Abstract / Summary
Abstract To develop and validate machine learning models for predicting acute and late radiation-induced skin reactions in patients with locally advanced rectal cancer who received neoadjuvant chemoradiotherapy. This retrospective analysis included 358 patients with locally advanced rectal cancer. Multimodal preprocessing data including clinical, anatomical, hematological and dosimetric data were collected. The univariate analysis was only used to describe the baseline correlation and was not used as the basis for feature selection. After evaluating multicollinearity by Spearman correlation analysis and VIF, the six algorithms were trained using a stratified 70/30 split and five-fold cross-validation of the training set. The independent test set was retained for the final evaluation. The performance metrics included AUC, F1 score, accuracy, sensitivity, specificity, balanced accuracy and calibration indicators. In the test cohort, the best-performing model for acute radiation-induced skin reactions was XGBoost, achieving an AUC of 0.617 (95% CI 0.510–0.719). For late reactions, LightGBM achieved the highest discriminative performance with an AUC of 0.707 (95% CI 0.563–0.845). Key clinical and dosimetric variables associated with skin toxicity included tumor distance from the anal verge, age, prealbumin level, and V5000cGy dose parameters. Although collinearity diagnostics revealed substantial collinearity, particularly for the late-onset endpoint, tree-based models maintained stable predictive performance across validation folds. Overall model performance indicated moderate discriminative ability, with AUC values below 0.75 across all models. Machine learning models integrating multimodal clinical and dosimetric features demonstrated moderate performance in predicting radiation-induced skin toxicity in rectal cancer patients. These findings suggest potential utility for risk stratification; however, further external validation in larger prospective cohorts is required prior to clinical application.