Artificial Intelligence-Based Electrocardiogram Model as a Predictor of Postoperative Atrial Fibrillation Following Cardiac Surgery: Retrospective Cohort Study

Abstract

Background: Postoperative atrial fibrillation (AF) after cardiac surgery is common and is associated with substantial clinical and economic repercussions. However, existing strategies for preventing postoperative AF remain suboptimal, limiting proactive management. Advances in artificial intelligence (AI) may improve the prediction of postoperative AF. Studies have shown that deep learning applied to electrocardiograms (ECGs) can detect subtle patterns in non-AF ECGs associated with a history of (or impending) AF (referred to as the AI-ECG-AF model). As a noninvasive test routinely performed throughout the perioperative period, the ECG presents a unique opportunity for additional risk stratification. Objective: We aimed to determine whether the AI-ECG-AF model can serve as an independent risk factor for postoperative AF after cardiac surgery, compare its predictive performance with existing postoperative AF prediction tools, and assess its additive value. Methods: This single-center retrospective cohort study included 2266 patients (5402 standard 12-lead ECGs) who underwent cardiac surgery at a tertiary hospital in South Korea between December 2018 and December 2023. The AI-ECG-AF model was trained on 4.05 million non-AF standard 12-lead ECGs (1.13 million patients) using a 1D EfficientNet-B0 architecture and achieved an area under the receiver operating characteristic curve (AUROC) of 0.901 (95% CI 0.900-0.902) in its held-out test set. Postoperative AF was defined as AF documented by ECG within 30 days after surgery. Using multivariable logistic regression, we assessed the association between the AI-ECG-AF model score and postoperative AF, adjusting for conventional clinical variables. We also investigated the additive or synergistic predictive value of the AI-ECG-AF model score when combined with an existing postoperative AF tool (the postoperative atrial fibrillation score) or other risk factors, based on the AUROC. Results: After adjusting for other clinical variables, a 10% absolute increase in the AI-ECG-AF model score was associated with a 1.197- to 1.209-fold increase in the odds of developing postoperative AF. The AI-ECG-AF model score significantly enhanced postoperative AF prediction: the AUROC of the existing postoperative atrial fibrillation score was 0.643; adding the AI-ECG-AF model score increased it to 0.680 (P<.001), and combining the AI-ECG-AF model score with other risk factors raised it to 0.710 (P<.001). Conclusions: The AI-ECG-AF model serves as a novel, robust, and independent risk factor for postoperative AF following cardiac surgery and provides additive or synergistic predictive value when integrated with existing postoperative AF prediction tools or other risk factors. By capturing atrial electrophysiological vulnerability not reflected in conventional clinical scores, the AI-ECG-AF model may function as a noninvasive biomarker for preoperative risk stratification for postoperative AF prediction in cardiac surgery patients, potentially enabling targeted prophylaxis and closer monitoring during the perioperative period.