AI model using 6 parameters including DVI and SVI differentiated severe low-gradient AS from moderate AS with AUC 0.853 (95% CI 0.788-0.919).
Can an AI model using echocardiographic parameters accurately differentiate severe low-gradient aortic stenosis from moderate aortic stenosis?
A semi-automatic AI model incorporating both automatically extracted echocardiographic parameters and manually calculated DVI and SVI can accurately differentiate severe low-gradient from moderate aortic stenosis.
Abstract Background Development of AI could help in diagnosing the most challenging subtype of aortic stenosis (AS) - low-gradient (LG) AS, which due to the discordance between the mean transvalvular gradient (mAG) and aortic valve area (AVA) poses challenges in determining stenosis severity. Purpose To develop an AI model capable of automatically differentiating severe LG AS from moderate AS based on echocardiographic images. Methods From 174 consecutive patients with suspected severe LGAS from National Institute of Cardiology in Warsaw, 158 (82 moderate AS, 72 severe LGAS) with a complete set of high-quality echocardiographic exams were included in the study. Image-derived parameters, including left ventricular internal diameter (LVID), interventricular septum thickness (IVS),left ventricular posterior wall thickness (LVPW) and left ventricular ejection fraction (EF)) were automatically extracted with the use of two established AI models: EchoNet Dynamic and EchoNet LVH and then used to train our AI model (extreme gradient boosting, XGBoost) to differentiate between low-gradient severe and moderate AS. In a similar manner, semi-automatic model was created by additionally incorporating the manually calculated doppler velocity index (DVI) and stroke volume index (SVI) parameters (available for 135 from 158 patients). The stratified cross-validation was used to assess the quality of the model. The in-build XGBoost’s gain measure was used to assess the feature importance of the parameters used in the model. The 95% confidence intervals (CIs) were calculated with a bootstrapping method. Results The performance for the semi-automatic model, based on 6 parameters, i.e., LVID, LVPW, IVS, EF, SVI, and DVI, was 0.853 (95% CI: 0.788, 0.919), and on the same dataset of 135 patients, the performance for fully-automatic approach (4 parameters used: LVID, LVPW, IVS, and EF) was 0.660 (95% CI: 0.568, 0.752); DeLong’s test result, p-value 00005; Figure 1). As for the semi-automatic approach, the results were: (1) EchoNet+SVI, AUC=0.718 (95% CI: 0.632, 0.805), (2) EchoNet+DVI, AUC=0.809 (95% CI 0.735, 0.882), and (3) EchoNet+SVI+DVI, AUC=0.853 (95% CI: 0.788, 0.919). When data from all 158 patients, for which these 4 parameters that could be derived automatically were used, the performance increased to 0.719 (95% CI: 0.64, 0.798), p=0.34; Figure 1).Of the features used in the model, the manually obtained DVI and SVI were the most important ones. They were followed by: EF, LVID, IVS, and LVPW (Figure 2). Conclusion AI enables automatic differentiate between low-gradient severe and moderate AS using only imaging parameters. The models developed in this study can serve as components of a diagnostic decision support system, offering additional insights for complex AS cases without necessitating further diagnostic testing.
Wrzosek et al. (2025) studied this question. AI model using 6 parameters including DVI and SVI differentiated severe low-gradient AS from moderate AS with AUC 0.853 (95% CI 0.788-0.919).