Predicting the death caused by cardiovascular and/or cerebrovascular disease within 7 years follow-up using machine learning models

Abstract

Abstract Background Cardiovascular and cerebrovascular disorders, which are now the leading and second causes of death in the majority of the world's countries, affect human life and health. Moreover, the prediction of death caused by these is deemed necessary to prevent it and reduce its rate. Purpose We employ machine learning models for the prediction of cardiovascular and/or cerebrovascular death within 7 years follow-up using clinical and laboratory features. Methods We analyzed the dataset of German epidemiological trial on ankle brachial index (getABI) study, including 5,587 patients (mean age 71.14 years, 40.7 % male, 359 deceased patients and 5,228 alive patients). Recursive feature elimination with cross-validation selected 11 features from a set of 55 features: diabetes mellitus, ABI index before exercise, vitamin D, troponin I, triglycerides, potassium, LDL cholesterol, HDL cholesterol, the flow noise of the right and left carotid artery and the pulse status of the right popliteal artery. Additionally, the handling of missing values was achieved by SimpleImputer. Random Forest (RF), Adaptive Boosting (AdaBoost), LightGBM classifiers were applied to predict death and were trained and tested using 5-fold cross-validation. Moreover, GridSearchCV tuned the hyperparameters of the models. Results LightGBM was the most accurate model, achieving 87.56 % mean balanced accuracy and 90.63 % mean value of the area under (AUC) the Receiver Operating Characteristic Curve (ROC). The mean ROC-AUC value of the RF, and AdaBoost were equal to 87.25 % and 89.52 %, respectively. Τhe mean ROC-AUC value, its standard deviation and the mean balanced accuracy of the models are presented in Figure 1. Moreover, Figure 2 indicates the ROC-AUC values for each k-fold and the mean value of them which achieved by the best classifier, the LightGBM. Conclusion Machine learning models achieved accurate prediction of cardiovascular and cerebrovascular death in 5,587 patients within 7 years follow-up utilizing basic medical data.