Artificial Intelligence-Enhanced Electrocardiography for Prediction of Occult Atrial Fibrillation in Patients With Stroke Who Undergo Prolonged Cardiac Monitoring.

Abstract

To evaluate the performance of an artificial intelligence (AI)-enhanced electrocardiography (ECG; AI-ECG) algorithm to predict atrial fibrillation (AF) detection on prolonged cardiac monitoring (PCM) after index stroke. This retrospective study included all adult patients with ischemic stroke evaluated at Mayo Clinic with baseline ECG and PCM from January 1, 2018, to December 31, 2020. We recorded demographic characteristics, clinical features, presumed stroke mechanism, PCM duration, and PCM outcome (AF vs no AF) and AF burden. Electrocardiograms were analyzed using the AI-ECG algorithm to determine likelihood of AF capture with PCM. Stroke etiology was adjudicated using TOAST (Trial of ORG 10172 in Acute Stroke Treatment) and embolic stroke of undetermined source (ESUS) definitions. The ability of the AI-ECG algorithm to predict AF detected by PCM was assessed via receiver operating characteristics analysis, calculating the area under the receiver operating characteristic curve (C statistic). Sensitivity and specificity analyses were performed for each tool using optimal cutoffs (using maximum Youden indices). We identified 863 patients for inclusion in the study. The median age was 69 years, 496 (57.5%) were male, 367 (42.5%) were women, and 561 patients (65.0%) were categorized as having ESUS. Prolonged cardiac monitoring detected AF in 85 patients (9.8%). Median duration of PCM was 30 days (IQR, 25 to 30 days). The AI-ECG algorithm identified a notable difference in probability of AF on PCM. For its optimal model output cutoff of 0.24, AI-ECG had a negative predictive value of 94.2% (95% CI, 92.2% to 95.9%) and a specificity of 81.8% (95% CI, 78.9% to 84.4%) for excluding AF on PCM. When evaluating for an AF burden of 6 minutes or longer, the AI-ECG had a negative predictive value of 96.7% (95% CI, 95.5% to 97.6%). There was no significant difference in the area under the receiver operating characteristic curve when comparing the ESUS vs non-ESUS subgroups (P=.42). This study found that AI-ECG may help identify patients unlikely to have AF on PCM and can predict the occurrence of longer episodes of AF. Thus, AI-ECG may be used to stratify which patients should undergo PCM after stroke. Future studies should compare the performance of AI-ECG and PCM for the clinical end point of stroke recurrence.