High-density lipoprotein-related inflammatory ratios and coronary heart disease: A cross-sectional machine learning analysis of NHANES 2009 to 2020

Figures

• 

  Figure 1.. Study flowchart and analytical pipeline. Participant selection from the NHANES 2009 to 2020 database and machine learning workflow. From 119,033 initial participants, 14,745 met inclusion criteria (840 CHD, 13,905 non-CHD) after excluding those aged <20 years or with missing HDL-C, blood cell count, CHD diagnosis, or covariate data. Participants were randomly split into training (70%, n = 10,321) and validation (30%, n = 4424) sets. Six machine learning models (KNN, random forest, logistic regression, GBDT, decision tree, XGBoost) were developed following LASSO-based feature selection with 10-fold cross-validation. AUROC = area under the receiver operating characteristic curve, CHD = coronary heart disease, GBDT = gradient boosting decision tree, HDL-C = high-density lipoprotein cholesterol, KNN = K-nearest neighbors, LASSO = least absolute shrinkage and selection operator, MCC = Matthews correlation coefficient, NHANES = National Health and Nutrition Examination Survey, XGBoost = eXtreme gradient boosting.

• 

  Figure 2.. LASSO regression for feature selection. (A) Ten-fold cross-validation curve showing cross-validation mean versus log-transformed lambda. Vertical dashed lines indicate minimum lambda (red, −7.141) and 1-standard-error lambda (blue, −4.536). Numbers at the top denote nonzero coefficients. (B) Coefficient trajectories for all features across lambda values. Each colored line represents 1 feature. As lambda increases, less important features are shrunk to 0. CVM = cross-validation mean, HDL-C = high-density lipoprotein cholesterol, LASSO = least absolute shrinkage and selection operator, LHR = lymphocyte-to-HDL cholesterol ratio, MHR = monocyte-to-HDL cholesterol ratio, NHR = neutrophil-to-HDL cholesterol ratio, PHR = platelet-to-HDL cholesterol ratio.

• 

  Figure 3.. Machine learning model performance on validation set. Comprehensive evaluation of 6 machine learning models (n = 4424). (A) ROC curves with AUC values. Diagonal dashed line represents random chance (AUC = 0.5). (B) Calibration plots comparing predicted probabilities to observed proportions. Diagonal line indicates perfect calibration. Brier scores are shown in parentheses. (C) Decision curve analysis showing net benefit across threshold probabilities. “Treat All” and “Treat None” serve as references. AUC = area under the curve, DCA = decision curve analysis, ROC = receiver operating characteristic.

• 

  Figure 4.. SHAP force plot for model interpretation. SHAP force plot showing feature contributions to CHD prediction for a single individual (f (x) = 0.118) relative to the base value (E[f(x)] = 0.0559). Red segments represent features increasing risk (positive contributions); blue segments represent protective features (negative contributions). Segment width indicates contribution magnitude. Selected feature values are annotated. HDL-C = high-density lipoprotein cholesterol, LHR = lymphocyte-to-HDL cholesterol ratio, MHR = monocyte-to-HDL cholesterol ratio, NHR = neutrophil-to-HDL cholesterol ratio, PHR = platelet-to-HDL cholesterol ratio, SHAP = SHapley Additive exPlanations.