Development of a machine learning-based predictive model for intraoperative hypothermia risk during radical surgery for oral cancer.

Abstract

Objective: To develop and validate a machine learning (ML)-based model for predicting the risk of intraoperative hypothermia in patients undergoing radical oral cancer surgery and to identify key contributing risk factors for clinical reference.

Methods: This retrospective study included 402 patients who underwent radical oral cancer resection, divided into training (n = 281) and validation (n = 121) cohorts. Demographic data, physiologic indicators, and intraoperative variables were collected. Predictive models were constructed using Least Absolute Shrinkage and Selection Operator (LASSO) regression, eXtreme Gradient Boosting (XGBoost), and Random Forest (RF) algorithms. Model performance was evaluated using receiver operating characteristic curves, calibration plots, and Shapley Additive Explanations (SHAP) analysis.

Results: The RF model demonstrated superior performance, achieving an area under the curve (AUC) of 0.821 (95% confidence interval [CI]: 0.783-0.856) in the training cohort and 0.807 (95% CI: 0.742-0.865) in the validation cohort, with 64.6% sensitivity. It outperformed both the XGBoost model (validation AUC = 0.721) and LASSO model (validation AUC = 0.738). SHAP analysis identified surgical duration > 441 minutes (odds ratio [OR] = 2.31), baseline temperature ≤ 36.5°C (OR = 3.12), and intraoperative fluid volume ≥ 4.6 liters (OR = 1.89) as the most important predictors. Calibration curves showed strong agreement between predicted and actual outcomes (mean absolute error = 0.17).

Conclusion: The ML-based RF model provides reliable prediction of intraoperative hypothermia risk in oral cancer surgery. Surgical duration and baseline temperature emerged as key risk factors, offering targets for perioperative risk stratification and intervention.