Diagnosis of cardiac conditions from 12-lead electrocardiogram through natural language supervision

Current AI approaches for cardiac diagnosis require condition-specific supervised learning with extensive labeled datasets, leading to fundamental scalability barriers. We developed an ECG-CLIP model, applying contrastive multimodal learning to enable zero-shot cardiac diagnosis from 12-lead ECGs using natural language supervision. Trained on 800,034 ECG-text pairs from MIMIC-IV-ECG, ECG-CLIP evaluated 18 cardiac conditions without condition-specific training. The model achieved superior performance for rhythm abnormalities (AUROC > 0.90) compared to morphological conditions. External validation demonstrated robust AUROC rank consistency (ρ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{\rho }}$$\end{document} = 0.934), including remarkable zero-shot performance for pediatric patients despite no pediatric training cases. Direct comparison showed ECG-CLIP approached supervised models while providing broader diagnostic coverage. Demographic analysis revealed U-shaped age-dependent performance and condition-specific sex-age patterns. By eliminating dependence on labeled data, ECG-CLIP enables diagnosis of various cardiac conditions via text-based queries. This paradigm shift from rigid task-specific models to flexible unified systems addresses critical deployment barriers, potentially expanding global access to expert-level ECG interpretation.