Uncertainty-Calibrated Self-Supervised Learning Model for Early Anomaly Detection in Wearable Health Monitoring Systems

Abstract

Background: Continuous multimodal physiological data generated by wearable health monitoring devices makes it easier to identify abnormal health issues early on. Although there is still progress in the field, identifying incipient anomalies remains an issue because of the limited availability of labelled data, the vast inter-individual variability, sensor noise and poor confidence estimation.

Methods: This study introduces a self-supervised learning framework with uncertainty calibration for early detection of anomalies in wearable health monitoring systems. The proposed methodology obtains strong physiological representations via self-supervised temporal reconstruction, eliminating the need for anomalous labeled data. An uncertainty-aware inference mechanism based on the Monte - Carlodropout is implemented to infer predictive confidence. Anomaly detection is defined as a statistical difference between normal physiological distributions expressed in a Mahalanobis distance weighted by uncertainty. A threshold optimization strategy is used to achieve tradeoffs between precision and recall. The framework is tested on the publicly available WESAD wearable stress dataset with multi-modality wrist worn physiological inputs.

Results: Experimental outcomes show AUROC = 0.994 and AUPRC = 0.940 values , substantially outperforms the conventional auto-encoder and LSTM reconstruction benchmarks. The model accurately classifies 98.18 per cent of cases, with precision 0.855,recall 0.890, and F1-score 0.872, and also has a low rate of false alarms at 1.13 per cent and calibration error of 0.066. The approach is sensitive to physiological anomalies indicating stress on average 822 temporal frames in advance of the annotated commencement.

Conclusion: These results affirm that self-supervised representation learning with uncertainty-calibrated deviation modeling brings reliable, fast, and credible anomaly identification.

Implication: The proposed approach demonstrates potential for proactive wearables health monitoring and preventive clinical decision support.