Implementation of a patient risk behavior monitoring system using real-time sensor data and AI

Incidents involving high-risk patient behaviors remain a critical challenge in medical institutions. While advancements in sensor technology and artificial intelligence (AI) have shown promise, existing monitoring systems often rely on single-modality sensors, facing limitations in accuracy, privacy, and coverage. This study proposes a multimodal patient risk behavior monitoring system that integrates data from CCTV, mmWave radar, wearable devices, and location sensors. The system employs specialized deep learning models, including VideoMAE for video-based anomaly detection and an LSTM-AutoEncoder for analyzing physiological signals, to identify high-risk behaviors such as self-harm, falls, and aggression in real time. By fusing data from heterogeneous sensors, the system enhances detection robustness while mitigating the privacy concerns associated with constant visual surveillance. When a risk is detected, it immediately alerts medical staff, enabling prompt intervention. The primary contributions of this work are the development of an integrated, multimodal architecture that improves upon single-sensor systems and the implementation of a privacy-conscious framework for continuous patient monitoring, thereby enhancing safety for both patients and healthcare providers.