Enhancing skin toxicity predictions in breast cancer radiotherapy through integrated CT radiomics, dosiomics, and machine learning models

Abstract

Objective

This study aimed to develop and test a multidimensional machine learning framework that combines radiomic, dosiomic, and dosimetric features to predict severe skin toxicity in breast cancer patients undergoing radiotherapy. The goal was to improve prediction accuracy and support personalized treatment planning.

Materials and methods

The study analyzed a retrospective cohort of 789 breast cancer patients. Pre-treatment computed tomography images were used to extract 215 radiomic features, while 78 dosiomic features were derived from radiation dose distributions. Additionally, 22 dosimetric features, such as dose-volume histograms and mean dose, were included. Feature selection was performed using three methods: Recursive Feature Elimination (RFE), Least Absolute Shrinkage and Selection Operator, and Analysis of Variance. Seven classifiers were trained and validated through 5-fold cross-validation. These classifiers included k-Nearest Neighbors, Random Forest, Support Vector Machines, Gradient Boosting Machine, Extreme Gradient Boosting, Voting, and Stacking. Model performance was evaluated using accuracy, sensitivity, and area under the receiver operating characteristic curve (AUC-ROC). SHapley Additive exPlanations (SHAP) analysis was applied to interpret the contributions of individual features to the model predictions.

Results

The best predictive performance was achieved by combining radiomic, dosiomic, and dosimetric features. The Stacking Classifier with RFE showed the highest metrics: 96% accuracy, 95% sensitivity, and 97% AUC-ROC. Among the individual feature types, radiomic features performed better than dosiomic and dosimetric features, achieving an AUC-ROC of 93% with RFE and the Voting Classifier. Dosiomic features were more predictive than dosimetric features alone, with an AUC-ROC of 94% using the Stacking Classifier. SHAP analysis highlighted that texture and spatial features were key predictors of skin toxicity.

Conclusions

Integrating radiomic, dosiomic, and dosimetric features significantly enhances the prediction of radiation-induced skin toxicity. These findings highlight the potential of machine learning to improve personalized treatment planning. Further validation in larger, multi-center studies is recommended.