Correlating Electrocardiograms with Echocardiographic Parameters in Hemodynamically-Significant Aortic Regurgitation Using Deep Learning.

Background: Of all electrocardiogram (ECG) deep-learning (DL) models used to detect left-sided valvular heart diseases, aortic regurgitation (AR) has been the hardest to detect. Moreover, to what extent ECGs could detect AR-related left ventricular (LV) remodeling and dysfunction is unknown. Objectives: We aimed to evaluate the ability of DL-based ECG models to predict LV remodeling parameters associated with hemodynamically significant AR. Methods: From 573 consecutive patients, 1457 12-lead ECGs close to baseline transthoracic echocardiograms confirming ≥ moderate-severe AR and before aortic valve surgery were retrospectively collected. A ResNet-based model was used to predict LV ejection fraction (LVEF), LV end-diastolic dimension (LVEDD), LV end-systolic dimension index (LVESDi), LV mass index (LVMi), LV end-diastolic volume index (LVEDVi), LV end-systolic volume index (LVESVi), and bicuspid aortic valve (BAV) from the ECGs. Five-fold cross-validation was used for model development (80%) with the held-out testing set (20%) to evaluate its performance. Results: Our DL model achieved area under receiver operating characteristic curves (AUROCs) of 0.77, 0.80, and 0.87 for discriminating LVEF < 55%, < 50%, and < 40%. For LVEDD > 65 mm, LVESDi > 30 mm/m Conclusions: Our DL model found correlations between ECGs and parameters indicating LV remodeling and dysfunction in patients with significant AR. Analyzing ECGs with DL models may assist in the timely detection of LV dysfunction and screening for the necessity of additional echocardiography exams, especially when echocardiography might not be readily available.