Biometric contrastive learning for data-efficient deep learning from electrocardiographic images.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Traditional supervised learning for AI in ECG analysis requires extensive labeled data.
• This limitation hinders the development of efficient AI models for detecting heart diseases from electrocardiograms (ECGs).

Purpose of the Study:

• To introduce Biometric Contrastive Learning (BCL), a self-supervised pretraining method for label-efficient deep learning on ECG images.
• To evaluate BCL's performance in detecting atrial fibrillation (AF), gender, and reduced left ventricular ejection fraction (LVEF < 40%) using ECG data.

Main Methods:

• Trained a convolutional neural network using ECG pairs from 78,288 individuals to identify variations within the same patient.
• Fine-tuned BCL-pretrained models for specific diagnostic tasks and compared performance against ImageNet initialization and simCLR.
• Externally validated models on independent cohorts from Germany and the United States.

Main Results:

• BCL demonstrated superior performance with limited labeled data, achieving equivalent results with 50% less data compared to other methods.
• With only 0.1% labeled data, BCL reached an AUROC of 0.88/0.79/0.75 for AF/Gender/LVEF < 40%, significantly outperforming ImageNet and simCLR.
• In external validation, BCL outperformed other methods even with 100% labeled data for Gender and LVEF < 40% detection.

Conclusions:

• A pretraining strategy leveraging biometric signatures from same-patient ECGs enhances AI model development efficiency.
• BCL represents a significant advancement for detecting disorders from ECG images, especially when labeled data is scarce.