Work

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

Year: 2026
Type: article
Study design: Observational Study
Abstract: 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. This dual-institution retrospective study included 94 breast cancer patients who received postoperative radiotherapy between March 2021 and December 2024. Four dose-volume-specific regions of interest (DVROIs) were automatically delineated within the ipsilateral lung based on the treatment plan. For each DVROI, 107 radiomic features were extracted using 3D slicer, along with nine clinical and dosimetric variables. Patients from Institution A comprised the training set, while those from Institution B formed an independent validation set. Preprocessing and feature selection were fitted solely on the training set and applied unchanged to the external set. Eleven machine learning (ML) algorithms were tuned via grid search within a repeated 5-fold cross-validation scheme and evaluated on the external set. The best algorithm for each DVROI was selected using a four-dimensional framework requiring a high external AUC, a small generalization gap, acceptable calibration, and a positive net benefit. Discrimination is reported with DeLong 95% confidence intervals, and SHAP was used for model interpretation. Boruta selected five features per DVROI, comprising 2–4 radiomic and 1–3 clinical/dosimetric variables; this count was consistent for V5. The final XGBoost model on DVROI5 achieved an external AUC of 0.776, an internal cross-validation (CV) AUC of 0.702, and a small generalization gap. The external Brier score was 0.203, the calibration slope was 1.255, and the net benefit improved by 0.113. Accuracy was 0.731, sensitivity 0.800, specificity 0.636, PPV 0.750, and NPV 0.700. Two DVROIs showed higher apparent AUC but failed the validation gates and were rejected. SHAP highlighted V5, GLCM, GLSZM, GLDM, and N3 as the top predictors. An interpretable XGBoost model using DVROI5 radiomic texture features and V5 demonstrated moderate discrimination for grade ≥ 1 radiation pneumonitis in breast cancer patients, with an external AUC of 0.776. The DVROI5 region, particularly the low-dose lung bath, was more informative than higher-dose regions. Limitations include a small cohort and the lack of formal model comparisons. These findings suggest that the low-dose irradiated lung is a key target for radiomic risk modeling.
Language: English
Field: Medicine
Open Access status: gold
Estimated APC paid: $1,690
OpenAlex ID: W7163711680
FWCI:
Cites:
Cited by:
Related to: 10