Wearable-Echo-FM: An ECG-echo foundation model for single lead electrocardiography

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Devices

Background:

• Artificial intelligence (AI) models show promise in detecting structural heart diseases (SHDs) from electrocardiograms (ECGs).
• Widespread adoption of AI for SHD detection is hindered by the limited availability of diagnostic labels for single-lead ECGs from wearable devices.
• Foundation models offer a potential solution for encoding diverse data types, including ECGs and clinical text.

Purpose of the Study:

• To develop and evaluate Wearable-Echo-FM, a foundation model designed to encode single-lead ECGs with information from echocardiographic text reports.
• To assess the model's performance in detecting left-ventricular systolic dysfunction (LVSD), diastolic dysfunction, and a composite SHD.
• To determine if contrastive pre-training on echocardiographic text can reduce the data labeling requirements for SHD screening on wearable ECG devices.

Main Methods:

• Utilized 274,057 single-lead ECG-echocardiogram pairs from 77,378 adults (2015-2019) for contrastive pre-training of convolutional neural network (CNN) and RoBERTa encoders.
• Fine-tuned the ECG encoder on progressively larger ECG datasets (250 to 250,260 ECGs) to detect specific cardiac disorders.
• Compared the performance of the fine-tuned Wearable-Echo-FM model against a randomly initialized CNN on an independent held-out test set.

Main Results:

• With the full training dataset, Wearable-Echo-FM achieved performance comparable to the baseline CNN across all metrics (AUROC for LVSD: 0.894 vs 0.884; diastolic dysfunction: 0.849 vs 0.843; composite SHD: 0.887 vs 0.869).
• Using only 0.5% of the training data (~1000 ECGs), Wearable-Echo-FM significantly outperformed the baseline CNN (AUROC for LVSD: 0.855 vs 0.548; diastolic dysfunction: 0.819 vs 0.582; composite SHD: 0.863 vs 0.496).
• Contrastive pre-training of single-lead ECGs using echocardiographic text effectively reduced the need for extensive labeled data.

Conclusions:

• Contrastive pre-training of single-lead ECGs with echocardiographic text is a viable strategy for developing effective AI models for SHD screening.
• Wearable-Echo-FM demonstrates the potential to significantly lower data annotation costs and accelerate the deployment of AI-powered cardiac diagnostics on wearable and portable devices.
• This approach holds promise for improving the accessibility and scalability of early detection for structural heart diseases.