[4Q2-OS-5] 医療におけるAIの社会実装に向けて

[4Q2-OS-5-03] Research for clinical application of electrocardiogram AI to determine left ventricular systolic dysfunction

〇Satoshi Kodera¹, Hirotoshi Takeuchi¹, Masataka Sato¹, Shinnosuke Sawano¹, Susumu Katsushika¹, Hiroki Shinohara¹, Masao Daimon¹, Issei Komuro¹ (1. The University of Tokyo )

Keywords:ECG, EF

Background: To classify artificial intelligence (AI) as a diagnostic tool, it must first be acknowledged as a programmed medical device. It is vital to clarify the eligibility criteria for patients, as well as those who are not eligible, throughout the application process. In this study, we investigated the application target of a deep learning model for interpreting left ventricular systolic dysfunction from an electrocardiogram (ECG). Methods: We created a model for interpreting left ventricular systolic dysfunction from ECG using deep learning. A stratified analysis was performed by age, sex, electrocardiographic findings, background heart disease, and test purpose. Results: ECG AI area under curve (AUC) 0.95 for the target patient. AUC by age and sex was 0.95 for males, 0.96 for females, 0.96 for those under 60 years of age, and 0.91 for those over 80 years of age. The AUC for each ECG finding was 0.95 for QRS duration less than120 msec, and 0.94 for QRS duration more than120 msec. The AUC for each background heart disease was 0.95 without left ventricular hypertrophy, 0.94 with left ventricular hypertrophy, 0.95 without pulmonary hypertension, and 0.94 with pulmonary hypertension. The AUC for each test purpose was 0.95 for workup purposes and 0.94 for screening purposes. CONCLUSIONS: ECG AI for determining left ventricular systolic dysfunction is widely applicable to patients.