Difference of Precision in Atrial Fibrillation Prediction by Convolutional Neural Network Model Using Sinus Rhythm Electrocardiogram According to Risk Profiles

Background: Artificial intelligence–enhanced electrocardiograms (AI-ECGs) can identify atrial fibrillation (AF) from sinus-rhythm ECGs, but a positive result is not always immediately confirmed. We evaluated AI-ECG performance in subgroups with different pretest probabilities of AF. Methods and Results: A convolutional neural network algorithm was applied to sinus-rhythm ECGs from 14,775 patients without prevalent (previously documented) AF at the index visit. To generate subgroups with different pretest probabilities, a study-specific clinical stratification index, the clinical diagnostic probability (CDP), was constructed from 22 baseline clinical variables; CDP was used solely for stratification, not as a prediction score. Patients were classified as low (CDP <0.1), middle (0.1–0.3), or high (≥0.3) CDP (baseline AF prevalence 3.5%, 11.3%, 16.5%), and by AI-ECG-derived diagnostic probability of AF (AIECG-AF-DP ≥0.8 or <0.8). The 3-year incidence of newly detected AF for high/low AIECG-AF-DP was 11.8/1.6, 20.5/2.6, and 30.4/4.6% per year in the low-, middle-, and high-CDP groups (all P<0.001). The area under the receiver operating characteristic curve was similar across CDP categories (0.783, 0.785, 0.731), whereas the area under the precision–recall curve increased (0.100, 0.158, 0.255) in parallel with no-skill baseline precision (0.022, 0.040, 0.108). Conclusions: AI-ECG discrimination was consistent across pretest probability strata, whereas its precision increased with background AF prevalence. AF was most frequently detected in patients with both a high AIECG-AF-DP and a high pretest probability, in whom repeated rhythm monitoring may have the highest yield.