Improving Cardiac Toxicity Prediction from PET Radiomics: Cross-Cohort Validation and Late-Fusion Ensemble Modeling
Abstract
Purpose
FDG-PET/CT scans obtained for lung cancer staging may encode underutilized cardiac information. This study aimed to: (1) assess the predictive power of PET-derived radiomics for pre-treatment cardiovascular conditions; (2) evaluate model performance across two clinically distinct cohorts; and (3) develop a late-fusion ensemble model to maximize predictive discrimination by integrating clinical and radiomics data.
Methods
We utilized two retrospective cohorts in a strict train–test design: Cohort A (SBRT, AJCC stage ≤3, n=95, 51% toxicity) for model development and Cohort B (mixed SBRT/standard fractionation with broader staging, n=89, 24% toxicity) for independent testing. We compared three strategies: (1) clinical-only (logistic regression); (2) radiomics-only (random forest with optimized feature selection); and (3) a late-fusion “router” ensemble trained on out-of-fold probabilities from the training set. The router integrates base model probabilities via constrained linear regression to learn optimal modality contributions. Performance was evaluated using Area Under the Curve (AUC) and Average Precision (AP).
Results
Cohort B patients were older (73.5 ± 9.4 vs. 67.9 ± 9.6 years) and had lower coronary artery calcium burden variability compared to Cohort A (246 ± 330 vs. 332 ± 522). Cardiac event prevalence differed significantly between cohorts (p=0.0004), confirming that Cohort B reflects a clinically distinct patient cohort. The clinical-only model achieved modest discrimination (test AUC 0.639, AP 0.358). Feature-selected radiomics improved discrimination (AUC 0.700, AP 0.475). The router achieved the best performance (AUC 0.718, AP 0.498). Notably, this AP represents a >2-fold improvement over the test cohort’s baseline prevalence (0.236). Learned router weights (Radiomics 0.72 vs. Clinical 0.31) indicated radiomics as the dominant source of discriminative signal.
Conclusion
PET-derived cardiac radiomics capture cardiovascular risk beyond routine clinical variables. A radiomics-dominant late-fusion ensemble maintained strong performance across clinically distinct cohorts. This approach supports the feasibility of robust cardiac risk profiling using standard oncologic PET/CT.