Abstract
INTRODUCTION: We propose a three-stage liberation decision framework (Stage-1 readiness, Stage-2 SBT success, Stage-3 extubation). While prior tools emphasize earlier stages, Stage-3-deciding whether to remove the tube after SBT-remains under-modeled. This study develops an AI model to predict successful extubation (no reintubation or non-invasive ventilation within 48 h) using routinely collected electronic medical record data, eliminating the need for additional manual bedside measurements. METHODS: Single-center retrospective analysis including 5,202 adults who underwent elective extubation after SBT success. Seven algorithms (Random Forest, LightGBM, XGBoost, Logistic Regression, multilayer perceptron, Voting, Stacking) were trained and evaluated by accuracy, sensitivity, specificity, PPV, NPV, and AUC; interpretability used SHAP; traditional indices (RSBI, etc.) served as comparators. We also implemented a working web-based prototype that verifies the model's usability and real-world feasibility, providing a foundation for future prospective clinical evaluation. RESULTS: LightGBM performed best (accuracy 0.797, sensitivity 0.800, specificity 0.763, PPV 0.977, NPV 0.231, AUC 0.861). XGBoost and Voting showed AUC 0.850 with slightly lower accuracies (0.783, 0.771); Stacking AUC 0.829; Random Forest AUC 0.818; MLP and Logistic Regression AUC 0.785 each. SHAP analysis identified SpO₂/FiO₂, department, bilateral lower-limb muscle strength, and dynamic compliance (Cdyn) as most influential predictors of extubation success. DISCUSSION: Within a three-stage liberation framework, a Stage-3 extubation-focused AI model-particularly LightGBM-outperformed traditional indices and offers explainable, EMR-based predictors to support timely tube removal. A web-based prototype has been developed for future prospective validation.