A Deep Learning Based System For a Long-term Elderly Behavioral Drift Detection

The detection of behavioral drifts by applications monitoring daily living activities, is very relevant in health care monitoring systems particularly for older persons. The detection of behavioral drifts can be crucial, as it may indicate the early stages of a disease. Our research objectives are focused on building a model to conduct a continuous and long-term analysis of elderly’s behavior in order to detect slow changes. The first originality of this work is to use rich contextual information such as weather conditions, holidays, seasons, etc. to identify and learn routine behavior patterns. These patterns are updated throughout the aging process of the elderly person by proposing a novel dynamic behavior clustering method through an iterative process using a moving-window mechanism. The routine behavior patterns are then used for comparing each observed day’s behavior to the associated routine behavior pattern in preceding period to define a similarity score. The second originality of this work is to exploit machine learning techniques on previously observed Activities of Daily Living (ADL) to forecast when the future ones should occur. We use this forecast to measure the actual gap with the future ADLs, as soon as they occur, and detect behavioral drifts of the monitored elderly people and thus specify its type (degradation or recovery). Experimental results are proposed to show the superiority of the proposed method compared to the existing state-of-the-art methods.