Dynamic Gating of Modality-Specific Experts for Interpretable CHD Diagnostic Prediction

Traditional multimodal approaches for Coronary Heart Disease (CHD) diagnostic prediction typically rely on static fusion strategies, assuming a fixed relative importance between structured biomarkers and unstructured clinical notes. This "one-size-fits-all" paradigm fails to capture the dynamic nature of clinical diagnosis, where the critical evidence shifts between modalities across different patients. To address this limitation, we propose a Compositional Mixture-of-Experts (MoE) framework. Unlike standard MoE models with homogeneous experts, our architecture is constructed from heterogeneous, role-specific reasoning modules: a Structured Expert, a Text Expert, and a Fusion Expert. A Dynamic Gating Network acts as a patient-specific arbitrator, implementing a routing inspired adaptive fusion mechanism. We further introduce an auxiliary loss mechanism to prevent expert collapse and ensure diverse specialization. Experiments on a real-world dataset of 60,339 EMRs demonstrate that our method not only achieves superior performance (AUPRC 0.9853) compared to static and attention-based fusion baselines (Cross-Attention Fusion, Multimodal Transformer; all p < 0.05) but also remains stable under moderate modal corruption and across the class-imbalance spectrum tested (1:1 to 20:1). Subgroup analysis suggests that the Compositional MoE's advantage is most pronounced in diagnostically challenging younger patients (<50 years: +0.0037 AUPRC over the strongest non-MoE model). The framework offers modality-level transparency: quantitative analysis of gating dynamics indicates that the model adaptively reweights information toward more reliable data sources when one modality is corrupted or ambiguous, which is qualitatively consistent with clinical intuition in representative cases.