Detecting abnormal behaviours of institutionalized older adults through a hybrid-inference approach

Abstract Residences for geriatric patients are usually understaffed, with each caregiver being in charge of several residents. A caregiver must assess the well being of residents and report to the medical staff if there is something unusual with a resident. Deviations from the routine will trigger an alarm, and automatic tools can help in making timely decisions. In this paper, we explore three visualization metaphors aiming at providing caregivers with an individualized overview of the activities carried out by residents in a given time frame. We postulate that this overview is sufficient to distinguish between normal and abnormal periods of time when visually compared in groups. We also present two automated approaches, data driven and knowledge driven respectively, to detect abnormalities. The visualization and the automated approaches are tested on a naturalistic dataset obtained from a long-term personalized sensing and annotation campaign in a residence for geriatric patients. Data is of two types, obtained from IoT infrastructure and wearables and from manual annotations made by the staff. Both approaches were empirically evaluated and validated in the paper. A side product of this research is a large repository of cleansed data from the sensing and annotation campaign for 45 older adults over a period of 39 months.