A Web-Based Early Risk Prediction Tool for Cardiovascular Disease

Abstract

Cardiovascular disease (CVD) is one of the leading causes of mortality around the world. The Sri Lankan community is also facing a major dilemma as the number of fatalities from cardiovascular disease rises. Due to multiple flaws identified and inadequate knowledge about this disease, traditional diagnosis procedures are prone to erroneous results. As a significant portion of individuals who are at risk are not correctly diagnosed as cardiac patients, the probability of producing precise results is slim. Another prevailing concern leading to the growth in CVD related mortality in developing countries like Sri Lanka is the shortage of medical personnel with the requisite training and skills. The goal of this study is to discover the flaws in traditional CVD diagnosis practices, as well as to assess, comprehend existing systems and the contributing risk factors of cardiovascular disease. Especially, it investigates how to incorporate artificial intelligence to predict the risk of CVD and also to enhance the performance of existing classifiers to provide an efficient early risk prediction for CVD. The Cleveland, Statlog, and Framingham datasets were used to test both conventional machine learning models (i.e., Logistic Regression, Naive Bayes, Random Forest, and Support Vector Classifier) and deep learning models (i.e., Deep Neural Networks and Long Short-Term Memory networks). Three home datasets were produced by modifying existing datasets to enhance usability. A stacked ensemble model integrating numerous heterogeneous classifiers was devised to increase the accuracy of weak classifiers. Stack machine learning and stack deep learning, two stacked ensemble models introduced in this research, are effective in enhancing the prediction accuracy of anomalistic classifiers. Stack machine learning models outperformed anemic classifiers by 84%, 78.6 %, and 82.7% for the Cleveland, Statlog, and Framingham datasets, respectively. Stack deep learning models surpassed feeble deep learning classifiers by 84.7%, 83%, and 81% for the Cleveland, Statlog, and Framingham datasets, respectively. The increased performance of the suggested ensemble learning approach is further validated by the Receiver Operating Characteristic (ROC) curves and confusion matrices. As the conclusion of the study, a web-based early risk prediction tool for cardiovascular disease was developed, incorporating the high-performing models.