Monitoring Health in Smart Homes using Simple Sensors Stewart Massie1 , Glenn Forbes1 , Susan Craw1 , Lucy Fraser2 Graeme Hamilton2 1 Robert Gordon University, Aberdeen, UK 2 Albyn Housing Society Ltd, Invergordon, UK s.massie@rgu.ac.uk, g.forbes6@rgu.ac.uk, s.craw@rgu.ac.uk Abstract specially-designed, technology-enabled “FitHomes”, target- ing specific activities identified as pre-cursors to falls. An We consider use of an ambient sensor network, in- outline solution is developed and initial experiments are un- stalled in Smart Homes, to identify low level events dertaken to evaluate alternative approaches to classifying ac- taking place which can then be analysed to gen- tivity with low level, raw data inputs from multiple sensors. erate a resident’s profile of activities of daily liv- ing (ADLs). These ADL profiles are compared to both the resident’s typical profile and to known 2 Related Work “risky” profiles to support evidence-based interven- Activities of Daily Living (ADLs) are events in daily life tions. Human activity recognition to identify ADLs which would be considered intrinsic to a person’s ability to from sensor data is a key challenge, a window- live independently. ADLs include being able to dress one- based representation is compared on four existing self, get out of bed, and feed oneself. Katz [Katz et al., 1963] datasets. We find that windowing works well, giv- originally proposed the term along with a scale for rating a ing consistent performance. We also introduce FIT- person’s independent ability using their performance in sim- sense, which is building a Smart Home environ- ple ADLs. The concept of losing ADLs as we age has in- ment to specifically identify increased risk of falls fluenced future research in the field by identifying that spe- to allow interventions before falls occurs. cific ADLs are more indicative of reduced capability than others. Observing variances in ADL performance can aid the identification of degenerative mental and physical capability. 1 Introduction Vestergaard [Vestergaard et al., 2009] identified a relation- Like most countries world-wide, the United Kingdom is fac- ship between performance in a 400-meter walk test and sub- ing an ageing population with many people living longer. Ten sequent mortality rates. This test is usually performed in hos- million people in the UK are currently over 65 with a fur- pital which allows a physician to also consider other metrics ther increase of 5.5 million projected over the next 20 years. from the test. These include but are not limited to, whether Three million people are aged over 80 which is expected to or not a break was taken, variation in lap times and existing double by 2030 [Population Division and Affairs, 2015]. An health conditions. However, lab-based testing is time con- ageing population puts additional strains on health and social suming, costly and impractical for many patients, especially services with both a smaller proportion of working popula- those with mobility issues. In addition, some studies have tion available to support services, and with the elderly having been able to identify risk of falling in the elderly using gait more complex medical needs. In this changing scenario it is velocity alone [Stone and Skubic, 2013; Jiang et al., 2011; important that we help people with mobility or social needs Montero-Odasso et al., 2005]. So while gait and other expres- to live independently for longer, and so reduce their reliance sions of movement are indicative of many underlying condi- on more expensive health care solutions. tions, measuring all aspects of gait, such as swing and stride In this paper we examine the potential for Smart Homes length, requires specialist equipment. Gait velocity, however, to support assisted living environments by monitoring health can be measured using simpler equipment and still provides trends. A mixed approach is proposed which exploits the excellent insight into subject movement. For instance, Rana patterns of activities identified by sensors to infer informa- [Rana et al., 2016] performed a study in which gait velocity tion about the health of the residents. We explore the use was estimated using simple infra-red motion sensors. We plan of everyday, low-cost ambient sensors installed in new-build to adopt this approach and, while lab-based testing can pro- Smart Homes with the aim of supporting tenants to live in- vide higher accuracy, envisage that accessible in-home testing dependently for longer. Specifically, we identify and discuss can contribute to early detection of health problems. the main challenges with an ongoing project to design and Housing installations with ubiquitous simple sensors offer deploy a real-world health monitoring system that senses and an opportunity to provide continuous behavioural and phys- predicts the level of risk of falling attributed to Smart Home iological monitoring of residents. These simple sensors can residents. Data is captured by a range of sensors installed in range from binary magnetic switches [Tapia et al., 2004], to IoT-monitored motion sensors [Suryadevara and Mukhopad- e.g. a door opening; however others output give continuous hyay, 2012], all of which can provide insight into behavioural readings provided at fixed polling rates. The data fusion task and physiological expressions. ADLs can then be modelled across multiple sensors with different output modes becomes by identifying temporal patterns in these sensor outputs. Sev- one of the main challenges in employing large numbers of eral manually annotated datasets taken from Smart Home in- sensors. We employ a two stage process. The first stage is stallations have been produced for the purpose of activity to generate activity profiles. This requires pattern identifica- recognition [Tapia et al., 2004; Van Kasteren et al., 2008; tion to create meaningful representations from the raw sensor Cook and Schmitter-Edgecombe, 2009]. We use these exist- data that capture the residents’ activities e.g. sleeping, dress- ing datasets in our experiments to evaluate the effectiveness ing, showering, cooking, and then to assemble these activities of alternative classification approaches. into personalised daily and weekly profiles. The second stage is the analysis of these profiles to enable both the identifica- 3 FitHomes & Predicting Falls tion of changing trends in the resident’s activities over time and to make comparisons with data collected from other sim- FitHomes is an initiative, lead by Albyn Housing Society Ltd ilar residents. Changes in the Smart Home resident’s own (AHS), that aims to support independent living with the sup- activity patterns over time can then be used to detect deterio- ply of custom-built Smart Homes fitted with integrated non- ration in health, while comparisons with the patterns of other invasive sensors. Sixteen houses are being built and near Smart Home residents can provide benchmark measures of completion at Alness near Invergordon. These houses are part health. The data thus supports evidence-driven intervention of a development cycle with further FitHomes planned to be tailored to the resident and their specific circumstances. built within the Inverness area within 3 years. FITsense 1 is a one year project that aims to exploit the sensor data to de- velop a prototype fall prediction system for the residents of 4 Generating Activity Profiles FitHomes. Human Activity Recognition (HAR) to identify ADLs is chal- A key consideration in designing a Smart Home focused lenging in Smart Home scenarios because large volumes of on health monitoring is the choice, mix and location of sen- data is generated from multi-modal sensors in real time mak- sors to use in order to provide a cost-effective solution that ing patterns associated with specific activities difficult to is also acceptable to residents. AHS have conducted initial identify. Figure 1 shows a diagram with example sensor ac- research and it is clear that their tenants want an unobtrusive tivations for motion sensors in a hall, kitchen and lounge to- system. Both video and wearables are considered too intru- gether with pressure sensors on the chair and bed. Simple sive for continuous in-home use; video due to privacy issues events can be inferred from this data to generate activities. A and wearables due to the ongoing overhead associated with mix of approaches will be adopted to identify activities and 24 hour operation. As a result, the focus in this project is on to then generate daily activity profiles. For the simple activ- simple everyday sensors, many of which are already widely ities shown (e.g. time sitting, time in bed, number of toilet used in security and automation applications. FITsense is an visits, number of room transitions) only one or two sensor applied project and with this approach we can establish the activations are required to identify the activity; a rule-based limits of existing technology now, rather than developing new approach with simple rules is sufficient and where effective solutions for the future. A further benefit is provision of a this approach will be adopted. low cost solution from the hardware perspective but with ad- More complex activities can only be recognised by the in- ditional challenges for the data analysis. teraction of several sensors e.g. food preparation, shower- FITsense aims to identify increased risk of falls and so a ing, disturbed sleep. For these more complex activities a key focus for monitoring is to identify activity levels, pat- Machine Learning (ML) approach will be adopted. HAR terns and speeds. However, monitoring can go beyond just typically employs a windowing approach to create a single movement to consider other factors that have been shown to aggregated vector representation on which ML (e.g. kNN, be related to falls, including dehydration, tiredness and men- Support Vector Machines or Naive Bayes) can be applied for tal health. Gaining information on these additional factors classification. These approaches can work well but are per- requires monitoring to also capture data on more general ac- haps less able to handle the data fusion scenarios from Smart tivities such as eating & drinking behaviours, sleep patterns, Homes because of difficulties in selecting appropriate time and toileting & grooming habits. With these criteria in mind windows for different activities; and due to the loss of infor- a range of sensors have been selected for the FitHomes, that mation when the sequence of events is not maintained, by ag- include: IR motion sensors that capture movement in each gregating within a window. We investigate the performance room; contact sensors to capture room, cupboard, and fridge of a windowing approach. door opening/closing; pressure sensors that identify use of the bed and chairs; IR beam break sensors used to identify 5 Reasoning with ADLs gait speed; electricity smart meters to identify power usage pattern; float sensors identifying toilet flushing; and humidity Identifying ADLs in themselves does not give an indication sensors to identify shower use. of health. However, it has been shown that functional assess- Most of the sensors chosen have a binary output that sim- ment is an effective way to evaluate the health status of older ply activate when the event they are monitoring takes place adults [Cook et al., 2015]; ADLs are lost as we age and in FITsense the plan is to monitor changes in ADL activity as 1 www.rgu.ac.uk/fitsense an indicator of deteriorating health and increased risk of falls. Figure 1: Identifying activities from sensor activations A Case-Based Reasoning (CBR) approach is adopted. In our periments: CASAS2 (adlnormal), Van Kasteren3 (kasteren) scenario, a set of ADL templates (together with contextual and two from the Massachusetts Institute of Technology4 information) is used as the problem representation to retrieve (tapia1/2). These datasets share similar properties to that ex- similar profiles from a case base of existing profiles. Solu- pected from FitHomes. They all capture binary sensor acti- tions will identify interventions, where required, and their vation data from the homes and have been labelled with class previous outcomes. information, i.e. the ADL identified during the specified time Figure 2 presents an overview of our approach. Low-level, period. time-stamped events identified by the sensors are transformed into a daily user profile. The profiles are a set of ADLs with Table 1: Overview of the datasets used. mixed data types: some ADLs are binary, e.g. disturbed sleep; some ADLs are counts, e.g. number of room transi- tions or stand up from seat count; some are cumulative daily Dataset Classes Attributes Instances time spans, e.g. time sitting; while others are numeric, e.g. adlnormal 5 39 120 average gait speed. Whatever the data type a similarity mea- sure is associated with each ADL so that comparison can be kasteren 7 14 242 made between them. A set of daily ADL profiles for a res- ident can then be compared with those in the case base, on tapia1 22 76 295 the right of Figure 2. Retrieval of similar profiles labelled as at risk identifies the need to recommend intervention, and tapia2 24 70 208 falling similarity with the user’s own previous profiles iden- tifies changing behaviours. Importance in determining simi- larity for FITsense is given to ADLs known to correlate with Table 1 gives an overview of the structure of the datasets. falls. For other health conditions the similarity knowledge These are relatively small datasets with between 120 and 295 could be refined to reflect specific conditions e.g. gait for instances, reflecting the high cost of manual labelling. The falls, erratic behaviour for Dementia, general physical activ- number of attributes varies between 14 and 76 reflecting dif- ity level for obesity, etc. ferences in the number of sensors present in different instal- A key challenge is to identify “risky” or deteriorating be- lation set ups. Likewise, there are differences in the number haviour. Labelled positive cases (identifying a fall is likely) of activities being monitored (i.e. classes) depending on the are rare because people don’t fall that often. The initial ap- focus of the particular study; tapia in particular has a large proach is to generate template solutions with guidance from number of different activity labels, some of which would not health care professionals. Then, as real data becomes avail- be relevant for predicting falls. Some activities are more pop- able, we can learn/refine/supplement these hand-crafted tem- ular than others and as a result most datasets do not have bal- plates with the addition of real experiences as they occur in anced class distributions. The window-based representation the data generated both by the user and by others. we use is a fixed-length vector which does not change with varying activity lengths. If we count the number of sensors in the installation there will be one attribute for each sensor. 6 Evaluation The attribute value being a count of the number of times the The initial task is to assess effectiveness at classifying ADLs 2 from raw sensor data. Live data is not yet available from http://casas.wsu.edu/datasets/adlnormal.zip 3 the FitHomes, so in this evaluation existing datasets are https://sites.google.com/site/tim0306/kasterenDataset.zip 4 used. Four publicly available datasets are used in our ex- http://courses.media.mit.edu/2004fall/mas622j/04.projects Figure 2: CBR Approach to Identifying ’Risky’ Behaviours sensor is activated during an activity timespan. The solution High accuracies, generally in excess of 90%, are achieved is a single class label, namely the labelled activity. on adlnormal and kasteren compared to highs of 61% and 47% on tapia1 and 2 respectively. The differences reflect 6.1 Experiment Set-Up that both tapia datasets present a much harder classification Popular ML algorithms that delivered good performance on task with over 20 fine grained activities, many of which are these datasets were selected from the Weka library to run with hard to distinguish even with over 70 sensors. Adlnormal default settings on the window-based representation of each and kasteren have fewer activities being identified (5 and dataset [Hall et al., 2009]. These were compared to Condi- 7 respectively) and fewer sensors (39 and 14 respectively). tional Random Fields (CRFs) also run with default settings on Kasteren in particular is more in line with the type of activi- the CRF++ toolkit. Both tools make use of different data for- ties and sensor network we plan for FITsense. mats, so each dataset was converted to ARFF (for Weka), and There is not a clear winner. BayesNet, k-NN and J48 CSV (for CRF++). Given the limited data available, Leave- provide good performance on the simpler datasets (adlnor- One-Out cross validation was applied on all experiments and mal and kasteren); k-NN gives highest accuracy on kasteren average accuracy results were recorded. which, having the fewest sensors and shortest activity se- • Bayes Network: Using the BayesNet bayes classifier. quences, is likely to have few noisy attributes. BayesNet • k-NN: Using the IBk lazy classifier (with k=3). gives highest accuracy on adlnormal which is distinguished by having long sensor sequences associated with activities. • SVM: Using the SMO function classifier. CRF gives highest accuracies on the more complex tapia • J48: Using the J48 tree classifier. datasets, which seems to indicate that the relationship be- • CRF: Using CRFs on the window-based representation. tween sensor activations becomes more important for distin- guishing similar activities from each other. Table 2: Experiment results (in accuracy %). 7 Conclusions Dataset BayesN k-NN SVM J48 CRF In this paper we have presented a Smart Home approach to adlnormal 98.3 91.6 92.5 92.5 95.0 monitoring health with a particular focus to predicting in- creased risk of falls for residents at 16 assisted living houses kasteren 92.6 94.2 81.0 93.4 80.6 being built in Scotland. Simple ambient sensors are employed to monitor activities of daily living. We propose a two stage tapia1 50.8 54.2 56.3 54.2 61.0 approach in which activities are first classified based on low level sensor data inputs. Daily/weekly activity profiles are tapia2 28.3 34.6 35.1 47.1 47.1 then assembled for each resident and compared to their own past data and known risky profiles. Key contributions of the work are: outlining a novel so- 6.2 Results and Discussion lution for identifying the risk of falls for Smart Home resi- The performance of BayesNet, k-NN, SVM, J48 and CRFs dents; and evaluating window-based representations for ac- are compared. The results can be seen in Table 2 with the tivity recognition from the low-level, data inputs delivered highest accuracy achieved on each dataset in bold. from sensors. Acknowledgements [Tapia et al., 2004] Emmanuel Munguia Tapia, Stephen S. This work was part funded by The Scottish Funding Coun- Intille, and Kent Larson. Activity recognition in the home cil via The Data Lab innovation centre. Thanks also to Matt using simple and ubiquitous sensors. Internation Confer- Stevenson at Carbon Dynamic and Angus Watson at NHS ence on Pervasive Computing, pages 158–175, 2004. Highland, Inverness for their support of the FITsense project. [Van Kasteren et al., 2008] Tim Van Kasteren, Athanasios Noulas, Gwenn Englebienne, and Ben Kröse. Accurate References activity recognition in a home setting. In Proceedings of [Cook and Schmitter-Edgecombe, 2009] Diane J Cook and the 10th international conference on Ubiquitous comput- Maureen Schmitter-Edgecombe. Assessing the quality of ing, pages 1–9. ACM, 2008. activities in a smart environment. Methods of information [Vestergaard et al., 2009] Sonja Vestergaard, Kushang V. Pa- in medicine, 48(5):480, 2009. tel, Stefania Bandinelli, Luigi Ferrucci, and Jack M. Gu- [Cook et al., 2015] Diane J Cook, Maureen Schmitter- ralnik. Characteristics of 400-meter walk test performance edgecombe, Diane J Cook, Maureen Schmitter- and subsequent mortality in older adults. Rejuvenation Re- edgecombe, and Prafulla Dawadi. Analyzing Activity search, 12(3):177–184, 2009. Behavior and Movement in a Naturalistic Environment Using Smart Home Techniques. 19(August):1882–1892, 2015. [Hall et al., 2009] Mark Hall, Eibe Frank, Geoffrey Holmes, Bernhard Pfahringer, Peter Reutemann, and Ian H. Witten. The WEKA data mining software: an update. SIGKDD Explorations, 11(1):10–18, 2009. [Jiang et al., 2011] Susu Jiang, Bofeng Zhang, and Daming Wei. The elderly fall risk assessment and prediction based on gait analysis. 2011 IEEE 11th International Conference on Computer and Information Technology, pages 176– 180, 2011. [Katz et al., 1963] Sidney Katz, Amasa B Ford, Roland W Moskowitz, Beverly a Jackson, and Marjorie W Jaffe. Studies of illness in the aged. the index of ADL: A stan- dardized measure of biological and psychosocial function. Journal of the American Medical Association, 185:914– 919, 1963. [Montero-Odasso et al., 2005] Manuel Montero-Odasso, Marcelo Schapira, Enrique R. Soriano, Miguel Varela, Roberto Kaplan, Luis A. Camera, and L. Marcelo May- orga. Gait velocity as a single predictor of adverse events in healthy seniors aged 75 years and older. Journal of Gerontology, 60A:1304–1309, 2005. [Population Division and Affairs, 2015] United Nations De- partment of Economic Population Division and Social Af- fairs. World population ageing, 2015. [Rana et al., 2016] Rajib Rana, Daniel Austin, Peter G. Ja- cobs, Mohanraj Karunanithi, and Jeffrey Kaye. Gait ve- locity estimation using time-interleaved between consecu- tive passive IR sensor activations. IEEE Sensors Journal, 16(16):6351–6358, 2016. [Stone and Skubic, 2013] Erik E. Stone and Marjorie Sku- bic. Unobtrusive, continuous, in-home gait measure- ment using the Microsoft Kinect. IEEE Transactions on Biomedical Engineering, 60(10):2925–2932, 2013. [Suryadevara and Mukhopadhyay, 2012] Nagender Kumar Suryadevara and Subhas Chandra Mukhopadhyay. Wire- less sensor network based home monitoring system for wellness determination of elderly. IEEE Sensors Journal, 12(6):1965–1972, 2012.