Objective: Cardiac dysfunction and valvular heart diseases progress slowly. Early detection of these diseases is challenging for non-cardiologists. Therefore, we developed an artificial intelligence (AI) model for Japanese to simultaneously detect cardiac dysfunction and valvular diseases using a 12-lead electrocardiogram (ECG) and recommend examination such as echocardiograms.
Methods: We conducted a retrospective cohort study, and enrolled 30,467 adult patients, who were exposed to both 12-lead ECG and echocardiography between 2006 and 2021. These patients included 945 medical checkup examinees (3.1%). We developed a model that can speculate decreased left ventricular ejection fraction (LVEF) <40% and/or moderate to severe valvular heart diseases from ECGs.
Results: The model achieved high specificity for estimating reduced LVEF (sensitivity 74.1%, specificity 95.3%), aortic stenosis (sensitivity 18.9%, specificity 95.2%) and aortic (sensitivity 36%, specificity 95.2%), mitral (sensitivity 35.6%, specificity 95.2%), and tricuspid valve regurgitation (sensitivity 39.4%, specificity 94.6%) by the ECG. If a patient has at least one of these statuses, the patient needs to receive echocardiography. The AI could determine the necessity of echocardiography by ECG (sensitivity 44.7%, specificity 92.9%).
Conclusions: AI-based 12-lead ECG analysis can be used in patient selection for precise examination, including an echocardiogram, without any expert consultation or professional ECG interpretation.
View full abstract