A foundation transformer model with self-supervised learning for ECG-based assessment of cardiac and coronary...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Limited labeled datasets hinder AI development for critical cardiac conditions like myocardial ischemia and coronary microvascular dysfunction.
• Existing AI models for electrocardiogram (ECG) diagnosis excel with abundant labeled data but struggle with specialized applications.
• A novel self-supervised ECG foundation model is proposed to address data scarcity challenges.

Purpose of the Study:

• To develop and evaluate a self-supervised learning (SSL) foundation model for ECG analysis.
• To improve diagnostic accuracy and generalizability for complex cardiac conditions using limited labeled data.
• To demonstrate the model's effectiveness in tasks requiring scarce, high-value labels.

Main Methods:

• A modified vision transformer was pretrained on a large unlabeled ECG dataset (MIMIC-IV-ECG, N=800,035).
• The model was fine-tuned using smaller, high-quality labeled datasets from positron emission tomography (PET) and clinical reports for 12 prediction tasks.
• Performance was validated across diverse external cohorts and cross-modality imaging databases (MRI, SPECT).

Main Results:

• The SSL foundation model significantly improved diagnostic accuracy in 11 of 12 prediction tasks compared to traditional supervised methods.
• Area under the receiver operating characteristic curve (AUROC) ranged from 0.763 (impaired myocardial flow reserve) to 0.955 (impaired left ventricular ejection fraction).
• The model demonstrated strong generalizability across external datasets and cross-modality evaluations, with AUROCs from 0.771 to 0.949.

Conclusions:

• Self-supervised learning pretraining enhances the diagnostic accuracy and generalizability of ECG foundation models.
• This approach effectively enables AI development for complex cardiac conditions where labeled data is scarce and costly.
• The versatile ECG foundation model supports AI-driven diagnosis for critical applications like myocardial ischemia and coronary microvascular dysfunction.