Artificial intelligence enabled ECG analysis for the detection of structural and electrophysiological left atrial remodeling

Atrial myopathy is central to the disease progression of atrial fibrillation (AF) and its structural and electrophysiological aspects are relevant to patient outcome. The aim of this study was to develop artificial intelligence-based models to predict left atrial enlargement (LAE) and left atrial low-voltage areas (LVA) from 12-lead electrocardiograms (ECG).

We developed unique prediction models for both endpoints using a cross-sectional, retrospective, single-center clinical routine database. ECGs from adult patients acquired between January 2016 and February 2023 were paired with corresponding data from cardiac imaging and electroanatomic mapping. Overall, 66,228 (LAE) and 6,955 (LVA) ECGs with paired clinical data were analyzed. The area under the receiver operating characteristic curve (AUROC) was 0.779 (95% confidence interval [CI] 0.771-0.787) for the prediction of any LAE, and 0.799 (95%CI 0.801-0.822) for the identification of severe LAE. AUROC for the model predicting LVA was 0.784 (95%CI 0.760-0.807), and P-waves were areas of particular interest to both models based on saliency maps. Factors associated with poorer model performance included tachycardia and presence of AF. Transfer learning and implementing model output alongside additional clinical variables in multivariate models did not lead to relevant improvements in predictive power.

In this study, we present well-performing artificial intelligence-based algorithms for predicting structural and electrophysiological aspects of atrial myopathy from 12-lead ECGs obtained from a single patient cohort. After external validation, these models may contribute to individualized therapy decisions in patients with atrial myopathy and atrial arrhythmias.

Graphical abstract