Transfer learning for predicting acute myocardial infarction using electrocardiograms.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Accurate and rapid identification of acute myocardial infarction (AMI) is crucial in emergency settings.
• Electrocardiograms (ECGs) are vital for AMI detection but manual interpretation is challenging.
• Machine learning for ECG analysis requires extensive, high-quality labeled data, which is often scarce.

Purpose of the Study:

• To investigate the effectiveness of transfer learning in improving machine learning models for AMI prediction from ECGs.
• To assess the impact of pre-training on non-chest-pain ECG data for subsequent AMI detection.
• To compare transfer learning models against traditional models trained without pre-training.

Main Methods:

• Utilized a large dataset of 840,000 ECGs from non-chest-pain patients for model pre-training (sex and age classification).
• Fine-tuned the pre-trained models using a dataset of 44,000 ECGs from chest-pain patients for AMI prediction.
• Evaluated performance across various state-of-the-art ResNet architectures and data sizes, comparing with non-transfer learning approaches.

Main Results:

• Transfer learning demonstrated substantial improvements in AMI prediction accuracy.
• The best performing model achieved an AUC increase from 0.79 to 0.85.
• Improvements were consistent across different ResNet architectures and dataset scales.

Conclusions:

• Simple transfer learning from non-chest-pain ECG data significantly enhances AMI prediction models.
• This approach effectively mitigates the data scarcity issue in developing accurate ECG-based diagnostic tools.
• Transfer learning offers a promising strategy for improving emergency cardiac diagnostics.