Construction and Temporal Validation of a Machine Learning Model for Predicting Clinically Significant (Grade ≥2) Acute Radiation Enteritis After Radiotherapy for Cervical Cancer

Purpose: To develop and temporally validate prediction models for clinically significant acute radiation enteritis, defined as Common Terminology Criteria for Adverse Events (CTCAE) grade ≥2 ARE and hereafter abbreviated SARE to remain consistent with the published literature on this endpoint, in patients with cervical cancer undergoing pelvic radiotherapy. Methods: A total of 363 patients diagnosed with cervical cancer were retrospectively included and received concurrent chemoradiation at the First Affiliated Hospital of Anhui Medical University. The training set (n = 203, 70%) and the internal test set (n = 87, 30%) were recruited from January 2020 to January 2024; an independent temporal validation cohort (n = 73; 72 with complete outcome and prediction data) was recruited from January 2024 to December 2025. Feature selection (Boruta plus random-forest recursive feature elimination) and hyperparameter tuning were performed exclusively within the training set. Six models were constructed—logistic regression (LR), a multilayer perceptron (MLP), a support vector machine (SVM), random forest (RF), XGBoost and LightGBM—and SHapley Additive exPlanations (SHAP) was used to interpret the final model. Results: Twelve optimal predictors were selected. The LR model demonstrated the most balanced performance, achieving AUCs of 0.855 (95% CI 0.792–0.910) in the training set, 0.830 (95% CI 0.721–0.918) in the internal test set, and 0.753 (95% CI 0.621–0.864) in the temporal external validation cohort. RF and LightGBM fitted the training data perfectly (AUC 1.000) but fell to 0.762 and 0.757 in validation, and XGBoost, the SVM and the MLP were lower (0.661, 0.599 and 0.616). In temporal validation LR had the lowest Brier score (0.176) but a sensitivity of only 0.40 (95% CI 0.176–0.619) at the 0.5 threshold. Decision curve analysis showed positive net benefit for LR across threshold probabilities of 0.05–0.80. SHAP ranked Rectal_Dmean, Bowel_V30 and SII as the leading contributors. Conclusions: Dosimetric parameters combined with composite inflammatory–nutritional indices show promising discrimination for grade ≥2 ARE. The LR model provides a simple, interpretable tool for pre-radiotherapy risk stratification, but prospective, multicenter validation is required before routine clinical implementation.