A study on predicting radiation pneumonitis in breast cancer patients undergoing radiotherapy using an interpretable machine learning model based on multi-parameter fusion

Purpose To develop an interpretable machine learning (ML) framework integrating CT radiomics from dose-specific lung regions with dosimetric parameters to accurately predict radiation pneumonitis (RP) in breast cancer patients undergoing radiotherapy. Methods This dual-institute retrospective study included 94 breast cancer patients. Four dose-volume-specific regions of interest (DVROI5, DVROI20, DVROI30, DVROI5–20) were defined on planning CT. A total of 107 radiomic features per DVROI were extracted alongside clinical and dosimetric variables. After feature selection (correlation filtering + Boruta), 11 ML models were trained (70% data, 5-fold cross-validation) and tested on 30% of data. SHAP analysis provided model interpretability. Results RP occurred in 36 patients (38.3%). Boruta selected only radiomic and dosimetric features (2–3 each per DVROI), excluding clinical covariates. XGBoost performed best for DVROI5 (test AUC=0.876), DVROI20 (AUC=0.750), and DVROI5–20 (AUC=0.821), while GBM excelled for DVROI30 (AUC=0.818). The optimal XGBoost model (DVROI5) achieved accuracy 0.852, precision 0.8, recall 0.8, and F1-score 0.8. SHAP identified V5, GLCM cluster prominence, surface-volume ratio, MLD, and V20 as top predictors. Conclusion An explainable XGBoost model combining DVROI5 radiomic features with dosimetric parameters predicted RP with high accuracy (AUC=0.876), outperforming traditional dosimetric thresholds. SHAP-enhanced interpretability supports clinical transparency and personalized risk mitigation. Multi-center prospective validation is needed before widespread adoption.