Cardiovascular Disease Prediction using Clinical Data

Cardiovascular disease (CVD) remains a major global health challenge due to its high morbidity and mortality rates and the difficulty of early diagnosis, highlighting the need for accurate and interpretable predictive tools to support clinical decision-making. This study developed an analytical machine learning pipeline to construct interpretable models for predicting CVD risk using open-source data from the National Health and Nutrition Examination Survey (NHANES) collected between 2017 and 2023. The dataset included 12,504 adult participants and 41 variables covering demographic, dietary, anthropometric, and health-related factors. A modular and reproducible machine learning framework was implemented to ensure data integrity and analytical sustainability. Missing values were handled using multiple imputation techniques, while feature scaling was applied to optimize model training. Recursive Feature Elimination (RFE) was used to identify the most informative variables, producing a reduced feature subset that improved predictive performance while enhancing interpretability. To address class imbalance, undersampling, oversampling, and hybrid resampling methods were applied. Seven machine learning models—XGBoost, Support Vector Classifier, Logistic Regression, Light Gradient Boosting Machine, Random Forest, Random Under-Sampling Boosting (RUSBoost), and a Voting Classifier—were trained and optimized through hyperparameter tuning. Model performance was evaluated using accuracy, precision, recall, specificity, F1-score, area under the precision–recall curve (AUPRC), and area under the receiver operating characteristic curve (ROC-AUC). To further improve model transparency and interpretability, multiple explainable artificial intelligence (XAI) techniques were employed, including SHapley Additive exPlanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), permutation feature importance, and traditional feature importance methods. The proposed framework demonstrates strong potential to support healthcare professionals during the CVD diagnostic process and highlights the importance of well-structured tabular data in developing interpretable, reliable, and reproducible healthcare analytics solutions.