Wearable-Echo-FM: an ECG echo foundation model for 1-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).
• Scaling AI for SHD detection is hindered by the limited availability of diagnostic labels for 1-lead ECGs, despite their ubiquity in wearable devices.
• Existing models require extensive labeled data, posing a challenge for developing effective screening tools for portable and wearable applications.

Purpose of the Study:

• To develop a foundation model, Wearable-Echo-FM, capable of encoding 1-lead ECGs with information from echocardiographic text reports.
• To reduce the data labeling requirements for developing SHD screening models using 1-lead ECGs.
• To establish a foundation for future validation and implementation of AI-based SHD detection on wearable and portable devices.

Main Methods:

• Contrastive pre-training of ECG convolutional neural network (CNN) and RoBERTa text encoders using 194,551 1-lead ECG-echo pairs from 77,378 adults.
• Fine-tuning the ECG encoder on progressively larger ECG datasets to detect left-ventricular systolic dysfunction (LVSD), diastolic dysfunction, and a composite SHD.
• Evaluating Wearable-Echo-FM against a randomly initialized CNN on an independent held-out test set.

Main Results:

• Wearable-Echo-FM demonstrated comparable performance to a baseline CNN when using the full training set for detecting LVSD (AUROC 0.894 vs. 0.884), diastolic dysfunction (0.849 vs. 0.843), and composite SHD (0.887 vs. 0.869).
• With only 0.5% of the data (approximately 1000 ECGs), Wearable-Echo-FM significantly outperformed the baseline CNN in detecting LVSD (0.855 vs. 0.548), diastolic dysfunction (0.819 vs. 0.582), and composite SHD (0.863 vs. 0.496).
• The results indicate a substantial improvement in label efficiency, with the model achieving strong performance using minimal labeled data.

Conclusions:

• Contrastive pre-training of 1-lead ECGs with echocardiographic text effectively reduces the need for extensive labeled data in SHD screening model development.
• Wearable-Echo-FM provides a robust foundation for label-efficient development of AI models for SHD detection on 1-lead ECGs.
• This approach facilitates the future validation and deployment of these models on widely available wearable and portable devices for broader cardiovascular health monitoring.