Abstract 14688: Identification of Patients With Atrial Fibrillation Using an Electrocardiogram in Sinus Rhythm: External Reproduction and Validation of a Deep Neural Network Algorithm

Abstract

Listen

Translate article

Introduction: Atrial fibrillation (AF) can occur asymptomatically and can only be diagnosed while the arrhythmia is present. Hence, it is underdetected. An artificial intelligence (AI) algorithm has been developed by Attia et al. to identify patients with AF based on a 12-lead electrocardiogram (ECG) in sinus rhythm (SR). We aim to reproduce these findings in an unrelated patient population and unrelated clinical center using the same deep neural network (DNN) architecture. Methods: All patients older than 18 years with at least one ECG in SR performed at our center between 10/1/2002 and 8/31/2022 were included in the study. Diagnostic labels were given by the GE-Marquette ECG system. For patients without an ECG in AF, all ECGs in SR were included. For patients with one or more ECGs in AF, all ECG in SR starting 31 days before the AF event were included. The ECGs were randomly allocated to a training, internal validation, and testing dataset in a 7:1:2 ratio. A DNN following the architecture of Attia et al. was trained to discriminate between patients with and without AF. Results: The dataset consisted of 494.042 ECGs in SR from 142.310 patients. AF was diagnosed in 8.9%. Testing the model on the first ECG of each patient (main analysis) resulted in a sensitivity of 78.9% (95% CI, 77.6 - 78.5), specificity of 78.0% (95% CI, 77.5 - 78.5), accuracy of 78.1 (95% CI, 77.6 - 78.5), area under the receiver operating curve (AUC) of 0.87 (95% CI, 0.86 - 0.87), and F1 score of 39.3% (95% CI, 38.1 - 40.6). Testing the model on all ECGs in the first 31 days and selecting the maximum probability of AF (secondary analysis) increased the sensitivity to 81.1 (95% CI, 78.4 - 83.7), specificity to 79.8 (95% CI, 79.3 - 80.3), accuracy to 79.8 (95% CI, 79.4 - 80.3), AUC to 0.88 (95% CI, 0.87 - 0.89), but lowered the F1 score to 20.8% (95% CI, 19.51 - 22.0). Conclusion: An AI-enabled ECG algorithm for the identification of patients with AF on a SR-ECG can be reproduced with external data in an external center at clinical grade accuracy.