Efficacy of Artificial Intelligence on 12-leads ECG to Detect Atrial Fibrillation during Sinus Rhythm

Abstract

 Objective: Atrial fibrillation (AF) increases the risk of stroke, heart failure, and death, making early diagnosis and initiation of anticoagulation therapy crucial. However, AF is often asymptomatic, making early detection challenging. Therefore, the objective of this research is developing artificial intelligence (AI) for screening Japanese patients at high risk of AF onset based on electrocardiogram (ECG) features of sinus rhythm and examining clinical effectiveness of this AI.

 Methods: Using retrospective cohort data of adult individuals who underwent 12-lead ECG examinations from 2000 to 2019 (total 30,467 individuals, including 3,528 cases of AF), we developed an AI model to predict the risk of AF onset. The generalization performance of the AI was evaluated. Additionally, we analyzed ECG findings of cases of the pre-AF state and ECG findings which are bases of the AI prediction. To assess the clinical significance of the AI, we simulated the diagnosis of atrial fibrillation by AI in an undiagnosed population.

 Results: The generalization performance of the AI showed a specificity 95.7% and a sensitivity of 44.0%. In the analysis of findings, when AI identified high-risk cases, ST-T abnormalities were most common (15.5%) among cases in the pre-AF state. The use of this AI in health check-ups for individuals aged 40 and above in Japan suggests the potential for preventing an estimated 9,554 strokes per year through this simulation.

 Conclusions: The development of AI for predicting the comprehensive risk of AF onset from ECG features of sinus rhythm in the Japanese population demonstrated practical enough accuracy, as validated by cohort data. Furthermore, multiple ECG abnormalities were found to influence the onset of paroxysmal atrial fibrillation, and the AI also based its risk assessment on multiple ECG abnormalities. Simulation in an undiagnosed population suggested the potential effectiveness of AF screening AI in health checks.