Presenter(s)
D. Panchal1, W. Cao2, N. Halder3, M. Dichmann3, P. Faherty4, E. Blomain5, W. Choi6, and Y. Vinogradskiy6; 1Department of Radiation Oncology, Thomas Jefferson University Hospital, Philadelphia, PA, 2Department of Radiation Oncology, Sidney Kimmel Medical College & Cancer Center at Thomas Jefferson University, Philadelphia, PA, 3Department of Radiation Oncology, Sidney Kimmel Medical College at Thomas Jefferson University, Philadelphia, PA, 4Thomas Jefferson University Hospital, Philadelphia, PA, 5Department of Radiation Oncology, Thomas Jefferson University, Philadelphia, PA, 6Dept. of Radiation Oncology, Sidney Kimmel Medical College and Comprehensive Cancer Center, Thomas Jefferson University, Philadelphia, PA
Purpose/Objective(s): Pre-therapy cardiac events in lung cancer patients are frequently unrecognized despite their impact on patient outcomes. Fluorodeoxyglucose PET scans obtained for routine staging may contain underutilized cardiac signal. This study aims to apply PET-derived radiomics features to detect pre-therapy cardiac events in lung cancer patients.
Materials/Methods: The study used two retrospective cohorts in a strict train-test design: Cohort A (SBRT, AJCC stage = 3, n = 95, 48% cardiac events) for model development and Cohort B (mixed fractionation, broader staging, n = 89, 24% events) for independent testing. Clinical features for model development included treatment variables (prescription dose, mean heart dose), patient characteristics (age, AJCC stage, functional status), and comorbidities (diabetes, hyperlipidemia, hypertension, smoking status). Radiomics features were extracted from the cardiac region of the PET scans; highly correlated features (|?| > 0.85) were removed, and the remaining features were reduced using family-wise principal component analysis, retaining three principal components per feature family. Cardiac event classification was compared across clinical-only, radiomics-only, and mid-fusion models, with the mid-fusion model combining clinical and radiomics features at an intermediate stage before final classification. Four machine learning classifiers (Logistic Regression, Support Vector Machine, Random Forest, and ExtraTrees) were evaluated. Models were trained and internally validated using 10-fold cross-validation, with performance evaluated by AUC and average precision (AP) to account for class imbalance.
Results: Cohort B patients were older (73.5 ± 9.4 vs 67.9 ± 9.6 years) with lower coronary artery calcium burden variability, and event prevalence differed significantly (p = 0.0004), confirming clinical distinctiveness. Across all feature sets, the ExtraTrees classifier achieved the highest performance. Clinical-only models showed limited predictive performance (best test AUC 0.550, AP 0.255). Incorporating PET-derived radiomics features improved prediction (radiomics-only: best test AUC 0.663, AP 0.398), highlighting the added value of imaging data. The mid-fusion model, which integrates clinical and radiomics features, achieved the best performance (best test AUC 0.666, AP 0.485), demonstrating synergistic gains from combining modalities.
Conclusion: The mid-fusion model demonstrated superior performance and robust generalizability in an independent cohort, accurately identifying pre-therapy cardiac events despite clinical differences and varying event prevalence. Integration of PET radiomics with clinical variables improved pre-therapy cardiac event classification compared to either modality alone, supporting multimodal modeling for baseline cardiac risk stratification. These predictions could guide early interventions or closer monitoring to mitigate cardiac risk prior to radiotherapy.