Yale School of Medicine (CarDS Lab)
Training code and model weights not public. Contact creators for more information.
Foundation model that encodes single-lead (lead I) ECGs with information from paired transthoracic echocardiography reports, aimed at label-efficient screening for structural heart disease (SHD) on wearable and portable single-lead ECG devices. A 7-layer 1D-CNN ECG encoder and a RoBERTa-based text encoder are contrastively pretrained (CLIP-style) on 194,551 ECG-echo report pairs from 77,378 adults in the Yale New Haven Health System, then the ECG encoder is fine-tuned on a temporally-distinct cohort to detect reduced LVEF, diastolic dysfunction, and a composite SHD label. Matches a randomly-initialized CNN at full training-data volume but substantially outperforms it in label-scarce regimes (e.g. with only 0.5% of labeled data).
Architecture
Hybrid
7-layer 1D-CNN single-lead ECG encoder (tapering kernel sizes 7->3, doubling filters 16/32/64) contrastively pretrained CLIP-style against a 6-layer, 12-head RoBERTa text encoder over echocardiography reports
Framework
PyTorch
Added to catalog
2026-08-10
Yale New Haven Health System 1-lead ECG-TTE Paired Cohort
194,551 lead-I ECG-echocardiography report pairs (pretraining, 2015-2018) plus temporally-distinct fine-tuning/test cohorts for LVEF<=40% (250,260 ECGs / 95,388 individuals), LV diastolic dysfunction (123,306 ECGs / 59,831 individuals) and composite structural heart disease (132,310 ECGs / 58,815 individuals), 2019-2023, from five hospitals in the Yale New Haven Health System.
LV systolic dysfunction (LVEF<=40%) detection from single-lead ECG
LV diastolic dysfunction detection from single-lead ECG
Composite structural heart disease detection (LVSD, moderate/severe valvular disease, and/or LV hypertrophy) from single-lead ECG