Temporal Variability and Influence of Measurement Conditions of AI-Based Atrial Fibrillation Risk Estimation.

Background: Although artificial intelligence (AI) has been developed to identify patients with paroxysmal atrial fibrillation (PAF) during sinus rhythm, information on its variability remains limited. We evaluated the reproducibility and effect of recording condition on the estimation of AF risk using an electrocardiography (ECG) machine equipped with an AI-based program. Methods: We extracted two ECG data from a single ECG test in 149 patients to evaluate reproducibility within 4 min. We also recorded ECG signals under 12 conditions (standard, two conditions shifting precordial electrodes, five conditions moving limb electrodes to the torso, three conditions contaminating noise, and reproducibility over 15 min) in 30 participants to evaluate changes from the standard. The results of the AF risk estimation are expressed at four levels. Results: The rate of participants within one level of error was 95% for reproducibility within 4 min and 87% for reproducibility over 15 min. Shifting the precordial electrodes upward or downward and replacing the left leg electrode with the torso electrode frequently caused a two- or three-level change. In clinical information, increased brain natriuretic peptide tended to increase the variability. Conclusions: The AF risk estimated by the AI-based program exhibited temporal variability. Shifting the precordial electrodes influenced AI-based AF risk estimation.