Dosimetric parameter-based machine learning to predict radiation pneumonitis following breast radiotherapy

Objective.Radiation pneumonitis (RP) is an important toxicity following breast radiotherapy. Although modern treatment techniques limit lung exposure, some patients still develop symptomatic RP, and predicting who is at risk remains difficult. This study aimed to develop and validate machine learning (ML) models based on ipsilateral-lung dosimetric parameters to predict clinically significant RP after breast radiotherapy.Approach.We retrospectively studied 184 women with breast cancer treated with three-dimensional conformal radiotherapy using either conventional or hypofractionated fractionation schemes, with or without regional nodal irradiation, including internal mammary node irradiation when indicated. Clinically significant RP was defined as Common Terminology Criteria for Adverse Events (CTCAE v5.0) grade ⩾2 within six months of treatment. Ipsilateral-lung dose-volume metrics (V5Gy-V45Gy), maximum lung dose (D_max), mean lung dose (MLD), and effective dose (D_eff) were extracted from treatment plans. Four ML models (LightGBM, random forest, support vector machine, and logistic regression) were trained and evaluated using cross-validation. Model performance and feature reduction were analyzed.Results.Fifty-two patients (28%) developed clinically significant RP. All four models showed good and consistent predictive performance, with cross-validated AUC values >0.8. Across all models, high-dose lung parameters were the most reliable predictors. Models with reduced features (using V45Gy,D_max,D_eff, and MLD) showed competitive performance metrics compared to the respective full models.Significance.ML models built from simple, routinely available lung dosimetric data can effectively predict clinically significant RP after breast radiotherapy. Our findings suggest that small regions of the lung receiving higher doses play a greater role in RP risk than low-dose exposure alone. With further validation, this approach could support more personalized treatment planning and earlier identification of patients who may benefit from closer monitoring or preventive strategies.