Foundation models for generalizable electrocardiogram interpretation: comparison of supervised and self-supervised...

Area of Science:

• Artificial Intelligence in Medicine
• Cardiovascular Diagnostics
• Machine Learning for Healthcare

Background:

• Existing artificial intelligence (AI) solutions for 12-lead electrocardiogram (ECG) interpretation often lack generalizability and are closed-source.
• Supervised learning (SL) limitations hinder adaptability of AI ECG interpretation across diverse clinical settings.
• Open-source foundational models are needed to address challenges in current AI ECG diagnostics.

Purpose of the Study:

• To develop and compare two open-source foundational ECG models: DeepECG-SSL (self-supervised learning) and DeepECG-SL (supervised learning).
• To evaluate the generalizability, fairness, and performance of these models across diverse clinical datasets.
• To establish self-supervised learning as a viable paradigm for ECG analysis, particularly in data-limited scenarios.

Main Methods:

• Trained two models (DeepECG-SSL and DeepECG-SL) on over 1 million ECGs with standardized preprocessing and automated report analysis.
• Pretrained DeepECG-SSL using self-supervised contrastive learning and masked lead modeling.
• Evaluated models on six multilingual private healthcare systems and four public datasets, assessing performance across 77 cardiac conditions and conducting fairness analyses for age and sex.

Main Results:

• Both models achieved high Area Under the Receiver Operating Characteristic Curve (AUROC) scores across internal, external public, and external private datasets (e.g., DeepECG-SSL internal AUROC: 0.990).
• Fairness analyses revealed minimal performance disparities across age and sex groups (true positive rate & false positive rate difference < 0.010).
• DeepECG-SSL demonstrated superior performance in digital biomarker prediction with limited labeled data, including 5-year atrial fibrillation risk (AUROC 0.742 vs. 0.720) and Long QT syndrome classification (AUROC 0.931 vs. 0.853).

Conclusions:

• Open-sourcing model weights, preprocessing tools, and validation code supports robust, data-efficient AI diagnostics.
• Self-supervised learning (SSL) is a promising paradigm for ECG analysis, enhancing accessibility, generalizability, and fairness.
• This work facilitates broader adoption of AI in cardiac diagnostics, especially in resource-limited clinical environments.