Domain-Adapted Fine-Tuning of ECG Foundation Models for Multi-Label Structural Heart Disease Screening

Duc Do, Minh Do, Dang Nguyen, Hung Huynh, Quan Le, Jacques Kpodonu, Phat Huynh, Khanh Le, Khoa Pham, Phi Pham-Van-Hoang, Quan K. Huynh, Ramez M. Odat, Perisa Ashar, Ethan Philip Lowder, Minh H.N. Le, Hoang Le, Phat V.H. Nguyen
Proceedings of the The 39th Canadian Conference on Artificial Intelligence, PMLR 318:678-692, 2026.

Abstract

Transthoracic echocardiography is the reference standard for confirming structural heart disease (SHD), but its use as a first-line screening modality is limited by cost, workflow burden, and specialist availability. This study investigated whether open pretrained electrocardiogram (ECG) models can support echo-confirmed multi-label SHD detection using the public EchoNext Mini-Model benchmark. We focused on six moderate-or-greater echocardiography-derived abnormalities spanning reduced left ventricular ejection fraction, increased left ventricular wall thickness, aortic stenosis, mitral regurgitation, tricuspid regurgitation, and right ventricular systolic dysfunction. Under a common experimental pipeline, we compared engineered ECG features with gradient boosting, end-to-end waveform learning from scratch, and transfer from open ECG foundation models. We then evaluated continued in-domain self-supervised adaptation of ECG-FM on EchoNext waveforms followed by selective supervised fine-tuning, with emphasis on the trade-off between discrimination and adaptation cost. Among the evaluated configurations, the adapted ECG-FM models achieved the strongest overall performance. Across adaptation depths, peak macro-AUROC and macro-AUPRC reached 0.8509 and 0.4297, respectively, while a more parameter-efficient operating point preserved nearly identical AUROC (0.8501) and achieved the highest fixed-threshold macro-F1 (0.3691). Late fusion of the release-provided covariates did not improve threshold-independent discrimination, and the evaluated low-rank adaptation (LoRA) configuration, alternative foundation backbones, and mixture-of-foundation-model strategies did not surpass the best adapted single-backbone operating points. These findings indicate that, for ECG-based case finding and echocardiography triage, the most effective transfer strategy is to combine target-domain self-supervised adaptation with selective supervised updating of a pretrained ECG backbone.

Cite this Paper

BibTeX
@InProceedings{pmlr-v318-do26a, title = {Domain-Adapted Fine-Tuning of ECG Foundation Models for Multi-Label Structural Heart Disease Screening}, author = {Do, Duc and Do, Minh and Nguyen, Dang and Huynh, Hung and Le, Quan and Kpodonu, Jacques and Huynh, Phat and Le, Khanh and Pham, Khoa and Pham-Van-Hoang, Phi and Huynh, Quan K. and Odat, Ramez M. and Ashar, Perisa and Lowder, Ethan Philip and Le, Minh H.N. and Le, Hoang and Nguyen, Phat V.H.}, booktitle = {Proceedings of the The 39th Canadian Conference on Artificial Intelligence}, pages = {678--692}, year = {2026}, editor = {Bouzar-Benlabiod, Lydia and Leung, Carson}, volume = {318}, series = {Proceedings of Machine Learning Research}, month = {25--29 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v318/main/assets/do26a/do26a.pdf}, url = {https://proceedings.mlr.press/v318/do26a.html}, abstract = {Transthoracic echocardiography is the reference standard for confirming structural heart disease (SHD), but its use as a first-line screening modality is limited by cost, workflow burden, and specialist availability. This study investigated whether open pretrained electrocardiogram (ECG) models can support echo-confirmed multi-label SHD detection using the public EchoNext Mini-Model benchmark. We focused on six moderate-or-greater echocardiography-derived abnormalities spanning reduced left ventricular ejection fraction, increased left ventricular wall thickness, aortic stenosis, mitral regurgitation, tricuspid regurgitation, and right ventricular systolic dysfunction. Under a common experimental pipeline, we compared engineered ECG features with gradient boosting, end-to-end waveform learning from scratch, and transfer from open ECG foundation models. We then evaluated continued in-domain self-supervised adaptation of ECG-FM on EchoNext waveforms followed by selective supervised fine-tuning, with emphasis on the trade-off between discrimination and adaptation cost. Among the evaluated configurations, the adapted ECG-FM models achieved the strongest overall performance. Across adaptation depths, peak macro-AUROC and macro-AUPRC reached 0.8509 and 0.4297, respectively, while a more parameter-efficient operating point preserved nearly identical AUROC (0.8501) and achieved the highest fixed-threshold macro-F1 (0.3691). Late fusion of the release-provided covariates did not improve threshold-independent discrimination, and the evaluated low-rank adaptation (LoRA) configuration, alternative foundation backbones, and mixture-of-foundation-model strategies did not surpass the best adapted single-backbone operating points. These findings indicate that, for ECG-based case finding and echocardiography triage, the most effective transfer strategy is to combine target-domain self-supervised adaptation with selective supervised updating of a pretrained ECG backbone.} }
Endnote
%0 Conference Paper %T Domain-Adapted Fine-Tuning of ECG Foundation Models for Multi-Label Structural Heart Disease Screening %A Duc Do %A Minh Do %A Dang Nguyen %A Hung Huynh %A Quan Le %A Jacques Kpodonu %A Phat Huynh %A Khanh Le %A Khoa Pham %A Phi Pham-Van-Hoang %A Quan K. Huynh %A Ramez M. Odat %A Perisa Ashar %A Ethan Philip Lowder %A Minh H.N. Le %A Hoang Le %A Phat V.H. Nguyen %B Proceedings of the The 39th Canadian Conference on Artificial Intelligence %C Proceedings of Machine Learning Research %D 2026 %E Lydia Bouzar-Benlabiod %E Carson Leung %F pmlr-v318-do26a %I PMLR %P 678--692 %U https://proceedings.mlr.press/v318/do26a.html %V 318 %X Transthoracic echocardiography is the reference standard for confirming structural heart disease (SHD), but its use as a first-line screening modality is limited by cost, workflow burden, and specialist availability. This study investigated whether open pretrained electrocardiogram (ECG) models can support echo-confirmed multi-label SHD detection using the public EchoNext Mini-Model benchmark. We focused on six moderate-or-greater echocardiography-derived abnormalities spanning reduced left ventricular ejection fraction, increased left ventricular wall thickness, aortic stenosis, mitral regurgitation, tricuspid regurgitation, and right ventricular systolic dysfunction. Under a common experimental pipeline, we compared engineered ECG features with gradient boosting, end-to-end waveform learning from scratch, and transfer from open ECG foundation models. We then evaluated continued in-domain self-supervised adaptation of ECG-FM on EchoNext waveforms followed by selective supervised fine-tuning, with emphasis on the trade-off between discrimination and adaptation cost. Among the evaluated configurations, the adapted ECG-FM models achieved the strongest overall performance. Across adaptation depths, peak macro-AUROC and macro-AUPRC reached 0.8509 and 0.4297, respectively, while a more parameter-efficient operating point preserved nearly identical AUROC (0.8501) and achieved the highest fixed-threshold macro-F1 (0.3691). Late fusion of the release-provided covariates did not improve threshold-independent discrimination, and the evaluated low-rank adaptation (LoRA) configuration, alternative foundation backbones, and mixture-of-foundation-model strategies did not surpass the best adapted single-backbone operating points. These findings indicate that, for ECG-based case finding and echocardiography triage, the most effective transfer strategy is to combine target-domain self-supervised adaptation with selective supervised updating of a pretrained ECG backbone.
APA
Do, D., Do, M., Nguyen, D., Huynh, H., Le, Q., Kpodonu, J., Huynh, P., Le, K., Pham, K., Pham-Van-Hoang, P., Huynh, Q.K., Odat, R.M., Ashar, P., Lowder, E.P., Le, M.H., Le, H. & Nguyen, P.V.. (2026). Domain-Adapted Fine-Tuning of ECG Foundation Models for Multi-Label Structural Heart Disease Screening. Proceedings of the The 39th Canadian Conference on Artificial Intelligence, in Proceedings of Machine Learning Research 318:678-692 Available from https://proceedings.mlr.press/v318/do26a.html.

Related Material