Machine Learning Implementations for Multi-class Cardiovascular Risk Prediction in Family Health Units

Abstract

Cardiovascular disease (CVD) risk prediction plays a significant role in clinical research since it is the key to primary prevention. As family health units follow up on a specific group of patients, particularly in the middle-aged and elderly groups, CVD risk prediction has additional importance for them. In a retrospectively collected data set from a family health unit in Turkey in 2018, we evaluated the CVD risk levels of patients based on SCORE-Turkey. By identifying additional CVD risk factors for SCORE-Turkey and grouping the study patients into 3-classes low risk, moderate risk, and high risk patients, we proposed a machine learning implemented early warning system for CVD risk prediction in family health units. Body mass index, diastolic blood pressures, serum glucose, creatinine, urea, uric acid levels, and HbA1c were significant additional CVD risk factors to SCORE-Turkey. All of the five implemented algorithms, k-nearest neighbour (KNN), random forest (RF), decision tree (DT), logistic regression (LR), and support vector machines (SVM), had high prediction performances for both the K4 and K5 partitioning protocols. With 89.7% and 92.1% accuracies for K4 and K5 protocols, KNN outperformed the other algorithms. For the five ML algorithms, while for the low risk category, precision and recall measures varied between 95% to 100%, moderate risk, and high risk categories, these measures varied between 60% to 92%. Machine learning-based algorithms can be used in CVD risk prediction by enhancing prediction performances and combining various risk factors having complex relationships.