Machine Learning Predicts Radiation Skin Reactions

MACHINE learning models demonstrated moderate performance in predicting acute and late radiation-induced skin reactions in patients with locally advanced rectal cancer receiving neoadjuvant chemoradiotherapy. The retrospective analysis included 358 patients and integrated clinical, anatomical, haematological and dosimetric data.

Six machine learning algorithms were trained using a stratified 70/30 split, with five-fold cross-validation applied to the training set. An independent test set was retained for the final assessment. Multicollinearity was evaluated using Spearman correlation analysis and variance inflation factor (VIF), while univariate analysis was used only to describe baseline correlations and was not used for feature selection.

Model performance was assessed using area under the curve (AUC), F1 score, accuracy, sensitivity, specificity, balanced accuracy and calibration indicators.

XGBoost and LightGBM Show Different Performance

For acute radiation-induced skin reactions, XGBoost produced the highest performance in the test cohort, with an AUC of 0.617 (95% confidence interval [CI]: 0.510–0.719). For late radiation-induced skin reactions, LightGBM achieved the highest discriminative performance, with an AUC of 0.707 (95% CI: 0.563–0.845).

Notably, several clinical and dosimetric variables were associated with radiation-induced skin toxicity. These included tumour distance from the anal verge, age, prealbumin concentration and V5000cGy dose parameters.

Although the collinearity analysis identified substantial collinearity, particularly for the late reaction endpoint, tree-based models maintained stable predictive performance across validation folds. Overall, however, AUC values remained below 0.75 across all models, indicating moderate discriminative ability.

Further Validation is Needed

The findings suggested that machine learning models incorporating multimodal clinical and dosimetric features may have potential for risk stratification among patients with locally advanced rectal cancer undergoing neoadjuvant chemoradiotherapy.

However, the reported performance did not establish these models for clinical use. The researchers indicated that further external validation in larger prospective cohorts would be required before clinical application.

Reference

Meng J et al. Machine-learning-based prediction of acute and late-onset skin reactions after neoadjuvant chemoradiotherapy for rectal cancer: a retrospective cohort study. Sci Rep. 2026;DOI:10.1038/s41598-026-72684-8.