Sensor-based remote healthcare monitoring for people with long-term conditions provides valuable insights into daily living patterns with minimal intrusion. Anomaly detection in sensor data is crucial for early identification of adverse health events, enabling timely intervention, preventing avoidable hospitalisations, and reducing the caregiving burden. An ideal anomaly detection model must be fast, personalised, robust to noise, minimally reliant on retraining and tuning, and, critically, explainable to clinicians and carers. This research develops and validates unsupervised Artificial Intelligence methods to detect anomalies and identify potential adverse health events using home sensor data in a lightweight, explainable, personalised yet generalisable manner for people living with dementia.

Existing statistical and deep learning approaches face limitations in smart home environments, which are characterised by low signal-to-noise ratios, irregular activity patterns, and scarce and often unreliable annotations. Statistical methods, based on predefined rules or distributions, struggle with noise and evolving patterns. Many deep learning models lack adaptability, personalisation, explainability, or computational efficiency. This research addresses these challenges by developing novel unsupervised and self-supervised methods for anomaly detection in IoT-based remote healthcare monitoring in a domain-agnostic way.

This work introduces the Multidimensional Contextual Matrix Profile, extending the Matrix Profile to multivariate sensor data for anomaly detection. The Contextual Matrix Profile is then leveraged to develop a self-supervised Graph Neural Network model, reframing anomaly detection as a graph-based problem. To improve noise robustness, Graph Barlow Twins contrastive learning is applied, aligning granular sensor data with a corresponding macroscopic view, and develop household-personalised alert thresholds for real-world deployment. Later, adaptive contrastive learning is proposed by dynamically learning macroscopic feature aggregation during training, outperforming prior models and introducing “spatiotemporal attention maps" for explainability. Finally, the deployment readiness of the dynamic contrastive anomaly detection framework is confirmed via operational validation in 90 patients living with dementia, supporting its integration into a live system for large-scale remote monitoring.