Remote Monitoring in Dementia Care - Lightweight, Explainable AI Validated for Early Warning of Health Events in the Home

Sensor-based remote health monitoring for people living with dementia (PLwD) enables early detection of adverse health events, reducing hospitalization risk. Identifying anomalies in real-world home activity data poses significant challenges due to noise, imprecise labels, inter-household variability, and the need for clinical explainability. We propose a lightweight, explainable AI pipeline for anomaly detection in home sensor data, aimed at early detection of health events, validated in an ongoing real-world dementia monitoring study. Our model generates noise-resilient daily representations to compute anomaly scores, compared against household-personalized thresholds to trigger alerts. Novel spatiotemporal attention maps uncover the source and timing of anomalies, offering household-specific and cohort-wide insights into atypical behavior patterns. Maximum typicality metrics provide a dynamic and continuous distinction between typical and atypical days, enabling real-time adaptation to incoming patient data. In addition, LLM-powered anomaly summaries support clinical monitoring teams by providing detailed descriptions of sensory observations. On a 65-patient internal validation cohort (18,800 person-days; Aug 2019-Apr 2022), the model achieved 84.64±2.36% sensitivity and 92.16±2.33% generalizability at a 7% maximum alert rate. In a larger 90-patient cohort (40,586 person-days; May 2022–Feb 2024), it achieved 77.04±1.35% sensitivity and 90.67±1.51% generalizability, a strong result given the inherent noise and variability of home sensor data. This AI-powered anomaly detection pipeline demonstrates high clinical utility for early in-home detection of health events in dementia care, and can be easily adapted to diverse remote monitoring settings.