Causal-Driven Severity Grading for Myocardial Ischemia Diagnosis Based on Magnetocardiography

Magnetocardiography (MCG) is currently a promising technique for non-invasive diagnosis of myocardial ischemia. Clinicians can evaluate the degree of ischemia and categorize subjects into three groups: mild ischemia, severe ischemia, and healthy individuals, by examining the spatiotemporal evolution of cardiac magnetic three-field maps. However, the complex spatiotemporal dynamics of MCG data make it challenging for a single modality to characterize ischemia severity. Consequently, we propose a causal-driven multimodal fusion framework that integrates handcrafted key features with MCG image representations. This framework systematically models two types of confounders using a Structural Causal Model (SCM), namely latent visual confounders and cross-modal fusion confounders. To mitigate spurious correlations and feature redundancy during representation learning, we design two causal-inspired modules based on front-door adjustment and counterfactual intervention. Experimental results on our dataset demonstrate the effectiveness of the proposed framework in improving MCG-based ischemia severity grading.