Unveiling Coronary Heart Disease Prediction through Machine Learning Techniques: Insights from the Suita Population-Based Cohort Study

Abstract We leveraged machine learning (ML) techniques, namely logistic regression (LR), random forest (RF), support vector machine (SVM), extreme gradient boosting (XGBoost), and LightGBM to predict coronary heart disease (CHD) and identify the key risk factors involved. Based on the Suita study, 7672 men and women aged 30 to 84 years without cardiovascular disease were recruited from 1989 to 1999, in Suita City, Osaka, Japan. Over an average period of 15 years, participants were diligently monitored until the onset of their initial cardiovascular event or relocation. CHD diagnoses encompassed primary heart attacks, sudden death, or coronary artery disease with bypass surgery or intervention. RF achieved the highest AUC (95% CI) of 0.79 (0.70–0.87), outperforming LR, SVM, XGBoost, and LightGBM. Shapley Additive Explanations (SHAP) on the best model identified the top CHD predictors. Notably, systolic blood pressure, non-HDL-c, glucose levels, age, metabolic syndrome, HDL-c, estimated glomerular filtration rate, hypertension, elbow joint thickness, and diastolic blood pressure were key contributors. Remarkably, elbow joint thickness was identified as a previously unrecognized risk factor associated with CHD. These findings indicated that ML methods accurately predict incident CHD risk. Additionally, ML has identified new incident CHD risk variables.