Predicting aortic stenosis progression using a video-based deep learning model of aortic stenosis built for single-view two-dimensional echocardiography

Abstract

Abstract Background We have recently described a deep learning (DL) model that relies on 2-dimensional parasternal long axis (PLAX) videos from transthoracic echocardiography (TTE) without Doppler imaging to identify severe aortic stenosis (AS) at the point-of-care. However, the timely diagnosis of AS has traditionally required skilled examination with Doppler imaging at regular time intervals. Purpose To explore the role of a DL model previously trained to detect severe AS based on 2D-PLAX videos in predicting the rate of AS progression among patients with no, mild, or moderate AS. Methods In this retrospective cohort study, a random sample of 2354 individuals undergoing serial monitoring with at least two separate TTEs and available Doppler parameters of AS severity (peak aortic valve velocity, AV Vmax) between 2016 and 2021 in the Yale-New Haven Health system were included. Eligible participants had a left ventricular ejection fraction (LVEF) >=50% and no evidence of severe AS at baseline. After automated selection of PLAX videos using an automated view classifier, the baseline deep learning-based AS probability index (DL-ASi) was calculated for each patient. The absolute change in the AV Vmax was divided by the time difference between each pair of consecutive TTE studies, and then averaged to derive the mean rate of AS progression. The association between the baseline DL-ASi and the mean rate of AS progression was assessed using ordinary least squares regression with adjustment for the patient’s sex, age, and baseline LVEF and AV Vmax. To visualize the interaction between the baseline DL-ASi and AV Vmax in predicting the rate of AS progression, an interaction term was included and a contour plot was plotted for both men and women. Results We included 8473 TTE studies from 2354 unique individuals (mean age 70±14 years, n=1138 [48.3%] women) performed over a median 2.5 (IQR 1.4-3.7) years (A). Higher DL-ASi at baseline was independently associated with a more rapid rate of AS progression (βadj 0.27 [95% CI 0.20-0.33, P<0.001]). The rates of AV Vmax progression ranged from 0.06±0.01 m/sec/year in the <0.2 DL-ASi group, to 0.29±0.04 m/sec/year in the ≥0.8 DL-ASi group (B). Contour plots illustrate how, for a given AV Vmax value, DL-ASi stratified the risk of AS progression in women and men (C). Conclusion In individuals with mild to moderate AS or no AS, a DL model trained to detect features of severe AS identifies individuals with a higher rate of AS severity progression, as assessed by the mean rate of AV Vmax change. This represents a promising approach for the longitudinal, point-of-care monitoring of patients with non-severe AS outside specialized centers.