GS-ECG: a generalist–specialist approach to structural heart disease prediction from electrocardiograms

Abstract 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.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-17
DOI
https://doi.org/10.21203/rs.3.rs-11566820/v1
Primary Topic
ECG Monitoring and Analysis
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

GS-ECG: a generalist–specialist approach to structural heart disease prediction from electrocardiograms

Clarisse Florence de Vries, Honghan Wu, Jiacong Mi, Derek T. Connelly
4 links
Research Square
ECG Monitoring and Analysis
preprint

GS-ECG: a generalist–specialist approach to structural heart disease prediction from electrocardiograms

Clarisse Florence de Vries, Honghan Wu, Jiacong Mi, Derek T. Connelly
preprint en

Abstract

Abstract 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.

Research Square
University of Glasgow (GB)
Canadian Institute of Steel Construction, National Institute for Health and Care Research, Shanghai Jiao Tong University, Medical Research Council, Engineering and Physical Sciences Research Council, School of Medicine, Shanghai Jiao Tong University
Good health and well-being
ECG Monitoring and Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon