ECGCLIP: A Foundation Model for Broad Cardiovascular Assessment from ECG
A team of researchers has introduced a novel model named ECGCLIP, which leverages signal-language contrastive learning to connect ECG waveforms with expert assessments. This model was developed using a substantial dataset comprising over 2.83 million ECG studies from more than 1.32 million patients. It underwent thorough evaluation through an internal validation set and nine external datasets, amounting to approximately 1.5 million ECGs. ECGCLIP successfully addressed 89 different tasks, including diagnoses and rare heart conditions, demonstrating superior performance compared to previous benchmarks, particularly excelling in atrial fibrillation and ST-segment elevation metrics.
Key facts
- ECGCLIP is a signal-language contrastive learning framework.
- It aligns ECG waveforms with expert diagnostic reports.
- Pre-trained on 2,837,962 ECG studies from 1,324,856 patients.
- Evaluated on internal test set and nine external cohorts (~1.5 million ECGs).
- Covered 89 downstream tasks: 45 ECG diagnoses, 39 echocardiographic targets, 5 rare cardiac diseases.
- Primary metric: PRAUC.
- Outperformed random initialization and Merl-R18 baselines.
- ECGCLIP-R34 achieved PRAUC 0.900 for atrial fibrillation and 0.870 for ST-segment elevation.
Entities
—