Wearable-Based Behavioural Monitoring for Dementia Care: From Dataset Development to Routine-Aware Anomaly Detection

INTRODUCTION: Disruption of daily routines, such as irregular sleep, skipped meals or reduced mobility, is often among the earliest observable markers of functional decline in dementia, yet continuous, unobtrusive monitoring of this disruption remains difficult to build and validate at scale, mainly due to the scarcity of realistic wearable datasets and of longitudinal behavioural data. OBJECTIVES: This paper reports two complementary contributions developed within the same wearable-monitoring project: the incremental construction of a real-world multimodal wearable dataset for human activity recognition, and a routine-aware behavioural monitoring framework that uses that dataset to train and evaluate a sequential anomaly detection model. METHODS: A dual-device wearable setup, combining a wrist-worn inertial prototype with an external spot-check physiological monitor, was used to collect labelled activities of daily living from twelve adult participants under supervised conditions. The resulting activity database seeds a knowledge-informed synthetic daily-routine generator and trains a Bidirectional Gated Recurrent Unit (BiGRU) sequential model for context-dependent anomaly detection, evaluated on synthetic routines with injected behavioural deviations. RESULTS: Activity-recognition baselines trained on the dataset reach an accuracy of 0.73 (LightGBM), while the anomaly detection model attains an AUROC of 0.999 and a PR-AUC of 0.989 on synthetic routines with injected behavioural deviations. CONCLUSION: These results support the feasibility of the acquisition pipeline and of the routine-modelling approach as a foundation for future validation with longitudinal, real-world dementia cohorts.