ESTRO 2025 - Abstract Book

S3813

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

predictions were validated on the hold-out prospective cohort to evaluate model robustness. Model performance was assessed using sensitivity, specificity, and area under the curve (AUC). Wilcoxon signed-rank tests were used as different models used identical training/testing data in every iteration. Results: Hs-cTnT elevation incidence was 31.9% in retrospective and 31.5% in prospective cohort. Table1 and Figure 1 demonstrate the quantitative performance and receive operating characteristics (ROC) curves of different models. The highest predictive performance in cross-validation was achieved by a model with clinical factors combined with substructure DVH metrics as inputs (AUC: 0.71 [0.70, 0.73], sensitivity: 0.64 [0.61, 0.57], specificity: 0.66 [0.64, 0.69]). While adding radiomic and dosiomic features in WH model achieved comparable cross-validation AUC (0.70 [0.68, 0.71], p =0.051), all WH-based models underperformed in hold-out validation (AUCs≤0.51), indicating limited robustness of WH-based models when evaluating different datasets. In contrast, the substructure DVH model achieved an AUC of 0.61 in hold-out validation, with a sensitivity of 0.47 and specificity of 0.76. Feature analysis identified the left anterior descending coronary artery V20Gy and right ventricle maximum dose as important predictors.