Leveraging ECG foundation models in critical care for sinus rhythm and atrial fibrillation classification

Q: What is the clinical evidence from this study?

Population: Atrial fibrillation and sinus rhythm (n=226). Intervention: Fine-tuned ECG foundation models vs. Off-the-shelf pre-trained ECG foundation models. Primary outcome: F1 score for atrial fibrillation detection.

Q: What does this research mean for the field?

Fine-tuned ECG foundation models achieve robust performance for detecting sinus rhythm and atrial fibrillation in ICU patients, significantly outperforming off-the-shelf models which show limited transferability. Novelty: ClaimNovelty.INCREMENTAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

Maria Galanty; Michiel Hulleman; Merijn C. Reuland; Björn van der Ster; Alexander P. J. Vlaar; Clara I. Sánchez; the AF Expert Panel

doi:10.1186/s40635-026-00949-7

August 26, 2026Intensive Care Medicine ExperimentalOpen Access

Leveraging ECG foundation models in critical care for sinus rhythm and atrial fibrillation classification

View Full Paper

Ask AI

Bookmark

Share

Key result

Targeted fine-tuning of ECG foundation models boosts ICU AF detection F1 score to 0.83.

n=226

Figures

Why the study?

It was unknown how well self-supervised ECG foundation models perform in detecting atrial fibrillation and sinus rhythm during continuous monitoring of intensive care unit patients.

Does targeted fine-tuning of ECG foundation models improve the classification of atrial fibrillation and sinus rhythm in ICU patients compared to off-the-shelf models?

Key Points

To evaluate the performance and generalizability of pre-trained ECG foundation models for detecting atrial fibrillation and sinus rhythm during continuous monitoring in intensive care unit patients.
Assessed four public ECG datasets (PTB-XL, Chapman-Shaoxing, Georgia, and CPSC) to identify datasets with high label consistency via 5-fold cross-validation.
Tested off-the-shelf fine-tuned ECG foundation models and models fine-tuned on selected public datasets (PTB-XL and Chapman-Shaoxing), both with and without weakly labelled ICU data, on an expert-annotated ICU cohort of 592 validated ECG segments from 226 patients.
The off-the-shelf ECG foundation model fine-tuned on MIMIC-IV-ECG demonstrated limited transferability on the local ICU expert dataset, yielding an F1 score of 0.523 (precision 0.359, recall 0.965) for atrial fibrillation and 0.677 (precision 0.610, recall 0.762) for sinus rhythm.
Fine-tuning on curated public datasets improved ICU expert evaluation performance to an F1 score of 0.814±0.021 for atrial fibrillation and 0.914±0.004 for sinus rhythm.
Incorporating weakly labelled local ICU data into fine-tuning further increased recall and achieved F1 scores of 0.830±0.009 for atrial fibrillation and 0.920±0.003 for sinus rhythm on the ICU expert dataset.

Structured PICO

Does targeted fine-tuning of ECG foundation models improve the classification of atrial fibrillation and sinus rhythm in ICU patients compared to off-the-shelf models?

P

Population

226 ICU patients with continuous ECG monitoring evaluated for atrial fibrillation and sinus rhythm to test the generalizability of ECG foundation models.

I

Intervention

Targeted fine-tuning of ECG foundation models (ECG-FMs) on curated public datasets, with and without weakly labelled in-house ICU data.

C

Comparator

Off-the-shelf pre-trained ECG-FMs (fine-tuned on MIMIC-IV-ECG).

O

Outcome

Classification performance (F1 score, accuracy, precision, recall) for detecting atrial fibrillation (AF) and sinus rhythm (SR).surrogate

Main Result

Absolute Event Rate: 0.83% vs 0.52%

Limitations

Imbalanced dataset with respect to AF and SR labels.
Small number of noisy samples limits conclusions on signal quality.
Misclassification of complex ICU rhythms such as atrial flutter and pacemaker rhythms.
Uneven and small subgroup sizes for demographic analyses.
Temporal uncertainty between documented and true time of rhythm assessment in weakly labeled data
Dataset imbalance with respect to AF and SR labels

Ask about this paper

Answers are grounded in this paper and cite their sources.

Implication

Off-the-shelf pre-trained ECG foundation models do not generalize well to continuous ICU monitoring, but targeted fine-tuning on high-quality data significantly improves their performance for detecting atrial fibrillation and sinus rhythm.

Abstract

Background Recent advances in deep learning have led to the development of ECG foundation models (ECG-FMs) trained with self-supervised learning, which can extract generalizable representations from large-scale data. In this study, we evaluated the performance of these ECG-FMs in detecting atrial fibrillation and sinus rhythm during continuous monitoring of intensive care unit (ICU) patients, using data from Amsterdam University Medical Center (Amsterdam UMC). Our three-stage study: (i) used within-dataset classification performance on public PhysioNet datasets as an indirect proxy for label consistency, (ii) tested off-the-shelf fine-tuned ECG-FMs on ICU data, and (iii) fine-tuned ECG-FM on curated public datasets, with and without weakly labelled in-house ICU data. Results Off-the-shelf model showed limited transferability (F1 = 0.68 for sinus rhythm (SR), 0.52 for atrial fibrillation (AF)). Fine-tuning on consistent public datasets achieved robust performance (F1 = 0.91 for SR, 0.82 for AF), while adding weakly labelled ICU data further improved recall at some cost to precision. Performance remained stable across varying ECG lead configurations, supporting adaptability to heterogeneous inputs. Conclusions These findings indicate that off-the-shelf pre-trained ECG-FMs do not generalise well to continuous monitoring in ICU populations; however, after targeted fine-tuning, they can achieve strong performance even in unseen cohorts, underscoring the importance of high-quality fine-tuning data and rigorous local validation.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Galanty et al. (2026) studied Atrial fibrillation and sinus rhythm (n=226). Fine-tuned ECG foundation models vs. Off-the-shelf pre-trained ECG foundation models was evaluated on F1 score for atrial fibrillation detection. Targeted fine-tuning of ECG foundation models improved atrial fibrillation detection in ICU patients from an F1 score of 0.52 with off-the-shelf models to 0.83.

synapsesocial.com/papers/6a914622d15324a1df3a9883 https://doi.org/10.1186/s40635-026-00949-7

View Full Paper

Ask AI

Bookmark

Share

Leveraging ECG foundation models in critical care for sinus rhythm and atrial fibrillation classification

Figures

Key Points

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Leveraging ECG foundation models in critical care for sinus rhythm and atrial fibrillation classification

Figures

Key Points

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading