Machine Learning-Based Risk Prediction of Heart Disease

Abstract

Cardiovascular diseases are now the main causes of death globally, and heart disease is one of the most common types. Early prediction of heart disease for early intervention and prevention in the general population has been achieved, but the current medical system’s scope of application is relatively small. A publicly available dataset with 918 records is used in this study. The five classification algorithms were logistic regression, K-nearest neighbors, decision tree, random forest and XGBoost. Repeatedly stratified cross-validation and grid search were used to optimise and evaluate the models, and recall and AUC were chosen as the main indicators. According to the above experiment results, random forest performs better than other models in all the above indices and achieves an AUC of 0.93, with both recall and F1 score exceeding 0.90. Feature Importance Analysis found that ST slope, ST depression and chest pain type were the main predictors and aligned with clinical guidelines. Based on the above analysis, the pipeline constructed in this experiment has been validated as a stable and explainable early heart disease screening aid.