Unsupervised Forecasting and Anomaly Detection of ADLs in single-resident elderly smart homes

As the aging population increases, predictive health applications for the elderly can provide opportunities for more independent living, increase cost efficiency and improve the quality of health services for senior citizens. Human activity recognition within IoT-based smart homes can enable detection of early health risks related to mild cognitive impairment by providing proactive measurements and interventions to both the elderly and supporting healthcare givers. In this paper, we develop and evaluate a method to forecast activities of daily living (ADL) and detect anomalous behaviour using motion sensor data from smart homes. We build a predictive Multivariate long short term memory (LSTM) model for forecasting activities and evaluate it using data from six real-world smart homes. Further, we use Mahalanobis distance to identify anomalies in user behaviors based on predictions and actual values. In all of the datasets used for forecasting both duration of stay and level of activities using duration of activeness/stillness features, the max NMAE error was about 6%, the values show that the performance of LSTM for predicting the direct next activity versus the seven coming activities are close.