A foundational vision transformer improves diagnostic performance for electrocardiograms.

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Machine Learning for Healthcare

Background:

• Electrocardiogram (ECG) analysis is crucial for diagnosing cardiac conditions.
• Current deep learning models like Convolutional Neural Networks (CNNs) require large datasets for effective ECG analysis.
• Transfer learning using natural image pre-training may not yield optimal results for biomedical data.

Purpose of the Study:

• To develop and evaluate HeartBEiT, a vision-based transformer model for ECG waveform analysis using masked image modeling.
• To compare HeartBEiT's performance against standard CNN architectures across varying training sample sizes.
• To assess the model's ability to improve diagnostic explainability by identifying relevant ECG regions.

Main Methods:

• Developed HeartBEiT, a transformer model utilizing masked image modeling for ECG analysis.
• Pre-trained HeartBEiT on a large dataset of 8.5 million ECGs.
• Compared HeartBEiT with CNNs on diagnosing hypertrophic cardiomyopathy, low left ventricular ejection fraction, and ST elevation myocardial infarction using independent validation datasets and varying sample sizes.

Main Results:

• HeartBEiT demonstrated significantly higher diagnostic performance compared to CNNs, particularly at lower training sample sizes.
• The model improved the explainability of diagnoses by highlighting biologically relevant regions within the ECG.
• Domain-specific pre-training for transformer models outperformed natural image pre-training in low-data scenarios.

Conclusions:

• Domain-specific pre-trained transformer models, like HeartBEiT, show superior performance over traditional CNNs for ECG analysis, especially in data-limited settings.
• HeartBEiT offers enhanced accuracy and granular explainability for cardiac diagnoses.
• This approach represents a significant advancement in applying AI to complex biomedical waveform analysis.