Multimodal Late Fusion Meta-Learning for Pre-Treatment Radiation Pneumonitis Prediction
Abstract
Purpose
Radiation pneumonitis (RP) remains a clinically significant dose-limiting toxicity in thoracic radiotherapy. Accurate RP prediction is challenging due to its multifactorial etiology and complex interactions among contributing factors. Although multimodal data including radiomic, dosimetric, and clinical variables are essential for improved prediction, effective learning is hindered by their high dimensionality and heterogeneity. This study proposes a robust multimodal learning framework to improve RP prediction.
Methods
Pre-treatment CT images, dosimetric data, and clinical variables from 421 NSCLC patients enrolled in NRG Oncology RTOG 0617 were analyzed. Features were partitioned into five predefined blocks: multilobe radiomics, ipsilateral lung dose metrics, contralateral lung dose metrics, treatment arm–specific variables, and general clinical factors. For each feature block, base learners including LR, SVM, RF, ET, XGB, KNN, and TabPFN were trained using stratified 5-fold cross-validation. Dimensionality reduction (none, neighborhood component analysis, or principal component analysis) was optimized per block and classifier by maximizing fold-level AUROC. Optimized block-level predicted probabilities were combined using a logistic regression meta-learner in a late fusion framework. Performance was evaluated using mean AUROC ± standard deviation and bootstrap-derived 95% confidence intervals (CI).
Results
Late fusion improved RP prediction compared with individual feature blocks. Relative to clinical-only models, multimodal late fusion generally achieved higher AUROC and reduced variability across classifiers. The highest performance was observed for RF base learners (AUROC = 0.687±0.067; 95% CI: 0.610–0.766), followed by KNN (0.678±0.063; 95% CI: 0.610–0.755). ET- and LR-based fusion achieved AUROC values of 0.653±0.052 (95% CI: 0.569–0.729) and 0.633±0.061 (95% CI: 0.569–0.717), respectively, while XGB, SVM, and TabPFN showed lower and more variable performance.
Conclusion
The proposed multimodal late fusion meta-learning enables robust and interpretable integration of radiomic, dosimetric, and clinical data for RP prediction. This approach improves pre-treatment risk stratification and shows promise for supporting personalized treatment planning in lung cancer radiotherapy.