Key result
GS-ECG outperforms EchoNext-Mini for predicting composite structural heart disease, achieving ~0.82 AUROC.
Why the study?
Structural heart disease often goes undiagnosed and echocardiography cannot be offered to every patient promptly. While ECG foundation models could help prioritize testing, SHD abnormalities are defined by echocardiography and fall outside pretraining labels, making it unclear how best to use these models.
Does a combined frozen generalist and specialist AI model (GS-ECG) improve the prediction of structural heart disease from routine electrocardiograms compared to fine-tuning an ECG foundation model or existing specialist models?
Observational (n=7,385)
Does a combined frozen generalist and specialist AI model (GS-ECG) improve the prediction of structural heart disease from routine electrocardiograms compared to fine-tuning an ECG foundation model or existing specialist models?
Answers are grounded in this paper and cite their sources.
Pairing a frozen ECG foundation model with a lightweight, task-specific specialist neural network (GS-ECG) significantly improves the automated detection of structural heart disease from routine ECGs compared to full model fine-tuning, aiding in prioritizing patients for echocardiography.
Structural heart disease (SHD) often goes undiagnosed, and echocardiography, the standard test, cannot be offered to every patient promptly. AI analysis of routine electrocardiograms (ECGs) could help prioritize who is scanned first. ECG foundation models promise broad applicability across ECG tasks, but SHD abnormalities are novel targets, defined by echocardiography and outside the labels on which these models were pretrained, so how best to use them is unclear. Rather than adapting the foundation model to these targets, we propose GS-ECG, a synergistic generalist–specialist approach that keeps it frozen and pairs it with a compact specialist, DGRNet, trained on novel SHD targets. In this setup, the generalist summarizes each recording in long-context features, while DGRNet focuses on SHD-specific waveform morphology at beat-level resolution, with a receptive field of about two cardiac cycles per feature. We evaluated GS-ECG in two health systems: MIMIC-IV-Ext-SHD (49,229 patients), a new public benchmark, and the EchoNext-Mini Dataset (36,286 patients). Full fine-tuning of the foundation model improved mean AUPRC by only 1.3–3.6% over freezing it, whereas DGRNet alone exceeded fine-tuning by 6.1–6.4%. Combining the frozen generalist with DGRNet in GS-ECG performed best, exceeding fine-tuning by 8.1% in both datasets while training 4.9% as many parameters; its composite SHD AUROC was 0.825 and 0.839. Against the published EchoNext-Mini Model on its own dataset, GS-ECG found 17–25% more patients with aortic stenosis, mitral regurgitation or tricuspid regurgitation among the top 5% ranked for echocardiography. Pairing frozen foundation models with purpose-built specialists offers an efficient route to clinical targets that lie beyond pretraining labels.
No takes yet. Share an insight, caveat, or question.
Mi et al. (2026) conducted an observational in Structural heart disease (n=7,385). GS-ECG (generalist-specialist fusion model combining ECGFounder-Frozen and DGRNet) vs. EchoNext-Mini Model was evaluated on Composite structural heart disease (SHD) prediction AUROC on the MIMIC-IV-Ext-SHD test set (AUROC 0.8248, 95% CI 0.8154-0.8339). GS-ECG outperformed the EchoNext-Mini Model for predicting composite structural heart disease on MIMIC-IV-Ext-SHD.
Synapse has enriched 10 related papers. Consider them for comparative context: