An ECG foundation model for generalizable cardiac function prediction across the lifespan

Area of Science:

• Cardiology and Artificial Intelligence
• Development of AI-ECG foundation models
• Paediatric and adult cardiac function assessment

Background:

• Artificial intelligence-enhanced electrocardiography (AI-ECG) offers scalable cardiac screening but faces challenges in paediatric generalizability due to data limitations and model derivation primarily from adult cohorts.
• Paediatric cardiac morphology and function exhibit significant variability, presenting an opportunity for developing more generalizable AI-ECG models.

Purpose of the Study:

• To develop and evaluate ECG-Fyler, an AI-ECG foundation model pretrained on a predominantly paediatric, all-age cohort.
• To assess the model's performance in predicting cardiac dysfunction and dilation using echocardiography and cardiac magnetic resonance data.
• To validate the model's generalizability across different age groups, institutions, and limited-data fine-tuning scenarios.

Main Methods:

• Pretrained ECG-Fyler on a large, predominantly paediatric cohort (1992-2023) at Boston Children's Hospital, using cardiologyspecific Fyler code annotations.
• Evaluated performance on internal paediatric cohorts using echocardiography (echo) and cardiac magnetic resonance (CMR) data, benchmarking against existing AI-ECG foundation models.
• Validated externally on an adult cohort from Columbia University Irving Medical Center, assessing performance across age groups, lesion types, and limited-data scenarios.

Main Results:

• ECG-Fyler demonstrated improved AUROC for biventricular dysfunction and dilation tasks, particularly in low-data settings.
• Internal validation showed ECG-Fyler detecting low left ventricular ejection fraction (LVEF ≤ 40%) with high accuracy (AUROC: 0.80) using only 100 fine-tuning samples, significantly outperforming other models.
• External validation on adults achieved an AUROC of 0.83 for LVEF ≤ 40%, and after fine-tuning on <10% of external data, outperformed a fully trained site-specific model.

Conclusions:

• Pretraining AI-ECG models on richly annotated, paediatric-dominant datasets enables efficient transfer learning across institutions and age groups.
• ECG-Fyler supports scalable AI-ECG screening and triage, especially in resource-limited settings where labels or imaging access are scarce.
• Clinically grounded supervised pretraining on a lifespan ECG corpus can generalize predictions of cardiac dysfunction across diverse populations, potentially improving screening and monitoring.