Paper Type
Short
Paper Number
PACIS2025-1632
Description
ICU extubation remains challenging due to traditional predictors’ limitations in capturing patients’ dynamic conditions. We present the Extubation Risk Prediction (ERP) algorithm, a deep learning model combining LSTM, CNN, and TCN to forecast extubation readiness and pneumonia risk from high-frequency ventilator data. Trained on objective outcomes, ERP achieved AUCs of 0.946 and 0.992, surpassing conventional metrics. In a simulation-based randomized study with 60 ICU physicians, ERP significantly improved decision accuracy (+46%) and reduced risk-averse behavior. Its explainable AI features further modulated clinical risk-taking by enhancing confidence in straightforward cases and promoting caution in complex ones.
Recommended Citation
Hu, Zhichao; Yao, Jiayu; Goh, Kim Huat; Yeow, Adrian; and D'Souza, Jared, "A Field Experiment Study on Artificial Intelligence-Assisted Extubation Decision" (2025). PACIS 2025 Proceedings. 19.
https://aisel.aisnet.org/pacis2025/ishealthcare/ishealthcare/19
A Field Experiment Study on Artificial Intelligence-Assisted Extubation Decision
ICU extubation remains challenging due to traditional predictors’ limitations in capturing patients’ dynamic conditions. We present the Extubation Risk Prediction (ERP) algorithm, a deep learning model combining LSTM, CNN, and TCN to forecast extubation readiness and pneumonia risk from high-frequency ventilator data. Trained on objective outcomes, ERP achieved AUCs of 0.946 and 0.992, surpassing conventional metrics. In a simulation-based randomized study with 60 ICU physicians, ERP significantly improved decision accuracy (+46%) and reduced risk-averse behavior. Its explainable AI features further modulated clinical risk-taking by enhancing confidence in straightforward cases and promoting caution in complex ones.
Comments
Healthcare