Development and validation of an interpretable machine learning model for acute radiation dermatitis in breast cancer

TL;DR

This study develops an interpretable machine learning model to predict acute radiation dermatitis in breast cancer patients, aiming to improve treatment planning and reduce severe side effects.

Contribution

The novel contribution is an interpretable machine learning model for predicting acute radiation dermatitis with key predictors identified via SHAP analysis.

Findings

A random forest model achieved an AUC of 0.84 in training and 0.748 in testing for predicting acute radiation dermatitis.

SHAP analysis identified CTVsc, CTVim, TNM stage II, and diabetic status as key predictors of radiation dermatitis.

The model showed better net benefits than 'treat-all' or 'treat-none' strategies at treatment thresholds of 25%–75%.

Abstract

Radiation dermatitis (RD), a common adverse reaction in breast cancer radiotherapy, impairs quality of life and increases healthcare burdens. Developing an effective risk prediction model is crucial for early high-risk patient identification and preventive interventions.

This study enrolled 691 breast cancer patients undergoing postoperative radiotherapy at our center from February 1 to December 19, 2024. RD severity and correlates were monitored during and 2 weeks after radiotherapy. The dataset was divided into training (n=552) and test (n=139) cohorts. Fourteen machine learning algorithms were evaluated via 10-fold cross-validation, with model selection based on Area Under the Curve (AUC) and other metrics. Model reliability was verified using an internal hold-out test set, and SHAP analysis ensured interpretability.

Among 691 patients,52.68% (n=364) developed grade ≥2 acute RD. The random forest model performed best, achieving an AUC of 0.84 (95% CI: 0.807–0.873) in training and 0.748 (0.665–0.831) in testing, with training/testing sensitivity/specificity of 0.811/0.747 and 0.877/0.576, respectively. Calibration curves confirmed prediction-observation consistency. Decision curve analysis indicated 0.2–0.4 higher net benefits than “treat-all” or “treat-none” strategies at 25%–75% treatment thresholds. Shapley Additive exPlanations (SHAP) analysis identified Clinical Target Volume-Supraclavicular (CTVsc), Clinical Target Volume-Internal Mammary (CTVim), TNM stage II, and diabetic status as key predictors.

This explainable machine learning model demonstrates robust discriminative power and clinical utility. Interpretability analysis revealed feature nonlinearities, providing a theoretical basis for personalized radiotherapy planning to reduce severe RD risk.

Figures