ECG-FM: an open electrocardiogram foundation model.

Area of Science:

• Artificial Intelligence
• Cardiology
• Biomedical Informatics

Background:

• Electrocardiogram (ECG) analysis is crucial for diagnosing cardiac conditions.
• Current deep learning models often require large labeled datasets, limiting their application.
• Developing generalizable and label-efficient ECG analysis tools is a significant challenge.

Purpose of the Study:

• To develop and evaluate ECG-FM, an open-weight foundation model for ECG analysis.
• To assess its performance on clinically relevant tasks like multi-label interpretation and LVEF prediction.
• To release ECG-FM and a public benchmark to foster transparent research.

Main Methods:

• A transformer-based foundation model (ECG-FM) was pretrained using hybrid self-supervision on 1.5 million 12-lead ECGs.
• ECG-specific augmentations, masked reconstruction, and contrastive learning were employed.
• Performance was evaluated on multi-label ECG interpretation and LVEF prediction using the MIMIC-IV-ECG dataset.

Main Results:

• Finetuned ECG-FM models outperformed baselines, especially in low-data regimes, showing strong label efficiency and generalizability.
• Achieved high AUROC for atrial fibrillation (0.996) and LVEF < 40% (0.929).
• The pretrained encoder demonstrated discriminative embeddings and competitive linear probing performance.

Conclusions:

• ECG-FM is a generalizable, label-efficient, and discriminative model for ECG screening, risk stratification, and monitoring.
• Its representations capture both morphological and semantic cardiac information.
• The open release of ECG-FM, code, and benchmark lowers barriers for reproducible and comparable ECG research.