Unsupervised Anomaly Detection for Psychiatric Inpatients Using Wearable Sensor Data : A Real-Time Monitoring Framework for Clinical Risk Management

The dynamic and erratic nature of patient behavior in psychiatric wards poses significant monitoring challenges, especially under high nurse‐to‐patient ratios where manual checks are labor‐intensive and error‐ prone. This paper proposes a machine and deep learning-based anomaly detection (AD) system using data from wearable devices to support real-time patient monitoring. Data such as activity levels, sleep patterns, and heart rate variability collected from wearable devices can provide early indicators of physiological deterioration in patients, as changes in these signals often precede noticeable symptoms. Our system leverages unsupervised algorithms, including One Class‐SVM, Isolation Forest, plain Autoencoder, Deep SVDD, LSTM Autoencoder, alongside supervised benchmarks Random Forest and XGBoost to continuously analyze patient data and detect abnormal patterns that may signal clinical risk. To validate effectiveness, we compared model -detected anomalies with clinician- assessed risk scores. Results showed strong alignment, with 75% of detected anomalies falling within the high- risk group marked by clinicians. Our system delivers roughly a six-to-sevenfold improvement in anomaly detection performance over a dummy classifier. Autoencoder achieved the highest AUC (0.74), with a 17% F1- score gain over the baseline (DC: F1 0.11). These findings demonstrate the feasibility of integrating wearable technologies and ML/DL for early risk detection, enabling scalable, non-invasive monitoring. The system offers a real-time, intelligent solution to improve safety, support clinical decisions, and reduce caregiver burden in psychiatric care environments.