The AutoAS deep learning model detected moderate to severe aortic stenosis with an AUC of 0.93 (95% CI 0.901-0.955), a specificity of 92.6%, and a PPV of 94.6%.
Observational (n=334)
Yes
Does the AutoAS deep learning model accurately grade aortic stenosis severity from B-mode echocardiography compared to expert interpretation?
An automated deep learning model can accurately grade aortic stenosis severity using only B-mode echocardiography, offering a potentially scalable alternative to expert Doppler interpretation.
Effect estimate: AUC 0.93 (95% CI 0.901-0.955)
Abstract Background Aortic stenosis (AS) is frequently under-diagnosed, particularly in the elderly and underserved populations 1. While Doppler echocardiography is the gold standard for diagnosis, its scalability is limited by the need for expert interpretation. In this work, we introduce AutoAS, a deep learning model trained on 30, 000 TTE studies to score AS severity on a continuous scale using only B-mode cine loops. Purpose To evaluate AutoAS as a stand-alone diagnostic tool and compare its aortic stenosis severity scoring capabilities to reference scores from a panel of 3 independent level III echocardiographers. Methods A validation dataset of 334 held-out patient studies (one study per patient) was assembled from both developmental and novel clinical sites. Each Doppler study was independently read by a panel of 3 level III echocardiographers that graded the severity of aortic stenosis on a 4-level scale (none, mild, moderate and severe). From panel severities, both a binary label and a continuous score were derived. The binary label referred to the presence or absence of "moderate to severe" aortic stenosis, while the score was curtailed to the interval 0, 1 and corresponded to the average of 3 individual scores (0 for none, 1/3 for mild, 2/3 for moderate and 1 for severe). Results In detecting "moderate to severe" aortic stenosis, AutoAS achieved an area under the ROC curve of 0. 93 (95% CI 0. 901–0. 955), with a specificity of 92. 6% and a PPV of 94. 6%. When comparing scores from the model to reference panel scores, Pearson and Spearman correlation coefficients were 0. 852 and 0. 855 respectively (Figure 1). Moreover, looking at median model scores associated with each set of concordant panel scores, the slope of the regression line was 1. 01 with an R² of 0. 95 (Figure 2). Conclusion An automated deep learning model using only B-mode cine loops from TTE was able to grade the severity of aortic stenosis on a continuous scale with high correlation to a reference severity score derived from adjudicated severities of a panel of 3 level III echocardiographers. Figure 1 Figure 2
Poilvert et al. (Thu,) conducted a observational in Aortic stenosis (n=334). AutoAS deep learning model vs. Panel of 3 independent level III echocardiographers was evaluated on Detection of moderate to severe aortic stenosis (AUC 0.93, 95% CI 0.901-0.955). The AutoAS deep learning model detected moderate to severe aortic stenosis with an AUC of 0.93 (95% CI 0.901-0.955), a specificity of 92.6%, and a PPV of 94.6%.