Interpretable Machine Learning Using Normal Lung Radiomics and Dosiomics to Predict Radiation Pneumonitis After Lung Cancer Radiotherapy

Abstract Objective This study aimed to develop and interpret machine learning models based on radiomics and dosiomics features extracted from normal lung tissue after excluding the planning target volume for predicting grade ≥ 2 radiation pneumonitis within 6 months after radiotherapy. Methods A total of 232 patients with lung cancer who received radiotherapy between February 2023 and June 2025 were retrospectively enrolled and divided into a training set (n = 162) and a validation set (n = 70). The region of interest was defined as the whole lung excluding the planning target volume. Radiomics features were extracted from pretreatment contrast-enhanced simulation CT images, and dosiomics features were extracted from three-dimensional RT Dose files using the same normal lung region. After preprocessing, LASSO regression and random forest were used for feature selection, and their intersecting features were retained. Eleven machine learning models were trained using five-fold cross-validation and evaluated in the validation set. Model performance was assessed using AUC, classification metrics, confusion matrix, and decision curve analysis. Kernel SHAP was applied for global model interpretation. Results Among all patients, 97 developed grade ≥ 2 radiation pneumonitis, with an overall incidence of 41.8%. From 214 extracted features, 18 radiomics and dosiomics features were selected for modeling. Validation-set AUCs ranged from 0.725 to 0.812. LASSO and linear discriminant analysis achieved the highest AUCs, both 0.812; the 95% CI for LASSO was 0.708–0.917. The final LASSO model showed an accuracy of 74.3%, sensitivity of 58.1%, and specificity of 87.2%. SHAP analysis indicated that first-order, shape, and texture features jointly contributed to model prediction. Conclusion Radiomics and dosiomics features from normal lung tissue may provide complementary information for predicting grade ≥ 2 radiation pneumonitis after lung cancer radiotherapy. The LASSO model showed favorable discrimination and interpretability in the validation set, but external validation is needed to further assess its stability and generalizability.