Abstract 12575: Subgroup Comparison of Deep-Learning Electrocardiogram Model Performance for Predicting New-Onset 1-Year Atrial Fibrillation Incidence

Abstract

Background: Surveillance of high-risk individuals as determined by an atrial fibrillation (AF) prediction model would be a good strategy to prevent stroke and other complications from AF. Recently, several deep learning (DL) models to predict new onset AF using 12-lead electrocardiography (ECG) have been introduced. However, their performance has not been compared among different patient groups. Methods : Data of 99,281 patients who underwent 12-lead ECG at Korea University Anam Hospital from January 2017 to December 2019, excluding patients already diagnosed with AF, were studied. The data was split into a 60% training set, 20% validation set and 20% test set. Using age, sex, and raw ECG waveforms as inputs, DL models of ResNet and a multi-layered perceptron module were developed to predict 1-year new onset AF incidence. Results: The DL model using both ECG waves and simple demographic information (age and sex) as input (ECG-AS model) showed better performance compared to the DL model using only ECG waves (ECG model) for AF prediction in the test dataset (area under the receiver operating characteristics curves [AUROCs] 0.79 for ECG model and 0.89 for ECG-AS model, respectively). In subgroup analysis divided by clinical features such as CHA2DS2-VASC score, there was little difference in the performance of the DL model. Interestingly, The DL model performed worse in patients with normal sinus rhythm (NSR) compared to patients with non-NSR (AUROC 0.70 vs 0.81 for ECG model, 0.82 vs 0.90 for ECG-AS model, Table 1). Conclusions: ECG of NSR may not have sufficient information to predict AF. In applying the AF prediction DL model using 12-lead ECG to real-world clinical practice, it may be helpful to add clinical information to the input along with appropriate patient selection.