An Electrocardiogram Foundation Model Built on over 10 Million Recordings

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Artificial intelligence (AI) shows promise in electrocardiogram (ECG) analysis for cardiovascular disease.
• Foundation models enhance medical AI, improving diagnosis and knowledge transfer.
• Challenges in ECG AI include limited data and poor generalization, especially for single-lead ECGs.

Purpose of the Study:

• To develop a general-purpose ECG foundation model (ECGFounder) for comprehensive cardiovascular disease diagnosis.
• To create a model that is effective out-of-the-box and easily fine-tunable for various downstream tasks.
• To extend ECG analysis capabilities to reduced-lead ECGs, including single-lead ECGs for mobile and remote monitoring.

Main Methods:

• Developed ECGFounder using 10,771,552 ECGs from 1,818,247 subjects with 150 diagnostic labels.
• Leveraged real-world ECG annotations from cardiologists for broad diagnostic capabilities.
• Extended model applicability to single-lead ECGs for diverse applications.

Main Results:

• ECGFounder achieved expert-level performance (AUROC > 0.95 for 80 diagnoses) on internal validation.
• Demonstrated strong classification and generalization across diagnoses on external validation sets.
• Outperformed baseline models by 3-5 AUROC points in demographic analysis, event detection, and cross-modality diagnosis after fine-tuning.

Conclusions:

• The ECG foundation model effectively generalizes across tasks, enhancing cardiovascular diagnostics.
• Facilitates integration with cloud systems for wearable ECG data analysis.
• Significantly advances AI in cardiology and aids in cardiac condition management.