Radiation dermatitis (RD) and superficial soft tissue fibrosis are frequent toxicities in patients receiving hypofractionated radiotherapy (HFRT) for breast cancer. Early identification of patients at high risk for acute RD could enable targeted supportive care and optimization of radiation planning. This study developed and evaluated multimodal machine learning models that integrate clinical variables, radiomic and dosiomic features, and deep learning (DS)–derived image features to predict acute RD following HFRT delivered with intensity-modulated radiation therapy (IMRT). Fibrosis analysis was treated as exploratory due to a limited number of events.
This single-center, pre-registered cohort study used retrospective data from early-stage breast cancer patients treated with IMRT-HFRT between 2017 and 2022. The final analysis included 160 patients. Outcomes assessed were multiclass RD (Grades 0–3) and a binary dermatitis endpoint (Grades 0–1 vs 2–3). Soft tissue fibrosis was recorded but had only 5 Grade 1 events and therefore was considered exploratory.
Model development compared two overarching frameworks: a conventional approach using clinical, radiomic, and dosiomic features; and an expanded framework that added DS features, contrast-limited adaptive histogram equalization (CLAHE) image enhancement, and SMOTE class rebalancing. Model types evaluated included Random Forest, XGBoost, LightGBM, and a combined DS-LightGBM pipeline.
Radiomic features were extracted from planning CT images using the planning target volume (PTV) as the region of interest. Dosiomic features were derived from three-dimensional dose distributions with 5 Gy dose-binning applied to the whole breast and tumor bed. Deep learning (DS) features were obtained by applying DenseNet-121 to resampled CT and dose images; these DS features were incorporated into the expanded modeling framework.
To address class imbalance and image contrast, the expanded pipeline applied CLAHE for image enhancement and used SMOTE for synthetic minority oversampling during training. SHAP (SHapley Additive exPlanations)–based analysis was used for feature selection to reduce dimensionality and identify the most informative predictors prior to model training.
All models were trained using 5-fold cross-validation within the training set. Performance was then evaluated on independent test data. The modeling comparison included both ensemble tree models (Random Forest, XGBoost, LightGBM) and the hybrid DS-LightGBM approach that combined DS-derived image features with LightGBM for classification. SHAP-guided feature selection aimed to both improve classifier performance and produce more interpretable models by prioritizing features with higher contribution to predictions.
The analysis included 160 patients. RD grade counts were: Grade 0 = 15, Grade 1 = 81, Grade 2 = 48, Grade 3 = 16. For binary classification (Grades 0–1 vs 2–3), 40% of patients had Grade ≥2 dermatitis. Only 5 patients exhibited Grade 1 fibrosis, limiting inference for fibrosis endpoints.
Across evaluated endpoints the DS-LightGBM model achieved the best internal performance. For multiclass RD prediction the DS-LightGBM test accuracy was 0.87 with an ROC-AUC of 0.95. For binary dermatitis classification the model attained accuracy of 0.89 and ROC-AUC values in the 0.88–0.90 range. Fibrosis prediction, interpreted cautiously because of rare events, reported accuracy 0.97 with ROC-AUC 0.85–0.89.
SHAP analysis identified that high-dose dosiomic subvolumes and DS-derived features were predominant predictors in the multimodal models. This suggests that localized dose distribution characteristics and complex imaging features captured by DenseNet-121 contributed substantially to model discrimination for RD. Clinical and conventional radiomic variables were included in models but were less prominent than the dosiomic and DS features according to SHAP rankings.
Key limitations reported by the authors include the single-center retrospective design and the lack of external validation; performance metrics reflect internal testing only. Fibrosis modeling is limited by the very small number of events (5 cases), so those findings are exploratory and not suitable to guide clinical decisions. The authors explicitly note that external validation on independent cohorts is required before clinical deployment. Additional work could also examine generalizability across institutions, different HFRT regimens, and prospective evaluation of the model’s impact on treatment planning and supportive-care interventions.
In this cohort of 160 early-stage breast cancer patients treated with IMRT-HFRT, multimodal integration of clinical variables, radiomic, dosiomic, and DS-derived features in a DS-LightGBM framework produced promising internal performance for predicting acute radiation dermatitis. SHAP-derived importance highlighted the role of high-dose dosiomic subvolumes and DS features as leading predictors. The models may support personalized planning and early toxicity mitigation to improve patient quality of life, but further external validation is warranted prior to clinical use. Fibrosis findings should be interpreted as exploratory because of the limited number of events.