Sensor-based remote health monitoring can be used for prompt detection
of adverse health in people living with dementia in the home. Current
anomaly detection approaches are challenged by noisy data, unreliable
event annotation and wide variability in home settings. We hypothesized
that a downturn in health would present as a discernible shift in
spatiotemporal patterns, which can be identified by monitoring the
temporal evolution of the household movement graph. We present a
lightweight contrastive learning approach to detect adverse events using
home activity changes, along with household-personalized alerting
thresholds based on the clinician-set target alert rate. Our
self-supervised Graph Barlow Twins model with aggregation-based node
feature masking is used to generate daily activity representations in
participant households taken from a real-world dataset collected by the
UK Dementia Research Institute. Daily graph differences represent the
anomaly score, which are compared to the householdpersonalized
threshold, and alerts raised to the clinical monitoring team. Attention
weights from the graph encoder support explainability and help focus on
the source of anomaly. Our model outperforms state-of-the-art temporal
graph algorithms in detecting agitation and fall events for three
distinct patient cohorts, with 81% average recall and 88%
generalizability at a target alert rate of 7%. To the best of our
knowledge, we offer the first use case of negative sample-free graph
contrastive learning for anomaly detection in a healthcare setting that
is domain-agnostic and can be applied to wider IoT settings.