Identifying patients with atrial fibrillation during sinus rhythm on ECG: Significance of the labeling in the artificial intelligence algorithm

• Published in: International journal of cardiology. Heart & vasculature Vol. 38; p. 100954
• Main Authors: Suzuki, Shinya, Motogi, Jun, Nakai, Hiroshi, Matsuzawa, Wataru, Takayanagi, Tsuneo, Umemoto, Takuya, Hirota, Naomi, Hyodo, Akira, Satoh, Keiichi, Otsuka, Takayuki, Arita, Takuto, Yagi, Naoharu, Yamashita, Takeshi
• Format: Journal Article
• Language: English
• Published: Ireland Elsevier B.V 01.02.2022; Elsevier
• Subjects: Artificial intelligence; Atrial fibrillation; Electrocardiography; Original Paper; Electrocardiography; Artificial intelligence; Atrial fibrillation
• ISSN: 2352-9067; 2352-9067
• DOI: 10.1016/j.ijcha.2022.100954

•High performance of AI algorithm to detect AF using SR-ECG was confirmed in patients without structural heart disease.•The performance of AI-enabled ECG to detect AF was high especially when the algorithm included SR-ECG taken after the index AF-ECG.•A similar tendency was observed when the performance was tested in patients with structural heart diseases. This study aimed to increase the knowledge on how to enhance the performance of artificial intelligence (AI)-enabled electrocardiography (ECG) to detect atrial fibrillation (AF) on sinus rhythm ECG (SR-ECG). It is a retrospective analysis of a single-center, prospective cohort study (Shinken Database). We developed AI-enabled ECG using SR-ECG to predict AF with a convolutional neural network (CNN). Among new patients in our hospital (n = 19,170), 276 AF label (having ECG on AF [AF-ECG] in the ECG database) and 1896 SR label with following three conditions were identified in the derivation dataset: (1) without structural heart disease, (2) in AF label, SR-ECG was taken within 31 days from AF-ECG, and (3) in SR label, follow-up ≥ 1,095 days. Three patterns of AF label were analyzed by timing of SR-ECG to AF-ECG (before/after/before-or-after, CNN algorithm 1 to 3). The outcome measurement was area under the curve (AUC), sensitivity, specificity, accuracy, and F1 score. As an extra-testing dataset, the performance of AI-enabled ECG was tested in patients with structural heart disease. The AUC of AI-enabled ECG with CNN algorithm 1, 2, and 3 in the derivation dataset was 0.83, 0.88, and 0.86, respectively; when tested in patients with structural heart disease, 0.75, 0.81, and 0.78, respectively. We confirmed high performance of AI-enabled ECG to detect AF on SR-ECG in patients without structural heart disease. The performance enhanced especially when SR-ECG after index AF-ECG was included in the algorithm, which was consistent in patients with structural heart disease.