As the prevalence of IoT sensor equipment in smart homes continues to rise, long term monitoring for personalised and more representative health tracking has become more accessible. The estimation of physiological health factors such as gait and heart rate can be captured using a range of diverse sensor equipment, while behavioural changes are now being monitored using simple binary sensors through activity classi cation and pro ling. Combining both physiological and behavioural monitoring in xed layout properties has already allowed us to e ectively consider fall risk. However, expanding application of the system to new layouts and conditions requires consideration of di ering retro t home layouts and sensor con gurations. A wider selection of sensors in varying con gurations could potentially allow for the identi cation of other health conditions such as heart disease and stroke.
Scotland and the UK are facing an aging population, as people live longer. 10 million people in the UK are currently over 65 years with a 5.5 million increase projected over the next 20 years. 3 million people are aged over 80 and the number is expected to double by 2030. This puts additional strains on the health and social services with both a smaller proportion of the population of working age available to support the service, and with the older population tending to have more complex medical needs. Furthermore, with modern lifestyles carers from within the family are less available. Generally, more people are tending to live alone and families are tending to live further apart as people are more likely to relocation for work or education either within the UK or internationally. In this changing scenario it is important that we help people with medical or social needs to live independently for longer and so reduce their reliance on more expensive health care solutions.
Smart Homes and accompanying IoT (Internet of Things) devices have become increasingly popular under the premise of home automation and security. Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). This recent push for wider acceptance of ubiquitous sensor devices in smart home environments provides a unique opportunity to perform long term monitoring of health in the home environment. While ongoing health conditions such as dementia and fall risk are typically diagnosed, monitored and addressed in hospital environments, the measurements taken in these tests can potentially be performative, and so not as representative of natural measurements. Additionally, performing tests in hospital environments can be expensive and di cult to access for remote communities or people with reduced mobility. By utilising smart home sensor networks to monitor resident behaviour and physiological expressions, the risk factors typically identi ed only through hospital testing can be monitored in casual home environments.
In prior work, a primarily rule-based system was designed and implemented to perform fall risk determination in a real smart home sensor environment. FitHomes is a social housing project lead by Albyn Housing Society Ltd focused on improving conditions and fostering independence for residents living with mobility issues. 16 purpose-built homes have been constructed at Dalmore near Invergordon in the UK; and an additional development of a further 32 homes is currently planned. FITsense is a data analysis system which uses human activity recognition to produce daily resident pro les in FitHome installations. A rulebased segmentation system is used to break long sequences of sensor activations down into smaller windows, upon which a rule-based activity classi er was used to identify ADLs (Activities of Daily Living). The instance and regularity of ADL expressions is used to produce ratings for an overall pro le of potential fall risk, considering factors such as sedentary time, bathroom usage and overall sleep time. These pro les are then be graphed over a given period of time to ag signi cant changes in relevant factors.
As additional FitHome installations are planned, challenges have arisen in expanding the existing FITsense system. The current rule-based systems require extensive domain knowledge with respect to the sensor con guration and layout in order to produce robust compatible rules, while the underlying work in generating rules for each additional unique layout is very time-consuming. Additionally, attempting to support exible sensor con gurations in variable house layouts increases the complexity and granularity of ADLs which can be captured and so increases the complexity of the rules required.
This project aims to address key issues facing the expansion of the FitHomes network, the FITsense system and its overall capabilities. By transitioning from a rule-based approach to ADL classi cation to a sustainable ML-based approach, the FITsense system needs to scale e ciently as the FitHomes network grows, especially to retro t home environments where unique con gurations and layouts will be commonplace. There may also be opportunities to introduce additional sensors to improve the behavioural modelling and physiological monitoring capabilities of the sensor network.
Long term monitoring in home environments has been a constantly evolving
eld due to the large range of pervasive sensor technologies being used, such as
infrared motion sensors and wearable accelerometers. A relatively simple and
high impact form of health monitoring in home environments is fall detection.
Fall detection is a well-developed area of research, with many iterative solutions
providing consistent strong performance [
The task of ADL classi cation from smart home sensor data can be split into two main areas: sequence windowing, by which long sequences of sensor activations are segmented into shorter sequences representative of an activity, and activity recognition, in which those sequences can then assigned an activity label. While the FITsense system works well as a proof-of-concept in the current FitHomes installation, it relies heavily on a rule-based approach to sequence windowing and activity classi cation. This was originally selected during development as only detailed domain knowledge on the house layout and sensor con guration were available. As there was no reliance on manually labelled sensor data, which would be required by a supervised model, development could progress alongside construction.
The initial FitHomes houses make use of a reference layout which is almost identical across all houses. However, as the FitHomes project expands to new environments including new constructions and existing retro t homes, additional intercompatible rulesets must be produced for each unique environment added to the sensor network. As more rules are added to the system it becomes more di cult to ensure e ective con ict resolution. As the overall supply of FitHomes data becomes more consistent and the work required to maintain the existing rule-based implementation increases, it has become clear that a data-driven MLbased approach to ADL classi cation would be a more sustainable method of supporting the expansion of the FitHomes network.
The challenges in producing a rule-based classi er stem from the requirement for extensive domain knowledge and the expense in producing a ruleset. However in developing an ML-based system, the key challenges are the acquisition of labelled training data and designing the representation used in the model. Few publicly available smart home activity recognition datasets exist, largely due to the cost and human involvement required in their production [5]. In practice, it was found in the FitHomes project that residents were most receptive to paperbased journalling over a 24 hour period. More training examples are required to produce a large enough dataset to train a deep model such an LSTM. An experimental form of manual bootstrapping is currently being performed by collecting the surrounding 13 days of unlabelled data from a resident's home and using their paper journal as a template for their behaviours and movements around the home in the manual labelling process. This compromise between data cost and quality requires further research to identify its e ect on the overall performance and scalability of an ML-based classi er and would be a useful area in which to receieve guidance. 3.2
The majority of sensor equipment used in the FitHomes network is comprised of simple infrared motion sensors producing binary activation data. The FITsense system is designed to extract behavioural information from movement and activity in each home, with sequential behaviour tracking capturing temporal dependencies between relevant activities. However, this coarse data stream limits the complexity of tracked activities. The pending expansion of the FitHomes network provides an opportunity to research additional sensor equipment which may be used to improve the behavioural and physiological monitoring capabilities of the FITsense system.
RF-based sensor technologies are becoming increasingly relevant due to their unintrusive nature [6]. Among the most useful demonstrated capabilities are indoor localisation, and heart and breathing rate estimation. However the highend equipment required to implement such solutions restrict their accessibility for ubiquitous deployment. An alternative hardware solution involves using commodity WiFi devices as a rudimentary Tx/Rx software de ned radio, through observation of CSI (Channel State Information). While commodity devices o er reduced control relative to enterprise-grade RF equipment, CSI-based solutions have shown promise in physiological and behavioural monitoring tasks [7].
Many current approaches to CSI-based physiological monitoring solutions take a signal processing approach rather than using ML methods. This has been shown to be successful with careful and time-consuming selection and preprocessing of relevant subcarrier signals, however it is possible a privileged learner could be trained to perform this task. I am seeking to learn more about data preprocessing for an ML-based solution, how it would factor in and whether it may o er better performance than current state of the art. 3.3
Researching risk factors for fall risk determination through both literature and discussion with healthcare professionals highlighted other conditions which could be diagnosed using similar methods. However establishing thresholds of risk factors which are both representative of medically signi cant metrics and personalised to suit the subject is a challenging task. The current thresholds used are produced through comparison to those found in hospital testing, however this does not factor in personalisation to residents and their speci c demographics.
Case-based reasoning could potentially have an application in the generation and comparison of resident behaviour pro les. I am seeking guidance in how initial risky prototype cases could be generated to populate the initial case base. A similarity metric for related risky behaviours and resident pro les will also require consideration.
Another area in which I am seeking guidance is my approach to the selection of health conditions for which to develop monitoring pro les, and whether it should be informed by the additional sensors added to the network, or vice versa. 4
My research methodology is still developing as the current literature is being reviewed. A review of pertinent sensor data analysis and fall risk determination procedures was published in 2019 [5].
Using FitHomes data, initial research outlining the e ect of highlighting existing temporal dependencies in sensor data sequences for ADL classi cation has been performed. This has shown promise in improving the performance of LSTM classi ers across a range of smart home datasets.
Preliminary data gathering using a rudimentary CSI collection setup has been performed, with several tools being developed to facilitate the preprocessing of generated data for realtime analysis.
Acknowledgements This work was part funded by The Scottish Funding Council via The Data Lab innovation centre. 3. Manuel Montero-Odasso, Marcelo Schapira, Enrique R. Soriano, Miguel Varela, and others. Gait Velocity as a Single Predictor of Adverse Events in Healthy Seniors Aged 75 Years and Older, 2005. 4. N. K. Suryadevara and S. C. Mukhopadhyay. Wireless sensor network based home monitoring system for wellness determination of elderly. IEEE Sensors Journal, 12(6):1965{1972, June 2012. 5. Glenn Forbes, Stewart Massie, and Susan Craw. Fall prediction using behavioural modelling from sensor data in smart homes. Arti cial Intelligence Review, pages 1{21, 2019. 6. Fadel Adib, Hongzi Mao, Zachary Kabelac, Dina Katabi, and Robert C Miller.
Smart homes that monitor breathing and heart rate. In Proceedings of the 33rd annual ACM conference on human factors in computing systems, pages 837{846.
ACM, 2015. 7. Jian Liu, Yan Wang, Yingying Chen, Jie Yang, Xu Chen, and Jerry Cheng. Tracking vital signs during sleep leveraging o -the-shelf wi . In Proceedings of the 16th ACM International Symposium on Mobile Ad Hoc Networking and Computing, pages 267{ 276. ACM, 2015.