HCM-Echo-VAR-Ensemble: Deep Ensemble Fusion to Detect Hypertrophic Cardiomyopathy in Echocardiograms

Abstract

Goal: To detect Hypertrophic Cardiomyopathy (HCM) from multiple views of Echocardiogram (cardiac ultrasound) videos. Methods: we propose HCM-Echo-VAR-Ensemble, a novel framework that performs binary classification (HCM vs. no HCM) of echocardiogram videos directly using an ensemble of state-of-the-art deep VAR architectures models (SlowFast and I3D), and fuses their predictions using majority averaging ensembling. Results: HCM-Echo-VAR-Ensemble achieved state-of-the-art accuracy of 95.28%, an F1-Score of 95.20%, a specificity of 96.20%, a sensitivity of 93.97%, a PPV of 96.46%, an NPV of 94.17%, and an AUC of 98.42%, outperforming a comprehensive set of baselines including other ensembling approaches. Conclusions: Our proposed HCM-Echo-VAR-Ensemble framework demonstrates significant potential for improving the sensitivity and accuracy of HCM detection in clinical settings, particularly by ensembling the complementary strengths of the SlowFast and I3D deep VAR models. This approach can enhance diagnostic consistency and accuracy, enabling reliable HCM diagnoses even in low-resource environments.

Keywords

• Cardiac assessment
• computer vision
• deep ensemble learning
• deep learning
• digital health
• echocardiogram