Authors - Mazin Alshamrani Abstract - Wearable physiological monitoring systems are increasingly deployed in health-critical contexts, yet existing approaches address state classification, transition detection, and signal safety as isolated problems. This paper introduces SafeWear, a unified signal quality-aware machine learning framework that jointly addresses (i) real-time and anticipatory detection of physiological state transitions, and (ii) multimodal anomaly detection and out-of-distribution (OOD) safety screening. The framework centres on a modality-level Signal Quality Index (SQI) acting as a front-end reliability gate distinguishing sensor-level failures from genuine physiological irregularities. Causal temporal models are evaluated for transition detection, while reconstruction-based and one-class detectors are benchmarked for anomaly safety under strict Leave-One-Subject-Out Cross-Validation (LOSO-CV) across 37 subjects and five wearable modalities. Preliminary results show causal models achieve median detection latencies below 5.2 s with early-detection rates exceeding 79%, and Deep SVDD with VAE yield the strongest anomaly discrimination (AUROC = 0.539 and 0.538).
Open Zoom