Towards abnormal behavior detection of elderly people using big data

Nowadays, smart living technologies are increasingly used to support older adults so that they can live longer independently with minimal support of caregivers. In this regard, there is a demand for technological solutions able to avoid the caregivers' continuous, daily check of the care recipient. In the age of big data, sensor data collected by smart-living environments are constantly increasing in the dimensions of volume, velocity and variety, enabling continuous monitoring of the elderly with the aim to notify the caregivers of gradual behavioral changes and/or detectable anomalies (e.g., illnesses, wanderings, etc.). The aim of this study is to compare the main state-of-the-art approaches for abnormal behavior detection based on change prediction, suitable to deal with big data. Some of the main challenges deal with the lack of "real" data for model training, and the lack of regularity in the everyday life of the care recipient. At this purpose, specific synthetic data are generated, including activities of daily living, home locations in which such activities take place, as well as physiological parameters. All techniques are evaluated in terms of abnormality-detection performance and lead-time of prediction, using the generated datasets with various kinds of perturbation. The achieved results show that unsupervised deep-learning techniques outperform traditional supervised/semi-supervised ones, with detection accuracy greater than 96% and prediction lead-time of about 14 days in advance.