Language-aligned ECG foundation modeling for data-efficient detection of cardiovascular diseases

Introduction Cardiovascular diseases (CVDs) are the leading cause of morbidity and mortality worldwide. CVD assessment often includes a 12-lead electrocardiogram (ECG), with over 300 million performed annually. Automated ECG interpretation tools have improved substantially, but ECG’s full diagnostic and predictive potential is limited due to an overreliance on task-specific labels in traditional supervised learning. To improve the generalizability and clinical applicability of ECG analysis, we propose a language-aligned ECG foundation model (FM) that leverages paired ECG waveforms and cardiologist overread reports for pretraining. Purpose This study evaluates whether language-aligned multimodal pretraining integrating ECG waveforms with cardiologist reports improve rare-disease detection and label efficiency in CVD identification, including acute myocardial infarction (MI), cardiac amyloidosis (CA), and hypertrophic cardiomyopathy (HCM). Methods We leveraged a retrospective dataset of over 1.7 million ECGs and reports from 542,288 patients to develop an ECG foundation model. Unlike existing approaches, this model integrates ECG data with textual annotations using contrastive learning. During pretraining, cardiologist reports provide natural language supervision to align ECG signal and text, enabling clinically meaningful representations (Fig. 1). Using an independent, external dataset of 800,000 ECGs (without annotations), we finetuned and evaluated our ECG FM for detection of acute MI, CA, and HCM, and compared it with a standard deep learning model (ResNet), a non-ECG foundation model (OpenAI CLIP) and a baseline model that uses age and gender. We assessed model performance and label efficiency, defined as the ability to achieve high predictive performance with less labelled training data. Results Our ECG FM achieved superior performance across all detection tasks against comparison models. For acute MI detection, the area under the curve (AUC) is 0.966 (95% CI=0.965-0.966), significantly outperforming OpenAI CLIP (AUC=0.951, CI=0.949-0.953, p<.001) and ResNet (AUC=0.964, CI=0.963-0.964, p=.012). For cardiomyopathies with low prevalence (<0.4%), the AUC for our model is 0.844 (CI=0.832-0.856) on CA detection (versus AUC=0.792 for the runner-up model), and 0.849 (CI=0.839-0.859) on HCM detection (AUC= 0.803 for the runner-up model). For all conditions (acute MI, CA, and HCM detection), our ECG FM consistently demonstrated superior label efficiency, achieving equivalent performance to the runner-up model using 83%, 95%, and 94% less data, respectively (Fig. 2). Conclusions Language-aligned pretraining integrating ECG waveforms and cardiologist reports enables learning of clinically meaningful representations rather than isolated signal patterns. This strategy improves detection of common and rare CVDs while markedly reducing labeled data requirements, supporting scalable and data-efficient ECG-based diagnosis.figure 1figure 2