Foundation models for electrocardiogram interpretation: clinical implications.

Area of Science:

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

Background:

• Existing AI for ECG interpretation often lacks generalizability and relies on supervised learning (SL) with extensive labeled data.
• Self-supervised learning (SSL) offers a potential solution by learning from unlabeled data, overcoming limitations of traditional methods.
• This study addresses the need for adaptable, open-source AI solutions in cardiac diagnostics.

Purpose of the Study:

• To develop and compare two open-source foundational ECG models: DeepECG-SL (supervised) and DeepECG-SSL (self-supervised).
• To evaluate the generalizability, fairness, and performance of SSL versus SL in ECG interpretation across diverse clinical settings.
• To provide accessible AI tools for robust and data-efficient cardiac diagnostics.

Main Methods:

• Trained two models, DeepECG-SL and DeepECG-SSL, on over 1 million ECGs to predict 77 cardiac conditions.
• DeepECG-SSL utilized contrastive learning and masked lead modeling on unlabeled data before fine-tuning.
• Performance was assessed across seven private and four public multilingual healthcare systems, including fairness and privacy evaluations.

Main Results:

• Both models achieved high performance (AUROCs ~0.98-0.99) across internal, public, and private external datasets.
• DeepECG-SSL showed superior performance in limited-data tasks like long QT syndrome genotype classification and atrial fibrillation risk prediction.
• Fairness analyses indicated minimal disparities across age and sex groups for both models.

Conclusions:

• Self-supervised learning (SSL) is a promising paradigm for ECG analysis, enhancing accessibility, generalizability, and fairness in AI-driven cardiac diagnostics.
• The open-source release of models, tools, and code aims to foster robust, data-efficient AI diagnostics.
• SSL models are particularly beneficial in clinical settings with limited annotated data.