Self-Extubation Early Warning via Autoencoder Anomaly Detection and a Novel Simulator for Quantitative Evaluation

This study proposes a self-extubation early warning system based on autoencoder anomaly detection with capacitive touch sensor arrays and introduces a novel simulator for quantitative evaluation. The detection system uses a smart dressing material embedded with 5 × 5 capacitive sensor array to continuously monitor tube displacement and contact conditions without compromising patient privacy. An unsupervised spatio-temporal autoencoder is trained on normal interaction patterns to detect deviations associated with self-extubation risk. Additionally, a heatmap-based visualization is employed to provide intuitive spatial feedback. To enable reproducible performance assessment, we developed a dedicated simulator that replicates realistic pulling scenarios under controlled conditions. The simulator incorporates an acceleration sensor to quantify pulling dynamics and supports repeatable evaluations of system behavior. Experiments demonstrated that the system achieved high detection accuracy, with a recall of 0.972 and an F1 score of 0.882 in simulated early warning scenarios. The combined approach provides a deployable detection system and a standardized evaluation framework, contributing to a proactive and privacy-preserving solution for self-extubation prevention in ICU and elderly care settings.