An Approach for Behavioral Drift Detection in a Smart Home

With the global population growing older and more vulnerable, healthcare systems are considering new approaches to maintain people autonomous in their own homes. Recent advances in pervasive computing technologies have opened up new opportunities to unobtrusively monitor human behavior at home. Sensors data related to the activities of daily living (ADLs) performed by the inhabitant can be collected and labeled manually or by using activity recognition algorithms. The purpose of this work is to propose an approach for detecting changes (drift) in the inhabitant behavior in order to detect potential changes in the health-state of the inhabitant. Based on unsupervised clustering, our approach use activity starting time and duration as key features to detect changes between time periods. Variations from one behavior to another one can be identified for subsequent review or intervention by a healthcare professional. The relevance and the nature of the change are asserted using the clustering validation metric called Silhouette. Case study experiments on real life and simulated datasets suggest that some user's behavior can go through smooth or abrupt changes and that these changes can be highlighted using our approach.