Comparative analysis of machine learning techniques for cardiovascular disease prediction

Early and accurate prediction of cardiovascular disease (CVD) is fundamental for reducing morbidity and mortality. Machine learning (ML) algorithms provide a data-driven, actionable foundation to strengthen clinical decision-making and enable more precise risk stratification. The purpose of this study is to identify the most significant risk factors for CVD and to compare the predictive performance of eight machine learning algorithms.

The study utilizes the Cardiovascular Disease dataset, an open-access resource from the Kaggle repository. It applies 5-, 10-, 15-, and 20-fold cross-validation (CV) to evaluate the performance of Logistic Regression, Decision Tree, Random Forest, Support Vector Machine (SVM), XGBoost, LogitBoost, Gradient Boosting, and LightGBM. Accuracy, sensitivity, specificity, precision, F1-score, false discovery rate (FDR), and area under the receiver operating characteristic curve (AUC) are used to evaluate the performance of the algorithms. The selection and ranking of relevant features are achieved through multiple methodologies, including Boruta, Regularized Random Forest, Recursive Feature Elimination, and LASSO.

All clinical and demographic characteristics except gender show significant differences between the CVD and non-CVD groups (p-values <0.001). While 5-fold cross-validation yields somewhat better metrics for most techniques, model performance remains stable across folds. With the best AUC of 0.800 (95% CI: 0.792-0.807) and balanced F1-score(0.77), LightGBM produces the most balanced findings while retaining a high sensitivity (0.791) under both 5-fold and 20-fold CV. Despite having low sensitivity (0.628), LogitBoost achieves the best accuracy (0.814) and specificity (0.911) in 20-fold CV. Considering all evaluation metrics, particularly AUC and sensitivity, LightGBM with 5-fold cross-validation is selected as the optimal model for CVD prediction. Gain analysis indicates that systolic blood pressure is the primary contributor, and feature selection techniques repeatedly identified systolic blood pressure, age, and cholesterol levels as prominent predictors.

LightGBM demonstrates modestly superior, more balanced performance compared to the other evaluated algorithms in this dataset. While the performance (AUC = 0.800) suggests potential utility as a decision support tool, external validation is needed before clinical deployment. The model is not yet ready for standalone clinical use.