Improving Predictions in the Era of Precision Medicine: A Comparison of Echocardiogram and Cardioproteomic Predictors of Post-Aortic Valve Replacement Outcomes

Abstract

Aortic stenosis (AS) is a serious form valvular heart disease characterized by stiffening and calcification of aortic valve leaflets. Aortic valve replacement (AVR) remains the only FDA-approved treatment for AS; however, up to 40% of patients can experience worsening of symptoms or death after intervention. Imaging measurements obtained by transthoracic echocardiogram (TTE) have been widely used to predict post-AVR outcomes; however, significant interobserver variability and lack of data standards limit their prognostic utility. The advent of Precision Medicine and cardioproteomics has enabled novel research into disease pathophysiology, given that proteins are proximal markers of disease state and severity. Thus, we hypothesized that modeling post-AVR outcomes with protein analytes would yield better accuracy than with TTE measurements. In this cohort study of patients who underwent AVR (N=75), we constructed two types of classification models—Logistic Regression and Convolutional Neural Networks—to compare areas under the curve (AUC) between protein analyte and TTE measurement models. Results showed that protein analyte models achieved better global accuracy; however, further works are needed to determine whether combining predictor types offers the best approach for predicting post-AVR adverse outcomes.