sensors

Article Electronic Wearable Sensors for Detecting –Pelvic Movements

Yuxin Zhang 1 , Pari Delir Haghighi 1 , Frada Burstein 1,* , Lim Wei Yap 2 , Wenlong Cheng 2,*, Lina Yao 3 and Flavia Cicuttini 4 1 Faculty of Information Technology, Monash University, Melbourne, VIC 3145, Australia; [email protected] (Y.Z.); [email protected] (P.D.H.) 2 Department of Chemical Engineering, Monash University, Melbourne, VIC 3800, Australia; [email protected] 3 School of Computer Science and Engineering, The University of New South Wales, Sydney, NSW 2052, Australia; [email protected] 4 Department of Epidemiology and Preventive Medicine, School of Public Health and Preventive Medicine, Monash University, Melbourne, VIC 3004, Australia; fl[email protected] * Correspondence: [email protected] (F.B.); [email protected] (W.C.); Tel.: +61-3-9903-2011 (F.B.); +61-3-9905-3147 (W.C.)  Received: 31 January 2020; Accepted: 5 March 2020; Published: 9 March 2020 

Abstract: Background: A nanomaterial-based electronic-skin (E-Skin) wearable sensor has been successfully used for detecting and measuring body movements such as finger movement and pressure. The ultrathin and highly sensitive characteristics of E-Skin sensor make it a suitable alternative for continuously out-of-hospital lumbar–pelvic movement (LPM) monitoring. Monitoring these movements can help medical experts better understand individuals’ low experience. However, there is a lack of prior studies in this research area. Therefore, this paper explores the potential of E-Skin sensors to detect and measure the anatomical angles of lumbar–pelvic movements by building a linear relationship model to compare its performance to clinically validated inertial measurement unit (IMU)-based sensing system (ViMove). Methods: The paper first presents a review and classification of existing wireless sensing technologies for monitoring of body movements, and then it describes a series of experiments performed with E-Skin sensors for detecting five standard LPMs including flexion, extension, pelvic tilt, lateral flexion, and rotation, and measure their anatomical angles. The outputs of both E-Skin and ViMove sensors were recorded during each experiment and further analysed to build the comparative models to evaluate the performance of detecting and measuring LPMs. Results: E-Skin sensor outputs showed a persistently repeating pattern for each movement. Due to the ability to sense minor skin deformation by E-skin sensor, its reaction time in detecting lumbar–pelvic movement is quicker than ViMove by ~1 s. Conclusions: E-Skin sensors offer new capabilities for detecting and measuring lumbar–pelvic movements. They have lower cost compared to commercially available IMU-based systems and their non-invasive highly stretchable characteristic makes them more comfortable for long-term use. These features make them a suitable sensing technology for developing continuous, out-of-hospital real-time monitoring and management systems for individuals with .

Keywords: E-Skin sensors; movement detection; wireless sensing technology; monitoring

1. Introduction Low back pain (LBP) is the leading cause of disability among individuals [1,2]. Lumbar–pelvic movements have been identified as major risk factors for developing LBP as well as indicators of LBP rehabilitation progress [3–5]. Assessment of how individuals perform lumbar–pelvic movements

Sensors 2020, 20, 1510; doi:10.3390/s20051510 www.mdpi.com/journal/sensors Sensors 2020, 20, 1510 2 of 28 during the rehabilitation progress could assist with identifying functional limitations and physical predictors of low back pain, and monitoring treatment outcomes [4]. The ability to assess these movements in a non-invasive and sensitive way in real-time and out-of-hospital could improve LBP prevention and treatment [3,4]. According to recent studies [6,7], there are significant differences between individuals with and without LBP for how the same lumbar–pelvic movement is performed, based on speed, angle range (range of movement), acceleration, and intensity. These features can be used to characterise each lumbar–pelvic movement for individuals with and without LBP. Assessment of the anatomical angle such as the angle of back extension or flexion is essential for the detection and measurement of lumbar–pelvic movements. Recent growth and advances in mobile and wireless sensing technologies present an unprecedented opportunity for collecting physiological signals and body movement data in mobile healthcare and diagnostic medicine. Due to this advancement, a number of sensing, systems including vision-based systems [8] and IMUs-based systems that can monitor the angle of body movements, have emerged in the market [9]. One common characteristic of these systems is to compute the angle by sensing coordinates of different points on a in a three-dimensional space. This data can also be used to calculate the speed, acceleration, and direction of the angle. Vision-based systems can provide precise and accurate measurements, but their use is limited to the laboratory setting. Additionally, their costs are relatively high. The IMU-based systems can be used for out-of-hospital monitoring. Yet, they require the application of multiple IMUs to provide a relatively accurate and precise measurement. They also need wired or wireless modules for the communication of each IMU. This can lead to a bulky and rigid packaging, which can be very uncomfortable for individuals to wear over a long period of time. Moreover, the accuracy and precision of measurement of the gyroscope in these systems cannot be guaranteed if the speed of movement is very slow. Electronic-skin (E-Skin) wearable sensors [10] have recently gained popularity due to their ultrathin, flexible, stretchable, self-healing, skin mimicking characteristics which enables them to perform mechanical, chemical, and biophysiological sensing while also being able to attach comfortably to human skin [10–12]. A plethora of materials, such as conductive polymers, metallic nanomaterials, ionic liquids, liquid metals, and conductive textile, have been exploited by researchers in fabrication of novel electronic [13,14]. Among the materials mentioned above, high aspect ratio one-dimension (1D) synthetic nanomaterials (such as metallic nanowire, metal oxide nanowire, and carbon nanotubes [14,15]) could be used to develop highly sensitive and highly stretchable soft electronic skin wearable gauge sensor [16–19]. A number of demonstrations on the capability of these high sensitivity nanomaterial-based E-Skin wearable sensors have been reportedly used for monitoring pulse signal, detecting , finger, and movement [13,16–18]. Since these sensors can detect minor strains and skin deformations [16–19], they are not only capable of measuring the angle of anatomical joints, but also of measuring minor movements of human body that cannot be detected by MEMS-based sensors such as inertial measurement unit (IMU). With the ability to attach conformally to human skin, E-Skin sensors can be used to detect subtle human motions such as lumbar–pelvic movements [19–22]. Lumbar–pelvic movements consist of flexion, extension, anterior posterior pelvic tilt, lateral flexion, and rotation. Detecting and measuring these movements can significantly help medical experts to collect detailed information to understand the relationship between physical movements and low back pain from a personalised perspective. This paper investigates the feasibility of using the state-of-the-art E-Skin sensors to detect five standard lumbar–pelvic movements including flexion, extension, pelvic tilt, lateral flexion, and rotation, and measure the change of angles during each movement. Compared to vision-based and IMU-based systems, E-Skin sensors are more comfortable to wear and offer a higher sensitivity to precisely detect subtle movements [22]. However, there is limited research reported on the use and validation of E-skin sensors for detecting and measuring lumbar–pelvic movements. In this paper, we introduce a method to measure and analyse the patterns of E-Skin sensor outputs (i.e., skin deformation data) for detecting the lumbar–pelvic movements. To the best of our knowledge, we are the first to carry out Sensors 2020, 20, 1510 3 of 28 a detailed experimental study on using E-Skin sensors for lumbar–pelvic movement detection and measuring. The E-Skin sensor we used has been developed by the Monash University NanoBionics Group (Melbourne, VIC, Australia) [20]. To validate the performance of E-Skin sensor on detecting and measuring lumbar–pelvic movements, ViMove (DorsaVi, Melbourne, VIC, Australia) was used as a golden standard. In previous report [6], 34 participants, including 18 with LBP and 16 without LBP, were recruited to perform different lumbar–pelvic movements with both ViMove and Vicon systems. The results showed that there was a clinically acceptable level of agreement between these two systems in terms of measuring lumbar–pelvic movements. The contributions of this paper can be summarised as the following: (1) a comprehensive review and comparative analysis of existing sensing technologies and systems for measuring and detecting lumbar–pelvic movements, (2) introducing a method to measure and detect lumbar–pelvic movements including flexion, extension, anterior posterior pelvic tilt, lateral flexion, and rotation, and their anatomical angles by using E-Skin sensors, and (3) providing recommendations for future research based on our experimental study. This paper is organised as follows. In Section2, existing wireless sensing technologies-based systems, which can be used for monitoring lumbar–pelvic movements, are reviewed based on our classification of existing sensing technologies. In Section3, the design of experiments is described. In Section4, results of the exploratory experiment are discussed. In Section5, the conclusion is presented.

2. Classification of Sensing Technologies for Body Motion Detection Existing sensing technologies for detecting and measuring body movements can be classified into three main categories: vision-based systems, IMUs, and flexible sensors.

2.1. Vision-Based Sensing Systems Vision-based systems normally consist of high-speed cameras and reflection markers or 3D cameras [6]. Their operation highly depends on adjusting camera settings, selecting proper lighting conditions, and the use of video/image processing algorithms. Vision-based systems can be further divided into two categories of marker-based systems and marker-less systems.

2.1.1. Marker-Based System Marker-based systems require reflection or transmitting markers for the measurement of physical movement. It uses infrared optical or high-speed cameras to detect the light reflection of the markers and measures activity using the computation of the markers’ trajectories in a three-dimensional space [23]. Based on different settings (e.g., the camera numbers and marker protocols), the accuracy of the marker-based system can be close to 0.1 mm and the sampling frequency can be as high as 1000 Hz [24]. This type of systems have been used for many studies such as measuring movements [25] and robotic motion monitoring [26]. Vicon is one of the world-renowned vision-based motion tracking systems which has been used in a variety of body motion related research and clinical studies, such as [8,27]. The advantage of the Vicon system is its ability to precisely reconstruct physical movements using markers. In [28], the Vicon system is proved to have a high accuracy in terms of measuring spinal and mobility. It is also the gold standard used to validate the performance of other motion tracking systems, including inertial measurement unit (IMU)-based measurement systems and mark-less vison systems [6]. Similar products have also been utilised for studies related to different physical movements, such as Optotrak Certus motion tracking system [29], MotionAnalysis system [30], and CODA motion analysis system [31]. Generally, the functions of these commercialised marker-based vision systems are similar. These systems provide a clinical-level solution for real-time measurement of physical activity. The high precision of the measurement outperforms other systems, but they have a number of limitations. Firstly, Sensors 2020, 20, 1510 4 of 28 the hardware and software are expensive. Secondly, installation of the system is time-consuming and complicated, particularly for the calibration of the camera position and light condition. Thirdly, it requires laboratory settings, which limit its use in measuring daily activities. Fourthly, the soft tissue artefacts, including incorrect marker positioning and skin sliding over the bones, may affect the measurement accuracy of marker-based systems [23].

2.1.2. Marker-Less System Marker-less vision-based systems use multi-cameras, IR sensors, or RGB-D cameras in the analysis of body movements or postures. The analyses could be done with a single image or video clips and do not require the attachment of markers to individuals. Microsoft Kinect is a well-known commercialised marker-less motion-capturing system [32]. It consists of a wide-angle motion sensing camera with an IR sensor to construct the depth information of the image/video. It is primarily designed for tracking joint movements of the human body [33]. Nevertheless, the performance of Kinect is limited by its latency and tracking range [34]. The organic motion system is a similar mark-less vision system with a wide tracking range compared to Kinect [33]. KinaTrax is another powerful marker-less vision-based posture analysis system with multiple camera set-ups that can adequately cover an area size of a baseball court [35]. In general, the marker-less based systems are relatively cheaper and easier to setup than the marker-based systems [35]. However, they provide lower accuracy for real-time movement tracking. Additionally, the real-time processing performance of the marker-less system is also not as good as the other sensing systems such as IMU-based systems [36]. With the development of the RGB-D cameras and image/video processing technology, the marker-less based vision systems may capture more complex movements and postures analysis in the future. Currently, these systems need sophisticated lighting conditions to achieve a better performance, so they are not suitable for continuous monitoring in different environments [36].

2.2. Inertial Measurement Unit (IMU) IMU consists of accelerometers, gyroscopes, and occasionally magnetometers [9]. With the development of microelectronic technology, these sensors have become smaller and easier to wear with high accuracy of measuring body movements [37]. Based on the number and type of the sensors, they can be further classified into IMU sensor systems and garment integrated sensor systems. IMU sensor systems include single and dual IMU-based systems. The systems which contain more than two IMU sensors are categorised under the multiple sensory systems.

2.2.1. IMU Sensor Systems Single IMU sensor-based systems are mostly designed to monitor the posture and movement of an individual by placing the sensor on the or lower back area [38]. Most of these systems are bio-feedback systems, such as Spineangel [39], Lumo Lift [40], and Upright posture trainer [41], designed to help improve the awareness of individuals of their sitting posture or inappropriate movements. One of their limitations is the difficulty in measuring complex physical movements and generating detailed information. Thus, combining IMU sensor with Vision-based sensors has become necessary for real-time sports training tasks [36]. Unlike the single IMU sensor system, the dual IMU sensor-based systems enhance the possibility of measuring complex movements. Considering that the single IMU device is a single point that indicates the lumbar–pelvic area, it can only reflect a single point movement in three-dimensional spaces. However, the lumbar–pelvic area is a soft surface which moves simultaneously with the spine. Thus, the movement of the two attached points in the dual IMU sensor-based device can be used to simulate the movement of the spine and indicate the movement of the lumbar–pelvic area. Sensors 2020, 20, 1510 5 of 28

Several dual IMU sensor-based systems, designed to measure the lumbar–pelvic movements, exist. Most of these systems developed for rehabilitation purposes include ViMove [42], Valedo Motion [43], and RIABLO [44]. ViMove is a commercialised movement monitoring device developed by DorsaVi [42]. It consists of two IMU sensors, two surface EMG sensors, and one remote recording and feedback device (RFD). It can measure multiple movements including the lumbar–pelvic area. Additionally, studies have been conducted on the concurrent validity of ViMove for measuring lumbar region inclination motion compared with Vicon [6]. ViMove is also used in several other lower back pain related clinical trials [45,46]. In contrast to Valedo and RIABLO, ViMove offers a means of monitoring daily activities, which highly enhances its mobility. The raw data of the sensors can be accessed through its application. Yet, it does not provide direct access to the data captured through APIs in real-time. In general, the single IMU sensor-based systems are useful for monitoring spinal postures and lumbar–pelvic movements. By using different advanced algorithms, the IMU can process human motion data in real-time with good accuracy [47]. Although, numbers of studies have proposed solution for solving the drift issues in IMU system, the drift error still exists due to the nature of IMU [48–50]. The dual IMU sensor-based systems are highly effective in monitoring lumbar–pelvic movements. Although these systems do not require complex laboratory settings like the vision-based systems, they still provide accurate measurement. Presently, the dual IMU sensor-based system is preferable because of its ability to monitor and measure complex lumbar–pelvic movements. Nonetheless, they are uncomfortable for long-time use because of their weight and rigidity.

2.2.2. Multiple Sensory System (Garment Integrated Sensor System) Multiple Sensory systems (Garment integrated sensor systems) use different types of sensors such as temperature, IMU, ECG, or EMG sensors, which are integrated into garments (e.g., tops, vests, or sportswear). The integration of these sensors usually creates a small wireless network, known as body sensor network (BSN), body area network (BAN), or wireless body area network (WBAN). The IMUs are the sensors, which are used to detect and measure physical movements. The multi-IMU sensor-based garment systems are designed for the measurement of three-dimensional human movement. They commonly have 8 to 12 IMU sensor nodes attached to garments and located on different parts of the human body including the head, , , , waist, , and foot [51–53]. Since these systems contain multiple nodes, they can measure different physical activities based on the movements of different body parts. For example, Zishi is a garment-based sensing system for trunk posture monitoring [54]. The garment is integrated with accelerometers and gyroscopes which make it capable of detect static and dynamic trunk movements with high accuracy (<1.5 degree). However, these garment integrated sensors are expensive, non-washable, and difficult to structure because of the quantity of the IMU sensors [53]. Furthermore, their design has to be personalised for different individuals such as different height and body measures. The purpose of garment sensing systems is to use multiple sensors together without compromising mobility. The advantage is their capacity for data fusion from distinct types of sensors which gather different data. However, the sweat and body movements from the daily use of garments may trigger repositioning of the garment and cause errors in the measurement. The calibration of this type of system is also very complicated [55].

2.3. Flexible Sensors Flexible sensors refer to sensors which can be flexed or stretched while the body moves. The flexible sensors are small and easy to integrate with a garment or other skin-contact substrates. There are two types of flexible sensors. One is plastic optical fibre sensor, which has been used in spine posture measurements [56,57]. The mechanism of these sensors is to monitor the bend of the structural beams. However, the plastic optical fibre is not soft enough to be attached to the human body and move as the Sensors 2020, 20, 1510 6 of 28 body moves. Therefore, it is better for measuring anatomical joint movements rather than detecting complex movements such as lumbar–pelvic movements. The other type flexible sensors are the resistive flex sensors [58]. The principle of resistive flex sensors is variable resistive characteristics of a conductive material [59]. This type of sensor consists of three components: the contact, sensor film, and substrate [60,61]. The contact is connected to the circuit board which detects the changing current. The sensor film is made of a conductive material. The resistance changes with the deformation of the conductive material. The substrate is the middle layer between sensor film and human skin. The intrinsic properties of the substrate determine the flexibility of the sensor. The E-Skin sensor is part of the resistive flex sensors. With the advance in nanotechnology, the sensitivity and stretchability of conductive material has been significantly improved. E-Skin sensors are now capable of measuring the slightest movement of the human body. With flexible integrated circuit boards (such as central processing unit and Bluetooth component), the sensory data can be wirelessly transmitted to a smartphone for further analysis [62]. Compared to the IMUs and Vision-based systems, the E-Skin sensor is cheaper, softer, and smaller, which makes it a suitable alternative for detecting and measuring physical movements. Because of its high sensitivity, most of studies have used E-Skin sensors to measure subtle body movements such as the movement of fingers [63,64]. Additionally, human skin is not a plane surface, but the E-Skin sensor can be easily attached to human skin unlike other existing sensors. Therefore, the soft tissue artefacts can be avoided by using E-Skin sensor. The commercially available stretch sensor, StretchSense, is a flexible strain gauge sensor like the E-Skin sensor we used in this study [65]. However, the StretchSense sensor utilizes capacitive strain sensing to measure the degree of strain [66]. Capacitive-based strain gauge sensor has a 5-layered sandwich structure, a dielectric layer sandwiched between two electrode layers and protected by two protective layers [66]. Compared to StretchSense, the E-Skin sensor utilizes piezoresistive strain sensing that consists of a 3-layered structure, in which a stretchable conductive layer is protected by two protective layers (see Figure1). Therefore, the E-Skin sensor is much thinner than the Stretch sense sensor and more comfortable to wear, with better accuracy and higher sensitivity. Additionally, the E-Skin sensor measures the change in electrical resistance so the circuit could be designed simpler compared to stretch sensors. It only needs a micro-control unit (MCU) with a 12-bit Analog-digital convertor (ADC) to read the changes in resistance. In [67], a multichannel conductive liquid-based strain sensor has been introduced. The results show the good performance (mean absolute errors <8 degree) in tracking joint angles during the gait cycles. However, the conductive liquid-based sensors may suffer from leakage problems, which makes them not suitable for wearable applications. The thickness of conductive liquid-based sensors is also much higher than E-Skin sensor. Moreover, it needs to be mounted on a sensing suit for motion measurement, which makes it less accurate comparing to the sensors which are directly attached on human body. Then, [68] introduced a similar strain sensor for human motion detection. Unlike E-Skin sensor, it uses a stretchable carbon nanotube to measure the 1D strain. The results show it is capable of detecting different human motions, but the performance of measurement was not reported. Additionally, this type of sensors also needs to be mounted on stockings or a sensing suit.

2.4. Summary Based on the review presented, it appears that vision-based systems provide the most accurate and precise monitoring tool for measuring human body motion. However, they are very expensive, and require proper camera settings and lighting conditions [27]. They can be only used in a lab environment, which limits their use for continuous out-of-hospital monitoring. IMU sensor-based systems are the most common used sensors for out-of-lab physical motion monitoring [69]. IMU sensors need to be attached to human body using stickers. However, the packaging material of the sensors are usually made of a rigid plastic that makes them inflexible and bulky [46]. The deformation of the skin may affect the repositioning of stickers when an individual performs physical activities, and this Sensors 2020, 20, 1510 7 of 28 could lead to an inaccurate measurement [70]. Additionally, for monitoring lumbar–pelvic movements, the placement of the sensor is around the waist of individual. The belt or edge of the clothes (such as jean trousers) has the potential to affect the measurement if it touches the sensors during the movements. To achieve accurate results, normally it is necessary to use at least two sensors to detect complex body movements such as lumbar–pelvic movements [42]. Therefore, the IMU sensor-based system can be problematic for long-term monitoring. Sensors 2020, 20, x FOR PEER REVIEW 8 of 28

Figure 1. (a) Schematic Illustration of E-Skin Sensor Fabrication; (b) E-Skin Sensor Dimension; (c) E- Figure 1. (a) SchematicSkin Electronic Illustration Module Description. of E-Skin Sensor Fabrication; (b) E-Skin Sensor Dimension; (c) E-Skin Electronic Module Description. The mechanical property of the fabricated E-skin sensor was studied by straining the sensor from 10% to 100% with increment of 10% using Thorlabs’ motorised linear translation stage and the Garment-integratedchange in electrical sensing resistance systems was measured are another and recorded option using to attach VersaSTAT multiple 4 Potentiostat sensors on human body. However,Galvanostat. they are As heavy shown andin Figure uncomfortable 2a, the signal record toed wear upon straining for along the E-skin period sensor ofis relatively time and are usually non-washable [71consistent.]. They The can strain-resistance also suffer re fromsponse thecurve same for E-skin repositioning sensor under different problem range that of physical stickers have when strain was also investigated and plotted in Figure 2b. The gauge factor (𝐺𝐹) of the E-skin sensor was an individual performscalculated using movements Equation (1), [53 wh].ere R is the electrical resistance of the E-skin recorded upon E-Skin sensorsstretching, (resistive Ro is the initial flex sensors)electrical resistance provide of E-skin a suitable sensor at solutionrest, and ɛ isfor the degree detecting of strain and measuring applied to E-skin sensor. The gauge factor was calculated based on the data plotted in Figure 2a. The lumbar–pelvic movements.gauge factor of the Compared strain sensor divided to vision-based into two sections. systems, At the lower they strain are of much10% to 30%, cheaper, the and do not require a great dealgauge of factor effort of the in sensor setting was upfound the to be equipment. 0.309 whereas at Unlike a higher other strain of flexible 30% to 100%, sensors the gauge [57 ], they are soft factor was 0.122, which is less sensitive than the E-skin sensor at lower strain. Based on the gauge and stretchable enoughfactor calculated, to attach the E-skin to any sensor human was found skin to surfacehave the ability conformally. of measuring This the strain reduces at a wide the repositioning problem of IMUrange sensors. of 10% Theyto 100%, can which also is particularly detect the useful slightest for application movement such as monitoring of the skin bending which angle can be used to determine the movementof human body and posture subcomponents [45,46]. The durability of theof the movements. E-Skin sensors was Based tested onby conducting this comparative 1000 analysis, cycles of strain test at 30% strain, which is shown in Figure 2c. The outputs of the sensor fluctuate we concluded thatslightly E-Skin in the initial sensors stage and have stabilised a greatafter 1000 potential cycles of strain. for After out-of-lab 1000 cycles, the lumbar–pelvic E-Skin sensor movement detection and measuringperformance remained which stable is a and focus was ofable this to fully paper. return to A its detailed original value comparison when being released of existing sensing after strain (shown in Figure 2d), which proves its durability and reliability. Hence, prior to the trial, systems for humanthe freshly motion fabricated detection E-Skin sensors and measurement will always be pre-strained is given for in 1000 Table cycles1. as a “learning” process for the E-Skin sensors. Since the E-Skin sensor uses graphite microflakes hybrid conductive network,Table its performance 1. Comparison will not be of affected Existing by an Humany chemical Motion reactions Sensing such as oxidation Systems. during the lumbar–pelvic movement detection [73]. However, to ensure the sensing accuracy, each E-Skin sensor was onlyVison-Based used once System on each participant for one Inertial experiment Measurement session. Unit Flex Sensor Marker-based Marker-less Multiple Conductive Technology Single IMU1 𝑅−𝑅 Dual IMUs Optical Fiber Systems Systems 𝐺𝐹 = ∙ Sensory (1) Material 𝜀 𝑅 Vicon; CODA; Lumo Lift; Strechsense; Products Kinect; KinaTrax ViMove; Valedo ZiShi [34,35] Optotrak Upright E-Skin Structural Skin Measurements 3D Coordinates beams Deformation Mobility In-lab Limited Range Out-of-lab Continuous Low Cost; Continuous Long-term out-of-lab High Precision Continuous out-of-lab Garment-integrable; Monitoring; Easy to Setup; monitoring; Pros and Accuracy; out-of-lab monitoring; Long-term Lower Cost; Less Expensive Measure Gold Standard monitoring; Multi-sensory monitoring Measure complex human Easy to Setup data fusion complex activity motions In-Lab Setting; Soft tissue Expensive; Limited Range; artefacts; Durability; Measure simple Measure simple Sensitive to Sensitive to Expensive; Non-washable; Complex Cons human motions; activity; lighting lighting Bulky Complex modelling Lower Accuracy Complex setting conditions; Soft conditions calibration process tissue artefacts process Sensors 2020, 20, 1510 8 of 28

3. Materials and Methods We have designed and fabricated E-Skin sensors for detecting and measuring all the five types of lumbar–pelvic movements, including flexion, extension, pelvic tilt, lateral flexion, and rotation. The following subsections discuss the details of these experiments. This research has been approved by Monash University Human Research Ethics Committee.

3.1. Instrumentation and Data Collection E-Skin sensors are fabricated based on a graphite microflake hybrid conductive network-based strain sensor and the illustration of the fabrication process is shown in Figure1a. In short, copper nanowires ink and graphite microflake ink were formulated and painted in U-shape on latex substrate. A silver paste is applied on both ends of the U-shaped strain sensor and attached to the electronic module. Lastly, the strain sensor is encapsulated with liquid latex and the whole device is packaged between two layers of Elastoplast stretchable sports kinesiology tape. The sports kinesiology tape is hypoallergenic and can be directly attached to human skin. The length and width of the fabricated device is 10 cm and 2.5 cm, respectively, with the maximum thickness of the device being 0.5 cm as shown in Figure1b. As shown in Figure1c, the core of the electronic module is an nRF51822 Bluetooth Low Energy (BLE) system-on-chip from Nordic semiconductor [72]. A voltage regulator is used to regulate input voltage to 3.3 V to power the BLE SoC (System on a Chip) and the strain sensor. The other terminal of the strain sensor is connected to the analog pin of the BLE SoC. The mechanical property of the fabricated E-skin sensor was studied by straining the sensor from 10% to 100% with increment of 10% using Thorlabs’ motorised linear translation stage and the change in electrical resistance was measured and recorded using VersaSTAT 4 Potentiostat Galvanostat. As shown in Figure2a, the signal recorded upon straining the E-skin sensor is relatively consistent. The strain-resistance response curve for E-skin sensor under different range of physical strain was also investigated and plotted in Figure2b. The gauge factor ( GF) of the E-skin sensor was calculated using Equation (1), where R is the electrical resistance of the E-skin recorded upon stretching, Ro is the initial electrical resistance of E-skin sensor at rest, and ε is the degree of strain applied to E-skin sensor. The gauge factor was calculated based on the data plotted in Figure2a. The gauge factor of the strain sensor divided into two sections. At the lower strain of 10% to 30%, the gauge factor of the sensor was found to be 0.309 whereas at a higher strain of 30% to 100%, the gauge factor was 0.122, which is less sensitive than the E-skin sensor at lower strain. Based on the gauge factor calculated, the E-skin sensor was found to have the ability of measuring the strain at a wide range of 10% to 100%, which is particularly useful for application such as monitoring bending angle of human body posture [45,46]. The durability of the E-Skin sensors was tested by conducting 1000 cycles of strain test at 30% strain, which is shown in Figure2c. The outputs of the sensor fluctuate slightly in the initial stage and stabilised after 1000 cycles of strain. After 1000 cycles, the E-Skin sensor performance remained stable and was able to fully return to its original value when being released after strain (shown in Figure2d), which proves its durability and reliability. Hence, prior to the trial, the freshly fabricated E-Skin sensors will always be pre-strained for 1000 cycles as a “learning” process for the E-Skin sensors. Since the E-Skin sensor uses graphite microflakes hybrid conductive network, its performance will not be affected by any chemical reactions such as oxidation during the lumbar–pelvic movement detection [73]. However, to ensure the sensing accuracy, each E-Skin sensor was only used once on each participant for one experiment session.

1 R R GF = − 0 (1) ε· R0 Sensors 2020, 20, 1510 9 of 28 Sensors 2020, 20, x FOR PEER REVIEW 9 of 28

FigureFigure 2.2. ((aa)) PlotPlot ofof resistanceresistance changechange ofof E-skinE-skin sensorsensor underunder anan appliedapplied strainstrain inin thethe rangerange ofof 10%10% toto 100%100% withwith incrementincrement ofof 10%10% strainstrain (b(b)) Strain-response Strain-response plots plots for for the the E-skin E-skin sensor. sensor. (c(c)) Durability Durability testtest ofof 10001000 cyclescycles 10%10% strainstrain atat aa frequencyfrequency ofof 0.40.4 HzHz andand (d(d)) the the performance performance ofof thethe strainstrain sensorsensor afterafter 10001000 strain strain cycles. cycles.

TheThe E-SkinE-Skin sensorsensor datadata isis collectedcollected wirelesslywirelessly viavia smartphone,smartphone, andand thethe samplesample raterate ofof 1515 HzHz forfor eacheach E-SkinE-Skin sensorsensor toto fitfit thethe throughput throughput of of smartphone smartphone BLE BLE transmission. transmission. TheThe user-interfaceuser-interface ofof thethe datadata collectioncollection mobilemobile applicationapplication isis shownshown inin FigureFigure3 .3. This This application application can can connect connect to to 3 3 E-Skin E-Skin sensorssensors atat thethe samesame time.time. TheThe mainmain functionsfunctions ofof thisthis applicationapplication includeinclude datadata transmissiontransmission controlcontrol module,module, data data processing processing module, module, local local storage storage control control module, module, and real-time and real-time data visualisation data visualisation module. Themodule. data transmissionThe data transmission control module control contains module two cont commands:ains two commands: start/end data start/end transmission data transmission and E-Skin sensorand E-Skin labelling. sensor The labelling. user can The use theseuser can two use commands these two to directlycommands control to directly all connected control E-Skin all connected sensors -Skin start andsensors end to the start data and transmission end the data simultaneously transmission simultaneously and label different and connectedlabel different E-Skin connected sensor withE-Skin user-designated sensor with user-designated name such as name left, right,such as and left, centre. right, and Once centre. the application Once the application receives the receives data streamsthe data (including streams (including timestamp timestamp and data value), and data data value), processing data moduleprocessing allocates module each allocates data stream each from data distreamfferent from E-Skin different sensor E-Skin into separated sensor into buff separateders (The length buffers of the(The data length buff ofer the is set data to bebuffer 15 bytes is set based to be on15 thebytes sample based rate) on basedthe sample on the rate) Bluetooth based identifiers on the Bluetooth (UUID). Foridentifiers this android (UUID). application For this aandroid simple Kalmanapplication filter a Javasimple class Kalman was implemented filter Java class for datawas smoothingimplemented [74 ].for Based data smoot on thehing empirical [74]. Based experiences, on the suitableempirical values experiences, for the initiation suitable of values the filter for were: the init measurementiation of the uncertainty filter were (: rmeasurement) = 4, estimate uncertaintyuncertainty ((p )𝒓)= =4, 4, processestimate variance uncertainty (q) = (𝒑0.05.) = 4, Theprocess current variance estimation (𝒒) = 0.05. and The estimate current uncertainty estimation updateand estimate were calculateduncertainty based update on thewere following calculated formulas, based on where the kfollowingstands for formulas, Kalman gain.where 𝒌 stands for Kalman gain. p k = (2) (p +𝒑r) 𝒌= (𝒑 𝒓) (2) current estimation 𝒄𝒖𝒓𝒓𝒆𝒏𝒕 𝒆𝒔𝒕𝒊𝒎𝒂𝒕𝒊𝒐𝒏 (3) = last estimation + k (measurement last estimation) (3) = 𝒍𝒂𝒔𝒕 𝒆𝒔𝒕𝒊𝒎𝒂𝒕𝒊𝒐𝒏∗ 𝒖𝒓𝒆𝒎𝒆𝒏𝒕 𝒌∗(𝒎𝒆𝒂𝒔 − − 𝒍𝒂𝒔𝒕 𝒆𝒔𝒕𝒊𝒎𝒂𝒕𝒊𝒐𝒏) p = (1 k) p + last estimation current estimation q (4) 𝒑=(𝟏−𝒌−) ∗𝒑∗ | 𝒍𝒂𝒔𝒕| 𝒆𝒔𝒕𝒊𝒎𝒂𝒕𝒊𝒐𝒏− −| 𝒄𝒖𝒓𝒓𝒆𝒏𝒕∗ 𝒆𝒔𝒕𝒊𝒎𝒂𝒕𝒊𝒐𝒏| ∗𝒒 (4)

Sensors 2020, 20, x FOR PEER REVIEW 10 of 28

Then the processed data buffers are sent to the local storage control module and real-time data visualisation module, respectively. The local storage control module creates text files for each data stream and stores the timestamps as well as the corresponding data values. The real-time data visualisation Sensors 2020, 20, 1510 10 of 28 moduleSensors displays 2020, 20, thex FOR processed PEER REVIEW data in a dynamic line chart with user-designated labels. 10 of 28

Then the processed data buffers are sent to the local storage control module and real-time data visualisation module, respectively. The local storage control module creates text files for each data stream and stores the timestamps as well as the corresponding data values. The real-time data visualisation module displays the processed data in a dynamic line chart with user-designated labels.

Figure 3. Data Collection Mobile Application User Interface. Figure 3. Data Collection Mobile Application User Interface. Then the processed data buffers are sent to the local storage control module and real-time data visualisationViMove (DorsaVi, module, Melbourne, respectively. Australia) The local was storage used controlas a golden module standard creates for text the files evaluation for each of data E-Skin sensors’stream performance and stores on the lumbar–pelviFigure timestamps 3. Datac asCollection movement well as Mobile the detection. corresponding Application ViMove User data Interface.is a values. clinically The validated real-time instrument data for monitoringvisualisation lower module back displays movements the processed [6]. As shown data in ain dynamic Figure 4, line ViMove chart with system user-designated consists of 4 labels. parts: the upper andViMove lower (DorsaVi,motion (DorsaVi, sensors Melbourne, Melbourne, (located Australia) Australia) at L1 wasand was usedPSIS), used as the aas golden left a golden and standard right standard surfacefor the for evaluation EMG the evaluation sensors of E-Skin (located of at eachE-Skinsensors’ side sensors’ ofperformance the spine performance aroundon lumbar–pelvi onL3), lumbar–pelvic recordingc movement and movement detection.feedback detection. ViMovedevice andis ViMovea clinically monitoring is avalidated clinically software instrument validated programs. The ViMoveinstrumentfor monitoring suite for alsolower monitoring includes back movements lower a low back back [6]. movements fitting As shown templa [6 in]. AsFigurete shownto help 4, ViMove in the Figure user system4 ,to ViMove ea consistssily position system of 4 parts: consists the thesensors upper and lower motion sensors (located at L1 and PSIS), the left and right surface EMG sensors (located on theof body 4 parts: according the upper to and their lower heig motionht. The sensors template (located is also at shown L1 and PSIS),in Figure the left4. The and EMG right surfacesensors EMG are used sensorsat each side (located of the at spine each around side of theL3), spinerecording around and L3), feedback recording device and and feedback monitoring device software and monitoring programs. to assess the muscle activities, and therefore, in our experiments we did not use these sensors. The sample softwareThe ViMove programs. suite also The includes ViMove a suitelow back also includesfitting templa a lowte back to help fitting the templateuser to ea tosily help position the user the to sensors easily rate of ViMove is 20 Hz. The outputs of ViMove can be collected using ViMove data collection software positionon the body the according sensors on to thetheir body height. according The template to their is also height. shown The in template Figure 4. isThe also EMG shown sensors in Figureare used4. on a ThePC.to assess EMGTable the sensors 2 musclecompares are activities, used E-Skin to and assess to therefore,ViMove the muscle basedin our activities, experimentson seven and criteria.therefore, we did not in use our these experiments sensors. The we didsample not userate theseof ViMove sensors. is 20 The Hz. sample The outputs rate of of ViMove ViMove is can 20 Hz.be collected The outputs using of ViMove ViMove data can collection be collected software using ViMoveon a PC. dataTable collection 2 compares software E-Skin onto ViMove a PC. Table based2 compares on seven criteria. E-Skin to ViMove based on seven criteria.

Figure 4. ViMove System Components. Figure 4. ViMove System Components. TableTable 2. 2.ComparisonComparison of of E-Skin sensorsensor and and ViMove. ViMove. Table 2. Comparison of E-Skin sensor and ViMove. E-SkinE-Skin ViMove ViMove E-Skin ViMove PhysicalPhysical Property Property Soft Soft Rigid Rigid MeasurementPhysicalMeasurement Property Skin Skin Deformation Deformation Soft 3D Coordinates 3D Rigid Coordinates ConnectivityMeasurementConnectivity Bluetooth SkinBluetooth Deformation 4.0 4.0 (BLE) 2.3 3D Ghz 2.3 Coordinates FrequencyGhz Frequency DataConnectivity Collection Device Bluetooth Smartphone 4.0/Tablet (BLE) 2.3 Ghz PC Frequency Data Collection DataClinical Collection Validation Smartphone/Tablet NO YES PC DeviceBattery LifeSmartphone/Tablet 1 week 24 hPC Device Price AUD $10~15 each module * (* not yet available for public sale) AUD $10,000 (Hardware and Software) ClinicalClinical Validation Validation NO NO YES YES BatteryBattery Life Life 1 1week week 24 h 24 h AUDAUD $10~15 $10~15 each each module module * *(* (* not not yetyet available for for public public AUDAUD $10,000 $10,000 (Hardware (Hardware and and PricePrice sale)sale) Software)Software)

Sensors 2020, 20, 1510 11 of 28

Sensors 2020, 20, x FOR PEER REVIEW 11 of 28 3.2. Study Design and Experimental Procedure 3.2. Study Design and Experimental Procedure Lumbar–pelvic movements refer to the lower back area movements that include flexion, extension, lateral flexion,Lumbar–pelvic anterior movements posterior pelvicrefer to tilt, the and lower rotation. back Thesearea movements movements that can include be performed flexion, by individualsextension, during lateral diflexion,fferent physicalanterior activities.posterior pelvic 6 participants tilt, and (includingrotation. These 2 males movements and 4 females can agedbe performed by individuals during different physical activities. 6 participants (including 2 males and 22–30 years with height 159–183 cm, weight 47–120 kg, and 2 of them have LBP history in the past 4 females aged 22–30 years with height 159–183 cm, weight 47–120 kg, and 2 of them have LBP history 6 months) were recruited in this experiment as our aim initially was to analyse skin deformation in the past 6 months) were recruited in this experiment as our aim initially was to analyse skin data and compare to ViMove data in order to develop a method for detecting the five lumbar–pelvic deformation data and compare to ViMove data in order to develop a method for detecting the five movements. The participants were instructed to perform the five standard lumbar–pelvic movements lumbar–pelvic movements. The participants were instructed to perform the five standard lumbar– basedpelvic on movements the low back based pain on patients’ the low back rehabilitation pain patien assessmentts’ rehabilitation process assessment [5]. These process five [5]. lumbar–pelvic These five movementslumbar–pelvic follow movements the biomechanical follow the standards biomechanical and widely standards used byand medical widely experts used by to medical assess the experts lumbar movementsto assess performedthe lumbar by movements different individuals performed [75 by]). Eachdifferent participant individuals performed [75]). eachEach movementparticipant for 5 timesperformed and repeated each movement the entire experimentfor 5 times for and 3 times. repeat Aftered the performing entire experiment each lumbar–pelvic for 3 times. movement After (LPM),performing participants each lumbar–pel had to returnvic movement to their static (LPM), position participants (i.e., standing had to inreturn a relaxed to their manner) static position for a few seconds(i.e., standing before continuing in a relaxed to manner) do the next for movement.a few seconds Participants before continuing could choose to do tothe withdraw next movement. or stop to restParticipants at any time could during choose the experiments. to withdraw or Altogether, stop to rest we at haveany time collected during 15 the times experiments. of each lumbar–pelvic Altogether, movements.we have collected In order 15 to times make of theeach experiment lumbar–pelvic comparable movements. to real In order life scenario, to make participantsthe experiment were instructedcomparable to perform to real life the scenario, LPMs at participants any speed withwere any instructed bending to orperform turning the angles. LPMs at We any had speed utilised with the regionsany bending of spine or as turning human angles. back landmarks We had utilised to indicate the regions the locations of spine of as sensor human placement. back landmarks As shown to in Figureindicate5, the adults locations have of 24 sensor separate placement. vertebrae As including shown in 7Figure cervical 5, adults vertebrae have from 24 separate the vertebrae (C1–C7), 12including thoracic vertebrae7 cervical vertebrae from the from upper the back neck (T1–T12)(C1–C7), 12 and thoracic 5 vertebrae from the from upper the back lower (T1– back (L1–L5).T12) and The 5 top lumbar E-Skin vertebrae sensor was from attached the lower vertically back (L1–L5). around T3–T7The top and E-Skin the bottom sensor one was was attached attached verticallyvertically around around L1–L5 T3–T7 as and shown the bottom in Figure one6 d.was The attached ViMove vertically lower motionaround sensorL1–L5 as was shown attached in Figure along the6d. line The of ViMove posterior lower superior motion iliac sensor spine was (PSIS). attached PSIS along of the the participant line of posterior was measured superior by iliac palpation. spine (PSIS). PSIS of the participant was measured by palpation. The placement of the ViMove motion The placement of the ViMove motion sensors were based on the ViMove templates. The placement of sensors were based on the ViMove templates. The placement of E-Skin sensor was informed by the E-Skin sensor was informed by the consultation with the medical experts involved in this research. consultation with the medical experts involved in this research. This sensor placement has been This sensor placement has been identified to produce the maximum E-Skin sensor output during the identified to produce the maximum E-Skin sensor output during the performance of flexion, performance of flexion, extension and pelvic tilts. In this paper, we present our analysis with one extension and pelvic tilts. In this paper, we present our analysis with one participant because there is participantno significant because difference there isbetween no significant participants difference with and between without participants LBP in terms with of andthe lumbar–pelvic without LBP in termsmovement of the lumbar–pelvicdata pattern [76]. movement The participant data patternwas instructed [76]. The to repeat participant each movement was instructed 5 times to at repeat the eachsame movement speed as 5possible. times at Figure the same 6 (a–c) speed show as the possible. flexion, extension, Figure6 (a–c) and anterior show the posterior flexion, pelvic extension, tilt, andrespectively. anterior posterior pelvic tilt, respectively.

FigureFigure 5.5. RegionsRegions of Spine.

Sensors 2020, 20, 1510 12 of 28 Sensors 2020, 20, x FOR PEER REVIEW 12 of 28

FigureFigure 6. 6. ExperimentExperiment Description: Description: (a ()a )Flexion; Flexion; (b (b) )Extension; Extension; (c ()c )Pelvic Pelvic Tilt; Tilt; (d (d) )Sensor Sensor Placement. Placement.

ViMoveViMove and and E-Skin sensorssensors useuse twotwo di differentfferent sampling sampling frequency. frequency. The The frequency frequency of ViMoveof ViMove data datais generally is generally slightly slightly lower lower than than 20 Hz 20 (19.0–19.8Hz (19.0–19.8 Hz), Hz), and and the the E-Skin E-Skin sampling sampling rate rate is around is around 15 Hz.15 Hz.In orderIn order to compareto compare the the outputs outputs of theseof these two two sensors, sensors, we we calculated calculated the the average average of dataof data points points per persecond second for for each each sensor. sensor. For For example, example, the the ViMove ViMove collected collected 19 data 19 data points points in 1 in s and 1 s and E-Skin E-Skin sensor sensor A, B, A,C receivedB, C received 14, 15, 14, and 15, 16 and data 16 points data points in the samein the second, same second, respectively. respectively. We summed We summed all data pointsall data in pointsthis second in this forsecond each for sensor each and sensor divided and divided the sums the by sums the numbersby the numbers of the dataof the points data points of each of sensor. each sensor.This allowed This allowed us to compare us to compare their average their valuesaverag usinge values the sameusing timestamp.the same timestamp. Since the participants Since the participantswere instructed were toinstructed perform to movements perform movements slowly, this slowly, downsampling this downsampling procedure procedure would not would affect not the affectsubsequent the subsequent analysis. analysis.

3.3.3.3. Results Results Analysis Analysis DuringDuring our our experiments, experiments, five five sets sets of ofoutput output sens sensoror data data was was collected collected from from both both ViMove ViMove and and E- SkinE-Skin simultaneously simultaneously while while the participant the participant was wasperforming performing flexion, flexion, extension, extension, pelvic pelvic tilt, rotation, tilt, rotation, and lateraland lateral flexion flexion (See Figure (See Figure 6d). 6d). TheThe E-Skin E-Skin sensor sensor raw raw signal signal output output range range from from 0 0to to 4096 4096 Analog-to-Digital Analog-to-Digital Units Units (ADU) (ADU) [77]. [77]. TheThe value value decreasesdecreases whenwhen the the E-Skin E-Skin stretches. stretches. In orderIn order to clearly to clearly depict depict the changes the changes in E-Skin in outputs,E-Skin outputs,we converted we converted the absolute the absolute E-Skin output E-Skin into output the change into the in change electrical in resistanceelectrical resistance for the E-Skin for the sensor E- Skinfor comparisonsensor for comparison rather than therather raw than signal the output. raw signal The output. absolute The E-Skin absolute output E-Skin was calculatedoutput was by calculatedusing E-Skin by using raw signal E-Skin output raw signal to determine output to a baselinedetermine value. a baseline The baseline value. The value baseline was calculated value was as calculatedan average as of an E-Skin average raw of signal E-Skin outputs raw signal for a staticoutputs standing for a positionstatic standing when E-Skinposition sensor when was E-Skin at its sensororiginal was length at its andoriginal not length stretched. and Innot this stretched. experiment, In this we experiment, used 5 s ofwe data used to 5 calculates of data to the calculate average theoutput average (i.e., output baseline (i.e., value). baseline The value). change The in electrical change in resistance electrical value resistance was calculated value was by calculated Equation by (5), Equationwhere reference (5), where voltage reference for 4096 voltage ADU for is 4096 3.3 V ADU and the is 3.3 current V and supplied the current is 5 supplied mA. is 5 mA.

𝑋X 𝑌=Y = 3.33.3 0.0050.005 (5) (5) 40964096 × ÷ Linear regression was applied to model the mapping between E-Skin and ViMove outputs because Linear regression was applied to model the mapping between E-Skin and ViMove outputs our results showed that the skin stretch increased as the lumbar–pelvic angle increased (during flexion because our results showed that the skin stretch increased as the lumbar–pelvic angle increased or extension), and there was a potential linear relationship between the outputs of E-Skin and ViMove. (during flexion or extension), and there was a potential linear relationship between the outputs of E- The linear relationship between the two variables can be presented as formula in Equation (6). Let X be Skin and ViMove. The linear relationship between the two variables can be presented as formula in the independent variable (E-Skin outputs in normalised resistance value), and Y be the dependent Equation (6). Let X be the independent variable (E-Skin outputs in normalised resistance value), and variable (ViMove outputs). In this work, we used the quadratic loss function in Equation (7) to calculate Y be the dependent variable (ViMove outputs). In this work, we used the quadratic loss function in the loss of the model and least-squares method in Equations (8) and (9) to minimize the loss. p refers Equation (7) to calculate the loss of the model and least-squares method in Equations (8) and i(9) to to the points on the regression line. x and y are the mean values of input X and the desired output Y. minimize the loss. 𝑝𝑖 refers to the points on the regression line. 𝑥 and 𝑦 are the mean values of input X and the desired output Y. Y = mX + c (6) X 𝑌=𝑚𝑋𝑐n 2 (6) L(x) = (yi pi) (7) i=1 − 𝐿(𝑥) = (𝑦 −𝑝) (7)

Sensors 2020, 20, 1510 13 of 28 Sensors 2020, 20, x FOR PEER REVIEW 13 of 28

Pn ∑= ((𝑥xi −𝑥x)(̅)(𝑦yi −𝑦y)) m = i1 − − (8) 𝑚= Pn 2 (8) ∑((𝑥x −𝑥x)̅) i=1 i − c𝑐== y −𝑚𝑥𝑦 mx ̅ (9)(9) − TheThe modelmodel is is trained trained based based on on 10 10 times times of of each each lumbar–pelvic lumbar–pelvic movements’ movements’ data data and and tested tested on the on otherthe other 5 times 5 lumbar–pelvictimes lumbar–pelvic movements’ movements’ data. Mean data. absolute Meanerror absolute (MAE) error of the (MAE) two measurement of the two systemsmeasurement (E-Skin systems sensor and(E-Skin ViMove) sensor is and used ViMove) to measure is used the to angular measure displacements. the angular displacements. ToTo furtherfurther illustrateillustrate thethe statisticalstatistical signific significance,ance, a a two-tailed two-tailed test test was was added added and and the the pp-Value-Value with with a asignificance significance level level of 0.01 of 0.01 was wasconsidered considered to determine to determine the statistical the statistical significance significance of our results. of our This results. paper Thisuses paperthe 15 times uses theof each 15 times lumbar–p of eachelvic lumbar–pelvic movement to calculate movement the p to-Value. calculate However, the p-Value. the data However, quantities theof each data participant quantities are of different each participant because they are conducted different the because experiment they conducteds at different the speed. experiments at differentThe speed.data analysis of each experiment is described in the following subsections. The data analysis of each experiment is described in the following subsections. 3.3.1. Flexion Result Analysis 3.3.1. Flexion Result Analysis The E-Skin measurement for the flexion movement is shown in Figure 7. The y-axis on the left The E-Skin measurement for the flexion movement is shown in Figure7. The y-axis on the represents the E-Skin sensor output. The y-axis on the right represents the ViMove output. The unit left represents the E-Skin sensor output. The y-axis on the right represents the ViMove output. of ViMove output is the degree of anatomical angles and the unit of E-Skin sensor output is Analog- The unit of ViMove output is the degree of anatomical angles and the unit of E-Skin sensor output is to-Digital Units (ADU) [77]. The x-axis represents time and the unit is second. The orange line Analog-to-Digital Units (ADU) [77]. The x-axis represents time and the unit is second. The orange line represents the outputs of ViMove. The green line represents the outputs of the bottom E-Skin sensor represents the outputs of ViMove. The green line represents the outputs of the bottom E-Skin sensor and the blue line represents the outputs of the top E-Skin sensor. and the blue line represents the outputs of the top E-Skin sensor.

140 50 120

R) 40

Δ 100

80 30

60 20 40 10 20 E-Skin Output ( 0 0 ViMove Output (Degree)

0 20406080100 Time Sequence (s) Bottom E-Skin Sensor Top E-Skin Sensor ViMove

Figure 7. Data Outputs of Flexion. Figure 7. Data Outputs of Flexion. For each flexion, as the orange line shows, the bending range of the participant is from 0 degrees For each flexion, as the orange line shows, the bending range of the participant is from 0 degrees (standing straight) to roughly 55 degrees (bending to the maximum position). The participant stayed (standing straight) to roughly 55 degrees (bending to the maximum position). The participant stayed at the maximum forward flexion position for approximately 10 s. In this work, we instructed the at the maximum forward flexion position for approximately 10 s. In this work, we instructed the participants to self-report standing straight when they are at their tallest position. It can be seen that participants to self-report standing straight when they are at their tallest position. It can be seen that outputs of both bottom (green line) and top (blue line) E-Skin sensors data fluctuated synchronously outputs of both bottom (green line) and top (blue line) E-Skin sensors data fluctuated synchronously with the changes in ViMove outputs. The E-Skin sensor outputs show a clear pattern associated with with the changes in ViMove outputs. The E-Skin sensor outputs show a clear pattern associated with each flexion. It can also be found that both bottom and top sensors were capable of detecting the each flexion. It can also be found that both bottom and top sensors were capable of detecting the start start of each flexion prior to the ViMove because the back skin stretched before the movement started. of each flexion prior to the ViMove because the back skin stretched before the movement started. This This period of time can be considered as the transition between the standing and flexion positions. period of time can be considered as the transition between the standing and flexion positions. In order to analyse the response time difference between ViMove and E-Skin outputs, we calculated In order to analyse the response time difference between ViMove and E-Skin outputs, we a threshold of 3 degrees for ViMove outputs and a threshold of 6.9 in resistance change for E-Skin calculated a threshold± of ±3 degrees for ViMove outputs and a threshold± of ±6.9 in resistance change for E-Skin output based on a relatively linear relationship between ViMove and the bottom E-Skin

Sensors 2020, 20, 1510 14 of 28 output based on a relatively linear relationship between ViMove and the bottom E-Skin sensor outputs. We calculated the changes between sensor outputs while the participant was standing straight and performing flexion. If the changes in sensor outputs exceeded the thresholds, we considered the point as the start of a flexion. The average difference between response times of two sensors was approximately 1 s. We measured the unit of the response time difference in seconds because we computed the average of the outputs of E-Skin and ViMove in seconds as discussed earlier. The results show that E-Skin sensor can measure flexion movement faster than ViMove does. This might be due to the slimline design (maximum thickness of 5 mm) and lightweight (2 g), which allows the soft and highly stretchable E-skin sensor to adhere conformally to the curvilinear skin surface [78,79], reducing the inertial effect during movements as compared to the bulkier (100.6 mm in width, 30.4 mm in length, and 9 mm in height) and 20-fold heavier ViMove. Besides, the conformal attachment of E-skin sensors to skin surface also avoid the drifting errors that IMU-based sensor has, which allows more accurate measurement of strain [80]. We applied linear regression to model the mapping between E-Skin outputs and ViMove outputs. Our results show that R square of linear regression for the data collected from the bottom E-Skin sensor is higher than the R square for the data from the top E-Skin sensor. The R square of linear regression using both sensors is 0.2% higher than the R square for the bottom E-Skin sensor. The results are the same considering the adjusted R square. Additionally, the p-Value for both top and bottom E-Skin sensor outputs is statistically highly significant (p < 0.001). This indicates that both top and bottom E-Skin outputs have relatively linear correlations with the trunk flexion angles, but the bottom ones are more appropriate for flexion angle measurement. The details of comparison results are shown in Table3. The data quantities of calculating these p-Values of flexion are 363.

Table 3. Linear regression comparison for flexion.

Independent R R2 Adjusted R2 Significance (2-Tailed) Top E-Skin Outputs 0.843 0.711 0.708 p < 0.001 Bottom E-Skin Outputs 0.954 0.910 0.909 p < 0.001 Both 0.955 0.912 0.911

Table2 results also show that the E-Skin sensor can be used for detecting the flexion with high accuracy based on clear and repeating patterns in the output. It also shows that the bottom E-Skin sensor’s position is the suitable position for measuring the flexion. This experiment was conducted on one participant, the mapping model can be improved by collecting data from a larger populated group and considering adding more features such as age, height, weight, BMI, and skin characteristics. As the experiments progressed, the differences between the maximum outputs of top and bottom E-Skin sensors and the participant’s maximum forward flexion angle for each flexion became relatively small. However, the differences between minimum outputs of top and bottom E-Skin sensors and the participant’s standing straight angle increased after each flexion. In Figure2, we had conducted 1500 times cyclic strain test and the result showed that after 1000 times strain training, the sensor performance appeared to be consistent. However, the drift appeared when the sensor was mounted on individuals, and this had been consistently occurring. We believe that this is an issue with the adhesives when sticking on individuals with different skin types, particularly dry skin, sweaty skin, and individuals with a lot of dead skins and this causes the sensor to not mount properly and slide against the skin when being stretched. Based on the model, the E-Skin sensor anatomical angle outputs of flexion are generated as shown in Figure8. It can be seen that E-Skin sensor angle outputs are very close to the actual angles (ViMove outputs). The MAE of the two measurement systems is 5.105 degrees. SensorsSensors2020, 202020, 1510, 20, x FOR PEER REVIEW 15 of 2815 of 28 Sensors 2020, 20, x FOR PEER REVIEW 15 of 28

FigureFigure 8. 8.Anatomical Anatomical AngleAngle Outputs Comparison Comparison (Flexion). (Flexion). Figure 8. Anatomical Angle Outputs Comparison (Flexion). 3.3.2.3.3.2. Extension Extension Result Result Analysis Analysis 3.3.2. Extension Result Analysis The resultsThe results of our of our experiment experiment to to detect detect andand measure extension extension are are presented presented in Figure in Figure 9. As9 .the As the The results of our experiment to detect and measure extension are presented in Figure 9. As the figure shows, each extension started at 0 degrees (standing straight) and ended at around −12 degrees figurefigure shows, shows, each each extension extension started started at at 0 0 degrees degrees (standing straight) straight) and and ended ended at around at around −12 degrees12 degrees (extend to maximum position). Unlike the results of flexion, the participant did not extend− to the (extend(extend to maximum to maximum position). position). Unlike Unlike thethe results of of flexion, flexion, the the participant participant did didnot notextend extend to the to the maximum position for every extension. maximummaximum position position for for every every extension. extension.

80 80 -2 -2

R) 60 -4 R) Δ 60 -4 Δ -6 -6 40 -8 40 -8 -10 20 -10 20 -12 -12 0 -14 E-Skin Output ( 0 -14 E-Skin Output ( -16

-16 ViMove Output (Degree) -20 ViMove Output (Degree) -20 0 20406080 0 20406080 Time Sequence (s) Time Sequence (s) Bottom E-Skin Sensor Top E-Skin Sensor ViMove Bottom E-Skin Sensor Top E-Skin Sensor ViMove

FigureFigure 9. 9.Data Data Outputs of of Extension. Extension. Figure 9. Data Outputs of Extension. TheThe lowest lowest extension extension angle angle is is from from −1010 degreesdegrees to to −1717 degrees. degrees. The The reason reason is because is because the the The lowest extension angle is from− −10 degrees to −−17 degrees. The reason is because the extensionextension requires requires an individualan individual to to put put their their headhead back and and look look at atthe the ceiling. ceiling. During During this movement, this movement, extension requires an individual to put their head back and look at the ceiling. During this movement, an individualan individual might might only only put put their their head head back back ratherrather than than extending extending the the lumbar lumbar spine spine (low (low back) back) to to an individual might only put their head back rather than extending the lumbar spine (low back) to the maximum position. Results show E-Skin sensor outputs do not closely fluctuate in line with the the maximumthe maximum position. position. Results Results show show E-Skin E-Skin sensor ou outputstputs do do not not closely closely fluctuate fluctuate in line in with line withthe the changes of ViMove outputs but there is a clear pattern for each extension. There are two regular changeschanges of ViMove of ViMove outputs outputs but but there there is a is clear a clear pattern pattern for for each each extension. extension. There There are are two two regular regular spikes spikes of top E-Skin sensor outputs (blue line) at the start and the end of each extension. There is also of topspikes E-Skin of sensortop E-Skin outputs sensor (blue outputs line) (blue at the line) start at the and start the and end the of eachend of extension. each extension. There There is also is aalso notable a notable variation in the output of the bottom E-Skin sensor when moving from the extension variationa notable in the variation output in of the the output bottom of E-Skinthe bottom sensor E-Skin when sensor moving when from moving the extensionfrom the extension position to a position to a standing straight position. The peak point of the bottom E-Skin sensor’s outputs at the standingposition straight to a standing position. straight The peak position. point The of the peak bottom point E-Skinof the bottom sensor’s E-Skin outputs sensor’s at the outputs second at extension the second extension is lower than the other extensions which could be related to the extension angle. second extension is lower than the other extensions which could be related to the extension angle. is lowerSimilar than patterns the other can extensionsalso be found which in other could extensions. be related Additionally, to the extension the highest angle. value Similar of the patternsbottom can Similar patterns can also be found in other extensions. Additionally, the highest value of the bottom also be found in other extensions. Additionally, the highest value of the bottom E-Skin sensor outputs changes according to the participant’s maximum extension angle in each extension. These regularities in data patterns can be used to detect an extension based on E-Skin outputs [81]. SensorsSensors2020, 202020, 1510, 20, x FOR PEER REVIEW 16 of 2816 of 28

E-Skin sensor outputs changes according to the participant’s maximum extension angle in each Similarextension. to These other movements,regularities in linear data patterns regression can was be used applied to detec to analyset an extension E-Skin andbased ViMove on E-Skin outputs. As shownoutputs in [81]. Table 4, the R square of linear regression for extension is considerably lower than results for flexion. ThisSimilar can to be other attributed movements, to the followinglinear regression reasons. was First, applied the relationship to analyse E-Skin between and the ViMove extension’s angleoutputs. and the As lower shown back in Table area 4, skin the R deformation square of line isar notregression linear. for Second, extension the is two considerably positions lower (top and bottom)than may results not for be flexion. the optimal This can positions be attributed for to detecting the following the reasons. extension. First, The the otherrelationship possible between positions the extension’s angle and the lower back area skin deformation is not linear. Second, the two positions could be tested to increase the accuracy of measurement. Despite this, the results still show that (top and bottom) may not be the optimal positions for detecting the extension. The other possible the p-value of top E-Skin outputs is statistically significant and E-Skin sensor can be used to detect positions could be tested to increase the accuracy of measurement. Despite this, the results still show extensionthat the based p-value on repeatingof top E-Skin patterns outputs in is the statistically E-Skin outputs. significant The and data E-Skin quantities sensor ofcan calculating be used to these p-Valuesdetect of extension extension based are 327.on repeating patterns in the E-Skin outputs. The data quantities of calculating these p-Values of extension are 327. Table 4. Linear regression comparison for extension. Table 4. Linear regression comparison for extension. Independent R R2 Adjusted R2 Significance (2-Tailed) Independent R R2 Adjusted R2 Significance (2-Tailed) Top E-Skin Outputs 0.318 0.101 0.090 0.004 Top E-Skin Outputs 0.318 0.101 0.090 0.004 Bottom E-Skin Outputs 0.183 0.033 0.021 0.107 BottomBoth E-Skin Outputs 0.320 0.183 0.0330.102 0.021 0.079 0.107 Both 0.320 0.102 0.079

BasedBased on theon the model, model, the the E-Skin E-Skin sensor sensor anatomicalanatomical angle angle outputs outputs of ofextension extension are generated are generated as as shownshown in Figure in Figure 10. The10. The MAE MAE of of the the two two measurement measurement systemssystems is 4.509 4.509 degrees. degrees. Although Although the the MAE MAE is loweris thanlower flexion, than flexion, the angle the angle outputs outputs of E-Skin of E-Skin sensors sensors for for extension extension are are not not close close to to the the actual actual angle outputsangle (ViMove outputs outputs),(ViMove outputs), as expected as expect baseded thebased linear the linear regression regression analysis. analysis.

Figure 10. Anatomical Angle Outputs Comparison (Extension). Figure 10. Anatomical Angle Outputs Comparison (Extension). 3.3.3. Anterior and Posterior Pelvic Tilt Result Analysis 3.3.3. Anterior and Posterior Pelvic Tilt Result Analysis The pelvic tilt consists of anterior and posterior pelvic tilt. The data outputs of the anterior and The pelvic tilt consists of anterior and posterior pelvic tilt. The data outputs of the anterior and posteriorposterior pelvic pelvic tilt tilt are are shown shown in in Figure Figure 11 11.. The fi figuregure shows shows that that there there is an is anobvious obvious rise and rise fall and fall patternspatterns in both in both sensor sensor outputs outputs when when the the participantparticipant performs performs pelvic pelvic tilt. tilt. Yet, Yet, the thedata data patterns patterns for for pelvicpelvic tilt aretilt diareff erentdifferent compared compared to to flexion flexion andand extensionextension results. results. The The top top E-Skin E-Skin sensor sensor outputs outputs fluctuatefluctuate according according to theto the changes changes of of pelvic pelvic tilt,tilt, while the the bottom bottom sensor sensor outputs outputs show show two twospikes spikes at at the startthe start and and the the end end of eachof each pelvic pelvic tilt. tilt. The repeatability of E-Skin sensor output is not high enough for detecting pelvic tilt compared to the previous two experimental results. The pelvic tilt requires the participant to tilt their anteriorly and posteriorly. These movements involve excessive subtle motions that are very difficult to repeat in the exact same way each time. Skin deformation data captured by E-Skin sensor reflect these minor changes and make it hard to achieve consistent repeatability. Sensors 2020, 20, 1510 17 of 28 Sensors 2020, 20, x FOR PEER REVIEW 17 of 28

15 50

R) 40 10 Δ

30 5

20 0

10 -5

E-Skin Output ( 0

-10 ViMove Output (Degree) -10 0 20406080100 Time Sequence (s)

Bottom E-Skin Sensor Top E-Skin Sensor ViMove

FigureFigure 11. 11.Data Data Outputs of of Pelvic Pelvic Tilt. Tilt.

The resultsThe repeatability of linear regressionof E-Skin sensor for modelling output is not the high mapping enough between for detecting E-Skin pelvic outputs tilt compared and ViMove outputsto the are previous similar totwo extension experimental (see results. Table5 ).The However, pelvic tilt the requires top E-Skin the participant sensor outputs to tilt their demonstrate pelvis a statisticallyanteriorly highly and posteriorly. significant These correlation movements with involve the ViMove excessive sensors subtle outputsmotions that with are an very R square difficult value at 62.9%to repeat (p < in0.001), the exact while same the way bottom each time. E-Skin Skin sensor deformation shows data significantly captured by lower E-Skin results. sensorThis reflect can be addressedthese minor by investigating changes and andmake determining it hard to achieve optimal consistent sensor repeatability. placementon the body to achieve higher sensing accuracy.The results The of linear data quantitiesregression for of calculatingmodelling the these mappingp-Values between of pelvic E-Skin tilt outputs are 476. and ViMove outputs are similar to extension (see Table 5). However, the top E-Skin sensor outputs demonstrate a statistically highly significant correlation with the ViMove sensors outputs with an R square value at Table 5. Linear regression comparison for pelvic tilt. 62.9% (p < 0.001), while the bottom E-Skin sensor shows significantly lower results. This can be addressed byIndependent investigating and determining R optimalR2 sensorAdjusted placement R2 Significance on the body (2-Tailed)to achieve higher sensing accuracy. The data quantities of calculating these p-Values of pelvic tilt are 476. Top E-Skin Outputs 0.793 0.629 0.625 p < 0.001 Bottom E-Skin Outputs 0.128 0.016 0.007 0.196 Table 5. Linear regression comparison for pelvic tilt. Both 0.796 0.633 0.626 Independent R R2 Adjusted R2 Significance (2-Tailed) Top E-Skin Outputs 0.793 0.629 0.625 p < 0.001 Based on the model, the E-Skin sensor anatomical angle outputs of pelvic tilt are generated as Bottom E-Skin Outputs 0.128 0.016 0.007 0.196 shown in Figure 12. TheBoth MAE of the two0.796 measurement 0.633 systems0.626 is 3.154 degrees. The MAE of the pelvic tilt is also lower than flexion. However, it can be seen that E-Skin sensors have a better performance on measuringBased posterior on the pelvicmodel, tiltthe thanE-Skin anterior sensor anatomical pelvic tilt. an Thegle outputs reason of for pelvic this tilt is becauseare generated the current as shown sensor placementinSensors Figure 2020 of 12. bottom, 20 The, x FOR MAE E-Skin PEER of REVIEW the sensor two me isasurement not capable systems of fully is 3.154 capturing degrees. Th anteriore MAE pelvicof the pelvic tilt movement. tilt 18is alsoof 28 lower than flexion. However, it can be seen that E-Skin sensors have a better performance on measuring posterior pelvic tilt than anterior pelvic tilt. The reason for this is because the current sensor placement of bottom E-Skin sensor is not capable of fully capturing anterior pelvic tilt movement.

Figure 12. Anatomical Angle Outputs Comparison (Pelvic Tilt). Figure 12. Anatomical Angle Outputs Comparison (Pelvic Tilt).

3.3.4. Lateral Flexion and Rotation Analysis Lateral Flexion and Rotation are the other two standard lumbar–pelvic movements. Lateral flexion involves the bending movement in the lateral direction (i.e., in sideward). Rotation refers to the up-torso rotation towards the right or left side while the pelvic area stays still. These movements are more complex than flexion, extension, and pelvic tilt. Based on the E-Skin sensor settings (e.g., the number of sensors and their placement) that we used in the previous three experiments, these two movements cannot be detected with acceptable precision. Therefore, as shown in Figure 13, we had to use two additional E-Skin sensors which were positioned diagonally at both sides towards the inferior angle of the participant’s right and left (ISA) while standing straight. This sensor placement is identified by the same procedures as the previous one.

Figure 13. New Sensor Placement.

Lateral Flexion The results of our experiments using E-Skin sensors and ViMove for left and right lateral flexion are shown in Figures 14 and 15, respectively. The orange line still represents the outputs of ViMove. The red and purple lines represent the outputs of left and right E-Skin sensors. The rise and fall patterns in ViMove data shown in Figure 14 represent a single left lateral flexion movement. Figure 15 shows an opposite pattern for the right lateral flexion movement. E-Skin sensors’ outputs and ViMove outputs clearly show repeating similar patterns. Both left and right E- Skin sensor outputs fluctuate according to the changes in the lateral flexion movement. Yet, for the

Sensors 2020, 20, x FOR PEER REVIEW 18 of 28

Sensors 2020, 20, 1510 18 of 28 Figure 12. Anatomical Angle Outputs Comparison (Pelvic Tilt).

3.3.4. Lateral Flexion and Rotation Analysis Lateral FlexionFlexion and and Rotation Rotation are are the the other other two standardtwo standard lumbar–pelvic lumbar–pelvic movements. movements. Lateral Lateral flexion involvesflexion involves the torso the bending torso bending movement movement in the in lateral the lateral direction direction (i.e., (i.e., in sideward). in sideward). Rotation Rotation refers refers to theto the up-torso up-torso rotation rotation towards towards the right the orright left sideor left while side the while pelvic the area pelvic stays still.area Thesestays movementsstill. These aremovements more complex are more than complex flexion, than extension, flexion, and extensio pelvicn, tilt. and Based pelvic on tilt. the Based E-Skin on sensor the E-Skin settings sensor (e.g., thesettings number (e.g., of sensorsthe number and their of sensors placement) and that their we placement) used in the previousthat we used three experiments,in the previous these three two movementsexperiments, cannot these betwo detected movements with acceptable cannot be precision. detected Therefore, with acceptable as shown precision. in Figure Therefore,13, we had as to useshown two in additional Figure 13, E-Skin we had sensors to use which two additional were positioned E-Skin diagonally sensors which at both were sides positioned towards thediagonally inferior angleat both of sides the participant’s towards the rightinferior and angle left scapula of the participant’s (ISA) while standingright and straight.left scapula This (ISA) sensor while placement standing is identifiedstraight. This by thesensor same placement procedures is identified as the previous by the one. same procedures as the previous one.

Figure 13. New Sensor Placement.

Lateral Flexion Sensors 2020, 20, x FOR PEER REVIEW 19 of 28 The results of our experiments using E-Skin sensors and ViMove for left and right lateral flexion flexion left lateral flexion, the range of the right E-Skin sensor outputs are higher than the left sensors’ are shown in Figures 14 and 15,15, respectively. The or orangeange line still still represents represents the the outputs outputs of of ViMove. ViMove. outputs, and for the right lateral flexion, the left E-Skin sensors show higher output values. The red and purple lines representrepresent thethe outputsoutputs ofof leftleft andand rightright E-SkinE-Skin sensors.sensors. The rise and fall patterns in ViMove data shown in Figure 14 represent a single left lateral flexion movement. Figure 15 shows an opposite pattern for the right lateral flexion movement.5 E-Skin 200 sensors’ outputs and ViMove outputs clearly show repeating similar patterns. Both left0 and right E- Skin sensor outputs fluctuate according to the changes in the lateral flexion movement. Yet, for the R) 150

Δ -5

-10 100 -15 50 -20

E-Skin Output ( 0 -25 ViMove Output (Degree)

0 20 40 60 80 100 120 140 160 180 200 Time Sequence (s) Left E-Skin Sensor Right E-Skin Sensor ViMove

Figure 14. Data Outputs of Left Lateral Flexion. Figure 14. Data Outputs of Left Lateral Flexion. The rise and fall patterns in ViMove data shown in Figure 14 represent a single left lateral flexion movement. Figure 15 shows an opposite pattern for the right lateral flexion movement.30 E-Skin sensors’ 150 outputs and ViMove outputs clearly show repeating similar patterns. Both left and right25 E-Skin sensor outputs fluctuate according to the changes in the lateral flexion movement. Yet, for the left lateral R) 20 Δ 100 15

10 50 5

0

E-Skin Output ( E-Skin Output 0

-5 (Degree) ViMove Output 0 20 40 60 80 100 120 140 160 180 200 220 240 Time Sequence (s) Left E-Skin Sensor Right E-Skin Sensor ViMove

Figure 15. Data Outputs of Right Lateral Flexion.

The linear regression results for modelling the mapping between E-Skin outputs and ViMove outputs shows that both left and right E-Skin sensor outputs have a statistically highly significant correlation (p < 0.001) with the actual anatomical angle (ViMove outputs). Tables 6 and 7 show the results for left and right lateral flexion respectively. The results of the mapping models for left and right E-Skin sensors are different for left and right lateral flexion in an opposite way. The data quantities of calculating these p-Values of left and right lateral flexion are 198 and 229, respectively.

Table 6. Linear regression comparison for left lateral flexion.

Independent R R2 Adjusted R2 Significance (2-Tailed) Left E-Skin Outputs 0.817 0.667 0.665 p < 0.001 Right E-Skin Outputs 0.832 0.692 0.690 p < 0.001 Both 0.893 0.798 0.796

Sensors 2020, 20, x FOR PEER REVIEW 19 of 28

left lateral flexion, the range of the right E-Skin sensor outputs are higher than the left sensors’ outputs, and for the right lateral flexion, the left E-Skin sensors show higher output values.

5 200 0

R) 150

Δ -5

-10 100 -15 50 -20

E-Skin Output ( 0 -25 ViMove Output (Degree)

0 20 40 60 80 100 120 140 160 180 200 Sensors 2020, 20, 1510 19 of 28 Time Sequence (s) Left E-Skin Sensor Right E-Skin Sensor ViMove

flexion, the range of the right E-Skin sensor outputs are higher than the left sensors’ outputs, and for the right lateral flexion, the left E-SkinFigure 14. sensors Data Outputs show of higher Left Lateral output Flexion. values.

30 150 25

R) 20 Δ 100 15

10 50 5

0

E-Skin Output ( E-Skin Output 0

-5 (Degree) ViMove Output 0 20 40 60 80 100 120 140 160 180 200 220 240 Time Sequence (s) Left E-Skin Sensor Right E-Skin Sensor ViMove

FigureFigure 15. 15.Data Data Outputs Outputs ofof RightRight Lateral Flexion. Flexion.

The linearThe linear regression regression results results for for modelling modelling the the mappingmapping between between E-Skin E-Skin outputs outputs and and ViMove ViMove outputsoutputs shows shows that that both both left left and and right right E-Skin E-Skin sensor sensor outputsoutputs have have a astatistically statistically highly highly significant significant correlationcorrelation (p < (0.001)p < 0.001) with with the the actual actual anatomical anatomical angleangle (ViMove outputs). outputs). Tables Tables 6 and6 and 7 show7 show the the resultsresults for left for and left right and right lateral lateral flexion flexion respectively. respectively. The The results results of theof the mapping mapping models models for for left left and and right E-Skinright sensors E-Skin are sensors different are for different left and for right left lateraland righ flexiont lateral in flexion an opposite in an way.opposite The way. data The quantities data of calculatingquantities these of p-Valuescalculating of these left andp-Values right of lateral left an flexiond right lateral are 198 flexion and 229,are 198 respectively. and 229, respectively.

TableTable 6. Linear 6. Linear regression regression comparison comparison for left left lateral lateral flexion. flexion. Independent R R2 Adjusted R2 Significance (2-Tailed) Independent R R2 Adjusted R2 Significance (2-Tailed) Left E-Skin Outputs 0.817 0.667 0.665 p < 0.001 LeftRight E-Skin E-Skin Outputs Outputs 0.817 0.832 0.6920.667 0.690 0.665 p < p0.001< 0.001 Right E-SkinBoth Outputs 0.8320.893 0.798 0.692 0.796 0.690 p < 0.001 Both 0.893 0.798 0.796

Table 7. Linear regression comparison for right lateral flexion.

Independent R R2 Adjusted R2 Significance (2-Tailed) Left E-Skin Outputs 0.798 0.637 0.635 p < 0.001 Right E-Skin Outputs 0.496 0.246 0.243 p < 0.001 Both 0.804 0.646 0.643

Based on the model, the E-Skin sensor anatomical angle outputs of left and right lateral flexion are generated as shown in Figures 16 and 17, respectively. The MAEs of the two measurement systems are 2.863 degrees and 5.039 degrees, respectively. The E-Skin sensor has a relatively good result on both left and right lateral flexion.

Rotation The results of the left and right rotations are shown in Figures 18 and 19. The ViMove outputs are represented by the orange lines. The patterns of the left and right rotation are similar to the left and right lateral flexion movements, but the difference is that there is a rise in the right E-Skin sensor outputs before the left E-Skin sensor outputs when the participant performs left rotation. In contrast, left E-Skin sensor outputs show a fall in the data values before the right sensor when the participant performs right rotation. Based on this clear and repeating trend, the rotation can be clearly detected. Sensors 2020, 20, x FOR PEER REVIEW 20 of 28

Sensors 2020, 20, x FOR PEERTable REVIEW 7. Linear regression comparison for right lateral flexion. 20 of 28

IndependentTable 7. Linear regressionR R 2comparison Adjusted for rightR2 lateralSignificance flexion. (2-Tailed) Left E-Skin Outputs 0.798 0.637 0.635 p < 0.001 Independent R R2 Adjusted R2 Significance (2-Tailed) Right E-Skin Outputs 0.496 0.246 0.243 p < 0.001 Left E-Skin Outputs 0.798 0.637 0.635 p < 0.001 Both 0.804 0.646 0.643 Right E-Skin Outputs 0.496 0.246 0.243 p < 0.001 Both 0.804 0.646 0.643 Based on the model, the E-Skin sensor anatomical angle outputs of left and right lateral flexion are generated as shown in Figures 16 and 17, respectively. The MAEs of the two measurement Based on the model, the E-Skin sensor anatomical angle outputs of left and right lateral flexion systems are 2.863 degrees and 5.039 degrees, respectively. The E-Skin sensor has a relatively good are generated as shown in Figures 16 and 17, respectively. The MAEs of the two measurement result on both left and right lateral flexion. systems are 2.863 degrees and 5.039 degrees, respectively. The E-Skin sensor has a relatively good Sensorsresult 2020on both, 20, 1510 left and right lateral flexion. 20 of 28

Figure 16. Anatomical Angle Outputs Comparison (Left Lateral Flexion). Figure 16. Anatomical Angle Outputs Comparison (Left Lateral Flexion). Figure 16. Anatomical Angle Outputs Comparison (Left Lateral Flexion).

Sensors 2020, 20, x FOR PEER REVIEW 21 of 28 outputs before the left E-Skin sensor outputs when the participant performs left rotation. In contrast, left E-Skin sensor outputs show a fall in the data values before the right sensor when the participant performs right rotation. Based on this clear and repeating trend, the rotation can be clearly detected. Figure 17. Anatomical Angle Outputs Comparison (Right Lateral Flexion). Figure 17. Anatomical Angle Outputs Comparison (Right Lateral Flexion). 250 Rotation Figure 17. Anatomical Angle Outputs Comparison (Right Lateral Flexion). 5 200 0 The results of the left and right rotations are shown in Figures 18 and 19. The ViMove outputs Rotation R) -5 are representedΔ 150 by the orange lines. The patterns of the left and right rotation are similar to the left The results of the left and right rotations are shown in Figures 18 and 19. The ViMove-10 outputs and right lateral flexion movements, but the difference is that there is a rise in the right E-Skin sensor are represented by the orange lines. The patterns of the left and right rotation are similar-15 to the left 100 and right lateral flexion movements, but the difference is that there is a rise in the right -20E-Skin sensor

50 -25 -30

E-Skin Output ( E-Skin Output 0 -35 ViMove (Degree) Output

0 20 40 60 80 100 120 140 160 180 200 Time Sequence (s) Left E-Skin Sensor Right E-Skin Sensor ViMove

Figure 18. Data Outputs of Left Rotation. Figure 18. Data Outputs of Left Rotation.

60 200 50

R) 150 40 Δ

30 100 20

50 10

0 E-Skin Output ( E-Skin Output 0

-10 ViMove (Degree) Output 0 20 40 60 80 100 120 140 160 180 200 Time Sequence (s) Left E-Skin Sensor Right E-Skin Sensor ViMove

Figure 19. Data Outputs of Right Rotation.

The results of linear regression of modelling the mapping between E-Skin outputs and ViMove outputs are consistent with our findings based on Figures 18 and Figure 19. As shown in Table 8, the right E-Skin outputs have a stronger relationship with the actual anatomical angle (ViMove outputs) while there is a relatively weak relationship between left E-Skin outputs and the anatomical angle when the participant performs left rotation. Similar results were also found in right rotation analysis (see Table 9 summary of the comparison). The data quantities of calculating these p-Values of left and right rotation are 193 and 209, respectively.

Table 8. Linear regression comparison for left rotation.

Independent R R2 Adjusted R2 Significance (2-Tailed) Left E-Skin Outputs 0.110 0.012 0.005 p < 0.001 Right E-Skin Outputs 0.747 0.558 0.555 p < 0.001 Both 0.768 0.590 0.584

Sensors 2020, 20, x FOR PEER REVIEW 21 of 28

outputs before the left E-Skin sensor outputs when the participant performs left rotation. In contrast, left E-Skin sensor outputs show a fall in the data values before the right sensor when the participant performs right rotation. Based on this clear and repeating trend, the rotation can be clearly detected.

250 5 200 0

R) -5 Δ 150 -10 -15 100 -20

50 -25 -30

E-Skin Output ( E-Skin Output 0 -35 ViMove (Degree) Output

0 20 40 60 80 100 120 140 160 180 200 Time Sequence (s) Left E-Skin Sensor Right E-Skin Sensor ViMove

Sensors 2020, 20, 1510 Figure 18. Data Outputs of Left Rotation. 21 of 28

60 200 50

R) 150 40 Δ

30 100 20

50 10

0 E-Skin Output ( E-Skin Output 0

-10 ViMove (Degree) Output 0 20 40 60 80 100 120 140 160 180 200 Time Sequence (s) Left E-Skin Sensor Right E-Skin Sensor ViMove

Figure 19. Data Outputs of Right Rotation. Figure 19. Data Outputs of Right Rotation. The results of linear regression of modelling the mapping between E-Skin outputs and ViMove The results of linear regression of modelling the mapping between E-Skin outputs and ViMove outputs are consistent with our findings based on Figures 18 and 19. As shown in Table8, the right outputs are consistent with our findings based on Figures 18 and Figure 19. As shown in Table 8, the E-Skin outputs have a stronger relationship with the actual anatomical angle (ViMove outputs) while right E-Skin outputs have a stronger relationship with the actual anatomical angle (ViMove outputs) there is a relatively weak relationship between left E-Skin outputs and the anatomical angle when the while there is a relatively weak relationship between left E-Skin outputs and the anatomical angle participant performs left rotation. Similar results were also found in right rotation analysis (see Table9 when the participant performs left rotation. Similar results were also found in right rotation analysis summary of the comparison). The data quantities of calculating these p-Values of left and right rotation (see Table 9 summary of the comparison). The data quantities of calculating these p-Values of left and are 193 and 209, respectively. right rotation are 193 and 209, respectively. Table 8. Linear regression comparison for left rotation. Table 8. Linear regression comparison for left rotation. Independent R R2 Adjusted R2 Significance (2-Tailed) Independent R R2 Adjusted R2 Significance (2-Tailed) LeftLeft E-Skin E-Skin OutputsOutputs 0.110 0.110 0.012 0.012 0.005 0.005 pp << 0.0010.001 Right E-Skin Outputs 0.747 0.558 0.555 p < 0.001 Right E-SkinBoth Outputs 0.747 0.768 0.558 0.590 0.555 0.584 p < 0.001 Both 0.768 0.590 0.584

Table 9. Linear regression comparison for right rotation.

Independent R R2 Adjusted R2 Significance (2-Tailed) Left E-Skin Outputs 0.843 0.710 0.709 p < 0.001 Right E-Skin Outputs 0.483 0.233 0.229 p < 0.001 Both 0.843 0.711 0.708

Based on the model, the E-Skin sensor anatomical angle outputs of left and right lateral flexion are generated as shown in Figures 20 and 21 respectively. The MAEs of the two measurement systems are 5.679 degrees and 7.001 degrees, respectively. The E-Skin sensors’ performances on measuring left and right rotations are relatively better than extension. However, the E-Skin sensors’ angle outputs have a larger MAE compare to left and right lateral flexion. The largest angular difference of the two measurement systems for both left and right rotations can be more than 15 degrees.

3.3.5. Summary (Cross Lumbar–Pelvic Movement Analysis) Based on the previous analytical results, the detection and measurement of flexion, extension, and pelvic-tilt are mainly based on the top and bottom sensors. Among which, each E-Skin sensor plays different roles: (1) the bottom E-Skin sensor is important in measuring flexion angle, (2) The top E-Skin sensor is essential for the detection and measurement of pelvic-tilt. The existing sensor placement is not suitable for measure extension. In lateral flexion and rotation analysis, it can be Sensors 2020, 20, x FOR PEER REVIEW 22 of 28

Sensors 2020, 20, x FOR PEER REVIEW 22 of 28 Table 9. Linear regression comparison for right rotation.

IndependentTable 9. LinearR regressionR2 comparisonAdjusted Rfor2 rightSignificance rotation. (2-Tailed) Left IndependentE-Skin Outputs 0.843R 0.710R2 Adjusted 0.709 R2 Significancep < 0.001 (2-Tailed) RightLeft E-Skin Outputs 0.483 0.843 0.233 0.710 0.229 0.709 pp < < 0.001 0.001 Right E-SkinBoth Outputs 0.843 0.483 0.711 0.233 0.708 0.229 p < 0.001 Both 0.843 0.711 0.708 SensorsBased2020 ,on20, 1510the model, the E-Skin sensor anatomical angle outputs of left and right lateral flexion22 are of 28 generatedBased as on shown the model, in Figures the E-Skin 20 and sensor 21 respecti anatomicalvely. The angle MAEs outputs of the of twoleft andmeasurement right lateral systems flexion are are 5.679generated degrees as andshown 7.001 in Figuresdegrees, 20 re spectively.and 21 respecti The vely.E-Skin The sensors’ MAEs performancesof the two measurement on measuring systems left and are seen that the right and left sensors are more important. According to these findings, we can draw right5.679 rotations degrees andare relatively7.001 degrees, better re thanspectively. extension. The However,E-Skin sensors’ the E-Sk performancesin sensors’ onangle measuring outputs lefthave and a a conclusion that vertical sensor placement on human back is suitable for detecting and measuring largerright rotationsMAE compare are relatively to left better and thanright extension.lateral flex However,ion. The thelargest E-Sk inangular sensors’ difference angle outputs of the have two a sagittal plane related movements. However, different sagittal plane movement may require different measurementlarger MAE systemscompare for to both left left and and right right lateral rotations flex canion. be The more largest than 15 angular degrees. difference of the two vertical positions for sensors. On the other hand, sloping or horizontal sensor placements are more measurement systems for both left and right rotations can be more than 15 degrees. suitable to detect and measure the movements which are performed in coronal and transverse planes.

Figure 20. Anatomical Angle Outputs Comparison (Left Rotation). Figure 20. Anatomical Angle Outputs Comparison (Left Rotation). Figure 20. Anatomical Angle Outputs Comparison (Left Rotation).

Figure 21. Anatomical Angle Outputs Comparison (Right Rotation). Figure 21. Anatomical Angle Outputs Comparison (Right Rotation). 4. Discussion Figure 21. Anatomical Angle Outputs Comparison (Right Rotation). 3.3.5. Summary (Cross Lumbar–Pelvic Movement Analysis) 4.1. Principle Results 3.3.5.Based Summary on the (Cross previous Lumbar–Pelvic analytical results, Movement the dete Analysis)ction and measurement of flexion, extension, and In this study, the potential of using E-Skin sensors (resistive flex sensor) to detect and measure pelvic-tiltBased are on mainly the previous based onanalytical the top results, and bottom the dete sensors.ction and Among measurement which, each of flexion, E-Skin extension, sensor plays and lumbar–pelvic movements was identified through a detailed classification and comparison of existing differentpelvic-tilt roles: are mainly(1) the bottombased onE-Sk thein topsensor and is bottomimportant sensors. in measuring Among flexionwhich, angle,each E-Skin (2) The sensor top E-Skin plays human motion sensing technologies. Our experimental results confirmed that these stretchable and sensordifferent is essential roles: (1) for the the bottom detectio E-Skn andin sensor measurement is important of pelvic-tilt. in measuring The existingflexion angle,sensor (2) placement The top E-Skinis not highly sensitive sensors were capable of measuring five lumbar–pelvic movements and could be used sensor is essential for the detection and measurement of pelvic-tilt. The existing sensor placement is not for detecting these movements based on their obvious repeating patterns. The performance of E-Skin sensors was evaluated against the ViMove outputs as a clinically tested gold standard. We applied linear regression to model the mapping of E-Skin outputs to the actual movement angles captured by ViMove. We found that E-skin sensor was capable of measuring simple anatomical joint movements such as trunk flexion. The results also showed the E-Skin had a quicker response to movement changes than the IMU-based sensor systems. It was also able to measure flexion movement faster than ViMove. Sensors 2020, 20, 1510 23 of 28

Overall, our results show that E-Skin sensors have the potential for continuous detection and measurement of lumbar–pelvic movements in out-of-lab settings. They can be used as part of low back pain management and monitoring systems to assist individuals during their rehabilitation.

4.2. Study Limitations In this study, we were able to model the mapping of the E-Skin outputs to actual anatomical joint angles in lumbar–pelvic movements, but this mapping can be further improved. The aim of this study is to explore the potential of using E-Skin sensor as an alternative for the lumbar–pelvic movement detection and measurement. Only limited numbers of participants were recruited in this study. Considering the individual variability in skin deformations when establishing the mapping model, different individuals may have different skin characteristics (e.g., loose and tight), age, height, weight, and BMI. These personalised factors need to be taken into consideration during the modelling. More experiments with different participants who have different skin characteristics and back shapes are needed to collect a wide range of real data for establishing a generic robust mapping model to accurately detect and measure different body movements.

5. Conclusions and Future Work E-Skin sensors have become a new trend in the field of wearable sensors and body motion detection [82]. However, current studies use the E-Skin sensors for measuring movements such as the motion of the throat muscles [83] or fingers [64]. There is a lack of studies that consider using E-Skin sensors for detecting and measuring lumbar–pelvic movements. In the meantime, the need for continuous and out-of-hospital monitoring and management of low back pain propels the use of low-cost and comfortable wearable sensors. The human body is not a rigid and plain surface, and the lumbar spine shape varies from individual to individual. Traditional IMU-based sensors (e.g., ViMove) are attached to the human skin by using medical tapes [42]. Because IMU-based sensors are not stretchable, there may be deviations from the original sensor placement after the individual performs body movements which may lead to measurement errors [70]. Compared to the IMU-based sensors, E-Skin sensors are stretchable and can be attached comfortably to different types and shapes of human skin. Comparing to vision-based sensing system, E-Skin sensor also was identified to be much more comfortable for long-term monitoring and does not have the soft tissue artefacts. These characteristics makes it a suitable alternative for continuous out-of-hospital monitoring of such movements. The experimental results of this paper prove that the vertical placement of the E-Skin sensors on human back is capable of measuring sagittal plane related movements such as flexion. The horizontal/inclined placements of E-Skin sensors on human back is suitable for measuring coronal and transverse plane related movements such as lateral flexion and rotation. Therefore, the E-Skin sensor had great potential to detect the features of different lumbar–pelvic movements and measure anatomical joint angles. For future work, we intend to conduct the experiments on a larger population. We also plan to consider different combinations of the multiple sensor placements. Because skin deformations do not provide any information about the dimension (i.e., the directions of body movements in three-dimensional space), multiple sensors with the optimal sensor placements may provide valuable information for selecting more effective data fusion techniques to model the three-dimensional movements. Different machine learning algorithms and data fusion methods such as SVM, Decision Tree, and kNN [84] need to be explored and tested for establishing a robust and accurate mapping model. We also plan to also extend our work to other types of body movements including knee bending, ankle rotation, and neck flexion.

Author Contributions: Conceptualization, Y.Z., P.D.H., and F.B.; methodology, Y.Z., P.D.H., F.B., and F.C.; validation, Y.Z., P.D.H., L.W.Y., and F.C.; formal analysis, Y.Z., P.D.H., L.Y., and L.W.Y.; investigation, Y.Z. and L.W.Y.; resources, L.W.Y. and W.C.; writing—original draft preparation, Y.Z.; writing—review and editing, L.Y., Sensors 2020, 20, 1510 24 of 28

P.D.H.and F.B.; supervision, P.D.H.and F.B.; project administration, P.D.H., F.B., W.C., and F.C.; funding acquisition, P.D.H., F.B., W.C., and F.C. All authors have read and agreed to the published version of the manuscript. Funding: This research is partly supported by Monash University Institute for Medical engineering (MIME) funding. Conflicts of Interest: The authors declare no conflict of interest.

References

1. Vos, T.; Flaxman, A.D.; Naghavi, M.; Lozano, R.; Michaud, C.; Ezzati, M.; Shibuya, K.; Salomon, J.A.; Abdalla, S.; Aboyans, V.; et al. Years lived with disability (YLDs) for 1160 sequelae of 289 diseases and injuries 1990-2010: A systematic analysis for the Global Burden of Disease Study 2010. Lancet 2012, 380, 2163–2196. [CrossRef] 2. Murray, C.J.L.; Vos, T.; Lozano, R.; Naghavi, M.; Flaxman, A.D.; Michaud, C.; Ezzati, M.; Shibuya, K.; Salomon, J.A.; Abdalla, S.; et al. Disability-adjusted life years (DALYs) for 291 diseases and injuries in 21 regions, 1990-2010: A systematic analysis for the Global Burden of Disease Study 2010. Lancet 2012, 380, 2197–2223. [CrossRef] 3. Hildebrandt, V.H.; Bongers, P.M.; Dul, J.; van Dijk, F.J.; Kemper, H.C. The relationship between leisure time, physical activities and musculoskeletal symptoms and disability in worker populations. Int. Arch. Occup. Environ. Health 2000, 73, 507–518. [CrossRef] 4. Steffens, D.; Ferreira, M.L.; Latimer, J.; Ferreira, P.H.; Koes, B.W.; Blyth, F.; Li, Q.; Maher, C.G. What Triggers an Episode of Acute Low Back Pain? A Case—Crossover Study. Arthritis Care Res. 2015, 67, 403–410. [CrossRef] 5. Guzmán, J.; Esmail, R.; Karjalainen, K.; Malmivaara, A.; Irvin, E.; Bombardier, C. Multidisciplinary rehabilitation for chronic low back pain: Systematic review. BMJ 2001, 322, 1511–1516. [CrossRef] 6. Mjøsund, H.L.; Boyle, E.; Kjaer, P.; Mieritz, R.M.; Skallgård, T.; Kent, P. Clinically acceptable agreement between the ViMove wireless motion sensor system and the Vicon motion capture system when measuring lumbar region inclination motion in the sagittal and coronal planes. BMC Musculoskelet. Disord. 2017, 18, 124. [CrossRef] 7. Laird, R.A.; Kent, P.; Keating, J.L. How consistent are , range of movement and lumbo-pelvic rhythm in people with and without back pain? BMC Musculoskelet. Disord. 2016, 17, 403. [CrossRef] 8. Claus, A.P.; Hides, J.A.; Moseley, G.L.; Hodges, P.W. Thoracic and lumbar posture behaviour in sitting tasks and standing: Progressing the biomechanics from observations to measurements. Appl. Ergon. 2016, 53, 161–168. [CrossRef] 9. Yang, C.-C.C.; Hsu, Y.-L.L. A review of accelerometry-based wearable motion detectors for physical activity monitoring. Sensors 2010, 10, 7772–7788. [CrossRef] 10. Jason, N.N.; Ho, M.D.; Cheng, W. Resistive electronic skin. J. Mater. Chem. C 2017, 5, 5845–5866. [CrossRef] 11. Larson, C.; Peele, B.; Li, S.; Robinson, S.; Totaro, M.; Beccai, L.; Mazzolai, B.; Shepherd, R. Highly Stretchable Electroluminescent Skin for Optical Signaling and Tactile Sensing. Science 2016, 351, 1071–1074. [CrossRef] [PubMed] 12. Wang, Y.; Wang, L.; Yang, T.; Li, X.; Zang, X.; Zhu, M.; Wang, K.; Wu, D.; Zhu, H. Wearable and highly sensitive graphene strain sensors for human motion monitoring. Adv. Funct. Mater. 2014, 24, 4666–4670. [CrossRef] 13. Ling, Y.; An, T.; Yap, L.W.; Zhu, B.; Gong, S.; Cheng, W. Disruptive, Soft, Wearable Sensors. Adv. Mater. 2019, 1904664, 1–13. [CrossRef] 14. Gong, S.; Cheng, W. One-Dimensional Nanomaterials for Soft Electronics. Adv. Electron. Mater. 2017, 3. [CrossRef] 15. Rogers, J.A.; Someya, T.; Huang, Y. Materials and Mechanics for Stretchable Electronics. Science 2010, 327, 1603–1607. [CrossRef] 16. Gong, S.; Lai, D.T.H.; Su, B.; Si, K.J.; Ma, Z.; Yap, L.W.; Guo, P.; Cheng, W. Highly Stretchy Black Gold E-Skin Nanopatches as Highly Sensitive Wearable Biomedical Sensors. Adv. Electron. Mater. 2015, 1, 1–7. [CrossRef] 17. Jason, N.N.; Shen, W.; Cheng, W. Copper Nanowires as Conductive Ink for Low-Cost Draw-On Electronics. ACS Appl. Mater. Interfaces 2015, 7, 16760–16766. [CrossRef] Sensors 2020, 20, 1510 25 of 28

18. Gong, S.; Schwalb, W.; Wang, Y.; Chen, Y.; Tang, Y.; Si, J.; Shirinzadeh, B.; Cheng, W. A wearable and highly sensitive pressure sensor with ultrathin gold nanowires. Nat. Commun. 2014, 5, 1–8. [CrossRef] 19. Gong, S.; Lai, D.T.H.; Wang, Y.; Yap, L.W.; Si, K.J.; Shi, Q.; Jason, N.N.; Sridhar, T.; Uddin, H.; Cheng, W. Tattoolike Polyaniline Microparticle-Doped Gold Nanowire Patches as Highly Durable Wearable Sensors. ACS Appl. Mater. Interfaces 2015, 7, 19700–19708. [CrossRef] 20. Jason, N.N.; Wang, S.J.; Bhanushali, S.; Cheng, W. Skin inspired fractal strain sensors using a copper nanowire and graphite microflake hybrid conductive network. Nanoscale 2016, 8, 16596–16605. [CrossRef] 21. Ho, M.D.; Ling, Y.; Yap, L.W.; Wang, Y.; Dong, D.; Zhao, Y.; Cheng, W. Percolating Network of Ultrathin Gold Nanowires and Silver Nanowires toward “Invisible” Wearable Sensors for Detecting Emotional Expression and Apexcardiogram. Adv. Funct. Mater. 2017, 27, 1–9. [CrossRef] 22. Ryu, S.; Lee, P.; Chou, J.B.; Xu, R.; Zhao, R.; Hart, A.J.; Kim, S. Extremely Elastic Wearable Carbon Nanotube Fiber Strain Sensor for Monitoring of Human Motion. Acsnano 2015, 9, 5929–5936. [CrossRef] 23. Ancillao, A. Stereophotogrammetry in Functional Evaluation: History and Modern Protocols. In Modern Functional Evaluation Methods for Muscle Strength and Gait Analysis; Springer: Cham, Switzerland, 2018; pp. 1–29. 24. Windolf, M.; Götzen, N.; Morlock, M. Systematic accuracy and precision analysis of video motion capturing systems-exemplified on the Vicon-460 system. J. Biomech. 2008, 41, 2776–2780. [CrossRef] 25. Ancillao, A.; Savastano, B.; Galli, M.; Albertini, G. Three dimensional motion capture applied to violin playing: A study on feasibility and characterization of the motor strategy. Comput. Methods Programs Biomed. 2017, 149, 19–27. [CrossRef] 26. Ancillao, A.; Vochten, M.; Aertbeliën, E.; Decré, W.; De Schutter, J. Estimating the instantaneous screw axis and the screw axis invariant descriptor of motion by means of inertial sensors: An experimental study with a mechanical hinge joint and comparison to the optoelectronic system. Sensors 2020, 20, 49. [CrossRef] 27. Müßig, J.A.; Brauner, T.; Kröger, I.; Varady, P.A.; Brand, A.; Klöpfer-Krämer, I.; Simmel, S.; Horstmann, T.; Augat, P. Variability in trunk and pelvic movement of transfemoral amputees using a C-leg system compared to healthy controls. Hum. Mov. Sci. 2019, 68, 102539. [CrossRef] 28. Seerden, S.F.L.; Dankaerts, W.; Swinnen, T.W.; Westhovens, R.; de Vlam, K.; Vanwanseele, B. Multi-segment spine and hip kinematics in asymptomatic individuals during standardized return from forward bending versus functional box lifting. J. Electromyogr. Kinesiol. 2019, 49, 102352. [CrossRef] 29. van den Hoorn, W.; Bruijn, S.M.; Meijer, O.G.; Hodges, P.W.; van Dieën, J.H. Mechanical coupling between transverse plane pelvis and rotations during gait is higher in people with low back pain. J. Biomech. 2012, 45, 342–347. [CrossRef] 30. DiCesare, C.A.; Bates, N.A.; Myer, G.D.; Hewett, T.E. The validity of 2-dimensional measurement of trunk angle during dynamic tasks. Int. J. Sports Phys. Ther. 2014, 9, 420–427. 31. Mitchell, T.; O’Sullivan, P.B.; Burnett, A.F.; Straker, L.; Smith, A. Regional differences in lumbar spinal posture and the influence of low back pain. BMC Musculoskelet. Disord. 2008, 9, 152. [CrossRef] 32. Pacione, C.; Menke, S.; Andre, D.; Teller, E.; Safier, S.; Pelletier, R.; Handel, M.; Farringdon, J.; Hsiung, E.; Vishnubhatla, S. System for monitoring and managing body weight and other physiological conditions including iterative and personalized planning, intervention and reporting capability. United States patent US 8,398,546, 546 March 2013. 33. Bucciero, A.; Santo Sabato, S.; Zappatore, M. A Biomechanical Analysis System of Posture. In Proceedings of the 8th International Conference on Pervasive Computing Technologies for Healthcare, Oldenburg, Germany, 20–23 May 2014. 34. Brooks, A.L.; Czarowicz, A. Markerless motion tracking: MS Kinect and Organic Motion OpenStage. In Proceedings of the International Conference Disability, Virtual Reality & Associated Technologies, Laval, France, 10–12 September 2012; pp. 10–12. 35. Balazia, M.; Plataniotis, K.N. Human gait recognition from motion capture data in signature poses. IET Biome. 2017, 6, 129–137. [CrossRef] 36. Yang, G.; Tan, W.; Jin, H.; Zhao, T.; Tu, L. Review wearable sensing system for gait recognition. Cluster Comput. 2019, 22, 3021–3029. [CrossRef] 37. Cornacchia, M.; Ozcan, K.; Zheng, Y.; Velipasalar, S. A Survey on Activity Detection and Classification Using Wearable Sensors. IEEE Sens. J. 2017, 17, 386–403. [CrossRef] Sensors 2020, 20, 1510 26 of 28

38. Intolo, P.; Carman, A.B.; Milosavljevic, S.; Abbott, J.H.; Baxter, G.D. The Spineangel: Examining the validity and reliability of a novel clinical device for monitoring trunk motion. Man. Ther. 2010, 15, 160–166. [CrossRef] 39. Ribeiro, D.C.; , G.; Abbott, J.H.; Milosavljevic, S. Validity and reliability of the Spineangel lumbo-pelvic postural monitor. Ergonomics 2013, 56, 977–991. [CrossRef] 40. Lumo Lift, Lumo Lift Posture Coach. 2016. Available online: http://www.lumobodytech.com/ (accessed on 14 June 2017). 41. Upright Tech UPRIGHT Posture Trainers. 2016. Available online: https://www.uprightpose.com/ (accessed on 14 June 2017). 42. Charry, E.; Umer, M.; Taylor, S. Design and validation of an ambulatory inertial system for 3-D measurements of low back movements. In Proceedings of the 2011 Seventh International Conference on Intelligent Sensors, Sensor Networks and Information Processing, Adelaide, Australia, 6–9 December 2011; pp. 58–63. 43. Schelldorfer, S.; Ernst, M.J.; Rast, F.M.; Bauer, C.M.; Meichtry, A.; Kool, J. Low back pain and postural control, effects of task difficulty on centre of pressure and spinal kinematics. Gait Posture 2015, 41, 112–118. [CrossRef] 44. Singh, A.; Klapper, A.; Jia, J.; Fidalgo, A.; Tajadura-Jiménez, A.T.; Kanakam, N.; Bianchi-Berthouze, N.; Williams, A. Motivating People with Chronic Pain to do Physical Activity: Opportunities for Technology Design. In Proceedings of the SIGCHI Conference on Human Factors in Computing, Toronto, ON, Canada, April 2014; pp. 2803–2812. 45. Haines, T.; Bowles, K.-A. Cost-effectiveness of using a motion-sensor biofeedback treatment approach for the management of sub-acute or chronic low back pain: Economic evaluation alongside a randomised trial. BMC Musculoskelet. Disord. 2017, 18, 18. [CrossRef] 46. Kent, P.; Laird, R.; Haines, T. The effect of changing movement and posture using motion-sensor biofeedback, versus guidelines-based care, on the clinical outcomes of people with sub-acute or chronic low back pain-a multicentre, cluster-randomised, placebo-controlled, pilot trial. BMC Musculoskelet. Disord. 2015, 16, 131. [CrossRef] 47. Sabatini, A.M. Estimating three-dimensional orientation of human body parts by inertial/magnetic sensing. Sensors 2011, 11, 1489–1525. [CrossRef] 48. Borenstein, J.; Ojeda, L.; Kwanmuang, S. Heuristic reduction of gyro drift in IMU-based personnel tracking systems. Opt. Photonics Glob. Homel. Secur. V Biom. Technol. Hum. Identif. VI 2009, 7306, 73061H. 49. Jiménez, A.R.; Seco, F.; Zampella, F.; Prieto, J.C.; Guevara, J. Ramp Detection with a Foot-Mounted IMU for a Drift-Free Pedestrian Position Estimation. In Proceedings of the International Conference on Indoor Positioning and Indoor Navigation, Guimarães, Portugal, 21–23 September 2011; pp. 21–23. 50. Ilyas, M.; Cho, K.; Baeg, S.H.; Park, S. Drift reduction in IMU-only pedestrian navigation system in unstructured environment. In Proceedings of the 2015 10th Asian Control Conference (ASCC), Kota Kinabalu, Malaysia, 31 May–3 June 2015; pp. 1–7. 51. Attal, F.; Mohammed, S.; Dedabrishvili, M.; Chamroukhi, F.; Oukhellou, L.; Amirat, Y. Physical Human Activity Recognition Using Wearable Sensors. Sensors 2015, 15, 31314–31338. [CrossRef] 52. Chhikara, A.; Rice, A.S.C.; Mcgregor, A.H.; Bello, F. Wearable device for monitoring disability associated with Low Back Pain. In Proceedings of the International Workshop of Wearable Micro Nanosystems for Personalized Health, Valencia, Spain, 21–23 May 2008. 53. Lin, W.-Y.; Chou, W.-C.; Tsai, T.-H.; Lin, C.-C.; Lee, M.-Y. Development of a Wearable Instrumented Vest for Posture Monitoring and System Usability Verification Based on the Technology Acceptance Model. Sensors 2016, 16, 2172. [CrossRef][PubMed] 54. Wang, Q.; Toeters, M.; Chen, W.; Timmermans, A.; Markopoulos, P. Zishi: A Smart Garment for Posture Monitoring. In Proceedings of the 2016 CHI Conference Extended Abstracts on Human Factors in Computing Systems, San Jose, CA, USA, 7–12 May 2016; pp. 3792–3795. 55. Yuan, Q.; Chen, I.M. 3-D localization of human based on an inertial capture system. IEEE Trans. Robot. 2013, 29, 806–812. [CrossRef] 56. Dunne, L.E.; Walsh, P.; Smyth, B.; Caulfield, B. Design and Evaluation of a Wearable Optical Sensor for Monitoring Seated Spinal Posture. In Proceedings of the 2006 10th IEEE International Symposium on Wearable Computers, Montreux, Switzerland, 11–14 October 2006; pp. 65–68. 57. Kam, W.; Sullivan, K.O.; Mohammed, W.S.; Lewis, E. Low cost portable sensor for real-time monitoring of lower back bending. In Proceedings of the 2017 25th Optical Fiber Sensors Conference (OFS), Jeju, Korea, 24–28 April 2017; Volume 10323, pp. 3–6. Sensors 2020, 20, 1510 27 of 28

58. Chandra, S.; Nag, A.; Mukhopadhyay, S.C.; Kosel, J. Wearable Flexible Sensors: A Review. IEEE Sens. J. 2017, 17, 3949–3960. 59. Saggio, G.; Riillo, F.; Sbernini, L.; Quitadamo, L.R. Resistive flex sensors: A survey. Smart Mater. Struct. 2015, 25, 13001. [CrossRef] 60. Khan, S.; Lorenzelli, L.; Dahiya, R.S. Technologies for printing sensors and electronics over large flexible substrates: A review. IEEE Sens. J. 2015, 15, 3164–3185. [CrossRef] 61. Wang, H.; Zhou, D.; Cao, J. Development of a skin-like tactile sensor array for curved surface. IEEE Sens. J. 2014, 14, 55–61. [CrossRef] 62. Mukhopadhyay, S.C. Wearable sensors for human activity monitoring: A review. IEEE Sens. J. 2014, 15, 1321–1330. [CrossRef] 63. Amjadi, M.; Pichitpajongkit, A.; Lee, S.; Ryu, S.; Park, I. Highly Stretchable and Sensitive Strain Sensor Based on Silver-Elastomer Nanocomposite. ACS Nano 2014, 8, 5154–5163. [CrossRef] 64. Dutta, D.; Champaty, B.; Pal, K.; Banerjee, I. movement based attender calling system for ICU patient management and rehabilitation. In Proceedings of the 2014 International Conference on Circuits, Power and Computing Technologies [ICCPCT-2014], Nagercoil, India, 20–21 March 2014; pp. 394–397. 65. StretchSense Stretch Sensors. Available online: https://www.stretchsense.com/stretch-sensors/ (accessed on 22 January 2019). 66. Xu, D.; Tairych, A.; Anderson, I.A. Where the Rubber Meets the Hand: Unlocking the Sensing Potential of Dielectric Elastomers. J. Polym. Sci. 2016, 54, 465–472. [CrossRef] 67. Menguc, Y.; Park, Y.L.; Martinez-Villalpando, E.; Aubin, P.; Zisook, M.; Stirling, L.; Wood, R.J.; Walsh, C.J. Soft wearable motion sensing suit for lower biomechanics measurements. In Proceedings of the 2013 IEEE International Conference on Robotics and Automation, Karlsruhe, Germany, 6–10 May 2013; pp. 5309–5316. 68. Yamada, T.; Hayamizu, Y.; Yamamoto, Y.; Yomogida, Y.; Izadi-Najafabadi, A.; Futaba, D.N.; Hata, K. A stretchable carbon nanotube strain sensor for human-motion detection. Nat. Nanotechnol. 2011, 6, 296–301. [CrossRef][PubMed] 69. Dominguez Veiga, J.J.; O’Reilly, M.; Whelan, D.; Caulfield, B.; Ward, T.E. Feature-Free Activity Classification of Inertial Sensor Data With Machine Vision Techniques: Method, Development, and Evaluation. JMIR mHealth uHealth 2017, 5, e115. [CrossRef][PubMed] 70. Beaudette, S.M.; Zwambag, D.P.; Bent, L.R.; Brown, S.H.M. Spine postural change elicits localized skin structural deformation of the trunk dorsum in vivo. J. Mech. Behav. Biomed. Mater. 2017, 67, 31–39. [CrossRef] [PubMed] 71. Yang, C.-M.; Yang, T.-L.; Huang, W.-T.; Chen, C.-H.; Hung, S.-H.; Cheng, C.-M.; Cheng, M.-H. A novel design and evaluation of wearable digital sensor for monitoring posture. Conf. Proc. IEEE Eng. Med. Biol. Soc. 2008, 2008, 1304–1307. [PubMed] 72. Raytac Corporation. Available online: https://www.raytac.com/ (accessed on 9 November 2018). 73. Yap, L.W.; Shi, Q.; Gong, S.; Wang, Y.; Chen, Y.; Zhu, C.; Gu, Z.; Suzuki, K.; Zhu, Y.; Cheng, W. Bifunctional Fe 3 O 4 @AuNWs particle as wearable bending and strain sensor. Inorg. Chem. Commun. 2019, 104, 98–104. [CrossRef] 74. Ozer, A.; John, E. Improving the accuracy of bluetooth low energy indoor positioning system using kalman filtering. In Proceedings of the 2016 International Conference on Computational Science and Computational Intelligence (CSCI), Las Vegas, NV, USA, 15–17 December 2016; pp. 180–185. 75. Da Gama, A.E.F.; Chaves, T.d.M.; Fallavollita, P.; Figueiredo, L.S.; Teichrieb, V. Rehabilitation motion recognition based on the international biomechanical standards. Expert Syst. Appl. 2019, 116, 396–409. [CrossRef] 76. Saraceni, N.; Kent, P.; Ng, L.; Campbell, A.; Straker, L.; O’Sullivan, P. To flex or not to flex? Is there a relationship between lumbar spine flexion during lifting and low back pain? A systematic review with meta-analysis. J. Orthop. Sport. Phys. Ther. 2020, 50, 121–130. [CrossRef][PubMed] 77. Martinec, E. Noise, Dynamic Range and Bit Depth in Digital SLRs; University Chicago: Chicago, IL, USA, 2008. 78. Gong, S.; Yap, L.W.; Zhu, B.; Cheng, W. Multiscale Soft–Hard Interface Design for Flexible Hybrid Electronics. Adv. Mater. 2019, 1902278, 1–28. [CrossRef] 79. Liu, W.; Li, F.; Stefanini, C.; Chen, D.; Dario, P. Biomimetic flexible/compliant sensors for a soft-body lamprey-like robot. Robot. Auton. Syst. 2010, 58, 1138–1148. [CrossRef] Sensors 2020, 20, 1510 28 of 28

80. Gong, S.; Yap, L.W.; Zhu, B.; Zhai, Q.; Liu, Y.; Lyu, Q.; Wang, K.; Yang, M.; Ling, Y.; Lai, D.T.H.; et al. Local Crack-Programmed Gold Nanowire Electronic Skin Tattoos for In-Plane Multisensor Integration. Adv. Mater. 2019, 31, 1–8. [CrossRef] 81. De Quadros, T.; Lazzaretti, A.E.; Schneider, F.K. A Movement Decomposition and Machine Learning-Based Fall Detection System Using Wrist Wearable Device. IEEE Sens. J. 2018, 18, 5082–5089. [CrossRef] 82. Wang, X.; Zhang, Y.; Zhang, X.; Huo, Z.; Li, X.; Que, M.; Peng, Z.; Wang, H.; Pan, C. A Highly Stretchable Transparent Self-Powered Triboelectric Tactile Sensor with Metallized Nanofibers for Wearable Electronics. Adv. Mater. 2018, 30, 1–8. [CrossRef][PubMed] 83. Hwang, B.U.; Lee, J.H.; Trung, T.Q.; Roh, E.; Kim, D.I.; Kim, S.W.; Lee, N.E. Transparent Stretchable Self-Powered Patchable Sensor Platform with Ultrasensitive Recognition of Human Activities. ACS Nano 2015, 9, 8801–8810. [CrossRef][PubMed] 84. Sun, Y.; Huang, R.; Zheng, J.; Dong, D.; Chen, X.; Bai, L.; Ge, W. Design and speed-adaptive control of a powered geared five-bar prosthetic knee using bp neural network gait recognition. Sensors 2019, 19, 4662. [CrossRef][PubMed]

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).