A multimodal deep learning model trained on 212,888 paired ECG–chest radiograph examinations from 116,380 patients across two Chinese centers substantially improved prediction of progression to moderate-to-severe regurgitant valvular heart disease (rVHD). The largest gain appeared for aortic regurgitation (AR), where combining both modalities raised the C-index from 0.616 (ECG alone) to 0.713, with AUROC reaching 0.729. Mitral and tricuspid regurgitation models also improved, achieving C-indices of 0.801 and 0.802 respectively. Grad-CAM visualizations confirmed the model attended to physiologically coherent features — relevant ECG leads and chamber-specific radiographic enlargement — lending biological credibility to its predictions.
Valvular heart disease affects roughly 2.5% of the general population and rises sharply with age, yet echocardiographic screening at population scale remains impractical. Repurposing ubiquitous, low-cost diagnostics like ECGs and chest X-rays as screening filters is an appealing strategy that this work advances meaningfully. The token-level cross-modal fusion and class-specific gating mechanism represent genuine architectural novelty, not just ensemble averaging. That said, important caveats apply: this is a retrospective single-country study, and generalizability to non-Chinese, community, or primary-care populations is unproven. The notably high AUPRC values (0.972 for both AR and MR) warrant scrutiny given typical class imbalance in incident disease datasets. Crucially, this is a preprint posted on medRxiv and has not yet undergone peer review — findings and performance metrics may shift after independent expert evaluation. If validated prospectively, this model could meaningfully triage high-risk patients toward echocardiography, optimizing resource allocation without sacrificing diagnostic sensitivity.
