Abstract 4368921: Artificial Intelligence Prediction of Atrial Fibrillation Using Pre-Cannulation Electrocardiograms in Patients Undergoing Veno-Venous Extracorporeal Membrane Oxygenation

Abstract

Backround: Atrial fibrillation (AF) is a common complication following veno-venous extracorporeal membrane oxygenation (VV-ECMO), leading to further complications from anticoagulation or stroke. The Artificial Intelligence AF Prediction Dashboard developed at Mayo Clinic generates two predictive outputs: (1) a continuous probability score and (2) a binary threshold-based signal to predict the likelihood of AF at an individual level after 12-lead electrocardiogram (ECG) is obtained. Using both continuous risk (%) and binary (yes/no) likelihood, we present a novel method to understand AF risk after VV-ECMO cannulation based on pre-ECMO 12-lead ECG. Methods: We reviewed a total of 147 patients on ECMO between 2018 and 2024, from which 85 VV-ECMO patients with complete data were included. AI Dashboard outputs were quantified as binary (1/0) or percent likelihood for AF. R was utilized for statistical analysis. Results: 85 patients met our inclusion criteria from which AF after cannulation occurred in 31 (36%). 17 patients had a history of AF prior to VV-ECMO initiation. The mean age of our cohort was 49 years with a mean BMI of 29.3. Baseline AI Dashboard continuous assessment identified 39 patients (46%) compared to the binary assessment identifying 31 patients (36%). Logistic regression in the binary assessment demonstrated an odds ratio of 1.08 (95% CI 1.04 – 1.12) with a p value of <0.001. The binary AF Prediction yielded a 58% sensitivity, 76% specificity, 58% PPV, and 76% NPV, with an accuracy of 67%. The continuous AF Probability score achieved 72% balanced accuracy with sensitivity 74%, specificity 70%, PPV 59% and NPV 83%, indicating fair overall performance and suggesting good utility for ruling out AF in low-risk patients. Although both outputs provided moderate predictive accuracy, the continuous score reported before ECMO cannulation demonstrated stronger overall performance with fewer false negatives and higher rule-out value. Conclusions: We demonstrate the novel use of an AI predictive model in the VV-ECMO population for the prediction of AF risk with both continuous and binary risk cutoffs from baseline ECG. This data highlights the need for further development in ECMO-specific AI models to help mitigate risk after the onset of atrial arrhythmias. As AI-driven tools become increasingly integrated into care, their performance must be continuously validated and recalibrated for specific populations to ensure their safe and effective use.