Estimated pulse wave velocity (ePWV) as a potential gatekeeper for MRI-assessed PWV: a linear and deep neural network based approach in 2254 participants of the Netherlands Epidemiology of Obesity study

Abstract

AbstractPulse wave velocity (PWV) assessed by magnetic resonance imaging (MRI) is a prognostic marker for cardiovascular events. Prediction modelling could enable indirect PWV assessment based on clinical and anthropometric data. The aim was to calculate estimated-PWV (ePWV) based on clinical and anthropometric measures using linear ridge regression as well as a Deep Neural Network (DNN) and to determine the cut-off which provides optimal discriminative performance between lower and higher PWV values. In total 2254 participants from the Netherlands Epidemiology of Obesity study were included (age 45–65 years, 51% male). Both a basic and expanded prediction model were developed. PWV was estimated using linear ridge regression and DNN. External validation was performed in 114 participants (age 30–70 years, 54% female). Performance was compared between models and estimation accuracy was evaluated by ROC-curves. A cut-off for optimal discriminative performance was determined using Youden’s index. The basic ridge regression model provided an adjusted R2 of 0.33 and bias of < 0.001, the expanded model did not add predictive performance. Basic and expanded DNN models showed similar model performance. Optimal discriminative performance was found for PWV < 6.7 m/s. In external validation expanded ridge regression provided the best performance of the four models (adjusted R2: 0.29). All models showed good discriminative performance for PWV < 6.7 m/s (AUC range 0.81–0.89). ePWV showed good discriminative performance with regard to differentiating individuals with lower PWV values (< 6.7 m/s) from those with higher values, and could function as gatekeeper in selecting patients who benefit from further MRI-based PWV assessment.