Predicting arterial stiffness with multilayer perceptron; a clinical data-driven approach

Abstract

Introduction: arterial stiffness is a decrease in the elasticity of the arteries, mainly those of large caliber. Objective: to predict the presence or absence of arterial stiffness through primary care examinations using multilayer perceptron. Materials and methods: analytical and cross-sectional study of a secondary database made up of 801 patients undergoing estimation of cardio-ankle vascular index (CAVI) using an atherosclerosis detector. The dependent variables were left CAVI (LCAVI), right CAVI (RCAVI). The independent variables were 11 tests obtained through clinical history. Multilayer perceptron-type neural networks were used, from which the classification table, predictive capacity, and diagnostic tests were evaluated. Results: for LCAVI, the model had an area under the curve (AUC) of 0.878. For RCAVI, the AUC was 0.869. The model test had a percentage of correct predictions for CAVI ≥ 9 of 70.70% and 67.10% for LCAVI and RCAVI, respectively. For CAVI < 9 it was 91.60% for LCAVI and 94.60% for RCAVI. The correlation between the presence and absence of arterial stiffness of the CAVI test of the atherosclerosis detector, and the CAVI obtained by multilayer perceptron was moderate for LCAVI (K = 0.579) and RCAVI (K = 0.587). Specificity values were 92% and 93% for LCAVI and RCAVI, respectively. The negative predictive value was 88% for LCAVI and 89% for RCAVI. Conclusions: the multilayer perceptron-type neural network, based on primary care examinations, had a high predictive value for LCAVI and RCAVI to exclude the presence of arterial stiffness.