PO-04-151 A NEW SEQUENTIAL ARTIFICIAL INTELLIGENCE ALGORITHM FOR PREDICTING ATRIAL FIBRILLATION USING SERIAL 12-LEAD ELECTROCARDIOGRAMS

Abstract

Because the early stage of atrial fibrillation (AF) is difficult to detect, a significant number of patients are exposed to the risk of stroke without forewarning. Artificial intelligence (AI) algorithms using 12-lead electrocardiograms (ECGs) are being developed to detect AF early. Unlike previous studies in which a single normal sinus rhythm (NSR) ECG was used, we hypothesized that the AI model could predict new-onset AF more accurately by learning the change between two NSR ECGs in an individual. In the present study, standard 12-lead ECGs of patients labeled by cardiologists from January 2010, to December 2021, at Samsung Medical Center were used for AI model development. All patients included in the study were classified into an NSR or AF group based on AF diagnosis. Two AI models (single ECG- and serial ECG-based) were developed using a light gradient boosting machine learning algorithm. The performance of AI models was evaluated based on the area under the receiver operating characteristic curve (AUROC) and precision-recall curve with sensitivity, specificity, accuracy, and F1 score. The AUROC results of the single ECG and serial ECG models were compared to evaluate the hypothesis. We trained 102,018 ECGs from 54,128 patients for the single ECG model and 126,106 ECGs from 106,722 patients for the serial ECG model. When testing the two AI models using test data sets, the performance for predicting new-onset AF was significantly higher with the serial AI model compared with the single AI model (single vs. serial AI model; sensitivity 0.68 vs. 0.86; specificity 0.95 vs. 0.98; accuracy 0.89 vs. 0.9; F1 score 0.74 vs. 0.90; AUROC 0.92 vs. 0.95; p<0.001). In addition, the contribution of various ECG parameters was scored using AI model developed with F-score ANOVA. In the serial ECG model, difference of P-wave amplitude between two ECGs was the leading variable for predicting AF. Comparing serial ECGs rather than a single ECG in an individual could predict future AF more accurately.Tabled 1Table 1. Baseline characteristicsSingle ECG modelOverall (n = 127,036)NSR group (n = 97,514)AF group (n = 29,522)p-valueAge, years64.3 ± 13.762.8 ± 13.570.1 ± 12.8< 0.001Male, n (%)62,136 (53.7)46,370 (50.7)15,766 (65.3)< 0.001Number of ECGs per patient1.9 ± 1.91.7 ± 1.43.3 ± 3.7< 0.001Serial ECG modelOverall (n = 157,440)NSR group (n = 129,721)AF group (n = 27,719)p-valueAge, years69.2 ± 24.367.9 ± 24.676.3 ± 20.8< 0.001Male, n (%)70,008 (50.6)56,448 (48.2)13,560 (63.4)< 0.001Number of ECGs per patient1.2 ± 1.517.5 ± 6.7< 0.001ECG=electrocardiography, NSR=normal sinus rhythm, AF=atrial fibrillation Open table in a new tab