Personalized Real-time Anomaly Detection and Health Feedback for Older Adults

The rapid population aging and the availability of sensors and intelligent objects motivate the development of information technology-based healthcare systems that meet the needs of older adults by supporting them to continue their day-to-day activities. These systems collect information regarding the daily activities of the users and potentially help to detect abnormal behaviors. Anomaly detection can subsequently be combined with real-time, continuous and personalized interventions to help older adults actively enjoy a healthy lifestyle. This thesis introduces a system that takes a novel approach to generate personalized health feedback. The proposed system models user's daily behavior in order to detect anomalous behaviors and strategically generates interventions to encourage behaviors conducive to a healthier lifestyle. To this aim, we propose a real-time solution which models the user daily routine using a task model specification and detects relevant contextual events occurred in their life through a Context Server (a middle-ware software). In addition, by a systematic validation through a system that automatically generates wrong sequences of tasks, we show that our algorithm is able to find behavioral deviations from the expected behavior at different times by considering the extended classification of the possible deviations with good accuracy. Later, the system uses a Mamdani-type fuzzy rule-based component to predict the level of intervention needed for each detected anomaly and a sequential decision-making algorithm, Contextual Multi-armed Bandit, to generate suggestions to minimize anomalous behaviour. We describe the system architecture in detail and we provide example implementations for corresponding health feedback.