Introduction: Left ventricular ejection fraction (LVEF) is routinely assessed by echocardiography (Echo), whereas electrocardiography (ECG) is acquired as part of standard care in most patients. Multimodal contrastive learning offers a promising strategy to enrich ECG representations with Echo-derived structural and functional information, enabling scalable LVEF estimation from ECG alone.
Objectives: Building upon the Echo2ECG framework, we propose 1) reformulating the downstream task from categorical LVEF classification to continuous ejection fraction estimation, using probabilistic modeling via Gaussian mixture regression (GMR), and 2) enhance model pretraining by proposing new loss terms that capture view-specific and cardiac-cycle–aware features relevant to LVEF estimation.
Methodology: Paired ECG-Echo studies were extracted from MIMIC-IV using a maximum temporal gap of one day, yielding 5,513 valid pairs. We adopt the original CLIP-style objective from Echo2ECG, aligning ECG embeddings with study-level Echo representations, obtained by attention pooling over multi-view embeddings, derived from modality-specific foundation models. The pretrained model was evaluated on LVEF estimation with EchoNext dataset for external validation using AUROC for categorical comparison and MAE for regression performance.
Results: We achieve an AUROC of 0.779 and MAE of 0.0776 on LVEF classification (reduced ≤ 40%, mildly reduced 40–50%, normal ≥ 50%) and regression, slightly outperforming unimodal ECG baselines (AUROC 0.776). When using a probabilistic GMR head, performance changed to an AUROC of 0.780. For reference, Echo2ECG achieved an AUROC of 0.723 on EchoNext under a different split, providing a rough performance benchmark.
Conclusion: Our results show that multimodal contrastive pretraining between ECG and multi-view Echo remains effective when transferred from the original private-data setting to public MIMIC-IV, achieving competitive external generalization. Moreover, extending the framework with GMR moves beyond categorical prediction toward clinically relevant continuous and uncertainty-aware LVEF estimation. Future work will investigate auxiliary pretraining objectives to further improve ECG-based EF estimation.