Machine learning with SHAP interpretation for modeling delayed extubation decisions in Atlantoaxial dislocation surgery

Atlantoaxial dislocation (AAD) surgery involves complex approaches that create a clinical dilemma: immediate extubation risks catastrophic airway obstruction, while delayed extubation increases morbidity. Data-driven frameworks to guide this clinical decision are currently lacking. This study aimed to develop and validate a machine learning framework to characterize the decision-making process for delayed extubation in this specific patient population. A single-center retrospective cohort of 1,352 adults undergoing elective AAD surgery between 2020 and 2025 was analyzed. The primary outcome was postoperative delayed extubation, defined as retention of the endotracheal tube upon admission to the intensive care unit or ward. Feature selection was performed using Least Absolute Shrinkage and Selection Operator regression within a 10-fold nested cross-validation framework. Five algorithms—Logistic Regression, Random Forest, Gradient Boosting, Support Vector Machine, and Extreme Gradient Boosting—were trained, calibrated, and compared. Model performance was evaluated using discrimination (area under the receiver operating characteristic curve [AUC]), calibration (calibration slope/intercept, Brier score), and clinical utility via decision curve analysis (DCA). Shapley Additive exPlanations (SHAP) values were used to interpret feature importance. The overall incidence of delayed extubation was 5.7% (77/1,352). Among the models evaluated, Logistic Regression achieved robust performance with an AUC of 0.816 (95% CI: 0.738–0.874) and the highest F1 score (0.386, 95% CI: 0.261–0.615). Calibration was adequate (slope: 0.953, intercept: −0.152; Hosmer–Lemeshow test, p = 0.435). DCA confirmed positive net benefit across threshold probabilities from 0.01 to 0.60. SHAP analysis identified the combined transoral and posterior approach, American Society of Anesthesiologists Physical Status classification, and difficult intubation as the dominant drivers of the decision. The proposed machine learning framework effectively characterizes institutional decision-making patterns for delayed extubation following AAD surgery. The model provides a data-driven basis for intraoperative extubation criteria, thereby offering a transparent framework to harmonize clinical practices and minimize unwarranted variability.