Enhancing Cardiovascular Disease Detection Through Dataset Integration and TVAE-Driven Synthetic Data Generation with 1D CNN

Abstract

Early diagnosis of cardiovascular disease (CVD) plays a vital role in disease prevention. This paper proposes to predict the risk of developing CVD´s using traditional clinical parameters and some cardiac biomarkers such as C-reactive protein (CRP), Homocysteine, and lipoprotein(a)(Lp(a)). This research works by augmenting two different datasets and make use of Tabular Variational Autoencoder (TVAE) to obtain 10,000 balanced samples. One-Dimensional Convolutional Neural Network (1D-CNN) has been trained to detect intricate relationships within said dataset. Our algorithm was rather efficient, reached accuracy of 91%, precision of 91%, recall of 90% and f1 score 90%. Feature Importance Permutation was employed to determine which features were the most important in making CVD prediction. The study found out that the most critical parameter was Hypertension, while Smoking and Diabetes came in a close second place. Although these conventional factors were responsible for making up the primary CVD prediction model, some specific biomarkers were indispensable for reaching high accuracy.