Radiomics Improves Prognostication Beyond Clinical Variables and Dose-Volume Metrics for Cardiac Events In Locally Advanced Non-Small Cell Lung Cancer
Abstract
Purpose
Cardiotoxicity remains a major concern in thoracic radiotherapy. Prior studies have associated whole heart (WH) and substructure dose-volume histogram (DVH) metrics with cardiac events; however, changing practice (e.g., immunotherapy consolidation, advanced treatment techniques) may limit DVH-based risk models. This study investigated whether CT-based radiomics and dosiomics added prognostic value beyond clinical variables and DVH features for cardiac events after definitive chemoradiotherapy in locally advanced non-small cell lung cancer (LA-NSCLC)
Methods
We performed a single-institution, multi-site retrospective analysis of 784 patients with LA-NSCLC treated between 2010 and 2021; 227 received consolidation immunotherapy (CIO) after radiotherapy. Clinical variables included demographics, cardiovascular risk factors, and treatment regimen. DVH features were extracted for the WH and multiple cardiac substructures, and corresponding radiomics and dosiomics features were computed. An imaging-perturbation approach was applied to remove unstable features. Endpoints were major adverse cardiac events (MACE; primary) and grade ≥3 cardiac events (G3+Cardiac; secondary). A standardized pipeline combining feature selection and elastic-net penalized Cox regression was developed and evaluated using fivefold nested cross-validation. Performance was assessed by the concordance index (C-index), and risk stratification was evaluated with log-rank tests. Subgroup analyses were applied to patients receiving CIO.
Results
Models incorporating clinical variables and WH radiomics achieved the highest prognostic performance for both MACE (C-index 0.75±0.01) and G3+Cardiac (0.73±0.01), significantly outperforming (P<0.001) clinical-only models (0.72±0.01 and 0.70±0.01, respectively). Adding WH or substructure DVH features to clinical variables did not improve performance. Substructure-based outcome models did not surpass WH-based models. In the CIO subgroup, the clinical+WH radiomics model showed a significant separation between low- and high-risk patients (log-rank P=0.006 for MACE; P<0.001 for G3+Cardiac).
Conclusion
Incorporating WH radiomics with clinical risk factors improves prognostic modeling of cardiac events in LA-NSCLC, enabling more accurate risk assessment and individualized cardiac-sparing strategies. Future studies will focus on prospective and multi-institution validation.