgemeinen Konsens gibt, vi Zusammenfassung dass Mehrspieler-Modi in gesundheitsbezogenen Disziplinen die Moti- vation und den Einsatz der Spieler im Vergleich zu Einzelspieler-Modi erhöhen. Ich habe systematisch nach Mehrspieler-Spielen in gesundheitsbezogenen Disziplinen gesucht und 13 Artikel gefunden, welche die Einschlusskriterien erfullten.¨ Die in diesen Artikeln beschriebenen Studien verwendeten alle kompetitive, kollaborative und kooperative Mehrspieler-Modi. Diese Modi repräsentieren jedoch nicht die volle Vielfalt an zur Verfugung¨ stehenden Mehrspieler-Modi. Neun Stu- dien zeigten positive Auswirkungen auf das Trainingserlebnis und sechs Studien zeigten eine verbesserte Trainingsleistung, wenn sie im Multiplayer-Modus gespielt wurden. Keine der dreizehn Studien zeigte positive Auswirkungen im Einzelspielermodus auf das Spie- lerlebnis und nur zwei Studien zeigten eine erhöhte Spieleleistung im Einzelspielermodus. Ich kam zu dem Schluss, dass Mehrspieler- Modi fur¨ die Gestaltung zukunftiger¨ Therapieeinheiten in Betracht gezogen werden sollten. Angeregt durch einige der eingeschlossenen Studien, konstatiere ich auch, dass Anpassung der individuellen Spiel- Schwierigkeit in Mehrspieler-Spielen eine Rolle in zukunftigen¨ The- rapieentwurfen¨ spielen wird. Diese Schlussfolgerungen werden vom Flow-Modell und dem Challenge Point Framework unterstutzt.¨ Die zweite Forschungsfrage war, ob Mehrspieler-Modi in der Thera- pie nach Schlaganfall das Trainingserlebnis und die Trainingsinten- sität durch gesteigerten Einsatz von (sub-)akuten Schlaganfallpati- enten erhöhen. Hierfuhr¨ fuhrte¨ ich eine Studie mit einem Airhockey- Computerspiel durch, dass von vierzig (sub-)akuten Schlaganfallpa- tienten im Einzelspieler- und Zweispielermodus gespielt wurde. Ich habe festgestellt, dass Zweispieler-Modi von 16 von 40 Patienten be- vorzugt wurden, obwohl ich keinen Unterschied zwischen dem Einzel- und Zwei-Spieler-Modus in Bezug auf Trainingserlebins und Einsatz gmessen habe. Außerdem habe ich einen signifikanten Unterschied zwischen kompetitiven und kooperativen Modi von Spielen festgestellt: Kompetitive Modi des Spiels waren motivierender und erhöhten die Trainingsintensität. Die dritte Forschungsfrage war, ob robotische Therapiegeräte Un- terschiede in den Fähigkeiten von einzelnen Spielern in Mehrspieler- Spielen ausgleichen können, indem sie die haptische Modalität ver- wenden. Fur¨ kompetitive Mehrspieler-Spiele habe ich eine Methode Zusammenfassung vii zur haptischen Schwierigkeitsanpassung entwickelt, um Unterschiede in den Fähigkeiten der einzelnen Spieler auszugleichen. Tests mit si- mulierten und gesunden Teilnehmern zeigten, dass der Ausgleich der Fähigkeiten wichtiger ist als die Bereitstellung von Mehrspieler-Modi als solches. Die vierte Forschungsfrage war, ob Patienten in einer Therapieumge- bung das gemeinsame Spielen mit ihren Partnern im Mehrspieler- Modi dem Spielen im Einzelspieler-Modi vorziehen, wenn beide – Patient und Partner – das jeweils passende Eingabegerät verwen- den. Mit dem Ziel, ein ausgewogenes Spiel zwischen dem Patienten und dem Partner zu erreichen, habe ich das Anpassen der Schwie- rigkeit in kompetitiven Mehrspieler-Modi fur¨ nicht-haptische Geräte ergänzt. Ich entwickelte außerdem eine virtuelle Kopplung eines nicht- haptischen Geräts mit einem haptischen Gerät, Ich habe sowohl die Anpassung als auch die virtuelle Kopplung von haptischen und nicht haptischen Geräten in einer Studie mit Patienten und deren Partnern erfolgreich hinsichtlich Machbarkeit getestet. Die Ergebnisse dieser Multiplayer-Studie zeigen eine Tendenz, dass Patienten in einem sol- chen Multiplayer-Modus motivierter sind, als wenn sie die gleichen Spiele im Einzelspieler-Modus spielen. Die Studie trägt daruber¨ hin- aus zu einer Untersuchung bei, die evaluiert, ob sich das Verhalten des Patienten zu Hause ändert, wenn der Partner in Mehrspieler-Modi an der Therapie teilnimmt. Insgesamt kann ich nach der Uberpr¨ ufung¨ aktueller Literatur hinsichtlich gesundheitsbezogener Mehrspieler-Spielen und der Realisie- rung meiner eigenen Studien, das Potential von Mehrspieler-Spielen fur¨ die robotergestutzte¨ Therapie nach Schlaganfall bestätigen. Dieses Potenzial wird nur dann ausgeschöpft, wenn die Mehrspieler-Spiele auch die unterschiedlichen Fähigkeiten der Spieler ausgleichen. Die- se Unterschiede in den Fähigkeiten der Spieler können durch eine Anpassung der Bedingungen fur¨ den Patienten und den Mitspie- ler uberwunden¨ werden. Ich habe zur Strukturierung der Elemente von patientenspezifischen robotergestutzten¨ Mehrspieler-Spielen bei- getragen, und dadurch die Diskussion verschiedener Strategien fur¨ die Schwierigkeitsanpasung erleichtert. Etablierte Modelle wie das Flow-Modell und das Challenge-Point-Framework unterstutzen¨ den Vergleich verschiedener Arbeiten an diesen Elementen hinsichtlich viii Zusammenfassung

Trainingserlebnis und -leistung. Die vorliegende Arbeit erweitert die bestehenden Modelle und erleichtert die Anwendung dieser Modelle fur¨ Studien mit patientenspezifischen robotergestutzten¨ Mehrspieler- Spielen. Mit dieser Arbeit habe ich das Spielen zwischen Spielern mit unterschiedlichen Fähigkeiten ermöglicht, wobei jeder Spieler frei das Gerät wählen kann, mit dem er spielt. ix

Acknowledgements

First of all, I would like to thank my wife Lea Br¨andle, my family, and my friends for their continuous support and patience. I want to express special thanks to my co-supervisor Verena Klamroth- Marganska and my colleagues Peter Wolf and Jaime Duarte. Their unwavering support, feedback, and creative thinking were fundamental and inspirational for my research. I want to thank my co-examiner Domen Novak for oﬀering his expertise and time to discuss my research and publications throughout my time as a PhD student. Furthermore, I want to express my gratitude to my supervisor Robert Riener. I want to thank him for creating and maintaining the possibility to work in such a wonderful team, on exciting research topics, and with excellent equipment. I should like to thank him for all of his advice and counsel as well as his patience towards the time I needed to understand, to value, and to implement them. I want to thank him for his trust and the possibility to grow on challenging responsi- bilities in the lab, in teaching, and at the Cybathlon. His relaxed and forgiving way of guidance by giving us not only a lot of freedom but responsibility. This was fundamental for me to commit to my work and that of my colleagues. x Acknowledgements xi

Preface

Experimental work, evaluation, and writing of this thesis was performed at the Sensory-Motor Systems (SMS) Lab, Institute of Ro- botics and Intelligent Systems (IRIS), Department of Health Science and Technology (D-HEST), ETH Zurich, Switzerland. The following chapters of this cumulative dissertation are based on publication manuscripts:

Chapter 3, based on the journal publication [16]: Kilian Baur, Alexandra Sch¨attin,Eling D. de Bruin, Robert Riener, Peter Wolf. Trends in robot-assisted and virtual reality- assisted neuromuscular therapy: A systematic review of health- related multiplayer games. Journal of NeuroEngineering and Rehabilitation, 2018 The open access articles published in BMC’s journals are made available under the Creative Commons Attribution (CC-BY) li- cense, which means they are accessible online without any re- strictions and can be re-used in any way, subject only to proper attribution (which, in an academic context, usually means citation).

Chapter 4, based on the manuscript in preparation [18]: Kilian Baur, Peter Wolf, Domen Novak, Dana Boering, Sabrina H¨orner,Corinna Dahlen, Joachim Berger, Robert Riener, Volker H¨omberg. Competitive versus cooperative forms of gaming in motor training with (sub-)acute stroke patients. 2019 xii Preface

Chapter 5, based on the conference publication [19]: Kilian Baur, Peter Wolf, Robert Riener, Jaime E. Duarte. Mak- ing neurorehabilitation fun: Multiplayer training via damping forces balancing diﬀerences in skill levels. Proceedings of the IEEE International Conference on Rehabilitation Robotics, 2017 ©2016 IEEE. Reprinted, with permission, from the original article. Chapter 6, based on the conference publication [17]: Kilian Baur, Peter Wolf, Verena Klamroth-Marganska, Walter Bierbauer, Urte Scholz, Robert Riener, Jaime E. Duarte. Robot- supported multiplayer rehabilitation: Feasibility study of haptically linked patient-spouse training. IEEE International Con- ference on Intelligent Robots and Systems, 2018 ©2016 IEEE. Reprinted, with permission, from the original article.

Appendix A, based on the conference publication [14]: Kilian Baur, Verena Klamroth-Marganska, Chiara Giorgetti, Da- niela Fichmann, and Robert Riener. Performance-based viscous force ﬁeld adaptation in upper limb strength training for stroke patients. Proceedings of the IEEE RAS and EMBS International Conference on Biomedical Robotics and Biomechatronics, 2016 ©2016 IEEE. Reprinted, with permission, from the original article.

Appendix B, based on the journal publication [19]: Kilian Baur, Florina Speth, Aniket Nagle, Robert Riener, Verena Klamroth-Marganska. Music meets robotics: A prospective randomized study on motivation during robot aided therapy. Jour- nal of NeuroEngineering and Rehabilitation, 2018 The open access articles published in BMC’s journals are made available under the Creative Commons Attribution (CC-BY) li- cense, which means they are accessible online without any re- strictions and can be re-used in any way, subject only to proper attribution (which, in an academic context, usually means citation). Appendix C, based on the submitted journal manuscript [15]: Kilian Baur, Nina Rohrbach, Joachim Hermsd¨orfer,Robert Riener, Preface xiii

Verena Klamroth-Marganska. The ”Beam-Me-In Strategy” - Re- mote Haptic Therapist-Patient Interaction with Two Exoskele- tons for Stroke Therapy. Journal of NeuroEngineering and Re- habilitation, 2018 The open access articles published in BMC’s journals are made available under the Creative Commons Attribution (CC-BY) li- cense, which means they are accessible online without any re- strictions and can be re-used in any way, subject only to proper attribution (which, in an academic context, usually means citation).

Each of these chapters is introduced with a short unnumbered section Foreword and Overview. This section bridges the previous with the current chapter and outlines the elements of patient-tailored robot- assisted multiplayer gaming. The manuscripts have been edited or shortened to reduce redundancy. xiv Preface xv

Contents

Abstract i

Zusammenfassung v

Acknowledgements ix

Preface xi

1 General Introduction 1 1.1 Intensive Rehabilitation Therapy after a Stroke . . . .1 1.2 Therapy Robots ...... 2 1.3 Virtual Environments ...... 3 1.4 Maintaining Patients’ Motivation ...... 4 1.5 The research gap: connecting people and devices . . .7 1.6 Outline of this work ...... 7

2 The Elements of Multiplayer Gaming in Robot-Assisted Training 9 2.1 Rehabilitation robots used in this thesis ...... 9 2.1.1 ARMin ...... 9 2.1.2 Armeo Spring ...... 10 2.2 State-of-the-art in robot-assisted rehabilitation games 11 2.2.1 Previous work with rehabilitation robots . . . . 11 2.3 Multiplayer in robot-assisted rehabilitation ...... 15 xvi Contents

3 Systematic Review of Health-related Multiplayer Games 19 3.1 Abstract ...... 21 3.2 Introduction ...... 22 3.3 Methods ...... 24 3.3.1 Protocol ...... 24 3.3.2 Eligibility criteria ...... 24 3.3.3 Study selection ...... 25 3.3.4 Data extraction process and data synthesis . . 26 3.3.5 Assessment of study quality ...... 28 3.4 Results ...... 31 3.4.1 Study selection ...... 31 3.4.2 Study characteristics ...... 31 3.4.3 Devices, games and multiplayer modes . . . . . 31 3.4.4 Assessment ...... 32 3.4.5 Influence of multiplayer games ...... 33 3.4.6 Quality evaluation ...... 34 3.5 Discussion ...... 35 3.5.1 Study selection ...... 35 3.5.2 Devices and games inconsistently discussed . . 35 3.5.3 Limited diversity of multiplayer modes . . . . . 36 3.5.4 The potential of playing with friends and relatives as co-players ...... 38 3.5.5 Balancing multiplayer interventions in the context of the flow model and the challenge point framework ...... 39 3.5.6 Conditional task difficulty: complementing the terminology ...... 40 3.5.7 Changing the individual conditions to optimize game experience ...... 40 3.5.8 Different assessments hinder inter-study comparison ...... 42 3.6 Conclusions ...... 44

4 Multiplayer Gaming in Robot-Assisted Training 47 4.1 Abstract ...... 47 4.2 Introduction ...... 49 4.3 Methods ...... 51 4.3.1 Participants ...... 51 Contents xvii

4.3.2 Ethics ...... 52 4.3.3 Apparatus ...... 52 4.3.4 Game ...... 52 4.3.5 Procedure ...... 55 4.3.6 Analysis ...... 56 4.4 Results ...... 58 4.4.1 Participants ...... 58 4.4.2 Game results ...... 59 4.4.3 Self-rated intrinsic motivation ...... 59 4.4.4 Therapist-rated intrinsic motivation ...... 59 4.4.5 Exercise intensity (actual eﬀort) ...... 59 4.4.6 Mode ranking and prediction ...... 59 4.5 Discussion ...... 62 4.5.1 (H1) Number of players (single- or two-player) 62 4.5.2 (H2) Interaction type (competitive or cooperative) ...... 63 4.5.3 (H3) Correlations and predictability ...... 64 4.6 Conclusion ...... 65

5 Balanced Multiplayer Gaming in Robot-Assisted Train- ing 69 5.1 Abstract ...... 69 5.2 Introduction ...... 71 5.3 Methods ...... 73 5.3.1 Framework ...... 73 5.3.2 Experimental setup ...... 73 5.3.3 Artiﬁcial intelligence of the computer-player . . 77 5.3.4 Study designs ...... 78 5.3.5 Data analysis ...... 81 5.4 Results ...... 81 5.4.1 Computer simulation ...... 81 5.4.2 Feasibility study ...... 82 5.4.3 Eﬀects on motivation ...... 82 5.5 Discussion ...... 83 5.6 Conclusion ...... 87 xviii Contents

6 Haptic Interaction in Multiplayer Games with Im- paired and Healthy Subjects 89 6.1 Abstract ...... 89 6.2 Introduction ...... 91 6.3 Methods ...... 92 6.3.1 Experimental setup ...... 92 6.3.2 Virtual reality environments ...... 92 6.3.3 Study protocol ...... 99 6.4 Results ...... 101 6.5 Discussion ...... 102 6.6 Conclusion ...... 106

7 General Conclusion 107 7.1 Major Contributions ...... 108 7.2 Outlook ...... 109

A Performance-based viscous force ﬁeld adaptation 111 A.1 Background ...... 111 A.2 Methods ...... 112 A.3 Results and Discussion ...... 117 A.4 Conclusion and Current State ...... 118

B Rehabilitation Game Without A Visual Display 119 B.1 Background ...... 119 B.2 Methods ...... 120 B.3 Results and Discussion ...... 124 B.4 Conclusion and Current State ...... 126

C Remote Haptic Therapist-Patient Interaction 127 C.1 Background ...... 127 C.2 Methods ...... 128 C.3 Results and Discussion ...... 132 C.4 Conclusion and Current State ...... 135 1

Chapter 1

General Introduction

1.1 Intensive Rehabilitation Therapy after a Stroke

There is roughly a one in six chance of suﬀering a stroke in middle aged adults [146, 25]. From 2000 to 2010, the number of deaths as a result from strokes declined by 22.8%. In addition, the relative rate of stroke death fell by 35.8% and the actual number of stroke deaths declined by 22.8% [3] implying that there are more people to rehabilitate following a stroke. Many stroke survivors do not regain full arm function. Although conventional neuromuscular therapy, i.e., occupational and physical therapy, is provided where caregivers are available, conventional neurorehabilitation appears to have little impact on impairment over and above that of spontaneous biological recovery [77, 171]. Thus, in 30–66% of the stroke survivors, long-term loss of arm function appears [92, 69, 174, 153]. By the end of one year post-stroke 27% of patients are totally dependent (including personal care)[28]. Due to the patients’ limits in body functions and cognitive functions, only 20% of patients return to gainful employment after four years post stroke [73]. The reintegration and independence of stroke patients is important to decrease the social and economic bur- den to society. Therefore, the limited success of conventional therapy has to be overcome by eﬀective training post stroke. 2 Chapter 1. General Introduction

The efficacy of training depends on the training intensity [155, 93, 182, 154]. Intensive training can be described in three dimensions: frequency of repetitions of desired movements [120, 180, 81, 2], amount of time that is dedicated to practice [93], and active participation of the patient [74, 102]. Frequency of repetitions and amount of time result in a total number of movements. The multiplication of both, number of movements and active participation results in a broader definition of intensity [91]. This definition incorporates both the number of movements and the training condition demanding for active participation. There exists an optimal condition at which the difficulty is just enough to provide information for motor learning but not more than a subject can handle. The relationship between learning and difficulty is discussed in the challenge point frame work [65]. If the tasks involves multiple players then the difficulty is dependent on the opponents skill level [44] (Figure 1.1). The more difficult a task is, the lower is the expectancy that an individual is fulfilling the task. The task performance or game performance is therefore an indirect measure of the difficulty of a task. However, the relationship between task performance and task difficulty is different for individual subjects.

1.2 Therapy Robots Intensify Neuromus- cular Rehabilitation

Rehabilitation robotics is the ﬁeld of research that is dedicated to intensifying neuromuscular rehabilitation. Rehabilitation robots usually haptically assist diﬀerent sensory-motor functions of neuromuscular patients in virtual tasks that can be motivated through audio- visual environments [27]. Most robotic systems for rehabilitation consist of a robotic device for haptic interaction with the patient and may be complemented by a monitor with loudspeakers for task presentation. Robotic systems for rehabilitation can increase training duration and active participation compared to conventional therapy [66]. Stud- ies showed that training with a therapy robot can enhance improvement of motor function in a chronically impaired paretic arm after 1.3. Virtual Environments 3

high high (doted)

Performance Lower skilled Higher skilled in practice (solid) opponent opponent learning Potential low low low Functional task di!culty high

Figure 1.1: Influence of an additional player (opponent) on task difficulty in a competitive task. In the challenge point framework, the skill level can affect the functional task difficulty thus moving the player closer to or farther from the optimal challenge point (peak of potential learning curve). The functional task difficulty is equal to the challenges provided by the task with respect to the skill level of the player and the conditions the game is played in. stroke more effectively than conventional therapy [85, 111, 90, 172]. However, the effectiveness is limited, particularly for patients with long-term arm deficits after stroke.

1.3 Virtual Environments Motivate in Robot-Assisted Training

Highly intensive robot-assisted training is most eﬀective when the motivation of patients to participate can be kept at a high level [105]. 4 Chapter 1. General Introduction

The necessity of appealing games in robot-assisted neuromuscular therapy to maintain the motivation of patients has already been identified [50]. Positive effects from virtual-reality-aided therapy in restoring arm motor impairments and motor related functional abilities was reported [165]. Therefore, number of movements and active participation contribute only to the amount of training. Neuromodulation in the brain is facilitated by the dopamine system which consists of several dopamine pathways. One of the dopamine pathways plays a major role in the motivational component of reward- motivated behavior. The other brain dopamine pathways are involved in motor control [20]. The level of dopamine in the brain increases when a reward can be anticipated [183]. Therefore, reward promotes neuromodulation and complements the amount of training to a intensive neuromuscular training. The necessity of appealing and challenging games for robot-assisted neuromuscular therapy to provide rewards for the patient has already been identified [50]. The challenges provided by the games should address the patient’s skill level to prevent boredom (too easy) and frustration (too difficult) and enable flow to be experienced by the patient. This is discussed in the flow model [38] (Figure 1.2). The experience of flow is associated with changes in neural activity and has the potential to contribute to neuromodulation through stroke therapy [168]. The number and difficulty of the challenges that are required to experience flow are different for individual patients and game settings. A different game experience profile exists for each patient and for each game or game mode. This highlights the dependency between game experience and challenges. Common tools for measuring game experience are questionnaires that measure intrinsic motivation [110].

1.4 Can Multiplayer Games Maintain Pa- tients’ Motivation?

Multiplayer games have emerged as a promising approach to provide social reward for patients in robot-assisted rehabilitation therapy. As known from entertainment games, the game experience is better in multiplayer mode compared to to games in single-player mode [55]. In 1.4. Maintaining Patients’ Motivation 5 high Higher skilled opponent

Player Frustration Flow

d Lower skilled

Challenge Boredomopponent

low Opponent low Skill high

Figure 1.2: Influence of an additional player on task experience in a competitive task. In the flow model, the skill level of the opposing player affects the perceived challenge (y-axis) of the task. This can move a player into or away from the flow zone of the task. The two players are on the opposite site of the flow zone. Both have the same distance (d) to the flow zone. health-related training and in a few studies of robot-assisted rehabilitation training, multiplayer games have been tested to ascertain their influence on game experience and game performance (Chapter 3). The combination of intensive patient-tailored robotic-assistance and multiplayer games that connect devices for differently skilled players appears to provide new possibilities in designing the therapy session of the future. To date, only single-player games have been used in clinics. Scientific evidence evaluating the benefits of these technologies, which could justify costs and effort, is still lacking [111]. Ther- apy robots may increase the training intensity, but intensity seems 6 Chapter 1. General Introduction to be multidimensional and its definition continues under discussion. Therapy robots have the potential to intensify therapy through multi- modal feedback, providing motivating games and patient-cooperative control strategies. Based on the robot’s multi-modal feedback, strategies are required that provide patient-tailored conditions addressing all dimensions of intensity. These strategies have to be designed, implemented and tested with patients.

Extending the applications to multiplayer games, the interaction between possible pairs of players changes the game experience and implicitly the training intensity regarding active participation [61] . Different skill levels of players may be involved in a multiplayer setting, ranging from healthy players to severely affected patients. Through the multiplayer game, two or more players are connected and indirectly affecting the task difficulty of each player. Difficulty adaptation for multiplayer games shall therefore be applied on an individual level. [5] Therapy robots provide an opportunity to use different modalities, including haptics, for difficulty adaptation. The haptic interaction with a robotic device remains hidden for co-players. Difficulty adaptation by changing the haptic condition for a player does not reveal differences in skill level to the other player. Further- more, game balancing by haptics does not require higher visual game complexity.

Therapy robots as input devices for multiplayer games complement the input devices for healthy players such as joysticks, keyboards or game pads. This broader set of input devices allows patients to play against their friends and relatives who are in good health. The relationship of the co-player to the patient might inﬂuence the behavior during the game play and might even inﬂuence the behavior outside of the clinic. The support of friends and relatives can facilitate behavior change in health related tasks such as using the impaired arm in daily life more often or stop smoking [145]. 1.5. The research gap: connecting people and devices 7

1.5 The research gap: connecting people and devices

Multiplayer games integrate an additional social dimension into the therapy. The incorporation of more than one player adds new possibilities to rehabilitation. The incorporated players add flexibility and the ingenuity of human thinking to the game environment, which is generally missed in single-player gaming against preprogrammed challenges or artificially controlled opponents. Furthermore, the multiplayer environment and related game mechanics can facilitate social interaction, ranging from conversation to haptic in-game interaction. Multiplayer games have not yet been designed for neuromuscular therapy. Games have to connect players with different skill levels - patient dyads or dyads of patient and healthy co-player - and need to be feasible to achieve the goals of rehabilitation. Furthermore, the effects of such games on patients’ game experience and game performance need to be investigated. In general, such applications need to be designed and tested targeting intensive and motivating therapy.

1.6 Outline of this work

For localization and investigation of challenges and opportunities in multiplayer games for robot-assisted arm therapy I framed conceptually the involved elements (Chapter 2). I discussed the challenges and opportunities considering the elements of this framework. In doing that, I integrated the framework in the foreword of each chapter: I reviewed the current literature in health-related multiplayer games (Chapter 3). Secondly, I designed and tested a multiplayer game played by stroke patients using an arm therapy device as game controller (Chapter 4). Finally, I designed and tested diﬃculty adaptation in multiplayer games using diﬀerently skilled players such as a stroke patient and the healthy spouse (Chapter 5). Moreover, I developed and tested further single-player and multiplayer applications that can be found in the Appendix. 8 Chapter 1. General Introduction 9

Chapter 2

The Elements of Multiplayer Gaming in Robot-Assisted Training

2.1 Rehabilitation robots used in this thesis

2.1.1 ARMin

The arm therapy robot ARMin was designed and evaluated by the groups of Riener and Dietz/Curt at ETH Zurich and University of Zurich [66, 117]. ARMin is used for the therapy of the affected arm and hand of neurological patients. The latest generation has seven degrees of freedom (DOF) allowing 3D shoulder rotation, elbow flexion/extension, pro-/supination and wrist flexion/extension (Figure 2.1). A hand actuation module supports and measures the opening and closing of the hand [82]. The ARMin exoskeleton is connected to the human arm with cuffs located at the upper arm, the lower arm and the fingers. The newest prototype of the ARMin robot (i.e., ARMin IV) is equipped with six-DOF force sensors placed on each cuff (Figure 2.2). These force sensors measure the interaction 10 Chapter 2. Conceptual Control Chart forces between the patient and the robot itself [97].

Figure 2.1: Seven DOF of the ARMin arm rehabilitation robot. Axis 1-3: 3D shoulder rotation. Axis 4: elbow ﬂexion/extension. Axis 5: pro-/supination. Axis 6: wrist ﬂexion/extension. Axis 7: Hand opening/closing.

2.1.2 Armeo Spring The Armeo Spring (https://www.hocoma.com/), a commercially available replica of the T-WREX [139], was used by subjects to play different modes of an air hockey computer game with the aﬀected upper limb. Armeo Spring is a 5 degree-of-freedom (3 in the shoulder, 1 in the elbow, 1 in the forearm) orthosis without robotic actua- tors, i.e., a passive system. The adjustable mechanical arm allows variable levels of gravity support by means of a spring mechanism. 2.2. State-of-the-art in robot-assisted rehabilitation games 11

Figure 2.2: Subject using the ARMin arm robot. The visual display is used to present task visualization.

The gravity support enables subjects, using residual upper limb function, to achieve a larger active range of motion (aROM) within a 3-dimensional workspace than possible without support.

2.2 State-of-the-art in robot-assisted rehabilitation games 2.2.1 Previous work with rehabilitation robots Previous work on ARMin and other rehabilitation devices mainly focused on single-player games consisting of one patient, one input device (interface) and one virtual environment (game) (Figure 2.3). The structure of these three subsystems is well-established in the design of games for entertainment [5]. The game features challenges, such as reaching tasks, that are displayed as output to a monitor attached to the patient. The patient interprets the challenges and 12 Chapter 2. Conceptual Control Chart reacts according to skill and motivation. In the case of a reaching task, the patient tries to move the arm towards the target position. The reactions are read by the interface, i.e. sensors on the robot, and cause an action in the game. The game is restricted to its own mechanics: providing challenges, and receiving and interpreting actions. The output in form of a display of the challenges, and the transla- tion of the input from the patient are considered in the interface, separately from the game mechanics.

Challenges Outputs Game Interface Patient Actions User Inputs

Figure 2.3: The relationships between the game, the interface, and the patient (adapted from Adams [5])

Depending on the interface used, the output may consist of a combination of three diﬀerent modalities: visual display, auditory display and haptic display. The state-of-the-art interface output is presented through a combination of visual display and auditory display usually implemented via a monitor and audio-loudspeakers. A haptic display can only be provided if the interface has been actuated. Rehabilita- tion robots allow the actuation. The patient’s input can be assessed through visual sensing, such as motion tracking systems, auditory sensing using microphones, or sets of position sensors and force sensors that are integrated into the rehabilitation robots. These integrated systems can measure kinematics of robotic structures and interaction forces between the robot and its environment. The patient is usually the key structure of the environment that the robot is interacting with. The therapist can be added as a second human in this game-interface- patient framework (Figure 2.4). The therapist selects the game to be played and implicitly the challenge to be faced by the patient. Game selection taking into account of the patients’ experience and their behavior. The therapist accompanies the patient throughout the therapy experience, and assesses and monitors the patient’s experiences and behavior. Although not always necessarily in a structured form, before selecting a next game and before starting a new session, 2.2. State-of-the-art in robot-assisted rehabilitation games 13 the therapist compares the target experience and behavior with the actual resultant experience and behavior to previous games and sessions. Know-how and clinical experience can support the therapist in the decision making process.

C Desired B Experience/ A Experience/ Behavior Challenges Outputs Behavior Therapist Game Actions InterfaceUser Inputs Patient

Figure 2.4: Simpliﬁed state-of-the-art concept: Single-player gaming in robot-assisted training

In a more detailed version of the framework, we see that the challenges provided by the game are transferred in two ways to the interface: the game complexity and the game reference position Gxref (Figure 2.5). The game complexity represents the option for the therapist to preset the difficulty of the task demanded by the game. The game complexity can influence the frequency of required actions or the task range, e.g., the speed and the range of the reference positions [179, 112, 36], influence the performance of virtual opponents [185], or adapt the tools within a game, such as sizes of avatars [62]. The game complexity can also influence parameters of the condition controller. The condition controller can adjust the haptic support or resistance provided by an actuated interface, such as a rehabilitation robot. The mapping between the actual position of the patient in a therapy setting, H xact to the measured position of the robotic structure Gxmeas in the game coordinate system usually remains unchanged. Therefore, the only way to change the level of complexity comes from the game itself. We complemented the framework with an assessment of the patient’s experience. This could be assessed during gameplay by using body- mounted sensors or after the gameplay using questionnaires. Body mounted sensors which measure electroencephalographic (EEG) or peripheral signals can assess the arousal dimension of the patient’s experience [34]. More commonly, questionnaires or inventories are used for post-game assessment of the patient’s experience. Recently, different versions of the intrinsic motivation inventory (IMI) have been 14 Chapter 2. Conceptual Control Chart used to describe the patient’s experience in the various subscales: interest/enjoyment, perceived competence, effort/importance, pressure/tension, perceived choice, value/usefulness, relatedness [110]. Further tools available are the Kids Game Experience Questionnaire [130], a questionnaire based on Malone’s theory of intrinsic motivations for learning (sensory immersion, control/choice, challenge/optimal difficulty, goal setting and feedback) [107], or self-designed questionnaires, e.g. asking for a preferred game or game mode.

C Desired B Experience/ A Behavior Experience/ Measured Therapist Game Interface Patient Behavior Assessment Experience

Game-/ Gxref Hxact Game mode-/ x F Complexity G meas H act Selection Game Complexity Complexity Condition Controller Game Performance

Figure 2.5: Detailed state-of-the-art concept: Single-player gaming in robot-assisted training

Examples to illustrate the state-of-the-art version of the game-interface- patient framework developed for the ARMin rehabilitation robot are gamified ADLs, fun games and visualized robot-assisted mobilization. [66]. Gamified ADLs require actions in home environments presented to the patients via a monitor. The robot provides assist-as- needed (AAN) support in multiple joints to enable movement initia- tion and successful target reaching in the 3D ADL environment even with patients with limited motor function. Fun games are usually limited to a one- or two-degrees-of-freedom movement of a patient. These fun games usually target specific movements at high repetition. Exoskeleton robots can also record and replay movements guided by a therapist and therefore, relieve the therapist in repetitively moving the patient’s arm for mobilization [117]. 2.3. Multiplayer in robot-assisted rehabilitation 15

Figure 2.6: Redesign of state-of-the-art rehabilitation games. Left: Kitchen environment, representing tasks from activities of daily living. – Center: ”Catch the thief”, a one-degree-of-freedom fun game where the patient (police man) has to follow a trajectory (thief). – Right: Visualized mobilization.

2.3 Multiplayer in patient tailored robot- assisted rehabilitation games

We can complement the framework to a multiplayer game by connecting a second player to the game (Figure 2.7). In addition, measurements from interface and game can be used for each individual player to optimize game performance and implicitly the game experience. These measurement can also be used to adjust game difficulty to enable a balanced multiplayer game. The data resulting from these measurements can be stored in a patient profile (Figure 2.8). The profile can facilitate the monitoring of the rehabilitation process of the patient together with the therapy schedule, task condition and game performance in multiplayer and single-player applications. The patient profile might provide lessons from previous therapy sessions that allow us to create new plans for future therapy sessions in order to optimize the therapy outcome. Alongside the interaction through the game itself the two players also interact verbally and visually with each other. This social ”human- human interaction environment” influences the behavior and experience of the patient. 16 Chapter 2. Conceptual Control Chart Behavior Experience/ Human-Human Interaction Environment Outputs1 Outputs2 User InputsUser 1 InputsUser 2 Interface 1Interface 2 Patient Co-Player Actions 1 Actions 2 Challenges 1 Challenges 2 Challenges Game C

B A Therapist Desired Desired Behavior Experience/ Figure 2.7: Simpliﬁedplayer targeted gaming concept in robot-assisted for training multiplayer training: Patient tailored and balanced multi- 2.3. Multiplayer in robot-assisted rehabilitation 17 Measured Experience Assessment Behavior Experience/ Patient Co-Player Interaction Environment Human-Human Human-Human act,2 act,2 act,1 act,1 x F F x H H Interface 1 Data Interface 2 Data Interface 1 Interface 2 Condition 1 Condition 2 Condition ref,1 H ref,2 H meas,1 meas,2 x x x x G G G G Condition Condition Controller Controller Integrated Integrated Complexity 1 Complexity Complexity 2 Complexity Interpretation Interpretation Game Performance 1 GamePerformance Game Performance 2 GamePerformance Game Game Game Complexity 1 Complexity 2 Complexity C

B A Patient Data Patient Patient Pro!le Patient Co-Player Data Co-Player Co-Player Pro!le Co-Player Game-/ Selection Therapist Therapist Complexity Gamemode-/ Desired Desired Behavior Experience/ Figure 2.8: Targeted conceptin for robot-assisted multiplayer training: training Patient tailored and balanced multiplayer gaming 18 Chapter 2. Conceptual Control Chart 19

Chapter 3

Systematic Review of Health-related Multiplayer Games

Foreword and Overview

The ﬁrst research question was whether there is a general consensus that multiplayer games in health-related disciplines increase the motivation and the eﬀort of the players compared to single player games (Figure 3.1). We systematically searched for multiplayer games in health-related disciplines and found thirteen articles that met our inclusion criteria. 20 Chapter 3. Systematic Review Measured Experience Assessment Behavior Experience/ Patient Co-Player Interaction Environment Human-Human act,2 act,2 act,1 act,1 x F F x H H Interface 1 Data Interface 2 Data Interface 1 Interface 2 Condition 1 Condition 2 Condition ref,1 H ref,2 H meas,1 meas,2 x x x x G G G G Condition Condition Controller Controller Integrated Integrated Complexity 1 Complexity Complexity 2 Complexity Interpretation Interpretation Game Performance 1 Game Performance Game Performance 2 Game Performance Game Game Game Complexity 1 Complexity 2 Complexity C

B A Patient Data Patient Patient Pro!le Patient Co-Player Data Co-Player Co-Player Pro!le Co-Player Game-/ Selection Therapist Therapist Complexity Game mode-/ Figure 3.1: Concept chart with highlighted (red) elements for robot-assisted multiplayer therapy Desired Behavior Experience/ 3.1. Abstract 21

3.1 Abstract

Multiplayer games have emerged as a promising approach to increase the motivation of patients involved in rehabilitation therapy. In this systematic review, we evaluated recent publications in health-related multiplayer games that involved patients with cognitive and/or motor impairments. The aim was to investigate the eﬀect of multiplayer gaming on game experience and game performance in healthy and non-healthy populations in comparison to individual game play. We further discuss the publications within the context of the theory of ﬂow and the challenge point framework.

A systematic search was conducted through EMBASE, Medline, Pub- Med, Cochrane, CINAHL and PsycINFO. The search was complemented by recent publications in robot-assisted multiplayer neurorehabilitation. The search was restricted to robot-assisted or virtual reality-based training.

Thirteen articles met the inclusion criteria. Multiplayer modes used in health-related multiplayer games were: competitive, collaborative and co-active multiplayer modes. Multiplayer modes positively affected game experience in nine studies and game performance in six studies. Two articles reported increased game performance in single- player mode when compared to multiplayer mode. The multiplayer modes of training reviewed improved game experience and game performance compared to single-player modes. However, the methods reviewed were quite heterogeneous and not exhaustive.

One important take-away is that adaptation of the game conditions can individualize the diﬃculty of a game to a player’s skill level in competitive multiplayer games. Robotic assistance and virtual reality can enhance individualization by, for example, adapting the haptic conditions, e.g. by increasing haptic support or by providing haptic resistance. The ﬂow theory and the challenge point framework support these results and are used in this review to frame the idea of adapting players’ game conditions. 22 Chapter 3. Systematic Review

3.2 Introduction

Multiplayer games provide diversified game play and incorporate social interaction to promote enjoyment of the involved players. The additional player adds new possibilities to the game environment, generally missed in single-player gaming against preprogrammed challenges or artificially controlled opponents. The multiplayer environment and related game mechanics can facilitate social interaction, ranging from conversation to haptic interaction. Due to the social interaction, the game experience is thought to be better in multiplayer compared to single-player gaming [55]. In this review, we investigated whether multiplayer environments have improved game experience or game performance in serious games of health-related disciplines and neuromuscular therapy. Furthermore, we compare different multiplayer game modes regarding game experience and game performance. The mode of the game specifies whether the players compete or cooperate with one another [78]. In line with flow theory, a competitive mode requires opponents of similar skill level to achieve enjoyment as the task difficulty experienced by one opponent is dependent on the other players’ skill level [156]. Comparable skill levels prevent boredom or stress and result in a meaningful challenge level that leads to a flow state when training [38]. Accounting for the individual’s skill level is not only important for game experience, but also for motor learning itself [65]. According to the challenge point framework, motor learning depends on the amount of interpretable information. An increase in task difficulty increases the amount of available information. However, the learning potential is only increased when information can still be interpreted and does not overload the individual, i.e. hamper task performance. As the ability to process information varies between individuals, the task difficulty should be adjusted to the individual in order to facilitate learning. The variability in processing information is even more striking in health-related gaming and in neuromuscular training in particular. In health-related gaming where players may have stronger differences in skill level, task difficulty adaptation in multiplayer games, i.e. accounting for the skill level of the opponents, bears a challenge for game developers: Adaptations of game mechanics – as commonly 3.2. Introduction 23 done for single-player games – affect the game conditions for all players. In health-related gaming environments – particularly in patients with neuromuscular deficits – this challenge is even more prominent due to large variability in information processing abilities and motor skills. This large variability leads to differences across active range of motion, muscular strength, interlimb coordination or spasticity, among others [84]. In multiplayer games, these deficits can make it so that two players cannot play an exciting match against one another [38]. It is therefore of interest to understand how to manipulate game conditions to balance the skill levels of patients and enable multiplayer gaming. Various solutions are offered that account for differences in skill level in multiplayer gaming in neurorehabilitation. These can be differen- tiated into adaptation of game mechanics (e.g. frequency of actions) and adaptation of the interface mechanics (e.g. robotic support). In multiplayer games, adaptations of the game mechanics affect the game condition for all players. In contrast, adaptation at the interface level allows for individual adjustment of the task difficulty. Thus, in settings with differently skilled players, individual interface mechanics can ensure that all players are challenged according to their skill level and can therefore increase game experience. The increased game experience leads to better game performances by all players involved [38, 124]. Increased game performance enabled by more energy expenditure of the patient facilitates the general idea of serious games, i.e., playing for a primary purpose other than pure entertainment [41]. If enhanced game performance is achieved by increased physical activity, training intensity is also increased. In neuromuscular therapy, training intensity – alongside early treatment, user-centered, and task-oriented training – is one of the key factors in neurorehabilitation [92, 155]. Therefore, multiplayer gaming has great potential to further increase the benefits of robot-assisted neuromuscular and virtual reality-assisted therapy [95, 29]. To facilitate the transfer and consolidation of multiplayer gaming from health-related gaming to robot-assisted neuromuscular therapy, we systematically reviewed the literature regarding multiplayer gaming in health-related games. In this section, we describe the available literature in the context of the flow model. To further facilitate a 24 Chapter 3. Systematic Review transfer to neuromuscular rehabilitation, this review discusses measures regarding perceived game experience and physical performance, which have been applied in studies using multiplayer modes offering social interaction. Furthermore, this review complements definitions of task difficulty with the term “conditional task difficulty” to facilitate the discussion of haptic difficulty adaptation in robot-assisted rehabilitation. All discussions refer to our research question: Do multiplayer games enhance experience and performance in robot-assisted and virtual-reality assisted neurorehabilitation?

3.3 Methods

3.3.1 Protocol The review protocol was based on an initial search review of the promising aspects of social interaction in virtual environments for rehabilitation interventions.

3.3.2 Eligibility criteria We developed a database adjusted electronic search strategy for EM- BASE, Medline, PubMed, Cochrane, CINAHL and PsycINFO in collaboration with a librarian from the Medicinal Library of the Uni- versity of Zurich. The search was restricted to English and German language literature. There was no limitation in publication date or restriction by study design. The first search was performed in April 2016 and was complemented by papers published since (until January 2018). We used medical sub-headings as search terms, including the following main terms for the population: motor impairments, cognitive impairments, stroke, elderly people; for interventions: competition, collaboration, cooperation, coopetition, competitive behavior, collaborative behavior, exergame, multiplayer-exergame, serious game, rehabilitative game, rehabilitative exergame, education game, com- puterized training, robot-assisted rehabilitation, robot-aided rehabilitation, virtual reality, virtual reality therapy, virtual world, social community, community integration, virtual community; for outcome: 3.3. Methods 25 persuasion, compliance, motivation, engagement, effort, adherence, therapy progress. The search strategy (additional file 1) was initially run in EMBASE and then adapted to the search format requirements of the other databases included in this review. The search results were supplemented by articles found through manual searches by scanning reference lists of identified studies.

3.3.3 Study selection

After duplicate citations were removed, two independent reviewers determined which articles should be included within this systematic review by scanning the titles, abstracts and keywords while applying the inclusion and exclusion criteria (Table 3.1). An article was included if both reviewers independently saw the potential that the article’s results might be of interest for multiplayer applications in robot-assisted neuromuscular therapy. To determine whether the full text should be retrieved for a given citation, the two reviewers marked each citation using a “yes,” “no,” or “unknown” (unsure whether yes or no) designation. The citations marked as “unknown” were discussed within the reviewers. The screening process involved simul- taneous title and abstract screening. The reviewers evaluated the same set of citations. A study was considered eligible for inclusion in the review when the studied intervention: (1) was a multiplayer intervention in a robot-assisted, video-game based or virtual reality based, health-related setting, and (2) examined the eﬀects of the intervention on perceived game experience or physical functioning of the players involved. This review includes a transfer to neurorehabilitation training gaming from related disciplines, such as health-related serious gaming. For this purpose, we considered typical types of exercise interfaces as described by Tanaka [158]. Feasibility studies with less than ten subjects were not included in the systematic assessment but relevant content was considered for the discussion. The full-text article was read if the title, abstract, or keywords provided insuﬃcient information to decide on inclusion (Figure 3.2). 26 Chapter 3. Systematic Review

Table 3.1: List of inclusion and exclusion details area inclusion details population healthy, obesity, motor impairments, cognitive impairments, stroke, elderly people study type intervention studies of any type, including case studies and non-randomized trials intervention competition, collaboration, cooperation, coopetition, competitive behavior, collaborative behavior, exergame, multiplayer- exergame, serious game, rehabilitative game/exergame, education game, computer- ized training, robot-assisted rehabilitation, robot-aided rehabilitation, virtual reality, virtual reality therapy, virtual world, social community, community integration, virtual community, video-game based outcomes persuasion, compliance, motivation, engagement, eﬀort, adherence, therapy progress, therapy success exclusion details reviews, animal studies, concept studies, feasibility studies, human-machine interaction (only one human involved), methodological, theoretical or discussion papers

3.3.4 Data extraction process and data synthesis

The following data were extracted from the studies: (1) characteristics of the studied population: number of participants, disease and age, (2) characteristics of the interventions: the design, frequency and duration of the intervention, co-interventions and control intervention; (3) characteristics of the outcomes: outcome measures and results (Tables 3.2 and 3.3). The included studies were divided into three groups according to their assessments: perceived game experience, physical performance, and personality factors. The review of perceived game experience included studies using inventories that 3.3. Methods 27

EMBASE Medline PubMed Cochrane CINAHL PsycInfo (n=608) (n=330) (n=77) (n=176) (n=159) (n=492)

Additional records identi!ed References identi!ed through Excluded duplicates (n=559) through other sources (n=47) database searching (n=1889)

Articles screened on basis of Excluded (n=1304) title and abstract (n=1330)

No intervention (n=11) Game concept (n=3) Additional records identi!ed Eligible for full text reading Single player mode (n=3) through citations and author (n=33) Out of scope (n=3) tracking (n=7) Totally excluded (n=20)

Studies included in qualitative synthesis (n=13)

Figure 3.2: Study selection flow chart. assessed the players’ game experience or game playtime expressing a positive game experience. The review of physical performance included studies assessing any physical fitness related quantity measured by the game score or physical quantities measured on the player. If a study monitored both perceived game experience and physical performance, the study appears in every relevant group, accordingly. Perceived game experience is discussed as the expressed opinions gathered from the activity. The opinion was assessed either during or after the task and either by answering a questionnaire or by assessing performance attributes (e.g. player’s decision on duration of task performance). Physical fitness is considered as a set of attributes that people have or achieve to perform a physical activity as defined by Caspersen [32]. Physical performance is considered as the performed physical activity represented by the amount of useful work accomplished within the task. Assessment of the usefulness of the work is defined by the task designers. Because we expected the interventions and reported outcome measures to be markedly varied, we focused on a description of the studies and their results, and on thematic synthesis as defined by Thomas and Barnett-Page [163, 12]. 28 Chapter 3. Systematic Review

3.3.5 Assessment of study quality

Our critical appraisal of the studies was based on a checklist designed for assessing the methodological quality of both randomised and non-randomised studies of healthcare interventions [43]. The checklist assesses biases related to reporting, external validity, internal validity and power. Six items were not considered in this review: adverse eﬀects (item 8), follow-up analyses (items 9 and 26), represen- tativeness of treatment locations and facilities (item 13), allocation concealment for participants (item 14), and blinding of investigators (item 15). These items were excluded because there were not any follow-up studies available that considered the given type of intervention (items 9 and 26), or because we considered these items as being of minor signiﬁcance for this review (items 8, 13, 14 and 15).

The remaining 21 items were applied by two reviewers (KB, AS) to assess the methodological quality of the selected full text studies (additional file 2). The total possible score was 22 points, where every item can be scored 0 or 1 point except for the confounding description (item 5; 0, 1 or 2 points). Scoring for consistency in recruitment time period (item 22) was, compared to the original checklist, changed to scoring for stating time period of intervention. Scoring for statistical power (item 27) was simplified to a choice between 0 or 1 points depending on the level of ability to detect a clinically important effect. The scale ranged from insufficient (β < 70% = 0 points) to sufficient (β ≥ 70% = 1 point). To assess the level of agreement between the investigators a Cohen’s kappa analysis was performed on all items on the checklist. In accordance with Landis and Koch’s benchmarks for assessing the agreement between raters a kappa-score of 0.81...1.0 was considered almost perfect, 0.61...0.8 was substantial, 0.41...0.6 was moderate, 0.21...0.4 was fair, 0.0...0.2 were slight and scores < 0 were poor [94]. The PRISMA-statement was followed for reporting items of this systematic review [100]. Therefore, eligible criteria, information sources, and search strategy were defined pre-search (30.10.2014) and remained unchanged (Additional File 1). Study selection, data collection, and data reporting are fully described within this paper However, the systematic review is not registered in any data base. 3.3. Methods 29 / / 9 7 . . 7 7 . . 6 7 . 13 2 14 11 51 ± ± ± ± ± 8 7 7 8 6 ...... Age* 19 5. . . 42 56 52 25 24 57 Subjects undergraduate students healthy subjects patients (ischemic stroke:rhagic stroke: 19; 3; brainder hemor- rotator tumor: cuff 4; tear: shoul- injury: 2; 1) traumatic with brain chronic15 arm impairment patients (ischemicorrhagic stroke: stroke: 1; 9;lowed ischemic by hem- stroke hemorrhagic fol- matic stroke: brain 2; injury:2) trau- with chronic 1; armtients impairment/20 cerebral in pa- the palsy acutestroke or subacute phase of 64 unimpaired friendsstudents) /5 arm (undergraduate impaired duelogical to injuries neuro- (ischemic stroke: 3;matic trau- brain injury:jury: 1) 1; spinal cord in- N 135 74 29 35 69 Design randomized within subject within subject within subject within subject one standard deviation (years). ± Study Feltz et al 2012 [47] Ganesh et al. 2014 [56] al.Gorˇsiˇcet 2017 [61] al.Gorˇsiˇcet 2017 [60] al.Gorˇsiˇcet 2017 [62] Table 3.2: Included studiesage reported by design and subject specifications, Part 1. *Age range or mean 30 Chapter 3. Systematic Review ± ± 3 . 70 / 7 3 . . 5 . 4 19 Age* 21. . . 62 26 25. . . 7322. . . 69 / 18. . . 23 18. . . 23 15. . . 19 15. . . 19 12. . . 16 Subjects healthy subjects 32 healthy / 16vivors hemiparetic using stroke their sur- impaired arm 30 unimpaired (noimpairment) / cognitive 8 or strokeundergraduate patients motor communication class undergraduate communication class at a large Midwestern university low-income, adolescents,school publicadolescents, high urban public high school seventh grade students exhibiting anactive in- lifestyle N 18 48 38 162 158 31 31 43 Design within subject within subject within subject assigned within subject randomized RCT within subject one standard deviation (years). ± Study Johnson et al. 2008 [79] Mace et al. 2017 [104] Novak et al. 2014 [119] Peng and Crouse 2013 [127] Peng and Hsieh 2012 [128] Staiano et al. 2012 [151] Staiano et al. 2013 [152] Verhoeven et al. 2015 [173] Table 3.3: Included studiesage reported by design and subject specifications, Part 2. *Age range or mean 3.4. Results 31

3.4 Results

3.4.1 Study selection The search provided a total of 1889 references. After adjusting for duplicates, 1330 remained. Of these, 1304 were discarded because they were out of scope of this review. The remaining 26 potentially relevant articles were supplemented by 7 additional references retrieved by a manual search. This resulted in a total of 33 articles eligible for full-text reading. After full-text reading, 19 articles were excluded because neither an intervention nor social interaction was presented (Figure 3.2). Thirteen articles were ﬁnally reviewed (Ta- bles 3.2 and 3.3).

3.4.2 Study characteristics All studies were published in English. The publication dates ranged from 2008 to 2017. In the selected studies, participants were seventh- graders [173], high-school students [151, 152], adults [56, 79, 104, 119], arm impaired [61, 60, 62, 104, 119], and undergraduate students [47, 62, 127, 128]. The age of the participants ranged from 12 years to over 90 years considering that four publications did not include full information about minimal and maximal age [47, 61, 62, 104].

3.4.3 Devices, games and multiplayer modes Six studies used a commercially available exergaming input device, such as a Nintendo Wii controller or a Kinect system, in combination with a commercially available game [127, 151, 152, 173], or an in-house developed game [47, 128]. Five studies used a haptic ma- nipulator or arm rehabilitation system with an in-house developed game [61, 60, 62, 79, 119]. One study used a grip transducer with an in-house developed game [104]. In six studies addressing multiplayer modes, a competitive mode was compared to a cooperative mode [61, 119, 127, 128, 151, 152]. Four of these studies also compared competitive or cooperative gaming against a single-player mode [61, 119, 127, 152]. Six studies compared either competitive or cooperative gaming against a single- 32 Chapter 3. Systematic Review player mode [47, 56, 60, 79, 104, 173]. Only one study compared a competitive mode to a single-player mode in different games [173] and one study compared multiple sessions of a competitive mode to a single-player mode [60]. One study investigated the influence of external forces, including forces from partner players, on performance and motor learning [56]. Included studies also compared discrepancy levels of a participant’s game performance to a competitor’s game performance [47], and different methods to decrease discrepancy in a multiplayer rehabilitation game [62]. Furthermore, included studies compared friend- and stranger-paired group multiplayer gaming [127], gaming in home and clinical environments [60], and different levels of interaction in a tele-rehabilitation game environment [47].

3.4.4 Assessment Assessment of perceived game experience Eleven out of thirteen articles assessed the effect of the intervention on the motivation of the player. To measure the players’ motivation, the Intrinsic Motivation Inventory (interest/enjoyment, perceived competence, effort/importance, pressure/ tension) [110] was used seven times [61, 60, 62, 79, 104, 119, 128], Malone’s theory of intrinsic motivations for learning (sensory immersion, control/choice, challenge/optimal difficulty, goal setting and feedback) [107] was used once [151], the Kids Game Experience Questionnaire [130] was used once [173], and an evaluation of motivation related adjectives (boring (reverse-coded), exciting, enjoyable, entertaining, fun, interesting, pleasant) [149] was used once [127]. A ranking regarding preferred game mode based on subsets of the Intrinsic Motivation Inventory (interest/enjoyment, perceived competence, effort/importance, pressure/tension) [110] was used in four studies [61, 79, 104, 119] and a weighted preference in one study [62]. Assessment of goal commitment [75, 167] was used once [128], and assessment of psychosocial attractiveness of game design regarding interpersonal communication in the four factors social interaction, collaboration, individual feedback and team feedback [80] once [151]. The Koehler effect [88], assessed by measuring the difference in per- sistence in executing the intervention task, was used once [47]. 3.4. Results 33

Psychosocial factors such as self-efficacy [137], efficacy self-esteem [134] and peer support [30] were assessed in one study [152]. Intervention specific questions for perception assessment of the external forces and fatigue were used in one study [56].

Assessment of physical performance Five out of thirteen articles assessed physical fitness by calculating energy expenditure: weight difference by measuring weight before and after the intervention [152], comparing the heart rate during an intervention with heart rate during individual control condition [47], calculating the metabolic equivalent of task [173] as defined by Ainsworth [6], and averaging the acceleration profile assessed during the intervention [127, 151] as introduced by Puyau [131]. Five out of thirteen studies, [56, 79, 104, 119, 128], assessed physical performance by comparing game performance scores. Three out of thirteen studies, [61, 60, 62], used the root-mean-square (RMS) of the velocity profile of the hand movement (introduced by Van Der Pas [170]); one out of these three studies also reported mean velocity [61]. One study, [79], assessed physical performance by extracting the peak velocity of the hand movement. One study, [47], assessed perceived exertion with the Borg scale [23].

Assessment of personality factors In three studies [61, 60, 119], assessments of perceived game experience and physical performance were compared with assessments of personality as suggested by Goldberg [59].

3.4.5 Influence of multiplayer games on game experience and performance The majority of the reviewed studies state that social interaction through multiplayer game settings improve both game experience and game performance. The profile of game experience and performance is different between individuals. However, the profile may be modeled in dependence on flow theory and challenge point framework (Figure 3.3). All studies that examined game experience stated 34 Chapter 3. Systematic Review that their results proved that multiplayer modes positively influenced game experience [47, 61, 60, 79, 104, 119, 127, 152, 173]. Most studies stated that multiplayer games led to better game performances or higher physical exertion in most of the measured dimensions [47, 56, 61, 104, 152, 173]. Two studies stated that single-player mode improved game performance when compared to multiplayer modes or increased physical exertion in certain dimensions [104, 127]. Two studies also found correlation of game experience and game performance [151, 79].

positive Player Less skilled Skilled high Single player Multiplayer Game experience Game performance negative low low Conditional task diculty high low Conditional task diculty high

Figure 3.3: Summary of the review. In comparison to single-player modes, multiplayer game modes have been shown to positively influence both game experience and game performance. The benefit of multiplayer modes is present for players of all skill levels and at all conditional task difficulties.

3.4.6 Quality evaluation The agreement on study quality between the two reviewers was almost perfect. The estimated Kappa value was 0.99 with a conﬁdence interval ranging between 0.96 and 1.00 (α = 0.05). The percentage of agreement between the two reviewers was 99.3%. The mean quality score was 12.1 points (maximum: 22 points, range: 10-14 points), the median value was 12 points and the mode was 11 points. The mean score for reporting was 6.2 points (maximum: 9 points; range: 5-8 points), for external validity 0.8 points (maximum: 2 points; range: 0-2 points), for internal validity (bias) 3.9 points (maximum: 3.5. Discussion 35

5 points; range: 3-4 points), for internal validity (confounding) 1.2 points (maximum: 5 points; range: 0-2 points) and for power 0.0 points (maximum 1 point; 0 points), see additional ﬁle 2.

3.5 Discussion

3.5.1 Study selection

Our search resulted in thirteen articles fulﬁlling the inclusion criteria. These articles evaluated the psychological and physiological experience of multiplayer gaming using various forms of player interaction in studies on neuromuscular patients, overweight adolescents, and students. The thirteen articles remain small compared to 1031 papers related to “robot-assisted training” and to 444 papers related to “exergaming” that were published until July 2018 [1]. This highlights that multiplayer gaming seems not yet systematically considered in neuromuscular therapy or other health-related training. The combination of both the small number of included papers and the hetero- geneity in their methods, prevents us from giving a clear answer to the research question.

3.5.2 Devices and games inconsistently discussed

The selection process of devices and of the commercially available games was consistently not discussed within any of the thirteen articles. The design process of the in-house developed games was discussed in three articles [47, 56, 104]. Feltz et al. [47] and Ganesh et al. [56] designed the games according to the paradigm to be tested, i.e., Koehler eﬀect [88] and reactive motor adaptation, respectively. Mace et al. [104] explained the game design process, and how their multiplayer game mode explicitly demands a collaborative behavior [78]. To better discuss and compare the study results, the targeted behavior and the actual game modes selected or designed, should generally be reported in more detail. 36 Chapter 3. Systematic Review

3.5.3 Limited diversity of multiplayer modes

The multiplayer modes in the reviewed studies have been commonly named competitive, collaborative and cooperative mode. According to the taxonomy proposed by Jarrasséet al. [78] and others (Fig- ure 3.4), some modes of training in the reviewed studies should be termed co-active (a task that can be solved by individual player) rather than cooperative (playing in the same team with different roles according to own individual skills, thus, a role being either “supported” or “supportive”; see also Sawers and Ting [142]). Al- though a cooperative mode in that sense has not been considered in the reviewed studies, this mode may have great potential in robot- assisted rehabilitation systems since the roles of “supported” and “supportive” are often given in a therapeutic setting with patient and therapist [97, 148]. Such a cooperative mode was tested with the ARMin arm rehabilitation robot (Appendix C). The collaborative mode (playing in the same team with equal roles) was used in four studies [47, 56, 61, 104]. Differences in effects of collaberative and competitive modes in health-promoting exergames have already been discussed by Marker and Staiano [109].

The distinction of competitive modes, as done by Mueller [115], seems reasonable to predict training outcome. For instance, in combat gameplay, physical eﬀort was higher than in object competition when both players controlled the same object [173].

Coopetition, i.e., competition and cooperation in combination, has been proposed as a new mode of multiplayer gaming [89]. Derived from a business concept, coopetition can be linked to a social health behavior where people compare their health behavior with others sharing the same health-related goal [89, 26]. Platforms such as social media enable this health behavior. This “coopetitive” behavior can be targeted by games that monitor the health progress in an under- standable and comparable representation, e.g. weight loss. Robot- assisted neurorehabilitation can provide similar representations using parameters such as range-of-motion assessments and performance measures in reaching tasks [82]. 3.5. Discussion 37 MODE Co-activity Coopetition Combat competition Combat yes TASK no co-player? Control the the Control Object competition coopetitive subtasks coopetitive non parallel subtasks non parallel yes no no goal? Same Each sub- Each dependently? superordinate superordinate parallel subtasks parallel task cantask be in- done Parallel competition Parallel competitive subtasks competitive Assistence Collaboration divisible subtasks divisible antagonistic subtasks antagonistic yes yes yes yes no no no no alone? Can each Can Can harm Can generous? Education Same roles? Is a co-player Is all subtasks the co-player? subtask be done agonistic subtasks agonistic cooperativ subtasks cooperativ interactive subtasks interactive Figure 3.4: Determinationert of [89], multiplayer and modes. Muellerteristics. [115], According The the task to appliedmultiplayer characteristics the multiplayer mode. and mode taxonomies the can The of players’therefore be full Jarrass´e[78], behavior needs determined variety Kon- deﬁne further based of investigation. the on multiplayer behavioral the modes characteristics task was of charac- not the covered by the included studies and 38 Chapter 3. Systematic Review

3.5.4 The potential of playing with friends and relatives as co-players

Another example to combine competition and cooperation has been tested in a game-based learning study [184]. In this study, a second level of social interaction was introduced. In the first level of interaction, the teams solved subtasks together in a cooperative manner. In the second level of interaction, the teams interacted with other teams in a competitive manner. Such combinations of competition and cooperation are applicable in clinical environments. Subgroups of patients in clinics or different clinics could compete against each other. Part of the competition could be a comparison of therapy progress, duration of device usage, or virtual points in rehabilitation games. The groups of patients can motivate their teammates to co- operatively contribute to their team’s score by training more. Such a setup is possible in telerehabilitation since participating in remote places does not seem to reduce the motivation to compete: game experience and game score have been shown not to be affected when competing in different rooms compared to cooperating in the same room [127]. Cooperating in a team might be even more motivating when the team member is a friend or a relative. Playing with a friend in a cooperative mode resulted in greater goal commitment compared to playing with a stranger [128]. Regarding rehabilitation games, this implies that in a cooperative game, a well-known game partner – such as a family member – may be preferred over a lesser-known patient or therapist. For competitive modes it was already shown that playing with well-known game partners does improve the game experience when compared to playing with lesser known game partners [61]. Such game play requires input devices for different skill levels, e.g. rehabilitation robots for a severely impaired patient and exergaming devices, such as a Nintendo Wii, for moderately impaired and unimpaired players. Tasks solved by players with different skill levels were discussed in the concepts of flow and of the challenge point framework[38, 65]. However, both concepts are not extensively discussed with regard to multiplayer games using different input devices for different skill levels. 3.5. Discussion 39

3.5.5 Balancing multiplayer interventions in the context of the flow model and the challenge point framework The flow model considers the immediate task difficulty and current skill level, thus, defines whether the player is experiencing boredom, flow or frustration. We use this model as a basis for a description of a player’s specific skill level: the game experience profile. The inverted-U-shape dependency of game experience on task difficulty has been confirmed in multiplayer health-related games [47]. The global maximum of the game experience profile represents the point of highest rated game experience. Flow is expected to be experienced close to this point. Different game modes result in differently shaped game experience profiles. These profiles are influenced by both the game mode and the players’ skills in the game, thus, resulting in different maxima. According to the flow model, the game experience of skilled players demands higher task difficulty to achieve maximum game experience [38]. In studies applying multiplayer modes, a cooperative mode can result in a better game experience compared to a competitive mode [128]. This increase in game experience could be represented by an overall shift upwards in the game experience profile or a game experience profile with less declination; both represented by a widened flow zone — a hypothesis that has yet to be confirmed. Novak et al [119] found that the player’s preferred mode may be predicted based on a correlation with the player’s personality. Therefore, a player-tailored selection of multiplayer mode may result in a better game experience, too. The challenge point framework, introduced by Guadagnoli [65], provides a theoretical basis to conceptualize the effects of task difficulty in motor learning. In competitive games, the game experience is linked to the skill level of the opponent [38]. A discrepancy in skill level makes the weaker player feel frustrated and the better player bored. Changing the conditions by adapting visual, auditory or haptic game elements can reduce the differences in skill level between players. The terminology of the challenge point framework can be used to describe the difficulty of any task, not only regarding the conditions, but also the characteristics of the task and the skill level 40 Chapter 3. Systematic Review of the subject.

3.5.6 Conditional task difficulty: complementing the terminology The challenge point framework defined the terms nominal task difficulty, i.e. the characteristics of the task only, and the functional task difficulty, i.e. the difficulty of the task relative to the skill level of the player and the conditions under which the task is performed. However, the term functional task difficulty does not allow for a distinction of the skill level of the player and the conditions under which the task is performed. Therefore, we propose an extended definition by introducing the term conditional task difficulty, i.e. the difficulty of the task relative to the conditions under which the task is performed. This extended definition is particularly relevant for environments where the conditions of the task can be adapted online, as when using haptic robots that can adapt the support of a patient based on his/her performance. In addition, this extended definition allows to collectively report task difficulty when a game is played with different input devices by the involved players.

3.5.7 Changing the individual conditions to optimize game experience By changing the condition individually, we can change the players’ game performance in relation to their individual skill level. In two included studies and five feasibility studies, game design features such as speed, frequency of actions, or avatar size were adapted in single and multiplayer rehabilitation games to account for differences in skill level [60, 62, 33, 106, 11, 112, 179]. Such design features may change the conditions generally for both players, instead of for each player individually. In addition, these features may visually reveal the different skill levels of the players which can be embarrassing for the worse performing player. Robotic devices offer unique design features to tailor the condition to the players’ motor abilities [36]. For instance, haptic force fields have been used to adjust the task difficulty of a training task. These 3.5. Discussion 41 force fields could be used in multiplayer games to individually set the level of difficulty (Chapter 5). In all difficulty adapting strategies, we have to consider that a change in condition for one player implicitly introduces a change in condition for the other one, since the players are connected over the nominal task itself. If we support the less skilled player, or hinder the more skilled player, in succeeding in the task we approach the highest game experience for both players (Figure 3.5). Performance-balanced games improved game experience compared to non-balanced games in one of the included studies and several feasibility studies [62, 19, 33, 106]. In therapy settings where the level of difficulty cannot be individually adapted, collaborative game designs may be preferred since they may require less similarity in skill level to achieve flow [104].

positive Player Less skilled Skilled high Multiplayer Initial state Target state Game experience Game performance negative low low Conditional task diculty high low Conditional task diculty high

Figure 3.5: Diﬃculty adaptation based on individual condition setting in multiplayer games. Game experience (left) can be optimized by balancing the game performance (right).

The motivation in cooperative modes can even be increased with conjunctive tasks design (i.e. collaborative according to Jarrassé[78]) facilitating the Koehler effect [72]. The Koehler effect occurs when an inferior team member performs a difficult task better in a team than one would expect from knowledge of his or her individual performance. The effect has been found to be strongest in conjunctive task conditions in which the group’s potential productivity towards a cooperative goal is equal to the productivity of its least capable member [48]. Experience with the Koehler effect implies that multiplayer 42 Chapter 3. Systematic Review difficulty adaptation targeting a moderate discrepancy compared to the player’s performance can optimise the game experience of the participants [47, 56]. Such small discrepancies can be achieved by robotic devices as they can individually assist or challenge players [29].

3.5.8 Different assessments hinder inter-study comparison Assessment of perceived game experience In general, measures of game experience can be acquired using physiological measures and questionnaires; only the latter were present in the included papers. Physiological measures, e.g. sensors measuring cardiac activity, significantly improve the assessment of game experience [13]. However, the required additional set of sensors may disturb game play and effort by the study operators. Among the presented questionnaires, the Intrinsic Motivation Inventory is widely used in various sports interventions and can therefore allow inter- study comparison [110]. However, the Intrinsic Motivation Inventory does not necessarily have good discriminative power as pointed out by one study [119]. That is why the participants have been asked to also rank the games regarding game experience related questions (e.g. “Which game mode do you prefer?”). Inventories regarding psychosocial attractiveness of game design in the dimensions of social interaction, collaboration, individual feedback and team feedback, can extend motivation inventories when different multiplayer modes are compared [151]. The perceived multiplayer mode, e.g. playing in a cooperative or competitive mode, will influence the answers regarding psychosocial attractiveness. However, the multiplayer mode might not be perceivable to all players. Thus, should be checked after each mode if the mode is not introduced accordingly [128]. The benefit of social interaction could be increased by integrating visual, auditory/verbal, and haptic elements. In the cognitive task study of Yu et al. [184] conversely, hiding of the opponent team (anonymous opponents) in a competitive mode showed improved game experience and satisfaction compared to visual presence of the opponent team. Hence, visual presence of the opponent seems to neg- 3.5. Discussion 43 atively influence game experience. In contrast, Johnson et al. [79] stated that the more modalities of social interaction integrated in motor tasks, the more enjoyable for the players. The perceived intensity of social interaction in different modalities seems to vary between task and player. In the study of Yu et al. [184] the intensity of social interaction within the team members might superpose any interaction with the opponent team. Therefore, assessing the perceived intensity of social interaction provided by different modalities and involved people is suggested. In the study of Gorsic et al. [60], the level of verbal interaction was assessed by a study operator. However, a standardized assessment for perception of interaction intensity has not been established yet.

Assessment of physical performance Energy expenditure is a common measure for physical fitness regarding physical performance [99]. Various methods to measure energy expenditure or their consequences include: the determination of acceleration profiles [127, 151], feature extraction of the velocity profile [61, 62, 79], measuring the weight drop [152], and measuring the heart rate [47]. Alternatively, the Borg scale, used as a measure of perceived physical performance in one study, is a widely accepted and valid measure of exercise intensity [23, 35, 47]. Both methods presented allow inter-game comparison or even comparisons regarding physical performance with non-gaming interventions. Physical performance corresponds to training intensity and, therefore, predicts improvement in motor function, too [123]. However, neuromuscular therapy and movement training in general are related to motor function and execution and rather than to physical exertion only [162]. Consequently, measures of physical fitness may not be sufficient to provide all relevant information regarding progress in physical functioning [133]. Therefore, game performance score or gamified assessments based on conventional motor functional assessments (e.g. Fugl-Meyer [54]) should be included to complement the evaluation of the influence of the game on physical functioning progress [82]. In games where the end effector of the input device controls the avatar, the end effector position affects the success or failure within 44 Chapter 3. Systematic Review the game task. Assessing the end effector position and its impact on task success does not demand functional movement patterns in arm joints. However, functional movements are important in upper limb training. In exoskeleton devices used for rehabilitation training it is possible to measure the movements of non-game controlling joints [85]. One solution is the assessment of quality of movements to measure the player’s functional physical performance. Assessments that are independent of task dimensions do not need a developer’s definition such as the spectral arc length metric [10]. However, the validity of movement quality assessments is under discussion [118].

Limitations We used a structured protocol to guide our search strategy, study selection, extraction of data and statistical analysis. However, limitations of this review should be noted: a publication bias may be present, as well as a language bias, given that we considered only interventions described in published studies and restricted our search to English and German language publications. A bias regarding research ﬁelds was generated since we mainly focused on neuromuscular therapy and health-related training.

3.6 Conclusions

Multiplayer modes can enhance the players’ perceived game experience and positively influence the players’ performance. Based on the small number of relevant studies published so far, a conclusion cannot yet be drawn about which multiplayer mode is best during neurorehabilitation training. A meta-analysis of game experience and game performance outcomes may be suggested as soon as more multiplayer studies with homogeneous outcome measures will be published. Nev- ertheless, this review demonstrated that the players’ individual skill levels and personalities, as well as their role in the game, must be taken into account when selecting and designing multiplayer modes. Based on the model of flow and the challenge point framework, we suggest an individual adaptation of game conditions, i.e. conditional task difficulty, to assimilate differently skilled players for an opti- 3.6. Conclusions 45 mal game experience. Furthermore, player specific selection of multiplayer modes may result in more robust interventions regarding game experience and requires less assimilation of differently skilled players. We suggest breaking the limited variety in multiplayer modes and fully exploring multiplayer modes and co-player’s characteristics such as the co-players presence, skill level, personality and relation to the player. We further suggest that future studies use a more stringent research design in which participants are allocated to either single play or multiplayer modes of exercise through randomised assignment. 46 Chapter 3. Systematic Review 47

Chapter 4

Multiplayer Gaming in Robot-Assisted Training

Foreword and Overview

The second research question was whether multiplayer games in stroke therapy increase motivation and the physical effort of (sub-)acute stroke patients (Figure 4.1). Considering the reviewed studies we focused on the specific patient group of (sub-)acute stroke patients and carried out a test using an air hockey computer game played with the ArmeoSpring device in different single- and two-player modes.

4.1 Abstract

Motivation is a crucial factor for the improvement of therapeutic outcome and compliance of patients in neurorehabilitation. Recent studies have proved that two-player games in neurorehabilitation can increase patient motivation. However, the impact of two-player games on the motivation of (sub-)acute stroke patients has not been extensively tested yet. We carried out a test using an air hockey computer game displayed on a monitor and controlled using an exoskeleton device. Four modes 48 Chapter 4. Multiplayer Gaming in Robot-Assisted Training Measured Experience Assessment Behavior Experience/ Patient Co-Player Interaction Environment Human-Human act,2 act,2 act,1 act,1 x F F x H H Interface 1 Data Interface 2 Data Interface 1 Interface 2 Condition 1 Condition 2 Condition ref,1 H ref,2 H meas,1 meas,2 x x x x G G G G Condition Condition Controller Controller Integrated Integrated Complexity 1 Complexity Complexity 2 Complexity Interpretation Interpretation Game Performance 1 Game Performance Game Performance 2 Game Performance Game Game Game Complexity 1 Complexity 2 Complexity C

B A Patient Data Patient Patient Pro!le Patient Co-Player Data Co-Player Co-Player Pro!le Co-Player Game-/ Selection Therapist Therapist Complexity Game mode-/ Desired Behavior Experience/ Figure 4.1: Concept chartspecific with patient highlighted group (red) elements for robot-assisted multiplayer therapy with a 4.2. Introduction 49 were tested in randomized order by (sub-)acute patients: single- player competitive (against a computer-controlled opponent), single- player cooperative (with a computer-controlled player), two-player competitive (against a patient-controlled opponent), and two-player cooperative (with a patient-controlled player). Exercise intensity was measured using the root-mean-square value of the hand velocity. Af- ter each game mode, the patients rated how true statements are out of the two Intrinsic Motivation Inventory subscales interest/enjoyment and effort/importance. After the session, the patients ranked all game modes with regard to interest/enjoyment and effort/importance. Forty (sub-)acute stroke patients played all four modes. Sixteen out of forty patients chose a two-player mode as their favorite mode. No significant effect of two-player as opposed to single-player in intrinsic motivation could be identified. However, we found a significant effect of competition as opposed to cooperation on interest/enjoyment (p=0.044) and on exercise intensity (p<0.001). Two-player modes are preferred by some of the (sub-)acute patients and are as motivating and intense as single-player modes. Competi- tion is more enjoyable and intense than cooperation for both acute and subacute stroke patients.

4.2 Introduction

Until recently, studies on two-player gaming were limited to subjective outcome measures collected from healthy subjects or a few chronic stroke patients [119, 79]. These limitations have been ad- dressed in several two-player studies. In these studies, the movement of the end-effector of a therapy device was analyzed to obtain quantitative outcome measures. Different therapy devices were incorporated: Bimeo arm rehabilitation system [61, 60, 62], force sensing hand grips [104], and the Rehabilitation Gaming system [106]. These studies with end-effector devices demonstrated that, compared to single-player modes, two-player modes can increase intrinsic motivation and are preferred by most of the players (healthy and patients). One study also showed that competitive two-player modes can also increase therapy intensity. [61] 50 Chapter 4. Multiplayer Gaming in Robot-Assisted Training

Two-player rehabilitation games have also been tested with the ARMin rehabilitation exoskeleton. Thirty healthy subjects and eight chronic stroke patients played three different modes of a virtual air-hockey game: single-player, i.e. one subject played against a computer without further player interaction; two-player competitive, i.e. two subjects played against each other; and two-player cooperative, i.e. two subjects played together against a computer opponent. This had the same result as the outcome of end-effector-based two-player studies, in that it was concluded that two-player games are a promising approach to increase patient motivation in robot-assisted upper limb therapy. Furthermore, the preferred two-player modes (competitive or cooperative) could be predicted from the individual personality subscales. However, the study with the ARMin robot had three major limitations: First, two different two-player modes were tested against only one single-player mode and two competitive modes were tested against only one cooperative mode; thus, an asymmetric study design was present. A single-player cooperative mode would complement the game modes for fair comparison. Secondly, no objective measures of exercise intensity were used. Thirdly, the study participants were only healthy subjects and chronic patients. Thus, no (sub-)acute patients were involved. In line with the earlier papers on two-player gaming of acute and subacute stroke patients [60, 104], we performed a study aiming to test two-player modes in acute and subacute stroke patients. In order to evaluate two-player modes and to address the limitations of the earlier studies with therapy devices, we tested in this study the following hypotheses:

(H1) In subacute and acute stroke patients, two-player modes increase, compared to single-player modes, (a) interest/enjoyment, (b) effort/importance self-assessed by the subjects, and (c) effort assessed based on the end-effector velocity profile.

(H2) In subacute and acute stroke patients, competitive modes increase, compared to cooperative modes, (a) interest/enjoyment, (b) effort/importance self-assessed by the subjects and (c) effort assessed based on the end-effector velocity profile. 4.3. Methods 51

(H3) A subject’s personality, age, and level of impairment predict the preferred game mode.

4.3 Methods

4.3.1 Participants Between January 2017 and September 2017, 44 inpatients were recruited for this single-center study in the SRH Gesundheitszentrum Bad Wimpfen (Germany). Subjects had to be at least 18 years old and were required to successfully complete a Fugl-Meyer Assessment (FMA) [54], a Mini Mental State Examination (MMSE) [51], and the initial assessment of the input device, the Armeo Spring. Based on the assessments, subjects were included in the study if the study operators rated the subject’s arm as mildly to severe- moderately impaired, characterized by ≥ 16 of a possible 66 points on the FMA [181], and the subject as not cognitively impaired to mildly cognitively impaired, characterized by ≥ 25 of a possible 30 points on the MMSE. We assessed age, sex, type of stroke, time since stroke, handedness, duration of training done with the exoskeleton device so far, and experience with visualization software in gaming and in worktime. Included subjects were matched into dyads by the therapists in the clinic. The therapists tried to build dyads of similar skill level. A subject was excluded if he or she suffered from serious mental or neurological illness (e.g., schizophrenia, dementia, strong depression), was pregnant or nursing, suffered from skin irritation, skin injury or orthopedic, rheumatologic or further illness/injuries affecting the impaired arm, had signs of osteoporosis, suffered from cyber-sickness, shoulder subluxation or the complex regional pain syndrome, or used a pacemaker or other implanted electronic device. Each subject’s personality was assessed within two days before or after the study session using the 50-item International Personality Item Pool (IPIP) [59] and the Revised Competitiveness Index (9 items) [76]. From the IPIP, the Big Five factor markers (i.e. extraversion, agreeableness, conscientiousness, emotional stability, and intellect/imagination) and from the Revised Competitiveness Index, 52 Chapter 4. Multiplayer Gaming in Robot-Assisted Training competitiveness of the patient were extracted. These markers were used in previous studies to analyze the effects of personality [119, 60, 181, 177].

4.3.2 Ethics The study was approved by the ethics committee of the Landes¨arztekam- mer Baden-W¨urttemberg (Germany) and all subjects gave written informed consent in accordance with the declaration of Helsinki.

4.3.3 Apparatus The two Armeo devices were placed parallel to each other about two meters apart, with both subjects looking in the same direction (Fig- ure 4.2A). The distance between the subject’s head and the screen was 1.2 meters. The individual device conﬁguration was set by a therapist based on a pre-study session. In the single-player modes, the subjects were separated by a room divider so that they could not see the other subject or the other subject’s screen, and verbal interaction between the subjects was prohibited.

4.3.4 Game We implemented four modes of a rehabilitation game that were all variations of the same air hockey game. The game design and game mechanics were based on the air hockey game implemented for our previous study with healthy subjects [119]. The game principle of the air hockey game was based on the classic game of Pong, which was also used in recent studies in neuromuscular rehabilitation [119, 61, 60, 62]. In all four modes of the air hockey game, the puck moved in a predefined game field. The game field is constrained by sidewalls on the left and right side, and by goals on the bottom and top side. The subject controlled a round mallet in front of the bottom goal by moving the entire arm in the exoskeleton. The movement of the mallet was restricted to move only laterally, i.e., a one degree of freedom (DOF) movement. The mapping of the hand position to the mallet position was normalized by the subject’s lateral aROM of the hand. The subject’s task was to score a goal on the top side 4.3. Methods 53

B 3:24

p,vmax

e 3 dtrigger 9.5 GDU : 4

Figure 4.2: Two-player setup. – A: Two impaired subjects playing the air hockey game. – B: Screenshot of the air hockey game in single-player competitive mode, with the goals, mallets, puck, and scoreboard. The display shown is for the single-player competitive mode. 54 Chapter 4. Multiplayer Gaming in Robot-Assisted Training

(offense) and not to receive a goal (defense). The subject blocked incoming pucks by hitting the puck with the mallet. If a puck was successfully blocked, it began moving toward the opponent’s side at an angle dependent on the collision angle of the puck with the round mallet. The dependency of the return direction on the collision angle facilitated offensive strategies for the subjects, trying to pass the mallets of the opposite team. In all four game modes, the opponent of the subject was always one computer-controlled or subject-controlled mallet. The score and the remaining playing time were displayed on the monitor. We designed four game modes: In the single-player competitive mode, the subject controlled the blue bottom mallet while the grey top one was controlled by the computer. In the single-player cooperative mode, the subject controlled the blue bottom mallet and was supported by a grey mallet controlled by a computer. These two mallets moved along the same line and could pass each other without any collision effects. The two mallets played against the computer- controlled grey top mallet. In the two-player competitive mode, the subject controlled the bottom mallet while the top one was controlled by the opponent subject. The subject sitting on the left controlled a blue mallet and the subject sitting on the right controlled a red mallet. In the two-player cooperative mode, the subject controlled a bottom mallet and was supported by a mallet controlled by the other subject. These two mallets moved along the same line and could pass each other without any collision effects. Both subjects played together against the computer-controlled grey top mallet. The puck speed was constant for all modes. The minimum time for the puck to travel from collision with the bottom mallet to collision with the top mallet (= 9.5 Game distance units (GDU)), or vice versa, was 1.1875 s (Figure 4.2B). The movement behavior of the computer-controlled mallet was based on three parameters: movement trigger distance dtrigger, proportional velocity control parame- ter p, and maximum mallet velocity vmax. By default, the computer stayed at its current position. As soon as the puck approached the computer’s side and passed dtrigger = 5.7 GDU (shortly before passing the centerline), the computer started to decrease, i.e., the lateral component of the distance to the puck. The velocity of the mallet 4.3. Methods 55 v = e · p was controlled by p = 10 s−1 and the velocity of the mallet was saturated at vmax = 7.2 GDU/s. The parameters were tuned by pre-study tests with acute stroke patients. For single-player cooperative mode, the parameters of the supporting computer-controlled mallet were designed to perform worse (dtrigger = −1 5.7 GDU, vmax = 5.4 GDU/s, p = 10 s ). These parameters were tuned targeting a balanced game when two supporting computer- controlled mallets played against the computer-controlled mallet. To prevent the game from getting stuck Gaussian distributed noise was added to the emergent angle of the puck. The game was managed as a network game with two personal computers involved. Both computers recorded the corresponding subject’s movements and gameplay by the sample rate of the game. For the hand velocity profile, movements before and after the 5 minutes gameplay and all the 5 seconds breaks when initially starting the gameplay and after each goal scored before relaunch of the game, were ignored.

4.3.5 Procedure The study protocol consists of three phases (Figure 4.3): a testing phase, a training phase, and a closing phase. In the testing phase, the subjects received the personality questionnaire and aROM was assessed on the device. Previously, FMA and MMSE were performed for checking inclusion criteria. In the training phase, every dyad played four rounds, every game mode once, in randomized order. The mapping of the patient’s hand position to the mallet was adjusted to the aROM assessed in the testing phase. Considering learning effects, the randomization was done by Latin Squares [178]: We randomized the order of the four modes once (multiplayer cooperative (MP-coop), single-player cooperative (SP-coop), multiplayer competitive (MP-comp), single-player competitive (SP-comp)) and changed only the mode to start with while keeping the order (e.g., SP-coop, MP-comp, SP-comp, MP-coop). Based on this procedure four different sets of modes were possible. We prepared twenty lots with five lots per set of modes. For the twenty dyads set of modes were pulled without replacing out of the twenty prepared sets. 56 Chapter 4. Multiplayer Gaming in Robot-Assisted Training

Each round started with a one-minute familiarization with the mode. The verbal introduction to the mode was the same for all subjects, including the emphasis in two-player modes that communication and, in the cooperative mode, agreements were allowed. The familiarization was followed by five minutes of gameplay. During the gameplay, the hand velocity profile of each subject was recorded. As an approximation of the exercise intensity, the hand velocity profile’s root-mean-square (RMS) value was extracted as it has been shown that the RMS value approximates energy expenditure of the subject [170, 164]. Furthermore, the result of the game (win, loss, draw) was extracted, and the score difference was calculated by subtracting the number of goals scored by the opponent from the number of goals scored by the subject or the subject’s team. After each game round, subjects rated ten statements from the Intrinsic Motivation Inventory (IMI) – five statements from the interest/enjoyment subscale and five from the effort/importance subscale – on a Likert scale from one to seven [110]. The therapist also rated each statement from the point of view of the patient. Once all four rounds had been played, the subjects answered the closing questionnaire. The closing questionnaire included a ranking of the four rounds played regarding perceived interest/enjoyment (“What game mode did you enjoy the most?”) and perceived effort/importance (“In what game mode did you put the most effort?”).

4.3.6 Analysis We calculated Pearson coefficients from product-moment correlation of score difference with age, FMA score, and MMSE score for each game mode. We analyzed effects on the intrinsic motivation subscales interest/enjoyment and effort/importance rated by subject and therapist, and the RMS value with linear mixed-effects models, using the lme4 pack- age in the R environment [42, 159]. In all models, subjects and therapists, and in two-player modes the dyads, were specified as a random factor to control for their associated intraclass correlation (i.e., random intercept models) [129]. We assigned rank points to each mode based on the two rankings from the closing questionnaires. A mode received four points for each 4.3. Methods 57

Round 1 Round 2 Round 3 Round 4 Testing F G Q F G Q F G Q F G Q Closing

- FMA - Game score - Mode ranking - MMSE Game play - RMS of end e ector - Personality One out of four modes: velocity questionnaire Singleplayer comp/coop - IMI (self-rated) - aROM Multiplayer comp/coop - IMI (therapist-rated) Questionnaire Familiarization 1 min 5 min

Figure 4.3: Study procedure. Every patient underwent ﬁrst a testing phase, secondly a training phase consisting of four rounds of familiarization, gameplay, and questionnaire, and thirdly a closing questionnaire. Analysis

first rank assigned by a subject, three points for each second rank, two points for each third rank, and one point for each fourth rank. The sum of rank points and the number of first ranks were reported within modes for both rankings: “What game mode did you enjoy the most?” and “In what game mode did you put the most effort?” No statistical test was applied on the sum of rank points. From the question “What game mode did you enjoy the most?” we grouped the participants into four groups according to four possible answers (i.e. the four game modes). We analyzed effects on the intrinsic motivation subscales interest/enjoyment and effort/importance within the four groups with linear mixed-effects models. We used the same models as in the general analysis of the intrinsic motivation subscales (see above). To determine if subjects’ favorite game mode can be predicted from personality scores, a stepwise forward/backward linear discriminant classifier was trained with nine possible input variables (age, FMA score, MMSE score and the six personality scores) and three output variables (preferred game mode, preferred game mode was single- player/two-player, preferred game mode was competition/cooperation). The classifier was validated with leave-one-out cross-validation, where all but one subject is used to train the classifier and one subject is then used to test its performance[57]. The cross-validation was performed as many times as subjects included, with each subject 58 Chapter 4. Multiplayer Gaming in Robot-Assisted Training serving as the test subject once. The stepwise algorithm includes or excludes variables if the output improved at least by 0.025 (i.e. one correct classified subject). The threshold for significance was set at p=0.05 in all tests.

4.4 Results

4.4.1 Participants

Four of the 44 recruited subjects had to be excluded: one and the matched subject due to an incomplete FMA, and one and the corresponding matched subject due to absence. Both dyads received no training and their data was not included in the analysis. The 23 men and 17 women who participated in the study were either suffering from an ischemic stroke (33) or from a hemorrhagic stroke (7). The subjects were between 24 and 91 years old, with a mean of 69.3 years and standard deviation of 12.5 years. In the FMA, the subjects scored between 27 and 66 points, with a mean of 58.9 points and a standard deviation of 9.7 points, and in the MMSE, the subjects scored between 25 and 30 points, with a mean of 27.6 points and a standard deviation of 1.6 points. Twelve subjects had some work experience where good perception of visual information was required and five subjects played computer games more than once a week. One subject was left-handed, all others right-handed, as self-reported. Twenty-five had an affected left arm while 15 had an affected right arm. In sixteen out of 40 subjects, the affected arm was the dominant one. Stroke onset was at least fifteen days before study session day, and was less than 70 days for 38 of 40 subjects. One subject had 156 days and one subject had 336 days since stroke onset. All subjects had already received between one and ten hours of Armeo training prior to the study, with a mean of 2.4 hours up to the day the study was performed. The dyad matching of the subjects led to 13 dyads with same sex (i.e. eight man-man and five woman-woman dyads) and 7 dyads with opposite sex. 4.4. Results 59

4.4.2 Game results On average, the cooperative modes resulted in a higher game score difference (4.4 in single-player and 4.3 in two-player) compared to single-player competitive (0.2) (Table 4.1). In single-player, game score difference correlated with age and in single-player cooperative, FMA score correlated with MMSE score. For the competitive two- player mode, the game score difference correlated with MMSE score difference and when cooperative the game score difference correlated with the sum of age and sum of MMSE score of the two subjects.

4.4.3 Self-rated intrinsic motivation We found a significant effect in the self-rated interest/enjoyment subscale for competition, χ2(1,N = 40) = 4.04, p = 0.044 (Table 4.2, Figure 4.4A). No effects were found in the self-rated effort/importance subscale.

4.4.4 Therapist-rated intrinsic motivation Therapist A operated the study on study day 1 (number of subjects: 4), day 6 (6 subjects) and day 8 (6 subjects), therapist B on study day 2 (6 subjects), day 4 (6 subjects) and day 7 (4 subjects), and therapist C on study day 3 (6 subjects) and day 5 (2 subjects). We found a signiﬁcant two-player eﬀect in the therapist rated subscales interest/enjoyment (Table 4.2, Figure 4.4B).

4.4.5 Exercise intensity (actual effort) We found a significant effect in the exercise intensity for competition, χ2(1,N = 40) = 19.64, p < 0.001, and a significant interaction, χ2(1,N = 40) = 5.03, p < 0.025 (Table 4.2, Figure 4.4C).

4.4.6 Mode ranking and prediction Single-player competitive was most frequently ranked ﬁrst place and received the most ranking points for both questions “What game mode did you enjoy the most?” and “In what game mode did you put the most eﬀort?” (Table 4.3). Index of the game mode did correlate 60 Chapter 4. Multiplayer Gaming in Robot-Assisted Training

30 30 28.8 28.4 28.1 27.3 27.5 27.0 27.2 27.1

20 20 self-rated e ort/importance interest/enjoyment 10 10 p=0.044

comp coop comp coop comp coop comp coop single-player two-player single-player two-player B

30 30 30.2 29.5 28.2 28.9 26.2 25.6 24.8 24.6

20 20 therapist-rated e ort/importance interest/enjoyment 10 p=0.046 10

comp coop comp coop comp coop comp coop single-player two-player single-player two-player C 0.6

0.4

0.32 0.33 0.29

RMS [m/s] 0.26 0.2

0.0 p<0.001 comp coop comp coop single-player two-player

Figure 4.4: Intrinsic motivation scores self-rated by the subjects (A) and rated by the therapist (B), and exercise intensity from the RMS value of the subjects’ velocity profiles (C) in all four modes. The mean values are indicated with a cross with the corresponding value attached. Only significant effects are indicated. 4.4. Results 61

Table 4.1: Game scores of all modes. In single-player competitive, single-player cooperative and two-player cooperative, the score differences were always calculated as subject(s) score minus the computer- controlled opponents score. In two-player competitive the differences were calculated as left playing subject score minus right playing subject score. game mode single-player two-player comp. coop. comp. coop. # of subjects 26 29 20 32 won # of draws 1 0 0 0 mean game 0.2 4.4 0.0 4.3 score difference Correlations between game score difference and f(age), f(FMA score), f(MMSE score) degrees of 38 38 18 18 freedom (df) f(age) subject’s subject’s subjects’ subjects’ age age age diff. age sum −0.58, r(df), p −0.45, 0.003 −0.29, 0.201 −0.61, 0.005 < 0.001 f(FMA score) subject’s subject’s subjects’ subjects’ FMA score FMA score FMA score FMA score diff. sum r(df), p 0.32, 0.045 0.37, 0.019 −0.05, 0.839 0.39, 0.086 f(MMSE subject’s subject’s subjects’ subjects’ score) MMSE MMSE MMSE MMSE score score score diff. score sum r(df), p 0.27, 0.087 0.41, 0.008 0.56, 0.011 0.45, 0.046

with ranking points of “What game mode did you enjoy the most?” (r(158) = −0.17, p = 0.037) and with ranking points of “In what game mode did you put the most effort?” (r(158) = −0.36, p < 0.001). The answer of the question “What game mode did you enjoy 62 Chapter 4. Multiplayer Gaming in Robot-Assisted Training the most?” was assumed to indicate the favored mode of each subject. For comparison of the four groups of subjects, each group favoring another game mode, we averaged the characteristics of the subjects belonging to each group and analyzed the answers of the Intrinsic Motivation Inventory for each group individually. The stepwise algorithm building up the linear discriminant classifier selected extraversion and emotional stability for best leave-one-out cross-validation correctly identifying preferred game mode for 15 out of 40 subjects (37.5%). When only identifying if the most preferred game mode was single-player or two-player 26 of 40 subjects (65%) were correctly identified with emotional stability as the only selected input variable. When only identifying if the most preferred game mode was a competitive or cooperative mode 23 of 40 (57.5%) subjects were correctly identified with agreeableness as the only selected input variable.

4.5 Discussion

4.5.1 (H1) Number of players (single- or two-player) We expected two-player modes to be preferred compared to single- player modes as a ranking of modes by subjects revealed in previous two-player studies [119, 79, 61, 104]. In our study, the two-player modes were preferred by 16 subjects whereas the other 24 subjects preferred the single-player modes. In contrary, according to the therapist’s ratings, the subjects enjoyed the two-player modes more than the single-player modes. However, the subjects rated all modes well in both IMI subscales in that the IMI subscales did not demonstrate that one number of subjects (single- or two-player) was superior to the other. More importantly, we can state that both single- and two- player modes were intrinsically motivating, i.e. enjoyed by and being of importance for the subjects. Based on the comments of the healthy subjects in a previous study, we expected the predictability of the computer-controlled mallet’s behavior to be boring for the subjects. Surprisingly, the subjects of the present study did not rate the computer-controlled mallet’s behavior negatively. Representative of comments made by patients 4.5. Discussion 63 about to the single-player competitive mode, one subject commented after about half of the game time: “Now I know how I can beat the opponent”. This expertise seems satisfying to the subjects and provides confidence. The predictable (and not human-like) behavior of the opponent in the single-player mode might have also positively influenced the game experience in single-player mode for acute and subacute stroke patients. That human-like behavior of opponents increases flow, enjoyment and presence was shown with healthy subjects [176]. For proper comparison of the single- and two-player modes, it is recommended designing computer-controlled players with more human-like and less predictable game behavior. According to motivation theories, verbal interaction that contributes toward feelings of competence during action can enhance intrinsic motivation for that action [39]. However, while in a previous study between healthy players extensive dialogs accompanied the gameplay [119], subjects in this study rarely commented on the gameplay. Although we clearly explained before each two-player mode that verbal interaction is allowed, most dyads did not talk, for instance to discuss strategies in the cooperative mode or verbally comment on gameplay events in the competitive mode. Previous studies incorporated friends [119, 60], while here the subjects were not familiar with each other when the first game mode started, which might be an explanation for the lack of interaction. In gameplay with healthy relatives or friends, there might be more verbal interaction between the subjects. To provide gameplay against healthy co-players, we see advantages in the use of therapy robots. Actuated therapy robots provide haptic support or haptic resistance changing the individual game condition [19]. Such individual adaptation of the game condition may facilitate balanced two-player gaming between unmatched dyads or gameplay between patients and healthy players such as therapists, relatives or friends of the patients.

4.5.2 (H2) Interaction type (competitive or cooperative) We expected that the subjects’ effort would differ between the cooperative modes and the competitive modes. Indeed, the RMS value was found to be higher in competitive modes compared to coopera- 64 Chapter 4. Multiplayer Gaming in Robot-Assisted Training tive modes. However, in the competitive modes every subject had to execute all defensive tasks, whereas cooperative modes allowed tasks to be shared between subjects. When comparing exercise intensity between competition and cooperation, we found an interaction effect between competition/cooperation and single/two-player. Specifically, the exercise intensity difference between competition and cooperation was larger in two-player modes than in single-player modes (Figure 4.4A). It appears that, in the cooperative modes, subjects relied on their co-player more if that co-player was human rather than computer-controlled, resulting in a greater decrease of exercise intensity when the cooperative partner was human. This drop in exercise intensity might therefore be reduced by reducing the reliance on the other player by, for example, making each player’s successes and failures explicit by assigning each puck hit or miss to a specific player and displaying this information on the screen. Such an explicit performance comparison has been previously shown to lead to increases in motivation and effort outside rehabilitation [109, 48]. The effect of competition as opposed to cooperation for all measures might also be limited by the fact, that the general game idea remains “competitive” even in cooperative modes. While cooperating with another human subject or computer-controlled player in cooperative modes, a competition against the opposing computer-controlled player remained. As long as a game provides comparison of individual performances, the game retains a competitive nature. This includes comparison with the own performance, as discussed in a study testing a therapy game for robot-assisted therapy for lower extremities [185]. The cooperative manner of a game can be emphasized by avoiding antagonistic subtasks [78]. A game design that is based on Pong but features no opposing player, is Breakout. Such game designs are favored to facilitate the advantages of cooperative gaming. Different two-player game concepts based on gamified activities of daily living, such as a cooperative cooking game, have recently been tested [63].

4.5.3 (H3) Correlations and predictability Younger age, higher FMA score and higher MMSE score correlate positively with game performance, i.e., score difference, as indicated 4.6. Conclusion 65 in all significant cases. Surprisingly, we report a tendency of negative correlation in the two-player competitive mode when correlating FMA score difference with score difference. Since range of motion of the subject is already considered in the mapping of the subject’s movement, parts of the functional differences between subjects are already eliminated. A prediction of a preferred game mode based on the suggested set of parameters (personality, age, FMA score, MMSE score) has to be discarded. Our evaluation showed that the predictability is moderate and therefore insufficient for clinical application.

4.6 Conclusion

Two-player games can be used as a new component thereby increasing the diversity in neuromuscular therapy sessions provided. Sixteen out of 40 patients preferred a two-player mode and over all patients, no significant difference regarding intrinsic motivation could be identified. Therefore, we recommend that two-player games in competitive and cooperative interaction forms to be considered for therapy design by therapists and clinics for those patients who favor two-player games over single-player games. It remains difficult to predict the preference of the patient by using personality questionnaires, clinical assessments or demographic data but the preference may be identified by asking the patients during the therapy process. Competitive game modes are preferred by acute and subacute stroke patients and lead to a more intensive gameplay than cooperative game modes. The design of the competitive modes was simple to understand and required non-stop awareness and physical effort from the patient. In contrast, each subject can solve each subtask in the cooperative modes, i.e. defending the approaching puck, individually and independently from the co-players actions. This redundancy, i.e., both subjects can solve the subtask, creates unclear roles if the subjects do not agree on a strategy beforehand. Therefore, other forms of game designs and of two-player interaction than the here presented cooperative mode, is suggested for non-competitive two-player therapy games. 66 Chapter 4. Multiplayer Gaming in Robot-Assisted Training - t, p 0.001 -0.8, 2.24, 0.025 0.344 -4.89, 17.84, < ) e ( RMS [m/s] ˆ µ 0.329 0.041 -0.064 -0.011 (0.018) (0.018) (0.013) (0.013) - t, p 1.45, -1.01 0.980 0.146 -2.28, ,0.076 21.62, ) e effort/ ( 1.38 importance ˆ -0.68 -1.95 µ 30.21 (0.67) (0.85) (1.40) (0.95) - t, p 0.35, -0.91 0.346 0.724 -1.80, 25.16, ,0.046 therapist-rated ) e intrinsic motivation ( interest/ enjoyment 0.33 ˆ -0.60 -1.35 µ 26.01 (1.03) (0.93) (0.66) (0.75) - t, p 0.43, 0.801 0.714 0.449 -0.28, -0.36, 29.26, ) e effort/ ( 0.28 importance ˆ -0.18 -0.33 µ 27.25 (0.93) (0.64) (0.64) (0.90) - self-rated t, p 0.78, 0.690 0.044 0.123 -0.37, -1.50, 36.24, ) e intrinsic motivation ( interest/ enjoyment 0.48 1.30 ˆ -0.23 µ 28.35 (0.78) (0.61) (0.87) (0.61) → → intercept two-player interaction (two-player cooperation competition single-player competitive) Table 4.2: Linear mixed-effectsintrinsic models motivation, analysis and of the the RMS self-rated value intrinsic of motivation, the the therapist-rated velocity profile approximating the exercise intensity. 4.6. Conclusion 67

Table 4.3: Mode Ranking. Game mode Single-player Two-player Comp. Coop. Comp. Coop. Rank “What game mode did you enjoy the most?” 1 15 9 7 9 2 9 12 7 12 3 6 11 19 4 4 10 8 7 15 Weighted sum 109 102 94 95 Rank “In what game mode did you put the most eﬀort?” 1 14 7 9 10 2 9 14 7 10 3 7 13 15 5 4 10 6 9 15 Weighted sum 107 102 96 95 68 Chapter 4. Multiplayer Gaming in Robot-Assisted Training 69

Chapter 5

Balanced Multiplayer Gaming in Robot-Assisted Training

Foreword and Overview

The third research question was whether multiplayer games therapy after stroke can account for diﬀerences in skill level by using the haptic modality provided by therapy robots (Figure 5.1). We used the same game that we tested with the (sub-)acute stroke patients and integrated a patient-tailoring adaptation of the haptics.

5.1 Abstract

Multiplayer environments are thought to increase the training intensity in robot-aided rehabilitation therapy after stroke. We developed a haptic-based environment to investigate the dynamics of two-player training performing time-constrained reaching movements using the ARMin rehabilitation robot. We implemented a challenge level adaptation algorithm that controlled a virtual damping coeﬃ- cient to reach a desired success rate. We previously designed and 70 Chapter 5. Balanced Multiplayer Gaming Measured Experience Assessment Behavior Experience/ Patient Co-Player Interaction Environment Human-Human act,2 act,2 act,1 act,1 x F F x H H Interface 1 Data Interface 2 Data Interface 1 Interface 2 Condition 1 Condition 2 Condition ref,1 H ref,2 H meas,1 meas,2 x x x x G G G G Condition Condition Controller Controller Integrated Integrated Complexity 1 Complexity Complexity 2 Complexity Interpretation Interpretation Game Performance 1 Game Performance Game Performance 2 Game Performance Game Game Game Complexity 1 Complexity 2 Complexity C

B A Patient Data Patient Patient Pro!le Patient Co-Player Data Co-Player Co-Player Pro!le Co-Player Game-/ Selection Therapist Therapist Complexity Game mode-/ Desired Behavior Experience/ Figure 5.1: Concept chart with highlighted (red) elements for robot-assisted patient-tailored multiplayer 5.2. Introduction 71 tested a performance-based adaptation of a virtual damping coefficient in a single-player application (Appendix A). We tested the algorithm’s effectiveness in regulating the success rate during game play in a simulation with computer-controlled players, in a feasibility study with six unimpaired players, and in a single session with one stroke patient. The algorithm demonstrated its capacity to adjust the damping coefficient to reach three levels of success rate (low [50%], moderate [70%], and high [90%]) during single-player and multiplayer training. For the patient – tested in single-player mode at the moderate success rate only – the algorithm showed also promising behavior. Results of the feasibility study showed that to increase the player’s willingness to play at a more challenging task condition, the effect of the challenge level adaptation – regardless of being played in single player or multiplayer mode – might be more important than the provision of multiplayer setting alone. Furthermore, the multiplayer setting tends to be a motivating and encouraging therapy component. Based on these results we will optimize and expand the multiplayer training platform and further investigate multiplayer settings in stroke therapy.

5.2 Introduction

The greatest benefits of therapy are seen in trainings with high level of intensity when patients train with high levels of mental and physical effort. To promote intensity during arm rehabilitation, researchers have developed multiplayer environments where two players can compete, or cooperate, with one another [119]. In these environments, a patient may interact with another patient, a therapist, a family member or a friend [62, 56]. This social context is thought to increase patients’ motivation and levels of mental and physical effort. However, when more than one person is involved in training, for example during a competitive game, differences in skill level can affect the experience of success and their motivation to train. To keep motivation high, the level of game difficulty can be adjusted to match the skill level. This approach has been previously discussed in the context of the model of flow [38]. Adjustments to the game’s difficulty can include the core mechanics (e.g. the speed at which a 72 Chapter 5. Balanced Multiplayer Gaming ball travels in a game of tennis), the conditions in which a player performs the task (e.g. restricting the movement of a player), or a combination of the two. These adjustments can be made by a therapist supervising the training, by the players, or by algorithms embedded in the game. The final option is especially attractive in rehabilitation settings because it does not require additional input from users, can be performed systematically, and can target specific training goals [106, 33]. In these multiplayer environments, one aspect that requires special attention is how can these games be played by highly impaired individuals, i.e., those who could not perform the task withouth external assistance. Rehabilitation games played with passive devices already provide different modulators that can be adjusted to account for different skill levels of impaired players [62]. However, for this population, robotic devices offer unique design features that allow a given motor task to be tailored to the players’ motor abilities [36]. In such setups, haptic force fields can be used to adjust the challenge level of a training task [108]. These force fields can be designed to support, or in some cases hinder, the player – thus changing the challenge level. In a multiplayer setting, this adaptation of the challenge level with haptic force fields can be applied individually for the different players. The focus of our work is on the use of robotic devices to promote social interaction in rehabilitation training. We thus developed a multiplayer robot-based environment, where players of various skill levels can interact with each other using different input devices. This environment is based on a conceptual framework, derived from the flow model, where we seek to maintain high levels of motivation and engagement in training by balancing the differences in skill level between players [38]. Our focus is on time-constrained reaching movements because reaching is an essential component for functional upper-limb tasks alongside grasping, moving/manipulating and releasing [96]. The reaching movements are constrained in time by a defined time window when the given target position has to be reached. We tested how an algorithm that adjusts the conditions of the task affects performance, effort and motivation during training. We first tested with simulated computer-players, then with a cohort of unim- 5.3. Methods 73 paired participants, and finally with a single stroke patient. We hypothesized that by adjusting the damping level of the force field we could control a player’s success rate. We further hypothesized that using this algorithm in a multiplayer setting would lead players to exhibit higher levels of motivation and effort.

5.3 Methods 5.3.1 Framework The conceptual framework is largely derived from the flow model. This model relates the ideas of skill level and challenge level to predict the mental state of a person engaged in a given task [38]. When considering a multiplayer task, for example a two-player competitive game, the challenge level of the task is related to the skill level of the opponent. Since the skill level between players can vary greatly, so can the challenge level experienced by each player. In cases where the difference in skill level is large, the flow model predicts this and the game will result in sub-optimal mental states for the players. That, of course, is undesirable for rehabilitation training where we want patients to be highly motivated and engaged during training. In our framework, we focus on the players’ abilities to perform the task – measured as success rate – and define it as being opposed to the challenge level (i.e. higher success means lower challenge; Figure 5.2). We then take a task condition – the damping coefficient µ in our current application – and adjust it to control the players’ success rate. By controlling the success rate, we assume that we can balance the differences in skill levels of players. For training under balanced conditions, we predict the players to be closer to a flow-like state.

5.3.2 Experimental setup Air hockey game The air hockey game was set up similarly to a previous study [119]. The side walls constrained the movement of the puck, but not of the mallets. At the beginning of each match – and after each goal – the puck appeared in the center of the ﬁeld. Both players then moved 74 Chapter 5. Balanced Multiplayer Gaming

Figure 5.2: Balancing the challenge level in robot assisted gaming. If we adjust the task conditions separately for each player, we can increase the damping for player A and decrease it for player B until they achieve comparable success rates. their mallets to the center of the field and a two-second countdown began. During the countdown, an arrow indicating the start direction of the puck appeared. After the countdown, the puck moved in the direction indicated by the arrow and the match started. The score of both players was displayed on the right side of the playing field. This version of the air hockey game can be played either in single- player mode or in multiplayer mode. The single-player mode features a computer opponent whose challenge level is adapted to the level of the player. The multiplayer mode features a human player using a computer mouse as the input device. The air hockey game can be broken down into two main subtasks: defense and offense. In this study, we focused on the defense task. This task requires the player to move the mallet left and right in order to reach the approaching puck in time. Success in the defense task is to hit the puck forward; failure is the puck passing the mallet and reaching the player’s goal zone. The success rate in this task is defined as the number of successful hits divided by the number of 5.3. Methods 75 total trials.

Rehabilitation robot In both single-player and multiplayer mode, the input device for the subject is the ARMin arm rehabilitation robot [66]. The elbow joint movement of the player was assessed by the position sensors in the robot and mapped to the player’s mallet (Figure 5.3).

Figure 5.3: Multiplayer setup. A: A subject uses the ARMin robot to play against a subject using the computer’s mouse. The monitor displays the computer game to both players. B: The air hockey game environment. The puck moves on a ﬁeld surrounded by four walls and two mallets defend two opposite goals.

We used a range of motion for the elbow between 30◦ and 90◦ and mapped from −12.3 GDU to 12.3 GDU. All other joints were initially fixed at a fully stretched pronated position: 90◦ lateral shoulder abduction, 90◦ horizontal shoulder adduction, 0◦ upper arm rotation, 90◦ lower arm pronation, 0◦ wrist extension and closed hand module. To address the challenge level of the game, we virtually adjusted the damping coefficient at the elbow joint of the robot. The lower limit of the damping coefficient was set to 0 Nms/◦. The upper limit of the 76 Chapter 5. Balanced Multiplayer Gaming damping was defined by the maximum torque that can be safely created by the motor and limited through voltage saturation at 59.5 Nm in elbow flexion and extension.

Challenge level adaptation algorithm

The challenge level of the game can be modulated by parameters of the game core mechanics and the players’ interface. However, changes in game core mechanics, such as puck speed, affect the difficulty level for both players generally. In contrast, changes at the player specific interface, such as damping forces, affect the game for each player individually. We used an algorithm that adjusted the challenge level of the task by adapting the damping coefficient, based on the performance of the player [150]. This algorithm was previously tested in a single-player rehabilitation game where a timing component of the game was adapted in order to regulate the challenge level [157]. However, its applicability in a multiplayer setting has not yet been tested. The algorithm adapted the challenge level of the task by changing the damping coefficient µ as shown in Equation 5.1. The damping affects the maximum velocity that a player can reach and this, in turn, affects the ability of the player using the damped device to successfully defend his or her goal.

◦ µi = βi · 0.01 Nms/ , β1 = 60 (5.1)

( βi + δ if success βi+1 = (5.2) βi − αδ if failure

Coefficients δ and α can be selected dependent on desired adaptation behavior. Coefficient δ defines the step size of the adaptation and coefficient α is defined by the desired success rates p∞ that is reached. The relation between α and p∞ is shown in Equation 5.3. p α = ∞ (5.3) 1 − p∞ 5.3. Methods 77

5.3.3 Artiﬁcial intelligence of the computer-player In our previous study (Chapter 4), the movements of the computer- controlled mallet were governed by a P-type controller. We updated the movement behavior of the computer-controlled mallet to follow a human-like reaching behavior. We believe this would allow us to better compare the results between the single-player and multiplayer conditions. We assessed the reaching behavior of the participants in the ARMin robot in a series of pilot studies. Based on these assessments we modeled the computer’s behavior using minimum jerk reaching movements [49]. We combined this minimum jerk model with normal distributions in reaction time tr, end-of-reach time te, and end-of-reach position pe to deﬁne the behavior of the computer opponent (Figure 5.4).

Figure 5.4: Assessment based approach to model reaching behavior for computer-players. A: Schematic figure of assessed reaching movement of the players. After the puck starts moving the player starts moving with a delay of reaction time tr. The puck can be reached within a restricted time with target position pe and target time te. B: The distributions for ∆tr, ∆pe and ∆te were assessed with three different angles Θ = [25◦, 35◦, 45◦] in both flexion and extension direction. C: The puck and mallet movements were presented in the same game environment. 78 Chapter 5. Balanced Multiplayer Gaming

5.3.4 Study designs Multiplayer computer simulation We tested the adaptive algorithm in a multiplayer setting by simulating human-like players of different skill levels. In this simulation, the algorithm adapted the challenge level of the task by changing the coefficient of a virtual damping following Equation 5.1. We simulated three players of different skill levels by defining three maximum virtual forces, Fv,max: 20, 60, and 100 virtual force units. These simulated players kept their maximum virtual forces constant throughout the game. The players competed against an opponent with an Fv,max of 100 virtual force units whose damping coefficient remained constant. We expected stronger players to reach higher damping coefficients than weaker ones for the same success rate, which we set to 50%, 70%, and 90%. Coefficient δ was set to one for all simulations. Each simulated match lasted 16 minutes.

Feasibility study with healthy players We tested the effect of the haptic field dynamically adapted to achieve specific success rates in a feasibility study with six healthy subjects (4 females; 25 to 31 years; right handed). All participants played with the dominant arm. Subjects were eligible if they met inclusion criteria:

• Minimum age of 18 years • Bodyweight under 120 kg • No serious medical or psychiatric disorder as reported by the participant • No orthopaedic, rheumatic or other disease restricting movements of the paretic arm • No shoulder subluxation of more than two ﬁngers width • No pacemaker or other implanted electronic device Subjects trained on two separate days. Both days started with an introduction phase of about four minutes. The subjects were given instructions and seated in the ARMin robot according to a standardized protocol [66]. They were then allowed to become familiar 5.3. Methods 79 with the ARMin and the game environment during a test game. As part of the introduction, the subjects were instructed to focus on the defensive part of the game.

On the first day of testing, the game was played in either single- player or multiplayer mode (Figure 5.5). On the second day, the game was played in the other mode. The order of modes was randomized. Each training mode was played in three different matches, each with a different desired success rate: low=50%, medium=70%, or high=90%. The order of the required success rate was randomized. Coefficient δ was set to one throughout all matches. Each match was played for eight minutes which was determined being a meaningful duration to play where the players did not express signs of physical fatigue. From each match, the temporal evolution of the damping coefficient was logged. The final damping coefficient provides a sur- rogate measure of the effort of patient in relation to his capability to apply forces. After each match, the subject rested for four minutes and responded to a motivation questionnaire. The motivation was measured with the Intrinsic Motivation Inventory (IMI) [110]. The IMI has previously been used with virtual environments for motor rehabilitation [36, 119]. We used four sub-scales of the IMI in this study: interest/enjoyment, effort/importance, pressure/tension, and perceived competence. For each sub-scale, we selected three statements including one inverse statement: 80 Chapter 5. Balanced Multiplayer Gaming

Interest/enjoyment • I enjoyed playing the air hockey game very much • Playing the air hockey game was fun to do. • Playing the air hockey game did not hold my attention at all. (Inverse) Effort/importance • I put a lot of effort into this. • I tried very hard on playing the air hockey game. • I didn’t put much energy into playing the air hockey game. (Inverse) Pressure/tension • I felt very tense while playing the air hockey game • I felt pressured while playing the air hockey game. • I was very relaxed in playing the air hockey game. (Inverse) Perceived competence • I think I am pretty good at playing the air hockey game. • I am satisfied with my performance at playing the air hockey game. • Playing the air hockey game, I couldn’t do very well. (Inverse)

Subjects rated how true each statement was on a 7-point scale, with 1 indicating ”not at all true”, and 7 indicating ”very true”.

Game players ARMin Mouse Computer

study day day 1 vs day 2 vs game mode, e.g. multiplayer singleplayer desired success rate, e.g. medlowhigh high med low Training phase IRGGGRRRRRI GGG duration [min] 48 4 8 4 8 44 8 4 8 4 8 4

Figure 5.5: Study protocol. The subjects played one day in single player mode and one day in multiplayer mode. Three different actions were performed: The subjects were instructed (I, 4 min), they played the game (G, 8 min) in defined mode and desired success rate, and they filled out the questionnaire while resting (R, 4 min) immediately after each play. 5.4. Results 81

Case study with one stroke patient

A 50-year-old female, four years after ischemic stroke in the left middle cerebral artery, and with moderate impairment of the right arm (Chedoke McMaster Stroke Assessment 3/7 for the arm and 4/7 for the hand [64]) agreed to participate and provided informed consent. She played the game in single-player mode during four minutes. We assessed the evolution of the damping coeﬃcient during the match. This study as well as the study with the healthy players was approved by the responsible institutions (KEK-ZH-Nr. 2015-0013, Zurich, Switzerland, and clinicaltrials.gov NCT02720341).

5.3.5 Data analysis

To evaluate the performance of the algorithm – with both simulation and study data – we used a moving window. The size of this moving window was set to 50 trials in order to reduce noise in the success rate data. The moving window was used to calculate the success rate and the average damping coeﬃcient experienced by the player during that window of trials. To evaluate the IMI data, the mean value of each sub-scale, for a given desired success rate in the multiplayer mode, was subtracted by the corresponding value in the single-player mode. The possible range for each score was therefore -6 to 6.

5.4 Results

5.4.1 Computer simulation

The algorithm adapted the damping coefficient for all three simulated players to reach close to the target success rates of 50%, 70% and 90% (Figure 5.6, Figure 5.7). All simulated players shifted away from the initial damping coefficients of 0.6 Nms/◦ and 1.5 Nms/◦ towards the desired success rates. Stronger simulated players achieved higher damping coefficients than weaker players. 82 Chapter 5. Balanced Multiplayer Gaming

Start End A C 100 1.8

1.2 90

0.6 Damping 80 0 coefficient (Nms/°) coefficient 100 200 300 400 500 Trial number 70 100

B 60 Success Rate (%)

70 50 Strong vs Moderate vs Weak vs

Success Rate (%) 40 40 100 200 300 400 500 0 0.6 1.2 1.8 Trial number (windowed) Damping Coefficient (Nms/°)

Figure 5.6: Dynamics of the challenge level adaptation in simulations. A-C: Exemplifying evolution of three simulated players’ data presented as (A) the damping coefficient over trial number, (B) the success rate over windowed trial number and (C) the success rate over the damping coefficient: The damping coefficient and the success rate evolve from that initial state to a final state targeting a desired success rate of 70%. We considered a window size of 50 trials.

5.4.2 Feasibility study For healthy players of different strengths, the damping coefficient was changed from the initial value of 0.60 Nms/◦. All healthy players shifted away from their initial success rate towards the desired success rate. Example data shows how the damping coefficient changes over time and how the players approach the desired success rate (Fig- ure 5.8, Figure 5.9).

5.4.3 Effects on motivation The difference of multiplayer and single player mode in final damping coefficient, for the total IMI score and for all sub-scales was tested 5.5. Discussion 83

90 weakmod.strong

V0 = 0.6 (Nms/°)

Final success rate Final (%) V0 = 1.5 (Nms/°) 50

50 70 90 Desired success rate (%)

Figure 5.7: Final success rate for the simulated matches. The players were simulated using two different initial damping coefficients of 0.6 Nms/◦(5) and 1.5 Nms/◦(4) targeting three different desired success rates of 50%, 70% and 90%, respectively.

(Figure 5.10). A one sided paired-samples t-test was conducted for the final damping coefficient, total IMI score and all sub-scales to compare the differences of multiplayer and single player to no difference after correcting for multiple tests (p∗=0.05/15) (Table 5.1).

5.5 Discussion

As a proof of concept, the majority of the human subjects in the study approached the desired success rates. Neither the challenge- adapted computer-player nor the multiplayer dynamics of playing against the human opponent playing with the mouse aﬀected the result. One particularly strong human subject reached the force limits of the robotic device when targeting the moderate and low success rates. The damping could not increase more since the motor torques 84 Chapter 5. Balanced Multiplayer Gaming

A C 100 Strong vs 1.2 Moderate vs Weak vs 90 0.6

Damping 80 0 coefficient (Nms/°) coefficient 50 100 150 Trial number 70 B 100

Success Rate (%) 60

70 50 Start End

Success Rate (%) 40 40 50 100 150 0 0.6 1.2 Trial number (windowed) Damping coefficient (Nms/°)

Figure 5.8: Dynamics of the challenge level adaptation in human players. Exemplifying evolution of three healthy players’ data presented as (A) the damping coefficient over trial number, (B) the damping coefficient over windowed trial number and (C) the success rate over the damping coefficient: We considered a window size of 50 trials. were limited due to safety reasons. This can be solved by the use of an additional challenge level modulator related to the interface device, e.g., change in joint angle mapping, being applied whenever reaching the limits of the robot. In the other extreme, at the desired high success rate, some subjects reached a ceiling in their success rate when approaching damping coefficients close to 0 Nms/◦. This could be the result of the lower viscosity forcing the player to control for higher accelerations of the reaching movements. Beyond the dynamics of the robot, another explanation could be that the subjects had to adapt to changes in damping that were better perceived at lower damping coefficients due to the fix step size δ that causes changes of a higher factor at this damping level. In particular, in the case of failure at high desired success rate the factor of change can be high (e.g. a change in damping coefficient of 0.09 Nms/◦, from 0.11 Nms/◦ to 0.02 Nms/◦. Thus, to 5.5. Discussion 85

D 90

Player vs Computer 50 Final success rate Final (%) Player vs Player Computer vs Player Patient vs Computer 50 70 90 Desired success rate (%)

Figure 5.9: Final success rate for all matches of all human players and computer-players. achieve higher success rates – and especially for patients whose inher- ent skill level may be lower – we suggest to migrate from the damping based challenge level modulation towards a supportive control challenge level modulation. Such supportive control can be based on an adaptive impedance controller along a minimum jerk trajectory as it implemented for the computer-player already. By visual inspection, a steady state at the desired success rate was not achieved for low and high desired success rate due to the small step size of the algorithm. This progress can be boosted by setting the initial challenge level based on the desired success rate and the force of the subject (e.g. based on a pre-evaluation of force). The step size δ could also be dynamically adapted to be higher at the start of the game and then decreased during play time, to enable fast adaptation at the beginning. There were no significant differences in the final damping coefficient reached when comparing multiplayer to single-player mode. How- ever, the results show the tendency that for low desired success rates 86 Chapter 5. Balanced Multiplayer Gaming

Figure 5.10: Difference in final damping coefficient and IMI scores of multiplayer and single-player mode. For each desired success rate (50%, 70% and 90%) the final damping coefficient and the mean scores of the total IMI and the IMI sub-scales are presented. No significant differences measured.

subjects approached a higher final damping coefficient in the single- player mode. That is contradictory to our hypothesis whereby we expected that higher motivation from social interaction would lead to higher physical effort exerted during training. Considering the limitation of our study that in the multiplayer game the challenge level was only adapted for the subject but not for the opponent, this emphasizes the importance of adapting the challenge level on both players in a multiplayer game. A further limitation is the assumption that the task can be reduced to a defensive task.

Regarding the subjective reports of motivation, effort, and competence (IMI questionnaire), we did not see significant differences between the single-player and multiplayer modes. Considering the total IMI-score, however, there is a tendency that the multiplayer mode can be more motivating, especially at lower desired success rates, than the single-player mode. 5.6. Conclusion 87

Table 5.1: p-values of multiplayer and single-player comparison Desired success rate 50% 70% 90% Final damping coeﬃcient; p= 0.73 0.40 0.82 Final Total IMI; p= 0.05 0.22 0.69 interest/enjoyment; p= 0.11 0.39 0.68 eﬀort/importance; p= 0.04 0.07 0.20 pressure/tension; p= 0.16 0.22 0.61 perceived competence; p= 0.72 0.50 0.91

5.6 Conclusion

In the presented feasibility study, the effort of a player could be optimized by adjusting the damping coefficient based on current success rate. This feasibility study in a multiplayer setting supports the general- izability of the algorithm beyond single-player application towards multiplayer games using a robotic exoskeleton as an input device. Based on these promising results we continued to test the algorithm on impaired participants and to further investigate its impact on motivation in a multiplayer setting in combination with difficulty adaptation for both human players. 88 Chapter 5. Balanced Multiplayer Gaming 89

Chapter 6

Haptic Interaction in Multiplayer Games with Impaired and Healthy Subjects

Foreword and Overview

The fourth research question was whether patients and their spouses enjoy playing using individual devices. We applied the diﬃculty adaptation for competitive games on non-haptic devices. These devices can be played by healthy co-players such as the spouse. We then further complemented the system with a new haptic approach by connecting diﬀerent players haptically in a cooperative therapy game (Figure 6.1).

6.1 Abstract

We developed two multiplayer games to link the game experience of two players: an Air Hockey game and a Haptic Kitchen game. In 90 Chapter 6. Haptic Interaction in Multiplayer Games Measured Experience Assessment Behavior Experience/ Patient Co-Player Interaction Environment Human-Human act,2 act,2 act,1 act,1 x F x F H H H H Interface 1 Data Interface 2 Data Interface 1 Interface 2 Condition 1 Condition 2 Condition meas,1 ref,1 ref,2 meas,2 x x x x G G G G Condition Condition Controller Controller Integrated Integrated Complexity 1 Complexity Complexity 2 Complexity Interpretation Interpretation Game Performance 1 Game Performance 2 GamePerformance Game Game Game Complexity 1 Complexity 2 Complexity C

B A Patient Data Patient Patient Pro!le Patient Co-Player Data Co-Player Co-Player Pro!le Co-Player Game-/ Selection Therapist Therapist Complexity Gamemode-/ Desired Behavior Experience/ Figure 6.1: Concept chart with highlighted (red) elements for robot-assisted patient-tailored multiplayer 6.2. Introduction 91 the competitive Air Hockey game, differences in skill levels between players were balanced by individualizing haptic guidance or damping forces. In the Haptic Kitchen game, a healthy player could support the patient’s movements using a virtual force field. The two players could control the haptic interaction since both the force field and the point of application were visualized. We tested the haptic performance balancing algorithm of the Air Hockey game and the spouse- controlled haptic support of the Kitchen game with patients post- stroke who trained both single- (i.e., alone) and multiplayer training (i.e., with spouse) in eight therapy sessions lasting 45 min each. Mean total rating in Intrinsic Motivation Inventory was 46.9 points (out of 63 points) for multiplayer modes, and 42.7 points for single player modes, respectively. The spouses applied the haptic support in the Haptic Kitchen game during 42% of the total game duration. We are currently testing more patient-spouse couples to better understand the effects of using these haptic approaches on the behavior and recovery of patients. We foresee this approach can improve the motivation during training and positively influence the at-home behavior of patients, an important goal of rehabilitation training efforts.

6.2 Introduction

We implemented and successfully tested an algorithm that controlled a virtual damping coefficient in an arm exoskeleton joint to balance game performance of healthy players (Chapter 5). However, a modulation of the damping coefficient is not suitable to manipulate the game experience of neuromuscular patients: Instead of damping forces, patients often need supportive forces to reach a desired game performance for a meaningful multiplayer gameplay. If the game performance of a patient and of a co-player could be controlled individually, a balanced multiplayer game could be achieved. Up to now, such a haptic performance balancing controller was not integrated in multiplayer therapy gaming. Besides entertaining games, such as single- or multiplayer competitive games, virtual environments for training of activities of daily living (ADL) are integrated in single player neuromuscular therapy after stroke [85]. Although the context of ADL environments might 92 Chapter 6. Haptic Interaction in Multiplayer Games increase verbal interaction in multiplayer gaming when playing with someone sharing daily life, e.g. spouse, multiplayer modes are not integrated in robot-assisted devices yet [63]. By using a haptic input device for one or two of the players, haptic interaction complements verbal interaction and therefore may boost effects of social interaction when jointly experiencing the ADL environment. We recently developed a novel multiplayer setting that facilitates haptic performance balancing in a multiplayer mode of a competitive game and haptic support controlled by a healthy co-player in a cooperative ADL environment. In the here presented study, we wanted to test with stroke patients and their spouses, whether such multiplayer modes result in a higher motivation compared to single- player modes. Furthermore, we aimed to test if haptic performance balancing decreases the discrepancy between patient’s success rate and pre-defined desired success rate during gameplay, and if haptic support is applied by the spouse players.

6.3 Methods

6.3.1 Experimental setup Two devices were selected for the study (Figure 6.2):

• For the patient: the ARMin arm rehabilitation robot for patients with neurological injuries (4th generation, [116, 66])..

• For the spouse: the HTC Vive’s handheld (High Tech Computer Corporation, Taoyuan, Taiwan).

A monitor, placed 1.2 m from the player’s shoulders, displayed the virtual reality environments. An occupational therapist monitored the players and devices during training.

6.3.2 Virtual reality environments We created two virtual reality environments with haptic elements for multiplayer interaction: an Air Hockey game and the Haptic Kitchen game (Figure 6.3). 6.3. Methods 93

Figure 6.2: Multiplayer setup with haptic interaction. The patient played with the ARMin robot (center) while the spouse played with the HTC vive handheld (left). The monitor displayed the virtual reality environment to both players. A therapist (right) guided and observed the therapy session. 94 Chapter 6. Haptic Interaction in Multiplayer Games

Familiarization Week 1 Week 2 Week 3 Week 4 I SP SP MP MP SP SP MP MP 6 days (d)≥1d ≥1d ≥1d≥1d ≥1d ≥1d ≥1d ≥1d

2-3 rounds of Air Hockey at 4 min 2-3 rounds of Haptic Kitchen at 4 min IMI Outcome measures

Air Hockey: - initial/"nal success rate - initial/"nal support/damping

Haptic Kitchen: - mean/RMS of end e#ector velocity - spouse support

t y per Session: - Intrinsic Motivation Inventory (IMI) tc (min. jerk) - interest/enjoyment q r - e#ort/importance, c p s - perceived competence t1 t0 qj x qc,j(t 0) qc,j(t c)

Figure 6.3: Study protocol. In day one of the study, patients were instructed (I) and their active range of motion was assessed. After one week of familiarization of the therapist-instructed at-home tasks, participants engaged in four weeks of training with two training sessions per week. Week one and three of the intervention sessions were single player sessions (SP). Week two and four of the intervention sessions were multiplayer sessions (MP). 6.3. Methods 95

Air Hockey game

The goal in the Air Hockey game was to defend one’s own goal (defensive task) while scoring on the opponent’s side (offensive task). For the ARMin, single-joint movements (e.g. elbow flexion/extension) were mapped to the mallet’s position along the horizontal axis. From the game’s dynamics, a player is required to do time-constrained reaching movements to meet the game’s goal. We extended an existing controller, designed to adapt the game’s difficulty by damping the player’s movement (Chapter 5), with a controller that can support the player. The support was needed for patients who otherwise would be unable to play the game competi- tively against an unimpaired player. The therapist was free to select the joint that was to be used by the patient before starting a round of the game. All other joints were fixed in place by a position controller. In single player mode, the opponent’s mallet was controlled by the computer using minimum jerk trajectories based on assessments with unimpaired participants [49]. In multiplayer mode, the opponent’s mallet was controlled by the spouse by moving the HTC Vive hand held controller. The spouse’s movement was mapped from a 0.6 m range (−0.3 m...0.3 m) in the lateral direction with the origin at the spouse’s shoulder. Movements in the vertical or anterior-posterior direction did not affect the mallet’s position. In the game environment, side walls constrained the position of the puck and the mallets. To balance gameplay, we focused on the defensive task of the game. The defensive task required the players to reach the approaching puck within a given time window. This embeds training of time- constrained reaching movements in a gamified environment. Success in this task was to reach the puck in time; failure was allowing the puck to pass by and reach the player’s goal zone. We defined the success rate at defending as the number of successful hits divided by the number of total defensive trials. We targeted a success rate, p∞, of 70 % for every player. For the patient we adapted, after each trial, the level βi, of damping or support of the movement. The adaptive algorithm used is described in detail in Chapter 5. The βi coefficient translated into forces that ranged from zero guidance (0 %) to full support (100 %). For a βi coefficient below 0, the forces turned from supportive to resistive. The initial level, β0, was set by the therapist 96 Chapter 6. Haptic Interaction in Multiplayer Games and constrained to only supportive forces. For a support coefficient βi between 0 % and 100 % the patient was guided haptically along a minimum jerk trajectory with an error reducing PD controller. The stiffness of the haptic guidance increased proportionally to the support coefficient βi. For support coefficients below 0 % the movement was damped by forces that increased proportionally to a decreasing βi. ( support, if β ≥ 0 Zi = i (6.1) damping if βi < 0;

Zi = support The support force reduced the reaching errors of the patient and thus increased the ability to successfully perform the reaching task.

eq,j(t) = qc,j(t) − qj(t) (6.2)

e˙q,j(t) =q ˙c,j(t) − q˙j(t) (6.3)

Fi,j(t) = −βi (kp,jeq,j(t) + kd,je˙q,j(t)) (6.4)

Joint j was selected by the therapist, and kp,j and kd,j were predefined control parameters. The angular position qj of joint j was controlled by the patient. The trajectory qc,j(t) was calculated for every defensive task as soon as the puck collided with the opponent’s mallet. The trajectory had a reaction time component and a minimum jerk component. We used Fitt’s law to estimate a reaction time of about 0.3 s from previous study data (Chapter 5). From the start position qc,j(t0) to the target position qc,j(tc) the trajectory was planned for a minimum jerk movement for the player’s mallet. The start position qc,j(t0) was the current position of the player’s mallet when the puck collides on the opponent’s side. The target position qc,j(tc) was the position for a collision with the center of the puck. The time of collision tc was known since the puck’s trajectory had known direction and speed. A time band and position band around tc and qc,j(tc) offered different locations for a successful collision. 6.3. Methods 97

Zi = damping The damping force limited the maximum velocity that a player can reach and thus aﬀected the ability to defend his or her goal [83].

◦ µi,j = βi · dj Nms/ (6.5)

Fi, j(t) = µi,jq˙(t) (6.6)

The damping factor dj was pre-defined for each joint j. The damping coefficient µi,j was defined as the damping factor dj multiplied with the current support coefficient βi. The adaptation of support, or resistance, for non-haptic interfaces was done virtually. For the computer player we iteratively adapted the maximum velocity of the mallet (i.e. simulating a damping force). For the spouse, we iteratively adapted the mapping between the handheld controller and the mallet’s position on the screen. The mapping ranged from assistive to disruptive using a mapping coefficient βs,i. In the assitive scenario (βs,i < 25), the mapping reduced the position errors between a minimum jerk trajectory and the spouse’s movements. The reduction of errors was maximum when βs,i = 0 %. In the disruptive scenario (βs,i ≥ 25), the mapping was visually dis- torted by increasing the discrepancy between the mapped player’s movement and the minimum jerk trajectory. For a βs,i coefficient value of 25 the mapping was unchanged. The therapist set the initial mapping coefficient βs,0 between 0 % and 100 %.

Haptic Kitchen game The Haptic Kitchen game was designed to train an activity of daily life. The new game consisted of several recipes that were split into a number of actions that the patient had to fulfill in series. The kitchen table and weight of the objects were haptically rendered. All objects’ positions, i.e. target position, were within the active range of motion (aROM) of the patient. The aROM of the patient was assessed in the first day of the study. Although both the ARMin robot and the HTC Vive are capable of mapping three dimensional movements, in the Haptic Kitchen game 98 Chapter 6. Haptic Interaction in Multiplayer Games we only considered movements in a two-dimensional plane. This plane was defined by a coordinate system (x, y). From the patient’s point of view, the x coordinate pointed in rightward direction and y coordinate in upward direction. For the ARMin, the origin of the coordinate system was situated in an assumed fixed shoulder joint of the patient. For the HTC Vive, the origin of the coordinate system was situated with outstretched arm and a total range of 0.6 m in horizontal direction (left-right) and 0.4 m in vertical direction (up-down). The anterior-posterior direction was integrated as a free degree-of- freedom not affecting the players’ avatar position for both players. End effector forces from the robot were only applied in x and y direction. In single player mode the Haptic Kitchen game supported the patient with an assist-as-needed task space path controller [67]. A haptic tunnel limited the movements perpendicular to the direct path of the current position of the patient’s hand to the target position. A haptic wall, moving with a five second delay from the start of the trajectory and moving along the path to the target, guided the patient’s hand to reach the target position within 13.5 seconds. We tested different values for delay and guiding time with stroke patients beforehand. Delay and guiding time were predefined targeting a meaningful gameplay for patients of different level in arm function A third force field, located near the target position, attracted the patient’s hand to the target position when the end effector was close by.

Fr(Fx,Fy) = f(p(px, py), r(rx, ry)) (6.7)

er = r − p (6.8)

q 2 2 er Fr = exp −0.2 er,x + er,y · − 0.1 · p˙ (6.9) |er|

Fr were the virtual forces applied by the robot at the end eﬀector based on the current target’s position. p was the current end eﬀector position and r was the target position. 6.3. Methods 99

The exponential function was selected to have a moderate attraction when a target was far and a strong attraction when the end effector was close to the target. This force field requires the patient to apply forces to reach the target while helping stabilize the end effector when it is close to the target to grasp the object. The therapist could choose the amount of support, i.e. stiffness of haptic tunnel and haptic wall and width of haptic tunnel, before starting the game. In multiplayer mode the assist-as-needed controller was adapted to enable haptic interaction with the spouse. The path control forces remained, but the moving haptic wall guiding the patient towards the target in single player mode was removed. Instead, the spouse’s movements were coupled to the patient’s hand by a magnet-like force field similar to the one applied at the target (see Equations 6.7 to 6.9). The spouse could enable the force field by pulling the handheld’s index trigger. Releasing the trigger disabled the force field again.

Fs(Fx,Fy) = f(p(px, py), s(sx, sy)) (6.10)

es = s − p (6.11)

q 2 2 Fs = max 0, 0.2 − 0.005 ex + ey · e s − 0.1 · p˙ (6.12) |es|

Fs were the virtual forces applied by the robot at the end effector based on the hand position of the spouse. These forces were translated into motor torques by the Jacobian. p is the current end effector position and s is the hand position of the spouse. Instead of using an exponential function, a linear function was used as it was more intuitive for the spouse. The influence of the force field was restricted to a distance of 0.4 m.

6.3.3 Study protocol The study was approved by the responsible ethical institution (KEK- ZH-Nr. 2015-0013, Zurich, Switzerland) and by clinicaltrials.gov (NCT02720341). 100 Chapter 6. Haptic Interaction in Multiplayer Games

In eight sessions (four single player and four multiplayer sessions) couples (stroke patient and their spouse) tested the two games using an A-B-A-B experimental design (Figure 6.3). A couple was included if the patient suffered from decreased motor function in one or two arms due to a stroke and met the inclusion criteria (aged at least 18 years, no excessive spasticity of the affected arm represented by modified Ashworth Scale greater than 3, no serious medical or psychiatric disorders assessed by the physician, no orthopaedic, rheumatological, or other disease restricting movements of the paretic arm, no shoulder subluxation, no skin ulcerations at the paretic arm, ability to com- municate effectively with the examiner such that the validity of the patient data could not be compromised, no cybersickness, no serious cognitive defects or aphasia preventing effective use of ARMin). Patients did both, single player robot-assisted therapy (A) and multiplayer robot-assisted therapy (B). One week after a familiarization session, where active range of motion was assessed and an at-home task for the patients was instructed, four weeks of training each were performed. The first and third week were single player sessions; second and fourth week were multiplayer sessions. Every session consisted of a standard 45 min therapy unit carried out by an occupational therapist. The session was designed to include at least two 4 min games of the Air Hockey game and at least two 4 min games of the Kitchen game. Every game was followed by a break of at least 1 min. The therapist was free to select the joints and initial support for the Air Hockey game, and the support level for the Haptic Kitchen game. The patient and spouse were free to interact during the multiplayer sessions. In the multiplayer sessions, patients trained in a team with their spouse. The therapist was free to select the initial difficulty level of the spouse. The spouses played with their dominant arm. To measure the technical feasibility of the intervention, we monitored the patient’s success rate and the level of support or damping of the movement in the Air Hockey game. In the Haptic Kitchen we assessed the patient’s velocity profile by extracting the mean and the root-mean-square values. We also assessed the amount of time the spouse triggered the haptic support. After each session the patients rated three statements from each of the 6.4. Results 101 three subscales interest/enjoyment, effort/importance, and perceived competence of the Intrinsic Motivation Inventory (IMI) [110]. Each statement had to be rated on a 7-point scale, with 1 indicating “not at all true”, and 7 indicating “very true”. From the ratings we calculated sums for all subscales and a total sum of the IMI ratings.

6.4 Results

Two chronic stroke patients (patient 1: female, 66 years old, two strokes, 12 and 27 months ago, right arm affected; patient 2: male, 34 years old, several strokes in the past 8 years, left arm affected) participated in the study, each with his/her spouse. None of the sessions was interrupted due to technical or motivational issues. All players reported that the games were fun to play and, in multiplayer settings, the players highly valued the new therapy experience. None of the patients reported any pain. Both spouses noted that the game was exhausting for them what indicates that the game was also challenging for them to some extend. The haptic balancing of the Air Hockey game used both support and additional damping (Figure 6.4A). The success rate in the Air Hockey game was calculated with a moving window of 25 trials. In the Haptic Kitchen game the spouse used the force field in phases (Figure 6.4B). Player 1 played the Air Hockey game 23 times (thirteen times with forearm supination/pronation, nine times with elbow flexion/extension, and once with horizontal shoulder ab-/adduction). The mean of all session’s success rate of patient 1 changed from initial 73 % to final 74 % in single player and from initial 52 % to final 70 % in multiplayer mode. Player 2 played the Air Hockey game 22 times (thirteen times with elbow flexion/extension and nine times with horizontal shoulder ab-/adduction). For player 2 the success rate remained unchanged from initial 79 % in single player and decreased from initial 83 % to final 78 % in multiplayer mode (Figure 6.5A). The final support was higher for every game in multiplayer conditions compared to what the therapist set pre-game. Two sessions of patient 2 (one single player and one multiplayer) had corrupt data from the Kitchen game which was excluded from the result plots (Figure 6.5B). The sum of the IMI ratings ranged from 30 to 57 points for single player and from 34 to 58 102 Chapter 6. Haptic Interaction in Multiplayer Games points for multiplayer sessions out of 63 possible points (Figure 6.6).

Figure 6.4: Example data of the two games. A: Support and damping values from the Air Hockey game data. S from a single player session of patient 1. The success rate is plotted against the desired success rate p∞ = 70 % . B: Velocity proﬁle in the Kitchen game. Phases where the spouse triggered the force ﬁeld are indicated.

6.5 Discussion

The combination of haptic and non-haptic approaches to balance skill levels in the Air Hockey game allowed for competitive gameplay for both players, i.e., patient and spouse. By using a haptic approach 6.5. Discussion 103

Figure 6.5: The mean results of the sessions of each patient in clouds (blue dots) and the mean for each training mode (red dots) are reported. A: Air Hockey results of games with at least 25 defensive trials. B:Mean and RMS values of velocity during the Kitchen game. 104 Chapter 6. Haptic Interaction in Multiplayer Games

Figure 6.6: Intrinsic Motivation Inventory (IMI) scores. The self- assessed sub-scales interest/enjoyment, effort/importance, and perceived competence of the IMI are presented. The sum of all sub-scales is presented in the total IMI sub-figure. 6.5. Discussion 105 with the patient, the modulation of the game conditions was not apparent to the patient. The haptic approach complements existing methods where the difficulty is adjusted non-haptically only by changing avatar size or ball speed [60]. The non-haptic manipulation of errors to adapt skill levels also conceals different skill levels. The algorithm that was used to adapt the condition for each player is therefore applicable in both haptic and non-haptic devices. Compared to the previous study (Chapter 5) we facilitated a within- game transition from conditions of patient support to patient challenge. In single sessions the algorithm seemed to work as the algorithm tends to adapt the condition towards desired success rate in particular when the initial condition was wrongly selected by the therapist (e.g. adapting from an initial success rate of 25 % to 52 %. However, the game time of 4 min and the selection by the therapist of the joint to be trained impeded full balancing of performance within game duration or transition of conditions between games. In the Haptic Kitchen game we offered the opportunity to the spouses to use a force field to support, or hinder, the impaired partner in a virtual activity of daily living. In nearly half of the game time (41 % by the spouse of patient 1, and 48 % by the spouse of patient 2) the opportunity was used to haptically support the impaired partner. In both games we noted an increased amount of verbal interaction compared to a patient-patient setting in another study we performed. In particular, the kitchen environment provided context for verbal interaction of the couples, including mentioning about their current at-home behavior. We considered a new type of game design only focusing on audio-haptic feedback which we expect to enhance social interaction during therapy games (Appendix B). We expect that including a close relative or spouse into the therapy process will positively affect the patients’ at-home behavior. Changing a health-relevant behavior, such as adhering to therapy recommendations or changing one’s diet is dependent on several self- regulatory skills as, for example, specified in the Health Action Pro- cess Approach [145, 21]. Stroke patients with arm impairments need to change their at-home behavior to support the occupational and physical training plan. To promote and maintain therapy progress, homework tasks designed by therapists need to be repeatedly trained 106 Chapter 6. Haptic Interaction in Multiplayer Games at home and the impaired arm needs to be involved in activities of daily living. However, good intentions are often not enough for a successfully maintained behavior change [147]. The social environment, e.g. the spouse, of the patient may facilitate (or hinder) this change in behavior. Social exchange processes, such as social support might play a prominent role in health behavior change [144]. We expect that the multiplayer sessions promote positive social ex- changes in spouses and thereby foster better adherence to therapy recommendations. The ultimate goal is to increase the use of the impaired arm in the patient’s everyday life. The health behavior change of the patients involved in the present study was monitored by wrist-worn accelerometry and related motivational, volitional and social variables self-reports. These data will be reported in future publications.

6.6 Conclusion

Rehabilitation robots in multiplayer games have the potential to haptically support patients in competitive games and facilitate haptic interaction in cooperative games. In competitive games, haptic guidance and damping forces can be used to offer difficulty adaptation for involved players. In cooperative games, haptic support can be used by the patient’s spouse to haptically interact with their partner. Based on the promising results we found with these two couples measured so far, we will continue to investigate the effect of multiplayer games using haptic therapy devices on patient’s kinematics and game experience. Our long-term goal continues to be the development of training environments that positively influence the patient’s at-home behavior. 107

Chapter 7

General Conclusion

Patients in the post stroke (sub-)acute phase enjoy playing multiplayer games, even though they might not always prefer multiplayer games comparison to single player games. Noting that comparison, playing a multiplayer game itself does not generally increase an individual’s motivation and nor improve game performance of the individual player. Other characteristics of the elements involved are just as important, such as tailoring games to the individual patient’s circumstances (patient-tailored) or the relationship to the co-player involved. Furthermore, the multiplayer mode influences game experience and performance. Competitive game play might be more motivating and provoke more intensive training than cooperative game play. These two modes are the most often used multiplayer modes but only represent two out of the various other modes available such as educative or assistive cooperation, or coopetition, i.e., competition with a common superordinate goal. Thus, this field of research is in its early stage and the possibilities have not yet been fully exploited. Patient-tailoring of games has the potential to facilitate more intensive training and at the same time, maintaining the patient’s motivation. Patient-tailoring might optimize both game experience and effort invested by the patient. I identified a tendency that the patient- tailoring might even outperform the more positive effects provided by multiplayer games. Therefore, patient-tailoring may not only improve the efficacy of multiplayer games by taking account of the different 108 Chapter 7. General Conclusion skill levels but also be the key factor in retaining patient motivation. Haptic devices such as rehabilitation robots provide new possibilities to tailor the therapy conditions for the patient. Adaptation to different levels of difficulty cannot only be performed in the game mechanics but also in the conditions provided by the devices. Hap- tic difficulty adaptation conceals skill differences and can therefore facilitate game play without revealing the weaknesses of the worse players in multiplayer games. Moreover, considering the diversity of multiplayer modes available, haptic interaction between players can provide new motivating multiplayer games towards highly intensive therapy sessions. In multiplayer games intensive training can be facilitated by co-players who haptically adapt the required effort of the patient by controlling the task difficulty: co-players – such as therapists, relatives, or other patients – can haptically assist as cooperator or challenge as a competitor. A multi-session study showed, that stroke patients enjoy playing in such an assistive cooperation. This assistive cooperation mode complements the sets of competitive and cooperative multiplayer modes.

7.1 Major Contributions

With my thesis a conceptual structure on the technological elements of patient-tailored multiplayer games in robot-assisted post-stroke therapy has been contributed to facilitate transfer between different games, devices and fields and the discussion of methods and results of different multiplayer studies (Chapter 2). I reviewed the field of robot-assisted multiplayer gaming and integrated the findings into established concepts such as the flow theory and the challenge point framework (Chapter 3). I carried out a multiplayer study with (sub-)acute stroke patients using an arm therapy device with a self- developed game and compared the methods and the results to the reviewed studies (Chapter 4). I enhanced and tested the multiplayer game with difficulty adaptation strategies, including haptic difficulty adaptation, to facilitate game play between differently skilled players using different devices (Chapter 5). Finally, I developed and tested a new cooperative therapy game that enables haptic support for the patient by a healthy co-player (Chapter 6). 7.2. Outlook 109

7.2 Outlook

This thesis’ contributions provide both structure and promising methods for future multiplayer applications in robot-assisted post-stroke therapy. Both might help to position future studies in the field. Fu- ture studies shall standardize their ways to measure motivation and exercise intensity to better compare how different characteristics of the elements in the structure affects the patient behavior and experience during the therapy. A continuous exchange and harmonization in this field of research among different research groups is therefore required. Furthermore, the possibilities in multiplayer games’ research field have not explored the different multiplayer modes available. Therefore, It is recommended that further multiplayer modes, games and combinations of players need to be investigated regarding feasibility, patient’s experience, and patient’s behavior. Furthermore, the integration of haptic interaction with multiplayer games should be tested more extensively. The studies conducted by us and other groups were investigating mainly short term effects focusing on motivation, game performance and exercise intensity. It remains unclear how these effects relate to established clinical measures. I suggest clinical long-term studies where all motivation, game performance, exercise intensity, and clinical assessments are monitored. Based on such studies, the potential of multiplayer games can be evaluated. The integration of both healthy players and patients into multiplayer games facilitates the investigation of social aspects of robot-assisted multiplayer games. I believe that playing with a friend or family member can influence the health behavior and therefore, bridges clinical therapy and the at-home-behavior of patients. I expect that changing the at-home behavior of stroke patients is required to cru- cially increase the number of movements performed with the affected arm and therefore increase the intensity of arm training. 110 Chapter 7. General Conclusion 111

Appendix A

Performance-based viscous force ﬁeld adaptation for arm therapy

A.1 Background

Muscle strength is considered as crucial for performance of tasks of daily living [126]. Muscle weakness is one of the major deficits after stroke [22] and may be substantially contributing to compromised functional performance [126]. Literature implies that specific strength training may yield to higher gains in muscle strength than conventional rehabilitation therapy and improve motor function [4]. We developed a new control approach for robot-assisted therapy. An adaptive viscous force field enables a customized strength training for stroke patients considering the individual strength profile. The control design was tested in a feasibility study which is discussed in the related paper. 112 Appendix A. Performance-based viscous force field adaptation

A.2 Methods

The control design is based on an adaptive viscous force field in a repetitive tracking task. The adaptation of the profile derives from the individual performance within the tracking task. The adaptation is applied not only at the general level (i.e., task level) considering inter-patient variability in strength. It also accounts for the kinematic dependency of strength of the individual patient by shaping the viscous force field within the trajectory, as a function of location and direction of the target object. The target object CT moves from the start point (transition point) PS along the trajectory back to the start point PS within the period T (Figure A.1). This trajectory is represented by the position of the target object at a specific normalized time s.

Figure A.1: Task design in a two-degree-of-freedom example. The ﬂashes represent the movement direction. The patient moves the cursor CP (grey dot) trying to track the target object CT (black dot). The control variable s (normalized time) starts at PS with s = 0.

t s = , 0 ≤ t ≤ T (A.1) T

tPs = 0, stPs = 0 (A.2) A.2. Methods 113

One round or repetition of the task is defined as the target trajectory starting and ending at PS, and the round index i is increasing incre- mentally by 1 whenever CT is passing PS. During the first round the index i is 1. In each round s is starting from 0. Throughout each round three profiles define task difficulty and performance of the patient:

velocity profile vi(s) viscous force field ri(s) performance profile pi(s)

The predefined velocity profile vi(s) refers to the movement of the target object CT . It does not change between the rounds (i.e. vi(s) = v(s) ).The participant is instructed to track CT during the entire task and moves the cursor CP with velocityx ˙ P (s). A viscous force field ri(s) is applied and challenges the participant in keeping track of CT . The initial viscous force field, used in the first round, is the predefined viscous force field r1(s).The viscous force field changes over time and generates velocity dependent forces that (in case of positive viscosity) counteract to the movement and thus require force from the participant. The performance profile pi(s) can be indicated by any quality of movement measure that can give an immediate performance measure at any value of s (e.g. position error, velocity error).

pi(s) = pi(xP (s), x˙ P (s)) (A.3)

The viscous force ﬁeld is updated according to the performance during the previous round.

ri+1 = ri+1(ri, pi) (A.4)

We propose two separate update sections (Figure A.2):

1. shape update

The viscous force field is shaped based on the normalized local performance valuesr ˜i+1(s), scaled and converted into viscous force field by the factor c1. It is represented by the shape of the viscous force field along the track. 114 Appendix A. Performance-based viscous force field adaptation

2. task level update

The general task levelr ¯i (mean level of the viscous force ﬁeld) is set according to the overall performancep ¯i of one entire round, scaled and converted to the resistance space by the factor c2. Both update sections are added to the viscous force ﬁeld of the previous round reduced by a forgetting factor (1 − L):

rˆi+1 = (1 − L)ri(s) + Lc1r˜i+1(s) + c2p¯i (A.5)

Figure A.2: The update of the viscous force field profile is separated into a shape of level section and a task level section. The shape of level defines difference in the viscous force field according to local abilities of the patient regarding the task execution. The task level defines the general level of the viscous force field regarding the task execution.

For the two update sections we propose the following functions: 1. shape update A.2. Methods 115

pi(s) − pi,min r˜i+1(s) = pi,N (s) = (A.6) pi,max − pi,min

2. task level update

Z 1 p¯i = f(pi,Σ) = f( pi(s)ds) (A.7) s=0 Since this round based update will end up in a discontinuous behavior at the transition point PS, transition conditions shall force the viscous force ﬁeld to be continuous during the transition phase.

rˆi+1(s = 0) =r ˆi(s = 1) (A.8)

0 0 rˆi+1(s = 0) =r ˆi(s = 1) (A.9)

To reduce the number of data points stored during the task and to smooth the proﬁle, the viscous force ﬁeld is reduced to a polynomial regressed function. The data points from both the shape section and the transition conditions are used as an input for the polynomial regression of order n.

0 ri+1 = pn(ˆri+1, ri(1), ri(1)) (A.10)

The task was implemented into the therapy robot ARMin. The trajectory selected for the feasibility study was a one-DOF movement in elbow ﬂexion-extension with constant speed. The tracking task was represented in a visualization of the target object and the patient cursor programmed in Unity 3D (Unity Technologies, 5.1) (Figure A.3). Summary of task parameters:

• Period of trajectory T = 18 s

• Trajectory length 180° (90° elbow extension, 90° elbow ﬂexion)

• velocity proﬁle v = 10°/s 116 Appendix A. Performance-based viscous force ﬁeld adaptation

Figure A.3: Task visualization. The target object (dark grey) follows the one-DOF trajectory starting from the left edge and moving with constant velocity to the right edge and back. The patient moves the patient cursor (light grey) to track the target object.

The task was performed for ﬁve minutes resulting in a total number of 16 completed rounds. The controller started initially with minimal viscous forces where only friction and inertia eﬀects were felt by the participant. The performance was measured as the position error ex,i(s) between target object and patient cursor. Summary of control parameters:

• initial viscous force ﬁeld r1(s) = 0 Nms/° 1 1 performance measure pi(s) = = • ex,i(s) |xt,i(s)−xp,i(s)| • update of viscous force ﬁeld 1. shape update

r˜i+1(s) = (0.5 − pi,N (s)) (A.11) 2. task level update

Z 1 1 pi,Σ = )ds (A.12) s=0 ex,i(s)

 −0.1 pi,Σ < 0.3/°  p¯i = 0.4 pi,Σ > 0.45/° (A.13)  10° (−1.1 + 3 pi,Σ) otherwise

• learning rate L = 0.5 A.3. Results and Discussion 117

• scaling factors c1, c2 = 1 Nms/°

The range of 0.30/° to 0.45/° was defined experimentally as a con- venient performance range for healthy subjects. The task level step sizes of −0.1 and 0.4 were selected to enable a fast increase of the viscous force field towards an individual maximum profile with option to reduce the viscous force field with decreasing performance. The polynomial regression was performed at an order of ten. This equals the highest possible order that was possible from a technical point of view and therefore, represents the performance of the patient with best accuracy.

0 ri+1 = p10(ˆri+1, ri(1), r (1)) (A.14)

A validation or optimization of the polynomial order was not part of this project. For this ﬁrst implementation the viscous forces were restricted to a minimum of 0 Nms/°.

A.3 Results and Discussion

Representative data of one subject is presented in Figure A.4. The viscous force field increased between first and last round in all participants. We could successfully implement a new control approach for robot- assisted therapy. The controller could increase the level of the viscous force field (i.e. the task level) towards the maximum of a healthy subject and adjusted the shape of the task level according to the individual strength of the subject. The selected velocity profile of the task leads to high positions errors at the positions were the tracking cursor changes direction due to infinite accelerations. For a future task design we propose finite accelerations that can be performed by patients, e.g. sinusoidal velocity profiles for one-DOF tracking tasks. The restriction of the viscous force field to a minimum of 0 Nms/° could be abolished to enable supporting forces fields for disabled participants. This has to be tested with stroke patients after a redesign of the update method. 118 Appendix A. Performance-based viscous force field adaptation

Figure A.4: Progress of the viscous force field of one individual participant. The viscous force field was not only shaped towards the local abilities of the patient (i.e., shape of each curve) but also with increasing global level of the viscous force field (relative position of each curve).

A.4 Conclusion and Current State

The successful implementation of a performance-based strength training using an adaptation of a viscous force ﬁeld showed great potential to enhance the outcome regarding strength and function in robot- assisted rehabilitation. The results of the feasibility study were used for a redesign of the implemented controller considering healthy and disabled participants. 119

Appendix B

Rehabilitation Game Without A Visual Display

B.1 Background

Music is a promising stimulator for intrinsic motivation in the context of rehabilitation [58, 52]. Music effectively promotes post-stroke recovery in motor and cognitive functions, and furthermore in emotional and social domains [160, 8, 24, 132, 140, 7, 86]. Studies that compared conventional therapy forms to therapy tasks embedded in active music making revealed that music-associated training increases the level of motivation significantly [52, 143]. Auditory displays have already been determined to be effective for navigation within complex systems [71]. Accordingly, sound is an au- dible source for navigation through the execution of a task in virtual scenarios without the need for a visual display unit, the advantage being that the visual focus can be on the trained limb rather than a graphical display, thus promoting visuo-motor control [87, 186]. We designed audio-haptic tasks in a way that they can be performed either with visual display (i.e., a monitor presenting the game workspace) 120 Appendix B. Rehabilitation Game Without A Visual Display as an audio-visuo-haptic environment or without a visual display as an audio-haptic environment only. To reduce the cognitive load of the participants and have more cognitive resources for creation and decision making processes, we designed the visual display and the haptic environment such that they both presented the same game workspace [114]. Accordingly, the visuals were not essential to complete the audio-haptic task. Given these related works, the primary hypothesis was that a gamified task promoting creativity embedded in a task for motor therapy increases intrinsic motivation more than a gamified task not promoting creativity. Our second hypothesis was that a gamified task in motor therapy without visual display increases intrinsic motivation more than a gamified task with visual display. Moreover, we hypothesized that promoting creativity and omitting a visual display would increase total training time, free movement time and perceived product value. We further hypothesized that promoting creativity and omitting a visual display would reduce energy expenditure, relative perceived training time and perceived effort. The detailed study protocol and results are reported in the related article.

B.2 Methods

The game environment was developed targeting horizontal movements at table height as this type of arm motion is required for activities of daily living, such as cleaning a table or moving objects on a table. Such training activities are already provided in several upper limb training devices for stroke therapy [66, 138]. Four conditions were developed in the Unity game engine [169]: A music task promoting creativity played without a visual display (C+V–), a music task promoting creativity played with a visual display (C+V+), a music task not promoting creativity played without a visual display (C–V–) and a music task not promoting creativity played with a visual display (C–V+). All four conditions used the same audio-haptic environment. In this environment, participants have to select one out of two sound samples positioned to the left and to the right, respectively, several times per condition. In the conditions with visual displays, the environment was complemented B.2. Methods 121 with visual feedback (audio-visuo-haptic environment). To provide task oriented training we haptically simulated an environment of moving objects on a table. Subjects could only move the end effector of the robot horizontally. Downward movements were restricted by a virtual table that was set at the level of the shoulder and provided arm support. The horizontal left-right movement served as game input. Therefore, the game can be considered as a one-degree-of-freedom task. As illustrated in Figure B.1, the haptic display included haptic walls at the end of the workspace (HWs). Sound zones (SZ) and center zones (CZ) were action zones where a vibrotactile haptic feedback (i.e., band-limited white noise) was provided at the hand module of the robot. The vibrotactile haptic feedback was induced via DC motors at the wrist joint of the robot. As a perceivable state indicator (i.e., being inside or outside of a SZ) for the participants, the vibrotactile haptic feedback was not breaking or stopping the participant’s movement. None of the participants commented on the vibrotactile feedback. While in C–V+ and C+V+, the task was visualized on a monitor, in C–V– and C+V–, the same movements were performed without visual feedback (i.e., without a monitor). The sounds used in the SZs consisted of fourteen different pairs of sound samples and two pairs of sound effects. These sixteen pairs of music creation elements were presented to the participant in fixed order. The sound samples consisted of synthetic piano, mallets, marimba, vibraphone, pads, drums, hi-hats, and claps. Harmonic elements were tuned in C-Major. Each sound sample lasted six seconds and was played in a loop with a tempo of 80 beats per minute. The two pairs of sound effects (i.e., modulators) were Reverb and Echo, and Resonance and Arpeggiator. To ensure a well-formed, aesthetic and pleasant music structure, the sound samples and the modulators were designed such that they suit to each other when playing simul- taneously. The underlying music composition was developed with the commercial music software Abelton Live 8 (Abelton). Subjects listened to the sounds using Sony MDR-7506® headphones. All conditions used the same audio-haptic environment wherein the haptic walls together with instrument sounds gave feedback about the position in the game. 122 Appendix B. Rehabilitation Game Without A Visual Display

30 cm 6 cm -6 cm -30 cm 40 cm 4 cm -4 cm -40 cm

HW SZ HWCZHW SZ HW

left game zone right game zone

Figure B.1: Screenshot of the tasks in V+ conditions with additional map of haptic elements: The visual environment displays sound icons (e.g., a bell or hands clapping) within sound zones (SZ), and a grey square that marks the center zone (CZ). SZs and CZ (wave-marked areas) are active zones where a vibrotactile haptic feedback is provided. The scene is limited by two haptic walls (HW, solid lines). HWs (dotted lines) limiting CZ are only turned on in the game phases where only the left or the right game zone is used. The red circle is the game “cursor”. All units are in centimeters.

The game rules for all four conditions are illustrated in Figure B.2. Each of the four conditions consisted of of three game phases in fixed order, namely an active movement phase, a listening phase, and a final phase). The three game phases were different in the conditions promoting creativity (C+) and in the conditions not promoting creativity (C–) but independent of the inclusion or omission of a visual display. The only difference between the V+ and the V– conditions was the B.2. Methods 123

Game phases Active movemement phase Listening phase Final phase

Condition Choose left or right SZ (e.g., “bell” or “clap”) Listen and feel free to move Move between SZs to keep sound playing C+V- ...... C+V+

......

1 2 34 ... 16 1 2 34 ... 16 1 2 34 ... 16

Condition “Go left!” or “Go right!” Listen and feel free to move Move between SZs to keep sound playing C-V- C-V+ ......

......

1 2 34 ... 16 1 2 34 ... 16 1 2 34 ... 16

Time 5-8 min 3-5min Open ended

Figure B.2: Game rules for all four conditions. C+ conditions and C– conditions for all three game phases (i.e., active movement phase, listening phase, final phase) are presented. In C+ conditions, four successively activated sounds add up to a music piece (upper field, from top to bottom: claps, drums, guitars, marimba, . . . ). In C– conditions (lower field), each consecutively activated sound is played alone.

workspace visualization on a monitor which was either provided (V+) or omitted (V–).

As the outcome measure the Intrinsic Motivation Inventory (IMI) was used. The IMI is a multidimensional measurement device for assessment of participants’ subjective experience related to a spe- ciﬁc activity. While there are many versions of the IMI, in this study, statements of the subscales interest/enjoyment (seven statements), perceived choice (seven statements), value/usefulness (two statements), man-machine-relation (ﬁve statements, in IMI original name of item: relatedness) were used [110]. The statement sentences for each sub-item were presented and subjects answered on a Likert scale ranging from 1 to 7 (1: do not agree at all; 7: fully agree). 124 Appendix B. Rehabilitation Game Without A Visual Display

Secondary outcomes (recorded)

During each task performance, the ARMin system recorded the total training time, time of free movement, i.e., duration of final phase, and root-mean-square (RMS) of the end effector velocity profile. RMS of the end effector velocity profile approximates the energy expenditure [170] and was recently used in neuromuscular therapy studies [60].

B.3 Results and Discussion

Linear mixed models analysis of the IMI subscales interest/enjoyment and perceived choice showed a significant positive effect of promoting creativity (C+ instead of C–). No significant effects for man-machine relation and value/usefulness were found. Linear mixed models of the total training time and RMS of the velocity profile showed a significant interaction effect in promotion of creativity and omission of visual display. In the conditions without a visual display, promotion of creativity leads to less total training time and higher RMS of velocity while the opposite effects are found in conditions with a visual display. No significant effects in free movement time were found. We could prove that intrinsic motivation is increased by tasks promoting creativity when performing a robot- assisted music task. The tasks promoting creativity were preferred and rated more enjoyable than those without creativity. These results are consistent with previous findings [53] and support the use of music to engage in robot-assisted training. In addition, the game modes that promoted creativity led to increased self-reported auton- omy (indicated by IMI subscale of perceived choice). Promoting creativity had no significant effect on usefulness. However, more subjects wanted to receive the music piece produced in modes where creativity was promoted. Therefore, process (i.e., creation of music piece) and product (i.e., music piece itself) must be discussed separately in game design regarding value. The omission of a visual display did not affect intrinsic motivation ratings. According to statements of participants, playing without a visual display has been a new and exciting experience for them. B.3. Results and Discussion 125

However, without a visual display the rules might not be intuitively understood. Non-intuitive game rules may lower the feeling of competence [136]. Although not systematically measured, subjects reported verbally that the conditions without visual display were challenging. Therefore, future studies should measure the IMI subscale perceived competence. Furthermore, the conditions without visual display might benefit from a more detailed audio guide. Accompany- ing the users throughout the first steps of the game verbally may increase the feeling of competence. For patients in neurorehabilitation therapy it might be even more challenging to understand the task [70]. For these patients, intuitive game design and clear game instructions are especially important. However, playing without a visual display facilitates the use of vision for movement observation which has a positive impact on recovery of motor functions [46]. Furthermore, vision could focus on human-human interaction. Seeing each other is particularly motivating while playing with training devices [79]. In our setup, we expected that the subjects would visually observe the human-robot interaction resulting in a higher man-machine relation. From observation we realized that subjects mainly looked straight ahead to the place where usually a monitor would be expected. As stated by the subjects, they were rather focusing on the auditory display of the game than on the arm movement. Likewise, no difference in man-machine relatedness was found. In further studies without visual displays, patients should be instructed, to visually focus on the arm movements [87, 186].

It is a principal objective of robot-assisted therapy to increase the intensity (i.e., energy expenditure, duration of therapy, number of repetitions) during therapy. In our study, creativity promoting tasks increased energy expenditure without visual display. However, (as shown by significant interaction) energy expenditure decreased with visual display. The possibility to explore the audio-haptic setting seemed to promote more intense moving behavior. On the contrary, when the target position was predefined and visualized (C–V+) the subjects moved fast with fewer slow movement phases for decision making. The subjects tried to reach the stated target positions (i.e., “go left”,“go right”) as fast as possible without resting phases for listening. However, in the condition with visual display where creativity 126 Appendix B. Rehabilitation Game Without A Visual Display was promoted (C+V+) the movements slowed down. In summary, we assume that when there is no visualisation of a reaching task ((V–), i.e., no extrinsically motivators for fast reaching movements), tasks promoting creativity may intrinsically motivate for more intensive movements. Regardless of the conditions, the subjects voluntarily played on average two additional minutes within each eight minutes task. Obvi- ously, the subjects were motivated to voluntarily extend time of arm movement after finishing the guided phases. Future work could provide haptic guidance or resistance [19]. The decision to use or omit a visual display could be to the user. Al- ternatively, therapy could start with visual display until proficiency increases and be omitted later, e.g.,as an adaptation in task difficulty.

B.4 Conclusion and Current State

The combination of music and activities promoting creativity in motor training promotes enjoyment, and thus intrinsic motivation of subjects performing robot-assisted training. As the audio-haptic environment is sufficient to create a meaningful gameplay, music tasks can be performed without a visual display. Promotion of creativity in a gamified task for neurorehabilitation may increase intrinsic motivation in patients but not training intensity in general. At the same time, omission of a visual display may not influence intrinsic motivation or training intensity. However, promotion of creativity differently influences training intensity dependent on the visual display of the task. When promoting creativity, audio- visuo-haptic environments lower training intensity while audio-haptic environments enhance training intensity. We demonstrated the feasibility of playing an audio-haptic music game and suggest a follow-up study on stroke survivors. 127

Appendix C

Remote Haptic Therapist-Patient Interaction with Two Exoskeletons for Stroke Therapy

C.1 Background

In telerehabilitation strategies robots are applied to enable therapy over distance [98]. Telerehabilitation refers to the remote provision of medical rehabilitation services, such as physical therapy, speech pathology, or occupational therapy by means of new communication and information technologies [31, 135]. Telerehabilitation might be helpful in patients where impaired mobility hinders access to medical experts. But even if telerehabilitation is seen as a promising service there is a slow uptake of technology in this field due to significant technical and conceptual challenges (i.e. the difficulty to assess a patient’s status remotely, also referred to as teleassessment [161]). 128 Appendix C. Remote Haptic Therapist-Patient Interaction

To assess biofeedback of a patient remotely, robot-assisted telerehabilitation systems need to be bidirectional (i.e. interaction forces caused by the patient haptically presented to the guiding therapist). However, robot-assisted telerehabilitation systems that provide such bidirectional communication are often limited in degrees of freedom [45, 121, 122, 125, 175]. We designed an application where an exoskeleton robot enables the therapist to feel the patient’s limitations in the own arm and, thus, provides a completely new way of patient-therapist interaction. We call it the “Beam-Me-In” strategy. “Beam-Me-In” is realized through use of two ARMin robots. Kinematic functions are assessed by the position sensors on one robot and are presented on the second robot (i.e., a unidirectional design of a master-slave system [37]). The ki- netic reaction in the second, guided robot can be assessed by force sensors and fed back to the first robot as an interaction force. We present a bidirectional master-slave system between two devices (i.e., two ARMins) with 7 degrees of freedom each that provides haptic reification of the patient’s impairments (ARMin 1) to the therapist’s arm (ARMin 2) and thus, provides technology that enables the therapist to be “beamed” into the patient [97]. We tested the application with therapists. In order to evaluate how far therapists can experience the patient’s disability, we determined how accurately, reliably, and confidently therapists can quantify patient’s motor impairments by having their arm actively or passively moved through the patient’s trajectory and then estimating outcomes based on the therapist’s own proprioception and vision. The study details can be found in the related article.

C.2 Methods

Four tasks for examination of a paretic arm were chosen to allow for assessment after stroke: active and passive ROM, resistance to passive movement (RPM), pathological muscle synergies (SYN) and quality of movement (QOM). For data acquisition for each of these four tasks, either recordings from a real subject were used, or subjects were simulated and then replayed during the study (Figure C.1). This ensured standardized conditions for each participant. C.2. Methods 129

Figure C.1: Recording and replaying the QOM assessment. (Left) Stroke patient recorded while performing the QOM task in transparent mode. (Right) Participant in slave mode, experiencing the replayed QOM performance (mirrored to the participant’s dominant side)

ROM

Data acquisition Three subjects with diﬀerent active (aROM) and passive ROM (pROM) in elbow joint were simulated (1. aROM 15°-110°, pROM 0°-120°; 2. aROM 50°-90°, pROM 20°-110°; 3. aROM 40°-85°, pROM 30°-105°).

Procedure To introduce the task, the participant was passive while the elbow joint was flexed and extended in ARMin by the experi- menter in intervals of 5 degrees from 0° to 120° and the participant was verbally informed about each 5°-step and could look at the arm position. Afterwards, the participant could freely move through the ROM for one minute to explore the limits. The participant was allowed to feel each of three simulated subjects (aROM: participant passive; pROM: participant active) ten times, and then quantified the aROM and pROM with a required 5° resolution. The three different ranges for aROM and pROM were used to differentiate severity among the different subjects. 130 Appendix C. Remote Haptic Therapist-Patient Interaction

RPM Data acquisition To evaluate muscle tone, the resistance to passive movement during passively induced flexion/extension was simulated in ARMin for three different subjects. Three subjects with varying degrees of impairment according to the “modified Tardieu Scale” (mTS) in the arm were simulated. The mTS is a clinically established test which assesses the muscle’s response to stretch at given velocities in degrees per second, and the quality of muscle reaction on an ordinal scale ranging from 0 to 4 (with “0” meaning “no spasticity”) [68]. Subject 1 represented a healthy person (mTS=0, pROM 0° to 120°, no speed threshold, no catch angle, no stiffness, no damping). Subject 2 represented a mildly affected person with a slight resistance of the elbow flexor muscles which was simulated by an increase in damping as soon as a certain speed threshold in extension was exceeded (mTS: 1, pROM: 20° to 110°, speed threshold: 80°/s, no catch angle, no stiffness, damping: 1 Nms/°). Subject 3 represented a severely affected person post-stroke where the movement was interrupted at a certain angle (“catch angle”) when a predefined speed threshold was reached (mTS: 2, pROM: 30° to 105°, speed threshold: 40°/s, catch angle: 60°, stiffness: 0,3Nm/°, no damping).

Procedure The participant was allowed to feel each of the three simulated subjects ten times. First, the participant quantified pROM (participant active) with a required 5° resolution. Then, the angle of muscle reaction, if present, was quantified and the quality of muscle reaction was rated following the common instructions of the mTS [9]. The assessment of the three different levels of resistance to passive movement was used to differentiate severity among the different subjects Since the same three pROMs as in the ROM task were assessed and range of motion is part of mTS, the results of ROM and RPM were compared to test for intra-rater reliability.

SYN Data acquisition An upper extremity flexor synergy can typically be observed in voluntary flexory arm movements [166]. Components of a flexor synergy were experimentally quantified in previous stud- C.2. Methods 131 ies [40, 103, 113]. While healthy subjects are able to selectively move one joint while keeping the other segments still (interjoint coordination), patients with post stroke commonly lose this capability and present a flexion synergy pattern with abduction and external of the shoulder together with flexion of elbow, hand and fingers [141]. To assess the ability of the participant to distinguish between a normal, selective movement and a loss of the inter-joint coordination resulting in a pathological muscle synergy, arm movements of three simulated subjects were presented to the participant. They were created based on movement profiles of a healthy subject (subject 1), and subjects post-stroke (subjects 2 and 3). For all three movements, the same starting position and a sinusoidal-type position-controlled movement with a period of 6 seconds duration was chosen (Figure C.2).

Synergy 1 Synergy 2 Synergy 3 150 150 150 Shoulder flexion Horizontal shoulder adduction Shoulder internal rotation 120 120 120 Elbow flexion

90 90 90

60 60 60 joint position [°] position joint joint position [°] position joint [°] position joint

30 30 30

0 0 0 0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000 5000 6000 Time [ms] Time [ms] Time [ms]

Figure C.2: Simulated movement of subjects 1 (left), 2 (middle) and 3 (right) for SYN task. Subject 1: 110° of pure shoulder flexion, no additional elbow movement. Subject 2 and 3: Reduced shoulder flexion with additional shoulder adduction, internal rotation and elbow flexion.

Procedure The participant behaved passively. First, all three arm movements were haptically presented to allow for comparison by the participant. Afterwards, each movement was presented three times and had to be rated for “selectivity” (i.e., ability to fractionate the movement) on a 6-point Likert scale (0=“not selective at all” to 5=“normally selective”) [101]. The assessment of the three different simulated arm movements was used to differentiate severity among the different subjects. 132 Appendix C. Remote Haptic Therapist-Patient Interaction

QOM Data acquisition Path accuracy and smoothness were used as in- dicators for quality of movement. To record the data, subjects were instructed to move a cursor (end-eﬀector of ARMin) as directly and smoothly as possible in a 2-DOF point-to-point reaching task on the graphic display. Path accuracy was calculated as the distance to path ratio [82]. A value of one represents a straight line; higher values im- ply a less accurate path. Movement smoothness was calculated as the arc length of the movement speed proﬁles’ normalized Fourier magnitude spectrum [10]. A smoothness value close to -2.8 was considered as “optimal”, lower values implied less smooth movement. An optimal trajectory was simulated and used as standard. Three trajectories of healthy subjects and a trajectory of a post-stroke subject with severe disability were recorded and haptically presented to the participant using the robot. The strong variance in duration of the healthy subjects’ movements is to be considered.

Procedure The participant was passive. First, the optimal trajectory was presented five times with visual feedback on the screen. Then, the subjects’ movements were presented haptically in randomized order, separated by a “washout”, presenting the optimal trajectory without visual feedback. The participant rated smoothness and movement accuracy on a 6-point Likert scale (0=“not at all” to 5=“normally smooth/accurate”). The assessment of the four different arm movements was used to differentiate severity among the different subjects.

C.3 Results and Discussion

We successfully tested the bidirectional control using two seven DOF exoskeleton robots in a teleassessment scenerio with therapists. The aim was not to enable the therapist to assess the patient’s motor function remotely. Our aim was to evaluate whether a therapist could feel the patient’s disability in his arm and used clinical assessment tools, to quantify this ”Beam-Me-In” strategy. We consciously limited the robotic feedback for the therapist to haptic feedback, not providing C.3. Results and Discussion 133 any numbers assessed by the ARMin rehabilitation robot. We showed that therapists could distinguish between different simulated movements of healthy subjects and patients post-stroke by means of the robot only, without directly touching the patient’s arm and regardless of the limited information provided by the robotic system. Thus, the “Beam-Me-In” strategy accounts for the therapist’s desire for haptic interaction as a component of hands-on therapy even with robotic technologies. The approach to “Beam-Me-In” was consistently rated positive. How- ever, most therapists only partly agreed on both that they could put themselves into the patients’ situation (i.e., reification) and that this allowed detecting the individual patient problems. The limited perceived reification may be explained by the mainly simulated performances in the four tasks. Nevertheless, “Beam-Me-In” was rated as a useful medium for assessment, therapy, teaching and learning during therapeutic education. It may give students insights into the clinical picture of a patient. Furthermore, ”Beam-Me-In” was seen as a suitable tool during telerehabilitation. Therefore, the ”Beam- Me-In” strategy has the potential to overcome reluctances towards robot-assisted rehabilitation that were presented in the introduction.

ROM

The number of correctly diﬀerentiated angles in ROM averaged 93.3 %. The mean absolute error in identifying each single angle averaged 4.9° with an absolute precision error of 6.5°. A robot can quantify ROM in a higher resolution than a therapist. A limitation of our study is that the limits of a subject’s movement in pROM were simulated by a simple spring-damper element at the patient limits, which did not consider biomechanical limitations, e.g., stretching of soft tissues and the resting tone of the muscles. Fur- thermore, the therapist could not inﬂuence the subject’s movement pattern during aROM assessment. By controlling the movement pattern (e.g. take more time to explore the limits) a therapist could have had more time to identify the angle. 134 Appendix C. Remote Haptic Therapist-Patient Interaction

RPM

The number of correctly scored mTS averaged 93.3 %. One participant did not identify the catch of the simulated severely affected subject 3 and was excluded for the evaluation of the catch angle quantification. The two way mixed effects model showed excellent intra-class correlation (according to Cicchetti (1994). The speed of movement is critical when assessing spasticity as both the joint angle and muscle reaction are velocity dependent. An increase in stretch velocity results in an increase in resistance to passive movement that we considered and implemented in our strategy [9]. Similar to a pROM assessment, guidance of the arm by the therapist and identification of a limitation in movement by the therapist is required for that assessment. Therefore, an automated interpretation by the robot is rather difficult, it demands therapist experience to react on the patients arm behavior. The “Beam-Me-In” strategy complements the clinical assessment with the possibility to assess RPM remotely.

SYN

All 15 participants could distinguish the severely affected, mildly affected and healthy subjects (all simulated). The quantification of the performances regarding severity illustrates the participants’ skill to distinguish between different movement synergies. The intra-class correlation was excellent (according to Cicchetti (1994)). Compared to end-effector based devices, exoskeleton rehabilitation devices provide measurements of single joints of a patient’s arm. Therefore, “Beam-Me-In” provides an excellent tool to measure and present arm synergies and further abnormal movement patterns.

QOM

The number of correctly diﬀerentiated QOM performances averaged 73.3 % for smoothness and 91.1 % for accuracy. The participants quantiﬁed the subjects’ smoothness and accuracy. The intra-class correlation was fair (according to Cicchetti (1994)). C.4. Conclusion and Current State 135

The two parameters smoothness and accuracy are hardly ever quantified in clinical routine. Unexpectedly, therapists were in average able to score smoothness and accuracy differentiating between slightly different movement patterns. Therefore, different movement patterns of different smoothness and accuracy can be haptically displayed by ARMin and interpreted by a therapist remotely using the ”Beam- Me-In” strategy. However, to increase the inter-rater reliability the backlash between human arm and the cuffs needs to be reduced. A therapist could not clearly say if the ”non-smooth” or ”non-accurate” movement is due to the subject’s performance or due to the participant’s own freedom to move within the robot. For optimal application of the ”Beam-Me-In” strategy, future redesigns of the ARMin robot should consider an undisturbed transfer of the movements between robot and human arm.

Intra-rater reliablity The diﬀerence was statistically not signiﬁcant (i.e., the null hypothesis could not be rejected) for all six angles of the pROM assessment.

C.4 Conclusion and Current State

The “Beam-Me-In” strategy allows for remote haptic interaction between the therapist and the patient. We could show that information about joint position, resistance to passive movement, inter-joint coordination, smoothness and accuracy during a point-to-point reaching task can be transferred to the therapist’s own arm and allows to assess these parameters. In particular, for the identification of abnormal movement patterns that need to be induced by passively moving the patient, “Beam-Me-In” offers a tool for remote assessment that is superior to the robot alone. For feasibility testing, we limited the resolution to provide patient impairments representing the entire patient population. As a next step, we would test the ”Beam-Me-In” strategy with higher resolution of abnormal movement patterns and also test the strategy with therapists and real patients in a clinical setting. However, for an clinical testing we had some safety consider- ations, such as instability, that could not have been solved with the 136 Appendix C. Remote Haptic Therapist-Patient Interaction existing hardware and control designs. Recent developments within our research group as the integration and investigation of further force sensors and the consideration of new motors may have built the basis to reconsider such a telerehabilitation system in arm therapy after stroke. We conclude that the “Beam-Me-In” strategy is a new opportunity to assess and train patients. The “Beam-Me-In” strategy offers a possibility to experience a new way of therapist-patient interaction. Therapists can subjectively assess movement characteristics of a subject via realistic haptic feedback through a seven degree-of-freedom exoskeleton. In combination with automated robot-assisted assessment the “Beam-Me-In” strategy may offer a complete tool to assess stroke patients remotely. The ”Beam-Me-In” strategy device has the potential to provide valu- able and sophisticated haptic feedback that will help address the barriers to implementing robot-assisted telerehabilitation. 137

Bibliography

[1] Web of Science [v.5.21] - All Databases Results. http://apps.webofknowledge.com/.

[2] Repetitive training of isolated movements improves the outcome of motor rehabilitation of the centrally paretic hand. Journal of the Neurological Sciences, 130(1):59–68, may 1995.

[3] Heart Disease and Stroke Statistics - 2014 Update: A report from the American Heart Association, volume 129. 2014.

[4] Louise Ada, Simone Dorsch, and Colleen G Canning. Strength- ening interventions increase strength and improve activity after stroke: a systematic review. The Australian journal of physio- therapy, 52(4):241–8, 2006.

[5] Ernest Adams. Fundamentals of Game Design. 2013.

[6] Barbara E. Ainsworth, William L. Haskell, Melicia C. Whitt, Melinda L. Irwin, Ann M. Swartz, Scott J. Strath, William L. O’Brian, David R. Bassett, Kathryn H. Schmitz, Patricia O. Emplaincourt, David R. Jacobs, and Arthur S. Leon. Com- pendium of Physical Activities: an update of activity codes and MET intensities. Medicine & Science in Sports & Exer- cise, 32(Supplement):S498–S516, 2000.

[7] Eckart Altenm¨ullerand Gottfried Schlaug. Apollo’s gift: New aspects of neurologic music therapy. Progress in Brain Re- search, 217:237–252, 2015. 138 Bibliography

[8] JuliàL. Amengual, Nuria Rojo, Misericordia Veciana de las Heras, Josep Marco-Pallarés, Jennifer Grau-Sánchez, Sabine Schneider, Luc´ıaVaquero, Montserrat Juncadella, Jordi Mon- tero, Bahram Mohammadi, Francisco Rubio, Nohora Rueda, Esther Duarte, Carles Grau, Eckart Altenmüller,Thomas F. Münte, and Antoni Rodr´ıguez-Fornells. Sensorimotor Plastic- ity after Music-Supported Therapy in Chronic Stroke Patients Revealed by Transcranial Magnetic Stimulation. PLoS ONE, 8(4):1–10, 2013.

[9] Noureddin Nakhostin Ansari, Soofia Naghdi, Scott Hasson, Mo- hammad Hasan Azarsa, and Somaye Azarnia. The Modified Tardieu Scale for the measurement of elbow flexor spasticity in adult patients with hemiplegia. Brain Injury, 22(13-14):1007– 1012, 2008.

[10] Sivakumar Balasubramanian, Alejandro Melendez-Calderon, Agnes Roby-Brami, and Etienne Burdet. On the analysis of movement smoothness. Journal of NeuroEngineering and Re- habilitation, 12(1):112, 2015.

[11] Bel´enRubio Ballester, Sergi Berm´udezI Badia, and Paul F M J Verschure. Including Social Interaction in Stroke VR-Based Motor Rehabilitation Enhances Performance : A Pilot Study. 21(4):490–501, 2012.

[12] Elaine Barnett-Page and James Thomas. Methods for the synthesis of qualitative research: a critical review. BMC Med Res Methodol, 9:59, 2009.

[13] Marie Christine Bastarache-Roberge, Pierre Majorique Léger, Fran¸coisCourtemanche, Sylvain Sénécal,and Marc Fredette. Measuring flow using psychophysiological data in a multiplayer gaming context. In Lecture Notes in Information Systems and Organisation, volume 10, pages 187–191. 2015.

[14] Kilian Baur, Verena Klamroth-Marganska, Chiara Giorgetti, Daniela Fichmann, and Robert Riener. Performance-based viscous force ﬁeld adaptation in upper limb strength training for Bibliography 139

stroke patients. Proceedings of the IEEE RAS and EMBS In- ternational Conference on Biomedical Robotics and Biomecha- tronics, 2016-July:864–869, 2016.

[15] Kilian Baur, Nina Rohrbach, Joachim Hermsd¨orfer, Robert Riener, and Verena Klamroth-Marganska. The ”Beam-Me-In Strategy” - Remote Haptic Therapist-Patient Interaction with Two Exoskeletons for Stroke Therapy. Journal of NeuroEngi- neering and Rehabilitation, 2018.

[16] Kilian Baur, Alexandra Sch¨attin,Eling D. de Bruin, Robert Riener, Jaime E. Duarte, and Peter Wolf. Trends in robot- assisted and virtual reality-assisted neuromuscular therapy: A systematic review of health-related multiplayer games. Journal of NeuroEngineering and Rehabilitation, 2018.

[17] Kilian Baur, Peter Wolf, Verena Klamroth-Marganska, Walter Bierbauer, Urte Scholz, Robert Riener, and Jaime E. Duarte. Robot-supported multiplayer rehabilitation: Feasibility study of haptically linked patient-spouse training. IEEE International Conference on Intelligent Robots and Systems, 2018.

[18] Kilian Baur, Peter Wolf, Domen Novak, Dana Boering, Sab- rina H¨orner,Corinna Dahlen, Joachim Berger, Robert Riener, and Volker H¨omberg. Competitive versus cooperative forms of gaming in motor training with (sub-)acute stroke patients. (in preparation), 2019.

[19] Kilian Baur, Peter Wolf, Robert Riener, and Jaime Duarte. Making neurorehabilitation fun: Multiplayer training via damping forces balancing diﬀerences in skill levels. Proceedings of the 2017 IEEE International Conference on Rehabilitation Robotics, 2017.

[20] Kent C. Berridge, Terry E. Robinson, and J. Wayne Aldridge. Dissecting components of reward: ’liking’, ’wanting’, and learning. Current Opinion in Pharmacology, 9(1):65–73, 2009.

[21] Walter Bierbauer, Jennifer Inauen, Sabine Schaefer, Maike Margarethe Kleemeyer, Janina L¨uscher, Claudia 140 Bibliography

König,Robert Tobias, Matthias Kliegel, Andreas Ihle, Lukas Zimmerli, Barbara M. Holzer, Klarissa Siebenhuener, Edouard Battegay, Christian Schmied, and Urte Scholz. Health Behav- ior Change in Older Adults: Testing the Health Action Process Approach at the Inter- and Intraindividual Level. Applied Psychology: Health and Well-Being, 9(3):324–348, 2017. [22] Richard W. Bohannon. Muscle strength and muscle training after stroke. Journal of Rehabilitation Medicine, 39(1):14–20, 2007. [23] Gunnar Borg. Borg’s perceived exertion and pain scales. Human kinetics, 1998. [24] Joke Bradt, Wendy L. Magee, Cheryl Dileo, Barbara L. Wheeler, and Emer McGilloway. Music therapy for acquired brain injury. Cochrane database of systematic reviews (Online), (7):CD006787, 2010. [25] Michael Brainin, Natan Bornstein, Gudrun Boysen, and Vida Demarin. Acute neurological stroke care in Europe: results of the European Stroke Care Inventory. European Journal of Neurology, 7(1):5–10, jan 2000. [26] Adam M. Brandenburger and Barry J. Nalebuff. Co-Opetition: A Revolution Mindset That Combines Competition and Coop- eration. 1996. [27] Bambi R. Brewer, Sharon K. McDowell, and Lise C. Worthen- Chaudhari. Poststroke upper extremity rehabilitation: a review of robotic systems and clinical results. Topics in stroke rehabilitation, 14(6):22–44, 2007. [28] J. C. Brocklehurst, P. Morris, K. Andrews, B. Richards, and P. Laycock. Social effects of stroke. Social science & medicine. Part A, Medical sociology, 15(1):35–9, jan 1981. [29] David A. Brown, Timothy D. Lee, David J. Reinkensmeyer, and Jaime E. Duarte. Designing Robots That Challenge to Op- timize Motor Learning, pages 39–58. Springer International Publishing, Cham, 2016. Bibliography 141

[30] William M. Bukowski, Betsy Hoza, and Michel Boivin. Bukowski et al. (1994).Measuring friendship quality during pre and early adolescence.pdf, 1994. [31] Craig R. Carignan and Hermano I. Krebs. Telerehabilitation robotics: Bright lights, big future? The Journal of Rehabilita- tion Research and Development, 43(5):695, 2006. [32] Carl J. Caspersen, Kenneth E. Powell, and Gregory M. Chris- tenson. Physical activity, exercise, and physical fitness: definitions and distinctions for health-related research. Public health reports, 100(2):126–131, 1985. [33] Glauco A. P. Caurin, Adriano A. G. Siqueira, Kleber O. An- drade, Ricardo C. Joaquim, and Hermano I. Krebs. Adaptive strategy for multi-user robotic rehabilitation games. Conference proceedings : ... Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineer- ing in Medicine and Biology Society. Conference, 2011(1):1395– 8, jan 2011. [34] Guillaume Chanel, Julien Kronegg, Didier Grandjean, and Thierry Pun. Peripheral Physiological Signals Emotion As- sessment : Arousal Evaluation Using EEG ’ s and Peripheral Physiological Signals Computing Science Center , University of Geneva. Mulimedia Content Representation, Classification and Security., 2006. [35] Michael J. Chen, Xitao Fan, and Sondra T. Moe. Criterion- related validity of the Borg ratings of perceived exertion scale in healthy individuals: a meta-analysis. Journal of Sports Sci- ences, 20(11):873–899, 2002. [36] Roberto Colombo, Fabrizio Pisano, Alessandra Mazzone, Car- men Delconte, Silvestro Micera, M Chiara Carrozza, Paolo Dario, and Giuseppe Minuco. Design strategies to improve patient motivation during robot-aided rehabilitation. Journal of neuroengineering and rehabilitation, 4:3, jan 2007. [37] Mario Cortese, Marco Cempini, Paulo Rogerio De Almeida Ribeiro, Surjo R. Soekadar, Maria Chiara Carrozza, and Nicola 142 Bibliography

Vitiello. A Mechatronic System for Robot-Mediated Hand Tel- erehabilitation. IEEE/ASME Transactions on Mechatronics, 20(4):1753–1764, 2015. [38] Mihaly Csikszentmihalyi. Play and intrinsic rewards. In Flow and the Foundations of Positive Psychology: The Collected Works of Mihaly Csikszentmihalyi, pages 135–153. 2014. [39] Edward L. Deci, Wayne F. Cascio, and Judith Krusell. Cogni- tive evaluation theory and some comments on the Calder and Staw critique, 1975. [40] Julius P.A. Dewald and Randall F. Beer. Abnormal joint torque patterns in the paretic upper limb of subjects with hemiparesis. Muscle and Nerve, 24(2):273–283, 2001. [41] Damien Djaouti, Julian Alvarez, and Jean-Pierre Jessel. Clas- sifying serious games: The G/P/S model. Handbook of research on improving learning and motivation through educational games: Multidisciplinary approaches, (2005):118–136, 2011. [42] Steve Walker Douglas Bates, Martin Maechler, Ben Bolker. Fit- ting Linear Mixed-Eﬀects Models Using lme4. Journal of Sta- tistical Software, 67(1):1–48, 2015. [43] Sara H Downs and Nick Black. The feasibility of creating a checklist for the assessment of the methodological quality both of randomised and non-randomised studies of health care interventions. J Epidemiol Community Health, 52(6):377–384, 1998. [44] Jaime E. Duarte, Kilian Baur, and Robert Riener. Flowing to the optimal challenge: An adaptive challenge framework for multiplayer games. Biosystems and Biorobotics, 15:381–385, 2017. [45] Minh Duc Duong, Kazuhiko Terashima, and Takanori Miyoshi. A novel stable teleoperation with haptic feedback by means of impedance adjustment via arbitrary time delay environment for rehabilitation. Proceedings of the IEEE International Confer- ence on Control Applications, pages 1744–1749, 2009. Bibliography 143

[46] Denis Ertelt, Steven Small, Ana Solodkin, Christian Dettmers, Adam McNamara, Ferdinand Binkofski, and Giovanni Buccino. Action observation has a positive impact on rehabilitation of motor deficits after stroke. NeuroImage, 36(SUPPL. 2):T164– T173, 2007. [47] Deborah L. Feltz, Brandon Irwin, and Norbert Kerr. Two- player partnered exergame for obesity prevention: using discrepancy in players’ abilities as a strategy to motivate physical activity. Journal of diabetes science and technology, 6(4):820–7, 2012. [48] Deborah L. Feltz, Norbert L. Kerr, and Brandon C. Irwin. Buddy Up: The Kohler Effect Applied to Health Games. Jour- nal of Sport & Exercise Psychology, 33(4):506–526, 2011. [49] Tamar Flash and Neville Hogan. The coordination of arm movements: an experimentally confirmed mathematical model. The Journal of neuroscience, 5(7):1688–1703, 1985. [50] Eletha Flores, Gabriel Tobon, Ettore Cavallaro, Francesca I. Cavallaro, Joel C. Perry, and Thierry Keller. Improving patient motivation in game development for motor deficit rehabilitation. Proceedings of the 2008 International Conference in Advances on Computer Entertainment Technology - ACE ’08, page 381, 2008. [51] Marshal F. Folstein, Susan E. Folstein, and Paul R. Mchugh. Mini-Mental State : A Practical Method for Grading the Cog- nitive State of Patients for the Clinician. International Journal of Geriatric Psychiatry, 12(3):189–198, 1975. [52] Nizan Friedman, Vicky Chan, Andrea N Reinkensmeyer, Ariel Beroukhim, Gregory J Zambrano, Mark Bachman, and David J Reinkensmeyer. Retraining and assessing hand movement after stroke using the MusicGlove: comparison with conventional hand therapy and isometric grip training. Journal of Neuro- Engineering and Rehabilitation, 11(1):76, 2014. [53] Nizan Friedman, Vicky Chan, Danny Zondervan, Mark Bach- man, and David J. Reinkensmeyer. MusicGlove: Motivating 144 Bibliography

and Quantifying Hand Movement Rehabilitation by using Func- tional Grips to Play Music. 33rd Annual International Confer- ence of the IEEE MBS, pages 2359–2363, 2011.

[54] Axel R. Fugl-Meyer, Lisbeth Jääskö,Ingegerd Leyman, Sigyn Olsson, and Solveig Steglind. The post-stroke hemiplegic patient. 1. a method for evaluation of physical performance. Scan- dinavian journal of rehabilitation medicine, 7(1):13–31, 1975.

[55] Brian J. Gajadhar, Yvonne A. W. de Kort, and Wijnand A. Ijsselsteijn. Shared Fun Is Doubled Fun : Player Enjoyment as a Function of Social Setting. pages 106–117, 2008.

[56] G. Ganesh, A. Takagi, R. Osu, T. Yoshioka, M. Kawato, and E. Burdet. Two is better than one: physical interactions improve motor performance in humans. Scientiﬁc reports, 4:3824, jan 2014.

[57] Maria Ursula Garczarek. Classiﬁcation rules in standardized partition spaces. page 234, 2002.

[58] Line Gebauer and Peter Vuust. Music interventions in health care. White Paper, 2014.

[59] Lewis R. Goldberg, John A. Johnson, Herbert W. Eber, Robert Hogan, Michael C. Ashton, C. Robert Cloninger, and Harri- son G. Gough. The international personality item pool and the future of public-domain personality measures. Journal of Research in Personality, 40(1):84–96, 2006.

[60] Maja Gorˇsiˇc,Imre Cikajlo, Nika Goljar, and Domen Novak. A multisession evaluation of an adaptive competitive arm rehabilitation game. Journal of NeuroEngineering and Rehabilitation, 14(1):128, 2017.

[61] Maja Gorˇsiˇc,Imre Cikajlo, and Domen Novak. Competitive and cooperative arm rehabilitation games played by a patient and unimpaired person : eﬀects on motivation and exercise intensity. Journal of NeuroEngineering and Rehabilitation, 14(1):1– 18, 2017. Bibliography 145

[62] Maja Gorˇsiˇc,Ali Darzi, and Domen Novak. Comparison of two difficulty adaptation strategies for competitive arm rehabilitation exercises. IEEE International Conference on Rehabilita- tion Robotics, 2017-July:1–6, 2017. [63] Maja Gorˇsiˇc,Minh Ha Tran, and Domen Novak. Cooperative Cooking : A Novel Virtual Environment for Upper Limb Re- habilitation *. In EMBC 2018 Conference Proceedings, 2018. [64] Carolyn Gowland, Paul Stratford, Maureen Ward, Julie More- land, Wendy Torresin, Sandra Van Hullenaar, Julie Sanford, Susan Barreca, Bernadette Vanspall, and Nancy Plews. Mea- suring physical impairment and disability with the chedoke- mcmaster stroke assessment. Stroke, 1993. [65] Mark A. Guadagnoli and Timothy D. Lee. Challenge Point: A Framework for Conceptualizing the Effects of Various Practice Conditions in Motor Learning. Journal of Motor Behavior, 36(2):212–224, 2004. [66] Marco Guidali, Alexander Duschau-Wicke, Simon Broggi, Ver- ena Klamroth-Marganska, Tobias Nef, and Robert Riener. A robotic system to train activities of daily living in a virtual environment. Medical and Biological Engineering and Computing, 49(10):1213–1223, oct 2011. [67] Marco Guidali, Alexander Duschau-Wicke, Melanie Büchel, Andreas Brunschweiler, Tobias Nef, and Robert Riener. Path Control – A Strategy for Patient-Cooperative Arm Rehabilita- tion. 8. Workshop: Automatisierungstechnische Verfahren für die Medizin 2009, pages 20–21, 2009. [68] A. Haugh, A. Pandyan, and G. Johnson. A systematic review of the Tardieu Scale for the measurement of spasticity. Disability and Rehabilitation, 28(15):899–907, 2006. [69] Andrew Heller, Derick T. Wade, Victorine A. Wood, Alan Sun- derland, Richard Langton Hewer, and Elizabeth Ward. Arm function after stroke: Measurement and recovery over the first three months. Journal of Neurology, Neurosurgery and Psychi- atry, 50(6):714–719, 1987. 146 Bibliography

[70] H. Hénon,I. Durieu, D. Guerouaou, F. Lebert, F. Pasquier, and D. Leys. Poststroke dementia: incidence and relationship to prestroke cognitive decline. Neurology, 57(7):1216–1222, 2001. [71] Thomas Hermann. Taxonomy and Definitions for Sonification and Auditory Display. Proceedings of the 14th International Conference on Auditory Display, pages 1–8, 2008. [72] Guido Hertel, Norbert L. Kerr, and Lawrence A. Messé.Moti- vation gains in performance groups: paradigmatic and theoretical developments on the Köhlereffect. Journal of personality and social psychology, 79(4):580–601, 2000. [73] Caisa Hofgren, A. Björkdahl,E. Esbjörnsson,and K. Stibrant- Sunnerhagen. Recovery after stroke: Cognition, ADL function and return to work. Acta Neurologica Scandinavica, 115(2):73– 80, 2007. [74] Neville Hogan, Hermano I. Krebs, Brandon Rohrer, Jerome J. Palazzolo, Laura Dipietro, Susan E. Fasoli, Joel Stein, Richard Hughs, Walter R. Frontera, Daniel Lynch, and Bruce T. Volpe. Motions or muscles? Some behavioral factors underlying robotic assistance of motor recovery. The Journal of Re- habilitation Research and Development, 43(5):605, 2006. [75] John R. Hollenbeck, Howard J. Klein, Anne M. O’Leary, and Patrick M. Wright. Investigation of the construct validity of a self-report measure of goal commitment. Journal of Applied Psychology, 74(6):951–956, 1989. [76] John Houston, Paul Harris, Sandra McIntire, and Dientje Fran- cis. Revising the Competitiveness Index Using Factor Analysis. Psychological Reports, 90(1):31–34, 2002. [77] Vincent S Huang and John W Krakauer. Robotic neurorehabilitation: a computational motor learning perspective. Journal of neuroengineering and rehabilitation, 6:5, 2009. [78] NathanaëlJarrassé,Themistoklis Charalambous, and Etienne Burdet. A framework to describe, analyze and generate interactive motor behaviors. PloS one, 7(11):e49945, jan 2012. Bibliography 147

[79] Michelle J. Johnson, Rui C. V. Loureiro, and William S. Har- win. Collaborative tele-rehabilitation and robot-mediated therapy for stroke rehabilitation at home or clinic. Intelligent Ser- vice Robotics, 1(2):109–121, jan 2008. [80] Jeanne D. Johnston, Lee Sheldon, and Anne P. Massey. Influ- encing physical activity and healthy behaviors in college students: lessons from an alternate reality game. In Serious Game Design and Development: Technologies for Training and Learn- ing, pages 270–289. 2010. [81] Kazumi Kawahira, Megumi Shimodozono, Atsuko Ogata, and Nobuyuki Tanaka. Addition of intensive repetition of facilita- tion exercise to multidisciplinary rehabilitation promotes motor functional recovery of the hemiplegic lower limb. Journal of Re- habilitation Medicine, 36(4):159–164, 2004. [82] Urs Keller, Sabine Schölch, Urs Albisser, Claudia Rudhe, Armin Curt, Robert Riener, and Verena Klamroth-Marganska. Robot-assisted arm assessments in spinal cord injured patients: A consideration of concept study. PLoS ONE, 10(5):e0126948, 2015. [83] Robert Kerr and Michael F. Walsh. Movement time in an un- derwater environment. Journal of Motor Behavior, 5(3):175– 178, 1973. [84] Verena Klamroth-Marganska. Three-dimensional, task-oriented robot therapy. Number Table 1. Elsevier Ltd., 2018. [85] Verena Klamroth-Marganska, Javier Blanco, Katrin Campen, Armin Curt, Volker Dietz, Thierry Ettlin, Morena Felder, Bernd Fellinghauer, Marco Guidali, Anja Kollmar, Andreas Luft, Tobias Nef, Corina Schuster-Amft, Werner Stahel, and Robert Riener. Three-dimensional, task-specific robot therapy of the arm after stroke: A multicentre, parallel-group randomised trial. The Lancet Neurology, 13(2):159–166, feb 2014. [86] E. J. Koeneman, R. S. Schultz, S. L. Wolf, D. E. Herring, and J. B. Koeneman. A pneumatic muscle hand therapy device. En- gineering in Medicine and Biology Society, 2004. IEMBS’04. 148 Bibliography

26th Annual International Conference of the IEEE, 4:2711– 2713, 2004.

[87] Chia Lin Koh, Shin Liang Pan, Jiann Shing Jeng, Bang Bin Chen, Yen Ho Wang, I. Ping Hsueh, and Ching Lin Hsieh. Predicting recovery of voluntary upper extremity movement in subacute stroke patients with severe upper extremity paresis. PLoS ONE, 10(5):1–15, 2015.

[88] Otto K¨ohler. Uber¨ den Gruppenwirkungsgrad der menschlichen K¨orperarbeit und die Bedingung optimaler Kollektivkraftreak- tion. Industrielle Psychotechnik, 1927.

[89] Johannes Konert, Stefan G¨obel, and Ralf Steinmetz. Video Game Personalization via Social Media Participation. In Games for Training, Education, Health and Sports, pages 35– 46. Springer, 2014.

[90] Hermano I Krebs, N Hogan, B T Volpe, M L Aisen, L Edelstein, and C Diels. Overview of clinical trials with MIT-MANUS: a robot-aided neuro-rehabilitation facility. Technology and health care: oﬃcial journal of the European Society for Engineering and Medicine, 7(6):419–423, 1998.

[91] Gert Kwakkel. Impact of intensity of practice after stroke: Issues for consideration. Disability and Rehabilitation, 28(13- 14):823–830, 2006.

[92] Gert Kwakkel, Boudewijn. J. Kollen, Jeroen van der Grond, and Arie J.H. Prevo. Probability of Regaining Dexterity in the Flaccid Upper Limb: Impact of Severity of Paresis and Time Since Onset in Acute Stroke. Stroke, 34(9):2181–2186, 2003.

[93] Gert Kwakkel, Robert C. Wagenaar, Tim W. Koelman, Gus- taaf J. Lankhorst, and Johan C. Koetsier. Eﬀects of intensity of rehabilitation after stroke: A research synthesis. Stroke, 28(8):1550–1556, aug 1997.

[94] J. Richard Landis and Gary G. Koch. An application of hi- erarchical kappa-type statistics in the assessment of majority Bibliography 149

agreement among multiple observers. Biometrics, 33(2):363, 1977.

[95] Catherine E. Lang, Keith R. Lohse, and Rebecca L. Birken- meier. Dose and timing in neurorehabilitation: Prescribing motor therapy after stroke. Current Opinion in Neurology, 8(6):549–555, 2015.

[96] Catherine E. Lang, Michael J Strube, Marghuretta D. Bland, Kimberly J. Waddell, Kenddra M. Cherry-Allen, Randolph J. Nudo, Alexander W. Dromerick, and Rebecca L. Birkenmeier. Dose-response of task-speciﬁc upper limb training in people at least 6 months post stroke: A Phase II, single-blind, randomized, controlled trial. Annals of Neurology, pages 1–13, 2016.

[97] Jessica Lanini, Toshiaki Tsuji, Peter Wolf, Robert Riener, and Domen Novak. Teleoperation of two six-degree-of-freedom arm rehabilitation exoskeletons. In IEEE International Conference on Rehabilitation Robotics, volume 2015-Septe, pages 514–519, 2015.

[98] Kate E. Laver, Daniel Schoene, Maria Crotty, Stacey George, Natasha A. Lannin, and Catherine Sherrington. Telerehabil- itation services for stroke. Cochrane Database of Systematic Reviews, 2013(12), 2013.

[99] James A Levine. Measurement of energy expenditure. Public health nutrition, 8(7A):1123–1132, 2005.

[100] Alessandro Liberati, Douglas G. Altman, Jennifer Tetzlaﬀ, Cynthia Mulrow, Peter C. Gøtzsche, John P.A. Ioannidis, Mike Clarke, P. J. Devereaux, Jos Kleijnen, David Moher, P C Gotzsche, John P.A. Ioannidis, Mike Clarke, P. J. Devereaux, Jos Kleijnen, and David Moher. The PRISMA statement for reporting systematic reviews and meta-analyses of studies that evaluate healthcare interventions: explanation and elaboration. BMJ, 339:b2700, 2009.

[101] Rensis Likert. A technique for the measurement of attitudes. Archives of psychology, 140:1–55, 1932. 150 Bibliography

[102] Martin Lotze, Christoph Braun, Niels Birbaumer, Silke Anders, and Leonardo G. Cohen. Motor learning elicited by voluntary drive. Brain, 126(4):866–872, 2003.

[103] Peter S. Lum, Charles G. Burgar, and Peggy C. Shor. Evi- dence for strength imbalances as a signiﬁcant contributor to abnormal synergies in hemiparetic subjects. Muscle and Nerve, 27(2):211–221, 2003.

[104] Michael Mace, Nawal Kinany, Paul Rinne, Anthony Rayner, Paul Bentley, and Etienne Burdet. Balancing the playing ﬁeld: collaborative gaming for physical training. Information Sci- ences, (September), 2017.

[105] Niall Maclean, Pandora Pound, Charles Wolfe, and Anthony Rudd. The Concept of Patient Motivation: A Qualitative Anal- ysis of Stroke Professionals’ Attitudes. Stroke, 33(2):444–448, feb 2002.

[106] Martina Maier, Bel´enRubio Ballester, Esther Duarte, Armin Duﬀ, and Paul F M J Verschure. Social Integration of Stroke Patients through the Multiplayer Rehabilitation Gaming Sys- tem. pages 100–114, 2014.

[107] Thomas. W. Malone. Toward a Theory of Intrinsically Mo- tivating Instruction. Cognitive Science: A Multidisciplinary Journal, 5(4):333–369, 1981.

[108] Laura Marchal-Crespo and David J Reinkensmeyer. Review of control strategies for robotic movement training after neurologic injury. Journal of neuroengineering and rehabilitation, 6:20, jan 2009.

[109] Arwen M. Marker and Amanda E. Staiano. Better Together: Outcomes of Cooperation Versus Competition in Social Ex- ergaming. Games for Health Journal, 4(1):141017093454003, oct 2014.

[110] Edward McAuley, Terry Duncan, and Vance V. Tammen. Psy- chometric properties of the Intrinsic Motivation Inventory in a Bibliography 151

competitive sport setting: a confirmatory factor analysis. Re- search quarterly for exercise and sport, 60(1):48–58, 1989. [111] Jan Mehrholz, Marcus Pohl, Thomas Platz, Joachim Kugler, and Bernhard Elsner. Electromechanical and robot-assisted arm training for improving activities of daily living, arm function, and arm muscle strength after stroke. The Cochrane database of systematic reviews, (11):CD006876, 2015. [112] MatjaˇzMihelj, Domen Novak, Maja Milavec, Jaka Ziherl, An- drej Olenˇsek,and Marko Munih. Virtual Rehabilitation En- vironment Using Principles of Intrinsic Motivation and Game Design. Presence: Teleoperators and Virtual Environments, 21(1):1–15, feb 2012. [113] Laura C. Miller and Julius P A Dewald. Involuntary paretic wrist/finger flexion forces and EMG increase with shoulder abduction load in individuals with chronic stroke. Clinical Neu- rophysiology, 123(6):1216–1225, 2012. [114] Roxana Moreno and Richard Mayer. Interactive Multi- modal Learning Environments. Educational Psychology Review, 19(3):309–326, 2007. [115] Florian ’Floyd’ Mueller, Martin R. Gibbs, and Frank Vetere. Taxonomy of exertion games. Proceedings of the 20th Aus- tralasian Conference on Computer-Human Interaction Design- ing for Habitus and Habitat - OZCHI ’08, pages 263–266, 2008. [116] Tobias Nef, Marco Guidali, and Robert Riener. ARMin III - arm therapy exoskeleton with an ergonomic shoulder actuation. Applied Bionics and Biomechanics, 6(2):127–142, jul 2009. [117] Tobias Nef, Matjaz Mihelj, and Robert Riener. ARMin: a robot for patient-cooperative arm therapy. Medical & biological engineering & computing, 45(9):887–900, sep 2007. [118] Nurdiana Nordin, Sheng Quan Xie, and Burkhard Wünsche. Assessment of movement quality in robot- assisted upper limb rehabilitation after stroke: a review. Journal of neuroengineering and rehabilitation, 11:137, 2014. 152 Bibliography

[119] Domen Novak, Aniket Nagle, Urs Keller, and Robert Riener. Increasing motivation in robot-aided arm rehabilitation with competitive and cooperative gameplay. Journal of neuroengineering and rehabilitation, 11(1):64, 2014. [120] Julie A. Nugent, Karl A. Schurr, and Roger D. Adams. A dose- response relationship between amount of weight-bearing exercise and walking outcome following cerebrovascular accident. Archives of Physical Medicine and Rehabilitation, 75(4):399– 402, 1994. [121] Roberto Oboe, Omar A. Daud, Stefano Masiero, Fabio Oscari, and Giulio Rosati. Development of a haptic teleoperation system for remote motor and functional evaluation of hand in patients with neurological impairments. International Workshop on Advanced Motion Control, AMC, pages 518–523, 2010. [122] Fabio Oscari, Roberto Oboe, Omar Andres Daud Albasini, Ste- fano Masiero, and Giulio Rosati. Design and construction of a bilateral haptic system for the remote assessment of the stiﬀ- ness and range of motion of the hand. Sensors (Switzerland), 16(10):1633, 2016. [123] Michelle M. Ouellette, Nathan K. Lebrasseur, Jonathan F. Bean, Edward Phillips, Walter R. Frontera, and Roger A. Fielding. High-Intensity Resistance Training Improves Muscle Strength, Self-Reported Function, and Disability in Long-Term Stroke Survivors. Stroke, 35:1404–9, 2004. [124] Joseph B. Oxendine. Emotional arousal and motor performance. Quest, 13(1):23–32, 1970. [125] Hyung Soon Park, Qiyu Peng, and Li Qun Zhang. A portable telerehabilitation system for remote evaluations of impaired el- bows in neurological disorders. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 16(3):245–254, 2008. [126] Carolynn Patten, Jan Lexell, and Heather E Brown. Weakness and strength training in persons with poststroke hemiplegia: rationale, method, and eﬃcacy. Journal of rehabilitation research and development, 41(3A):293–312, 2004. Bibliography 153

[127] Wei Peng and Julia Crouse. Playing in parallel: the eﬀects of multiplayer modes in active video game on motivation and physical exertion. Cyberpsychology, behavior and social networking, 16(6):423–7, jun 2013.

[128] Wei Peng and Gary Hsieh. The inﬂuence of competition, cooperation, and player relationship in a motor performance centered computer game. Computers in Human Behavior, 28(6):2100–2106, nov 2012.

[129] Jos´eC Pinheiro and Douglas M Bates. Mixed-Eﬀects Models in S and S-PLUS. 2000.

[130] Karolien Poels, Wijnand Ijsselsteijn, and Yvonne De Kort. De- velopment of the Kids Game Experience Questionnaire. Pro- ceedings of Meaningful Play . . . , 2008.

[131] Maurice R. Puyau, Anne L. Adolph, Firoz A. Vohra, Issa Zak- eri, and Nancy F. Butte. Prediction of activity energy expenditure using accelerometers in children. Medicine and Science in Sports and Exercise, 36(9):1625–1631, 2004.

[132] Alfredo Raglio. Eﬀects of music and music therapy on mood in neurological patients. World Journal of Psychiatry, 5(1):68, 2015.

[133] Brandon Rohrer, Susan Fasoli, Hermano Igo Krebs, Richard Hughes, Bruce Volpe, Walter R. Frontera, Joel Stein, and Neville Hogan. Movement smoothness changes during stroke recovery. The Journal of Neuroscience: The Oﬃcial Journal of the Society for Neuroscience, 22(18):8297–8304, 2002.

[134] Morris Rosenberg. Conceiving the self. RE Krieger, 1986.

[135] Trevor G Russell. Physical rehabilitation using telemedicine. Journal of telemedicine and telecare, 13(5):217–220, 2007.

[136] Richard M. Ryan and Edward L. Deci. Intrinsic and Extrinsic Motivations: Classic Deﬁnitions and New Directions. Contem- porary Educational Psychology, 25(1):54–67, 2000. 154 Bibliography

[137] James F. Sallis, Robin B. Pinski, Robin M. Grossman, Thomas L. Patterson, and Philip R. Nader. The development of self-eﬃcacy scales for healthrelated diet and exercise behaviors. Health Education Research, 3(3):283–292, 1988.

[138] R. Sanchez, D. Reinkensmeyer, P. Shah, J. Liu, S. Rao, R. Smith, S. Cramer, T. Rahman, and J. Bobrow. Monitoring functional arm movement for home-based therapy after stroke. The 26th Annual International Conference of the IEEE Engi- neering in Medicine and Biology Society, 4:4787–4790, 2004.

[139] Robert J. Sanchez, Jiayin Liu, Sandhya Rao, Punit Shah, Robert Smith, Tariq Rahman, Steven C. Cramer, James E. Bobrow, and David J. Reinkensmeyer. Automating arm movement training following severe stroke: Functional exercises with quantitative feedback in a gravity-reduced environment. IEEE Transactions on Neural Systems and Rehabilitation Engineer- ing, 14(3):378–389, 2006.

[140] Teppo Särkämöand David Soto. Music listening after stroke: Beneficial effects and potential neural mechanisms. Annals of the New York Academy of Sciences, 1252(1):266–281, 2012.

[141] K. Sathian, Laurel J. Buxbaum, Leonardo G. Cohen, John W. Krakauer, Catherine E. Lang, Maurizio Corbetta, and Susan M. Fitzpatrick. Neurological Principles and Rehabilitation of Ac- tion Disorders: Common Clinical Deﬁcits. Neurorehabilitation & Neural Repair, 25:21S–32S, 2011.

[142] Andrew Sawers and Lena H Ting. Perspectives on human- human sensorimotor interactions for the design of rehabilitation robots. Journal of neuroengineering and rehabilitation, 11(1):142, 2014.

[143] S. Schneider, P. W. Schönle,E. Altenmüller,and T. F. Münte. Using musical instruments to improve motor skill recovery following a stroke. Journal of Neurology, 254(10):1339–1346, 2007.

[144] Urte Scholz, Sibylle Ochsner, Pamela Rackow, Rainer Hornung, Gertraud Stadler, and Nina Knoll. Examining the relationship Bibliography 155

between daily changes in support and smoking around a self-set quit date. Health Psychology, 35(5):514–517, 2016. [145] Ralf Schwarzer. Modeling health behavior change: How to predict and modify the adoption and maintenance of health behaviors. Applied Psychology, 57(1):1–29, 2008. [146] Sudha Seshadri, Alexa Beiser, Margaret Kelly-Hayes, Carlos S. Kase, Rhoda Au, William B. Kannel, and Philip A. Wolf. The lifetime risk of stroke: Estimates from the framingham study. Stroke, 37(2):345–350, 2006. [147] Falko F. Sniehotta, Urte Scholz, and Ralf Schwarzer. Bridging the intention-behaviour gap: Planning, self-efficacy, and action control in the adoption and maintenance of physical exercise. Psychology and Health, 20(2):143–160, 2005. [148] Aiguo Song, Changcheng Wu, Dejing Ni, Huijun Li, and Hongyun Qin. One-Therapist to Three-Patient Telerehabili- tation Robot System for the Upper Limb after Stroke. Inter- national Journal of Social Robotics, 8(2):319–329, 2016. [149] Hayeon Song, Jihyun Kim, Kelly E. Tenzek, and Kwan Min Lee. The effects of competition and competitiveness upon intrinsic motivation in exergames. Computers in Human Behav- ior, 29(4):1702–1708, 2013. [150] Steven James Spencer. Movement training and post-stroke rehabilitation using a six degree of freedom upper-extremity robotic orthosis and virtual environment. University of Cali- fornia, Irvine, 2011. [151] Amanda E. Staiano, Anisha A. Abraham, and Sandra L. Calvert. Motivating Effects of Cooperative Exergame Play for Overweight and Obese Adolescents. Journal of Diabetes Sci- ence and Technology, 6(4):812–819, jul 2012. [152] Amanda E. Staiano, Anisha A. Abraham, and Sandra L. Calvert. Adolescent exergame play for weight loss and psychosocial improvement: A controlled physical activity intervention. Obesity, 21(3):598–601, 2013. 156 Bibliography

[153] A Sunderland, D Tinson, L Bradley, and R L Hewer. Arm function after stroke. An evaluation of grip strength as a measure of recovery and a prognostic indicator. Journal of neurology, neurosurgery, and psychiatry, 52(11):1267–72, 1989. [154] A Sunderland, D J Tinson, E L Bradley, D Fletcher, R Lang- ton Hewer, and D T Wade. Enhanced physical therapy improves recovery of arm function after stroke. A randomised controlled trial. Journal of neurology, neurosurgery, and psychiatry, 55(7):530–5, jul 1992. [155] Heidi Sveistrup. Motor rehabilitation using virtual reality. Journal of NeuroEngineering and Rehabilitation, 1:1–8, 2004. [156] Penelope Sweetser and Peta Wyeth. GameFlow: A Model for Evaluating Player Enjoyment in Games. Comput. Entertain., 3(3):3–3, 2005. [157] Hossein Taheri, Justin B Rowe, David Gardner, Vicki Chan, Kyle Gray, Curtis Bower, David J Reinkensmeyer, and Eric T Wolbrecht. Design and preliminary evaluation of the FINGER rehabilitation robot: controlling challenge and quantifying ﬁn- ger individuation during musical computer game play. Journal of neuroengineering and rehabilitation, 11(1):10, 2014. [158] Kazumoto Tanaka, Jim Parker, Graham Baradoy, Dwayne Sheehan, John R Holash, and Larry Katz. A Comparison of Exergaming Interfaces for Use in Rehabilitation Programs and Research. Loading, 6(9):69–81, 2012. [159] R Team. R Development Core Team. R: A Language and Environment for Statistical Computing, 55:275–286, 2013. [160] Michael H. Thaut, Gerald C. McIntosh, and Volker Hoem- berg. Neurobiological foundations of neurologic music therapy: Rhythmic entrainment and the motor system. Frontiers in Psy- chology, 6(FEB):1–6, 2015. [161] Deborah Theodoros and Trevor Russell. Telerehabilitation: Current Perspectives. Studies in health technology and infor- matics, 131:191–209, 2008. Bibliography 157

[162] Gergory T. Thielman, Catherine M. Dean, and a.M. Gen- tile. Rehabilitation of reaching after stroke: Task-related training versus progressive resistive exercise. Archives of Physical Medicine and Rehabilitation, 85(10):1613–1618, oct 2004.

[163] James Thomas and Angela Harden. Methods for the thematic synthesis of qualitative research in systematic reviews. BMC Med Res Methodol, 8:45, 2008.

[164] Kunio Tsurumi, Toru Itani, Norihide Tachi, Toshimasa Takan- ishi, Hatsuko Suzumura, and Hidemaro Takeyama. Estima- tion of energy expenditure during sedentary work with upper limb movement. JOURNAL OF OCCUPATIONAL HEALTH, 44(6):408–413, 2002.

[165] Andrea Turolla, Mauro Dam, Laura Ventura, Paolo Tonin, Michela Agostini, Carla Zucconi, Pawel Kiper, Annachiara Cagnin, and Lamberto Piron. Virtual reality for the rehabilitation of the upper limb motor function after stroke: a prospective controlled trial. Journal of neuroengineering and rehabilitation, 10(1):85, jan 2013.

[166] Thomas E. Twitchell. The restoration of motor function following hemiplegia in man. Brain, 74(4):443–480, 1951.

[167] Aharon Tziner, Richard E. Kopelman, and Neomi Livneh. Ef- fects of performance appraisal format on perceived goal characteristics, appraisal process satisfaction, and changes in rated job performance: A ﬁeld experiment. Journal of Psychology: Interdisciplinary and Applied, 127(3):281–291, 1993.

[168] Martin Ulrich, Johannes Keller, Klaus Hoenig, Christiane Waller, and Georg Gr¨on. Neural correlates of experimentally induced ﬂow experiences. NeuroImage, 86:194–202, 2014.

[169] Unity. Unity - Game Engine, 2017.

[170] Sanne C. Van Der Pas, Jeanine A. Verbunt, Dorien E. Breuke- laar, Rachma Van Woerden, and Henk A. Seelen. Assessment of arm activity using triaxial accelerometry in patients with 158 Bibliography

a stroke. Archives of Physical Medicine and Rehabilitation, 92(9):1437–1442, 2011.

[171] R P S Van Peppen, G Kwakkel, S Wood-Dauphinee, H J M Hendriks, Ph J Van der Wees, and J Dekker. The impact of physical therapy on functional outcomes after stroke: what’s the evidence? Clinical rehabilitation, 18(8):833–862, 2004.

[172] Janne M. Veerbeek, Anneli C. Langbroek-Amersfoort, Erwin E. H. van Wegen, Carel G. M. Meskers, and Gert Kwakkel. Eﬀects of Robot-Assisted Therapy for the Upper Limb After Stroke. Neurorehabilitation and Neural Repair, 31(2):107–121, 2017.

[173] Katrien Verhoeven, Vero Vanden Abeele, Brent Gers, and Jan Seghers. Energy Expenditure During Xbox Kinect Play in Early Adolescents: The Relationship with Player Mode and Game Enjoyment. Games for Health Journal, 4(6):444–451, 2015.

[174] D T Wade, R Langton-Hewer, V A Wood, C E Skilbeck, and H M Ismail. The hemiplegic arm after stroke: measurement and recovery. Journal of neurology, neurosurgery, and psychiatry, 46(6):521–524, 1983.

[175] Xibo Wang and Jiting Li. A tele-rehabilitation system with bilateral haptic feedback to both the therapist and the patient via time-delay environment. 2013 World Haptics Conference, WHC 2013, pages 331–334, 2013.

[176] David Weibel, BartholomäusWissmath, Stephan Habegger, Yves Steiner, and Rudolf Groner. Playing online games against computer- vs. human-controlled opponents: Effects on presence, flow, and enjoyment. Computers in Human Behavior, 24(5):2274–2291, sep 2008.

[177] David Weibel, Bartholom¨ausWissmath, and Fred W Mast. Immersion in Mediated Environments: The Role of Person- ality Traits. CyberPsychology, Behavior & Social Networking, 13(3):251–256, 2010. Bibliography 159

[178] B. J. Winer. Statistical Principles in Experimental Design. 1962.

[179] Frieder Wittmann, Olivier Lambercy, Roman R. Gonzenbach, Mark A. Van Raai, Raphael Hover, Jeremia Held, Michelle L. Starkey, Armin Curt, Andreas Luft, and Roger Gassert. Assessment-driven arm therapy at home using an IMU-based virtual reality system. IEEE International Conference on Re- habilitation Robotics, 2015-Septe:707–712, 2015.

[180] Hartwig Woldag, Guido Waldmann, Gabriele Heuschkel, and Horst Hummelsheim. Is the repetitive training of complex hand and arm movements beneﬁ cial for motor recovery in stroke patients? Clinical Rehabilitation, 17(7):723–730, 2003.

[181] Elizabeth J. Woytowicz, Jeremy C. Rietschel, Ronald N. Good- man, Susan S. Conroy, John D. Sorkin, Jill Whitall, and Sandy McCombe Waller. Determining Levels of Upper Extremity Movement Impairment by Applying a Cluster Analysis to the Fugl-Meyer Assessment of the Upper Extremity in Chronic Stroke. In Archives of Physical Medicine and Rehabilitation, volume 98, pages 456–462, 2017.

[182] Xiaotian Wu, Peter Guarino, Albert C. Lo, Peter Peduzzi, and Michael Wininger. Long-term Eﬀectiveness of Intensive Ther- apy in Chronic Stroke. Neurorehabilitation and Neural Repair, 35, 2015.

[183] Gabriele Wulf and Rebecca Lewthwaite. Optimizing performance through intrinsic motivation and attention for learning: The OPTIMAL theory of motor learning. Psychonomic Bul- letin & Review, 23(5):1382–1414, 2016.

[184] Fu-Yun Yu, Chialing Han, and Tak-Wai Chan. Experimen- tal comparisons of face-to-face and anonymous real-time team competition in a networked gaming learning environment. Cy- berpsychology & behavior : the impact of the Internet, multime- dia and virtual reality on behavior and society, 11(4):511–514, 2008. 160 Bibliography

[185] Lukas Zimmerli, Mario Jacky, Lars L¨unenburger, Robert Riener, and Marc Bolliger. Increasing patient engagement during virtual reality-based motor rehabilitation. Archives of physical medicine and rehabilitation, 94(9):1737–46, sep 2013. [186] Andrea Zimmermann-Schlatter, Corina Schuster, Milo A Puhan, Ewa Siekierka, and Johann Steurer. Eﬃcacy of motor imagery in post-stroke rehabilitation: a systematic review. Journal of NeuroEngineering and Rehabilitation, 5(1):8, 2008.