309 - Beyond DVH: Radiomics and Medical Foundation Models Improve Cardiac Event Prediction in Locally Advanced Non-Small Cell Lung Cancer

Purpose/Objective(s): Recent studies suggest that evolving practice in locally advanced non-small cell lung cancer (LA-NSCLC), particularly the adoption of consolidation immunotherapy (CIO), may limit the prognostic utility of whole-heart (WH) and cardiac substructure dose-volume histogram (DVH) metrics for cardiac events. This work explores whether radiomics, dosiomics, and state-of-the-art medical foundation models improve prediction of cardiac events after definitive chemoradiotherapy in LA-NSCLC.

Materials/Methods: We retrospectively analyzed 784 patients with LA-NSCLC treated between 2010 and 2021 across four hospitals (Sites 1-4) in our institution. Patient data from Site 1 (n=628) were used for model development and internal testing through fivefold nested cross-validation, whereas data from Sites 2-4 (n=156) were held out for independent testing. Candidate predictors included 53 clinical variables (demographics, cardiovascular risk factors, medications, and treatment regimen) as well as WH/substructure DVH metrics (132 total), radiomics (planning CT; 530/structure), and dosiomics features (516/structure). Additionally, CT volumes were cropped to the WH region and processed by an open-source foundation model, MedImageInsight (MII), to generate image-only embeddings (d=1024). Time to major adverse cardiac events (MACE) was modeled using elastic net-regularized Cox regression after feature selection. Model performance was quantified by the concordance index (C-index), and risk stratification was assessed using log-rank tests. Subgroup analyses were performed among patients receiving CIO (n=224).

Results: In internal testing, the clinical+WH radiomics model achieved the highest prognostic performance for MACE (C-index 0.77±0.02), compared with the clinical-only model (0.74±0.02). When applied to the hold-out test set, performance decreased for both models (0.67±0.01 for clinical+WH radiomics; 0.65±0.01 for clinical-only). Feature-importance analysis identified WH gray-level non-uniformity as the primary radiomics contributor. Adding DVH or dosiomics features to clinical variables did not improve performance, and substructure-based outcome models did not outperform WH-based models. In contrast, MII-derived embeddings provided added prognostic value beyond clinical variables and demonstrated stronger performance on independent testing (internal 0.75±0.02; hold-out 0.71±0.01). In the CIO subgroup, both the clinical+WH radiomics and clinical+MII models significantly stratified patients into low- and high-risk groups (log-rank P=0.037 and 0.007, respectively).

Conclusion: Incorporating WH radiomics and medical foundation model embeddings with clinical risk factors improved prediction of cardiac events in LA-NSCLC, supporting advanced imaging representations to inform individualized cardiac-sparing strategies. Future studies will focus on prospective validation and interpretability of foundation model embeddings.