Integrating Clinical, Dose, and Image Data to Predict Radiation Pneumonitis before Treatment
Abstract
Purpose
Radiation pneumonitis (RP) is a dose-limiting toxicity following thoracic radiotherapy. Existing RP prediction models have largely focused on clinical factors and conventional dose-volume histogram (DVH) metrics, with limited consideration of higher-order dosiomic and CT radiomic features within integrated models. Herein, we developed and evaluated predictive models of RP using combinations of clinical, DVH or dosiomic, and CT radiomic features.
Methods
We retrospectively analyzed 226 lung cancer patients treated with radiotherapy at BC Cancer. Clinical variables were obtained through chart review, and planning CT images and 3D dose distributions were extracted from Aria. Radiomic features were computed from CT volumes, while DVH and dosiomic features were derived from 3D dose distributions. RP prediction was evaluated using three supervised learning models: elastic net, random forest, and sparse partial least squares-discriminant analysis. Model performance was assessed by AUC using 10-fold cross-validation. Data integration was performed using early fusion (feature-level integration) and late fusion (prediction-level integration).
Results
In single-modality analyses, radiomic models achieved the highest predictive performance (AUC=0.68–0.69), significantly outperforming clinical-based models (AUC=0.57–0.62, p0.06). No significant performance difference was observed between DVH- and dosiomic-based models (p>0.2). Among multi-modality approaches, early fusion of clinical, DVH, and radiomic features yielded the strongest RP prediction (AUC=0.70–0.71), significantly outperforming two of three dose-based models (p0.3). Early and late fusion strategies showed comparable performance (p>0.05).
Conclusion
Integrating patient- and treatment-specific features improves RP prediction beyond dose-based modeling alone. Conventional DVH metrics performed comparably to dosiomic features, while radiomic features emerged as the most predictive single modality.