An AI-driven alert system for preventing unplanned extubation via arm movement monitoring in the ICU
Chen Chen, Qi Qian, Yun Yu, Yunyan Su, Yan Wang, Zheyun Wang
Abstract Unplanned extubation (UEX) in the intensive care unit (ICU) is a serious adverse event. Current prevention strategies relying on staff vigilance have limitations. Artificial intelligence (AI)-based computer vision offers a new approach for real-time, non-contact monitoring. Employing design science, the system was built using an open-source library (MediaPipe). Core modules included human tracking, hand detection, dynamic risk-zone management, and hierarchical alerting. Functional testing simulated ICU patient postures and movements under various occlusion scenarios. Metrics included tracking accuracy, hand detection coverage, risk judgment accuracy, false alarm rate, and real-time performance. Simulation tests showed a target localization accuracy of 92.5% and dynamic risk-zone adaptation accuracy of 90.0%. Hand recognition coverage was 95.8%, and overall two-level risk judgment accuracy was 93.3%. The false alarm rate was optimized to 2.8%. The average system latency was approximately 35 ms, enabling real-time processing. An AI-driven prototype for UEX prevention was developed and validated. Tests confirmed its technical feasibility in patient tracking, hand movement recognition, and tiered alerting. This system may serve as a potential tool to shift prevention from passive restraint to active alerting, forming a basis for future clinical studies.