Anatomy-Derived Hemodynamic-Proxy Features Improve Image-Based Classification of Congenital Heart Disease

Congenital heart disease (CHD) encompasses many types of cardiovascular abnormalities, making image-supported diagnosis particularly challenging. In this study, we evaluated whether anatomy-derived hemodynamic-proxy features could improve machine learning classification of CHD subtypes (healthy, conotruncal defects, septal defects, and complex) when physiological measurements were unavailable. Publicly available human cardiac segmentation datasets were harmonized into a single dataset containing 194 cases spread into four classes. Fifteen anatomy-derived hemodynamic-proxy features and fifteen conventional anatomical features were extracted from the segmentation masks. Four machine learning models were optimized and evaluated using repeated stratified five-fold cross-validation. Machine learning models were assessed on accuracy, class-balanced accuracy, and macro F1 score for the anatomy-derived hemodynamic-proxy feature set, conventional anatomical feature set, and combined feature set. Across all four models, the combined anatomical plus hemodynamic-proxy feature set outperformed the anatomical-only feature set, with the best combined models reaching approximately 0.64–0.66 macro F1-score. These exploratory findings suggest that anatomy-derived hemodynamic-proxy features may provide complementary information for CHD classification, but validation in larger, source-balanced external cohorts and against direct hemodynamic measurements is required before generalizability or clinical utility can be established.