EchoGraph-FusionNet: Adaptive Disease-Graph and Mixture-of-Experts Learning for Predicting Echocardiographic Abnormalities from ECGs

Background/Objectives: Cardiovascular diseases remain a major global health burden, while echocardiography, despite its diagnostic value, is not always readily accessible. This study proposes EchoGraph-FusionNet, a multimodal deep learning framework for predicting echocardiographic abnormalities and structural heart disease (SHD) from 12-lead electrocardiograms (ECGs) and clinical variables. Methods: The framework integrates ECG and clinical representations through cross-modal fusion, followed by a disease-aware mixture-of-experts module and a two-layer adaptive graph attention network that jointly models 11 echocardiographic abnormalities. Five continuous echocardiographic measurements were incorporated as auxiliary regression tasks. Prevalence-aware logit adjustment and asymmetric loss were employed to address class imbalance. Performance was evaluated on the predefined EchoNext training, validation, and held-out test partitions across three independent random seeds. Results: On the held-out test set, EchoGraph-FusionNet achieved a macro AUROC of 0.8066 ± 0.0025 and macro AUPRC of 0.2838 ± 0.0056 across 11 abnormalities. For SHD detection, AUROC and AUPRC reached 0.8345 ± 0.0001 and 0.8039 ± 0.0031, respectively. Conclusions: EchoGraph-FusionNet effectively combines multimodal representations, disease relationships, and multitask learning for ECG-based prediction of echocardiographic abnormalities, supporting its potential as a screening and prioritization approach for structural cardiac assessment.