Deep learning enables diagnosis of atrial cardiomyopathy from routine 12-lead electrocardiogram

Abstract Background and Aims Atrial cardiomyopathy (AtCM) contributes to the development of atrial fibrillation (AF), heart failure (HF) and stroke. Imaging-derived measures of left atrial (LA) structure and function are used to diagnose AtCM. Considering the tight coupling of heart structure and rhythm generation, this information might also be derived from 12-lead electrocardiogram (ECG), which is low-cost and readily available. Methods We finetuned a deep learning (DL) ECG foundational model to predict LA imaging indices based on 26’134 ECGs from the UK Biobank cohort. We then investigated if the ECG-predicted imaging features improved risk stratification of AF beyond the CHARGE-AF Score on a test set from the UK Biobank. We repeated this analysis for diagnosis of HF. We then externally validated our model and applied it to a cohort of stroke patients. Results Our DL model successfully predicted LA imaging indices from 12 lead ECG. In the UK Biobank test set, the ECG-predicted LA imaging features significantly improved risk stratification for AF beyond the CHARGE-AF score. ECG-predicted imaging markers showed superior test performance compared to established ECG markers of AtCM and an alternative DL approach This also held on external validation sets. Importantly, the predicted imaging features improved risk stratification for HF in the UK Biobank, even when excluding patients with AF, suggesting that our model captures AtCM beyond AF. Discussion We established a novel DL approach for the diagnosis of AtCM from 12 lead ECG. Due to the wide availability of ECG, our approach has the potential to improve screening and diagnosis of AtCM. Structured Graphical Abstract Key Question Can left atrial imaging markers of atrial cardiomyopathy predicted from 12 lead electrocardiogram (ECG) using a deep learning (DL) model improve risk stratification for atrial fibrillation (AF) and heart failure (HF)? Key Finding Our DL model successfully predicted left atrial imaging features from 12 lead ECG. The predicted left atrial imaging markers improved risk stratification for AF and HF outperforming a deep learning model trained to identify patients with AF directly. Take-home message DL allows improved diagnosis of atrial cardiomyopathy from ECG. Structured graphical abstract showing the model training and evaluation procedure. We first trained a deep learning model to predict left atrial imaging markers of atrial cardiomyopathy. We then added the predicted imaging features to clinical risk scores, creating multivariate regression models for risk stratification of different clinical outcomes. Finally, we externally validate the deep learning model and derived multivariate regression models on a diverse set of clinical cohorts.