This is a preprint. Please cite this work as: A.L. Alfeo, P. Barsocchi, M.G.C.A. Cimino, D. La Rosa, F. Palumbo, G. Vaglini, "Sleep behavior assessment via smartwatch and stigmergic receptive fields", Personal and Ubiquitous Computing, Springer, Vol. 22, Issue 2, Pages 227-243, 2017, (ISSN 1617- 4909) Machine Learning and Process Intelligence Research Group Paper draft - please export an up-to-date reference from http://www.iet.unipi.it/m.cimino/pub ORIGINAL ARTICLE Sleep behavior assessment via smartwatch and stigmergic receptive fields Antonio L. Alfeo1 · Paolo Barsocchi2 · MarioG.C.A.Cimino1 · Davide La Rosa2 · Filippo Palumbo2 · Gigliola Vaglini1 Abstract Sleep behavior is a key factor in maintaining reliable, and innovative computational approach to provide good physiological and psychological health. A well-known per-night assessment of sleep behavior to the end-user. We approach to monitor sleep is polysomnography. However, exploit heartbeat rate and wrist acceleration data, gathered it is costly and intrusive, which may disturb sleep. Conse- via smartwatch, in order to identify subject’s sleep behav- quently, polysomnography is not suitable for sleep behavior ioral pattern. More specifically, heartbeat rate and wrist analysis. Other approaches are based on actigraphy and motion samples are processed via computational stigmergy, sleep diary. Although being a good source of information a bio-inspired scalar and temporal aggregation of samples. for sleep quality assessment, sleep diaries can be affected Stigmergy associates each sample to a digital pheromone by cognitive bias related to subject’s sleep perception, deposit (mark) defined in a mono-dimensional space and while actigraphy overestimates sleep periods and night-time characterized by evaporation over time. As a consequence, disturbance compared to sleep diaries. Machine learning samples close in terms of time and intensity are aggregated techniques can improve the objectivity and reliability of into functional structures called trails. The stigmergic trails the observations. However, since signal morphology vary allow to compute the similarity between time series on dif- widely between people, conventional machine learning is ferent temporal scales, to support classification or clustering complex to set up. In this regard, we present an adaptive, processes. The overall computing schema includes a para- metric optimization for adapting the structural parameters to individual sleep dynamics. The outcome is a similarity between sleep nights of the same subject, to generate clus- MarioG.C.A.Cimino ters of nights with different quality levels. Experimental [email protected] results are shown for three real-world subjects. The resulting Antonio L. Alfeo similarity is also compared with the dynamic time warping, [email protected] a popular similarity measure for time series. Paolo Barsocchi [email protected] Keywords Sleep monitoring · Smartwatch · Stigmergy · Davide La Rosa Neural receptive field [email protected] Filippo Palumbo [email protected] 1 Introduction Gigliola Vaglini [email protected] One of the most important markers of a healthy lifestyle is represented by the quality and quantity of sleep. These 1 Department of Information Engineering, University of Pisa, factors directly affect the waking life, including pro- Largo L. Lazzarino 1, 56122 Pisa, Italy ductivity, emotional balance, creativity, physical vitality, 2 National Research Council, Institute of Information Science and the general personal health. Indeed, poor long-term and Technologies, via G. Moruzzi 1, 56124 Pisa, Italy sleep patterns can lead to a wide range of health-related Pers Ubiquit Comput problems, such as high-blood pressure, high stress, anxiety, smartphone sensors only, have not been validated by sci- diabetes, and depression [15]. In this context, the moni- entific literature or studies [50]. This is even more evi- toring of sleep patterns becomes of major importance for dent when considering long-term analysis. Many people various reasons, such as the detection and treatment of sleep track their sleep through mobile and wearable technology, disorders, the assessment of the effect of different med- together with contextual information that may influence ical conditions or medications on the sleep quality, and sleep quality, like exercise, diet, and stress. However, there the assessment of mortality risks associated with sleeping is limited support to help people make sense of this wealth patterns in adults and children [46]. of data, i.e., to explore the relationship between sleep data Traditionally, the polysomnographic (PSG) recordings and contextual data. In [41], authors try to bridge the have been widely used in order to infer the sleep quality [1]. gap between sleep-tracking and sense-making through the In particular, the PSG recordings include the electroen- design of a web-based tool that helps individuals understand cephalography (EEG), the electrooculogram (EOG), and the sleep quality. However, an automatic tool able to moni- electromyogram (EMG) data of the patient [16]. In this tor the sleep over the long period and give a user-tailored regard, the quality measure is usually captured with self- quality measure is still missing. reports via paper-based surveys and diaries that, although Indeed, most sleep scoring algorithms provide a being difficult and tedious to be collected, represent a reli- threshold-based analysis of subject’s activeness during the able source of information [12]. Nevertheless, sleep diaries whole night. Unfortunately, due to peculiarities of each sub- can be affected by cognitive bias related to the subject’s ject’s sleep, same thresholds cannot be effective for any sleep perception. user nor exhaustive for a sleep behavior analysis. As an In recent years, because of the development of ubiqui- example, same REM sleep ratio values can be obtained tous technology in health care, the research effort involving with nights characterized by different number and duration non-invasive sensors to assess and report sleep patterns is of REM-NonREM cycles, which is an important behav- actively progressing. In [46], contact-based pressure mat- ioral difference. In contrast with more traditional scoring tress and a non-contact 3D image acquisition device have algorithms, novel machine learning approaches can provide been used for sleep monitoring. In [37], motion detection, greater accuracy due to their ability to generate nonlin- sound, and vibration sensor give information about sleeping. ear classification borders [56]; moreover, they can improve A relevant source of information on sleeping is repre- the objectivity and reliability of the observations [39]. On sented by motion data coming from worn inertial sensors the other hand, the use of machine learning techniques (i.e., accelerometers) embedded in smartphones or wrist- often requires a careful tuning of their structural parameters, bands [50]. Indeed, data coming from worn devices has been which can be provided by employing an expert in the field thoroughly exploited in different scenarios aiming at moni- or even via brute-force search [60]. toring human activities [51], falls [30], or gait analysis [20, In this paper, we present an automatic tool for monitoring 21]. The use of this kind of information is the basis of the sleep behavior that uses a commercially available smart- so-called actigraphy. watch, in order to sample heart rate and wrist inertial data, Actigraphy is a non-invasive method of monitoring and a novel detection technique based on stigmergic recep- human rest/activity cycles. In 1995, the Standards of Prac- tive fields (SRFs). The SRF transforms an input time series tice Committee of the American Sleep Disorders Associa- into a stigmergic trail and provides a (dis-)similarity mea- tion (ASDA) commissioned a task force to evaluate the role sure against another signal trail. The dissimilarity measure of actigraphy in sleep medicine. The term actigraphy refers can be parametrically optimized according to sleep qual- to methods using wristband-like devices to monitor and ity annotations on a set of reference nights. The proposed collect data generated by movements. This involves a per- approach is able to handle the continuous availability of son wearing an accelerometer-based device on their wrist. data samples from many subjects, possibly using differ- ASDA’s effort on actigraphy led to a review paper on the ent models of devices among them. This can be effectively topic [56, 57] and a set of guidelines [4]. The acknowledge achieved by using a middleware communication platform for actigraphy as a valid tool by ASDA was an impor- which enables the possibility to connect different kinds tant landmark in its acceptance by sleep-related researchers of smart wristbands, thus supporting many technologies. and clinicians. The use of actigraphy is continuously ris- The presence of the middleware also enables the long- ing in sleep research and medicine, as demonstrated by the term scenario and the use of environmental/domotic sensors increasing number of publications over the years [11, 56]. for future developments in data fusion algorithms for the While commercially available activity trackers based on identification of correlations between sleep quality, daily wearable devices can be considered valid for measuring activities, and the characteristics of the surrounding envi- sleep phases and heart rate (HR) during sleep [63, 64], ronment.