Radiogenomics improves Classification of Genetic Idiopathic Cardiomyopathy

Background: Idiopathic non-ischemic cardiomyopathy (iNICM) is a heterogeneous disorder with likely genetic determinants. Radiomics using cardiovascular magnetic resonance (CMR) offers the potential to capture subtle imaging phenotypes linked to genetic variation. We sought to determine the ability of integrated radiogenomics to improve the classification of pathogenic vs. variant of unknown significance (VUS) in iNICM compared to models relying on clinical or radiomic features alone.

Methods: We analyzed 54 patients who underwent clinical CMR and genetic evaluation (Invitae Cardiomyopathy Comprehensive Panel, San Francisco, CA; Ambry CMNext Panel, Aliso Viejo, CA) for iNICM. CMR included pre- and post-contrast T1 mapping (modified Look-Locker inversion recovery) and 2D phase-sensitive inversion recovery late gadolinium enhancement (LGE), all in the short-axis orientation. The myocardium (T1 map: 3 slices; LGE: 8 mm slices, 10 mm gap from base to apex) was segmented using semi-automated techniques (cvi42 version 6.2.1 [4569], Circle Cardiovascular Imaging, Canada). Radiomic features from the myocardium were extracted (PyRadiomics, Python). We considered the following clinical variables in the models: age, sex, body surface area, systolic and diastolic blood pressure, heart rate, left ventricular (LV) and right ventricular (RV) mass index, LV and RV end-diastolic volume index, and LV and RV ejection fraction. Three machine learning models were developed for the classification of genotype: pathogenic vs. VUS: clinical parameters alone, radiomics alone, and a combined clinical+radiomics approach (Figure 1). Features were normalized and reduced using mutual information and random forest feature importance. Models were trained and evaluated using 5-fold cross-validation (CV) with repeated runs, classifying pathogenic variants versus VUS and reporting sensitivity, specificity, accuracy, and area under the curve (AUC).

Results: The radiomics-only model for classifying pathogenic vs. VUS in iNICM achieved a mean CV accuracy of 0.87 and an AUC of 0.83, outperforming the clinical-only model (0.61 accuracy, 0.47 AUC). The combined clinical+radiomics model showed intermediate performance (0.79 accuracy, 0.79 AUC) (Figure 2), suggesting that clinical parameters added noise rather than improving classification. The combined model yielded a net re-classification improvement of 68.7% (Figure 3).

Conclusion: Radiomics from cardiac MRI shows promise in classifying iNICM genetic variants, outperforming clinical model. External validation will be essential to confirm these findings and support future clinical application.

Study pipeline showing processing of clinical and radiomic features, feature selection, Bayesian network model training, and performance evaluation by cross-validation.
Cross-validation results comparing Clinical, Radiomics, and Combined models, and presenting sensitivity, specificity, accuracy, and ROC-AUC performance.
Net reclassification improvement (NRI) analysis demonstrating improvement in genotype discrimination when combining clinical and radiomics features compared with clinical features alone.