Interpretable machine learning-based prediction of intraoperative hypothermia in endoscopic resection of gastric submucosal tumors

Intraoperative hypothermia (IOH) is a prevalent and preventable complication associated with general anesthesia. This study aimed to develop and validate interpretable machine learning (ML) models for predicting the risk of IOH in patients undergoing endoscopic resection (ER) of gastric submucosal tumors (SMTs). We retrospectively analyzed data from 814 patients who underwent ER for gastric SMTs across six medical institutions between January 2018 and December 2025. Patients were divided into a training cohort ( n = 419), an internal validation cohort (IVC; n = 179), and an external validation cohort (EVC; n = 216). The Least Absolute Shrinkage and Selection Operator (LASSO) regression was employed for feature selection, and seven ML algorithms—support vector machine (SVM), decision tree (DT), extreme gradient boosting (XGBoost), random forest (RF), light gradient boosting machine (LGBM), k-nearest neighbors (KNN), and logistic regression (LR)—were subsequently developed using the selected predictors to predict IOH risk. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals (CIs), along with sensitivity, specificity, accuracy, and F1 score. Furthermore, Shapley Additive Explanations (SHAP) were applied to enhance model interpretability by quantifying the contribution of each predictor. IOH occurred in 168 patients (20.6%). The RF model demonstrated good predictive performance, achieving an AUC of 0.868 (95% CI: 0.793–0.932) in the IVC and 0.783 (95% CI: 0.698–0.861) in the EVC. It exhibited high specificity (IVC: 0.957; EVC: 0.972) and moderate sensitivity (IVC: 0.579; EVC: 0.250). Accuracy was 0.877 in the IVC and 0.838 in the EVC. SHAP analysis identified body mass index, age, tumor size, and tumor location as the most influential predictors of IOH. We developed an interpretable ML-based predictive model for IOH in gastric SMTs patients undergoing ER. By providing insights into key risk factors, this model may inform individualized intraoperative warming strategies, pending prospective validation.