Abstract
High-density lipoprotein (HDL)-related inflammatory ratios (monocyte-to-HDL cholesterol ratio [MHR], lymphocyte-to-HDL cholesterol ratio, neutrophil-to-HDL cholesterol ratio [NHR], platelet-to-HDL cholesterol ratio) represent composite biomarkers integrating lipid metabolism and inflammatory pathways. We developed machine learning models to evaluate their utility in coronary heart disease (CHD) classification using a large population-based dataset. We analyzed data from the National Health and Nutrition Examination Survey 2009 to 2020, including 14,745 US adults aged ≥20 years (mean age 51.8 ± 17.6 years). Self-reported CHD diagnosis was the outcome variable. Machine learning models (eXtreme gradient boosting, random forest, logistic regression) were developed to evaluate cross-sectional associations between HDL-related inflammatory ratios and CHD prevalence. Self-reported CHD prevalence was 5.7% (n = 840). All HDL-related inflammatory ratios were significantly elevated in CHD patients: MHR (0.54 ± 0.35 vs 0.42 ± 0.23, P < .001), lymphocyte-to-HDL cholesterol ratio (2.05 ± 3.12 vs 1.55 ± 1.02, P < .001), and NHR (4.06 ± 2.89 vs 3.11 ± 1.77, P < .001). eXtreme gradient boosting demonstrated optimal performance with an area under the receiver operating characteristic curve of 0.8892, accuracy of 96.55%, and precision of 86.00%. SHapley Additive exPlanations analysis identified age as the most important predictor, with MHR and NHR ranking among the top 5 features. Machine learning models incorporating HDL-related inflammatory biomarkers achieved high discrimination (area under the receiver operating characteristic curve = 0.8892) for identifying cross-sectional associations with CHD prevalence. These findings reveal significant cross-sectional associations between HDL-related inflammatory ratios and CHD prevalence, rather than predictive relationships for incident events. These readily available biomarkers from routine blood tests provide substantial value for cardiovascular risk stratification. Prospective validation is warranted to establish their utility for predicting incident CHD events.
Keywords: coronary heart disease, high-density lipoprotein cholesterol, inflammatory biomarkers, machine learning, monocyte-to-HDL cholesterol ratio, neutrophil-to-HDL cholesterol ratio, NHANES, XGBoost
1. Introduction
Coronary heart disease (CHD) remains the leading cause of cardiovascular morbidity and mortality worldwide, with cardiovascular disease responsible for approximately 928,700 deaths in the United States in 2021.[1] Despite significant advances in therapeutic interventions and preventive strategies, CHD incidence continues to rise globally.[2] Early identification of individuals at high risk for CHD is crucial for implementing timely preventive measures and optimizing clinical outcomes. However, current risk prediction models demonstrate suboptimal performance in diverse populations, highlighting the need for improved risk stratification tools.[3]
Traditional cardiovascular risk assessment relies primarily on established clinical factors, including age, sex, smoking status, blood pressure, and lipid profiles, as exemplified by the Framingham Risk Score and Pooled Cohort Equations.[4,5] While these conventional models have provided valuable clinical guidance, they often fail to capture the complex pathophysiological mechanisms underlying atherosclerosis.[6] The inflammatory paradigm of atherosclerosis has emerged as a critical component of CHD pathogenesis, with chronic low-grade inflammation playing a pivotal role in plaque initiation, progression, and rupture.[7] High-sensitivity C-reactive protein (hs-CRP) has been most extensively studied, yet its clinical implementation remains limited by cost and nonspecific nature.[8]
High-density lipoprotein cholesterol (HDL-C) has long been recognized as protective against cardiovascular disease through reverse cholesterol transport and anti-inflammatory properties.[9] However, recent evidence suggests that HDL functionality may be more important than absolute HDL-C levels.[10] Concurrently, circulating inflammatory cells participate in various stages of atherosclerosis development.[11] HDL-related inflammatory biomarkers – lymphocyte-to-HDL cholesterol ratio (LHR), monocyte-to-HDL cholesterol ratio (MHR), neutrophil-to-HDL cholesterol ratio (NHR), and platelet-to-HDL cholesterol ratio (PHR) – represent composite indices integrating inflammatory cell counts with HDL-C levels.[12] Recent studies demonstrated associations between these biomarkers and cardiovascular outcomes, including acute coronary syndromes and in-stent restenosis.[13] Importantly, while neutrophil-to-lymphocyte ratio and platelet-to-lymphocyte ratio reflect inflammatory cell balance, HDL-related ratios uniquely integrate anti-inflammatory lipid metabolism with pro-inflammatory cellular activation. These 4 HDL-related ratios were selected based on distinct pathophysiological mechanisms: MHR reflects monocyte-mediated foam cell formation and plaque progression, as monocytes promote inflammatory response and reduce plaque stability through pro-inflammatory cytokines,[14] LHR represents lymphocyte-driven adaptive immune responses in atherosclerosis, with lymphocytes participating in antigen recognition and antibody production throughout atherogenesis,[15] NHR captures neutrophil-mediated endothelial dysfunction and thrombotic risk, as neutrophils aggravate tissue injury through inflammatory mediators and promote thrombosis via neutrophil extracellular traps,[16] and PHR integrates platelet activation with HDL-mediated inflammation modulation, reflecting the balance between prothrombotic states and anti-inflammatory lipid metabolism.[17] Although PHR demonstrated weaker performance in preliminary analyses, its inclusion was necessary to comprehensively evaluate all major cellular components (monocytes, lymphocytes, neutrophils, and platelets) and their interactions with HDL metabolism. This comprehensive approach ensures a complete assessment of the immune-lipid interface in CHD pathogenesis. However, systematic evaluation using machine learning algorithms for CHD classification remains unexplored.
The application of machine learning techniques in cardiovascular medicine has gained considerable momentum due to their ability to handle complex data and identify nonlinear relationships. Machine learning algorithms, particularly ensemble methods such as eXtreme gradient boosting (XGBoost), have demonstrated superior performance in medical prediction tasks.[18] The integration of novel biomarkers with machine learning offers promising opportunities to enhance CHD risk assessment while maintaining interpretability through techniques such as SHapley Additive exPlanations (SHAP) analysis.[19]
We hypothesized that HDL-related inflammatory ratios would be significantly elevated in individuals with CHD compared with those without CHD, reflecting the integration of pro-inflammatory cellular activation and impaired HDL function; that machine learning models incorporating these composite biomarkers would demonstrate superior discriminative performance for identifying cross-sectional associations with CHD prevalence compared with traditional statistical methods; and that these readily measurable biomarkers would provide complementary predictive value beyond conventional cardiovascular risk factors.
Using the National Health and Nutrition Examination Survey (NHANES), a nationally representative population-based survey with comprehensive data spanning 2009 to 2020,[20] this study aimed to: compare HDL-related inflammatory biomarkers between participants with and without CHD; develop and validate machine learning models to identify cross-sectional associations between these biomarkers and CHD prevalence; benchmark performance against multiple algorithms; quantify feature importance through SHAP analysis; and assess the potential of these accessible biomarkers for integration into cardiovascular risk stratification.
2. Methods
2.1. Study design and participants
This cross-sectional study analyzed data from NHANES 2009 to 2020, a nationally representative survey assessing the health and nutritional status of the noninstitutionalized US civilian population through standardized interviews, physical examinations, and laboratory assessments.[20] The NHANES protocol was approved by the National Center for Health Statistics Research Ethics Review Board, and all participants provided written informed consent.
From 6 consecutive NHANES cycles, we included 14,745 adults aged ≥20 years after excluding participants with missing CHD status (n = 8742), incomplete blood cell counts (n = 7234), missing HDL-C measurements (n = 3891), or missing covariates (n = 1774) from the initial 59,842 participants (Fig. 1). This study adhered to the Strengthening the Reporting of Observational Studies in Epidemiology guidelines.
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.
2.2. Outcome and exposure variables
CHD was defined by self-reported physician diagnosis through standardized questionnaires. While this approach may introduce misclassification bias, it enables assessment in large population-based samples and demonstrates reasonable concordance with medical records.[21]
HDL-related inflammatory biomarkers were calculated as:
MHR = monocyte count (×103/μL)/HDL-C (mmol/L)
LHR = lymphocyte count (×103/μL)/HDL-C (mmol/L)
NHR = neutrophil count (×103/μL)/HDL-C (mmol/L)
PHR = platelet count (×103/μL)/HDL-C (mmol/L)
Complete blood counts were measured using automated hematology analyzers (Beckman Coulter MAXM, HmX, DxH 800, Beckman Coulter, Inc., Brea). HDL-C was determined by direct enzymatic methods.[22] All measurements were performed on fasting blood samples following standardized protocols.
2.3. Covariates
Demographic variables included age, sex, race/ethnicity, education level, marital status, and poverty income ratio. Lifestyle factors included smoking status and alcohol consumption. Physical examination data included body mass index and systolic and diastolic blood pressure. Laboratory parameters included fasting glucose, total cholesterol, and triglycerides.
Hypertension was defined as blood pressure ≥130/80 mm Hg or antihypertensive medication use.[23] Diabetes was defined as fasting glucose ≥126 mg/dL, hemoglobin A1c ≥ 6.5%, self-reported diagnosis, or antidiabetic medication use.[24] Cancer history was self-reported.
2.4. Data partitioning and feature selection
To prevent data leakage, the dataset was randomly split into training (70%, n = 10,321) and validation (30%, n = 4424) sets using stratified sampling before any analysis. The validation set remained unseen until final model evaluation. All feature selection and model development procedures were performed exclusively on the training set.
Univariate analysis identified variables associated with CHD in the training set (chi-square tests for categorical variables, t tests for continuous variables, significance threshold P < .05). Least absolute shrinkage and selection operator (LASSO) regression with 5-fold cross-validation was then applied to the training set to select optimal features, with the regularization parameter (λ) determined by the minimum cross-validation error.[25] This procedure selected 15 features: age, sex, race/ethnicity, education level, marital status, diabetes, cancer history, smoking status, hypertension, HDL-C, and all 4 HDL-related inflammatory ratios (Fig. 2).
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.
2.5. Data preprocessing
All continuous features were standardized using Z-score normalization (mean = 0, standard deviation = 1), fitted on the training set, with the same transformation parameters applied to the validation set. Extreme outliers were managed using Winsorization at the 1st and 99th percentiles to preserve sample size while reducing outlier influence.
2.6. Machine learning model development
Six algorithms were trained on the training set using LASSO-selected features: K-nearest neighbors, random forest, logistic regression, gradient boosting decision tree, decision tree, and XGBoost.[26] Hyperparameter optimization used grid search with 5-fold cross-validation on the training set, maximizing the area under the receiver operating characteristic curve (AUROC). To address class imbalance (CHD prevalence of 5.7%), the Synthetic Minority Oversampling Technique (SMOTE) was applied only within each cross-validation fold during training to prevent data leakage. The validation set remained in its original imbalanced state to reflect real-world prevalence.
2.7. Model evaluation and interpretability
Model performance was assessed on the independent validation set using AUROC, accuracy, precision, recall, F1-score, specificity, and Matthews correlation coefficient. The optimal classification threshold was determined using the Youden index. Calibration was evaluated through calibration plots, and clinical utility through decision curve analysis (DCA).[27]
Confidence interval (CI) estimation: 95% CIs for AUROC were calculated using the Hanley–McNeil method, which accounts for both positive and negative sample sizes. CIs for accuracy, precision, recall, and specificity were calculated using the Wilson score interval method, which provides better coverage properties than the normal approximation, especially for proportions near 0 or 1.
SHAP analysis quantified feature contributions in the best-performing model. SHAP values calculate the marginal contribution of each feature across all possible feature coalitions, providing both local and global interpretability.
2.8. Statistical analysis
Descriptive statistics are presented as mean ± standard deviation for continuous variables and frequencies with percentages for categorical variables. Between-group comparisons used independent t tests for continuous variables (or Mann–Whitney U tests for non-normally distributed data) and chi-square tests (or Fisher exact test when expected cell counts <5) for categorical variables. All analyses used unweighted data. Missing data were handled through complete-case analysis after applying exclusion criteria. To ensure reproducibility, random seed values were set (seed = 2024) for all stochastic procedures. Statistical analyses were performed using R version 4.3.0 (R Foundation for Statistical Computing, Vienna, Austria) with packages including survey, caret, and presentation and analysis of receiver operating characteristic curves (an R package). Machine learning implementations used Python version 3.9 (Python Software Foundation, Beaverton) with scikit-learn, XGBoost, imbalanced-learn, and SHAP. All statistical tests were two-tailed with a significance level α = 0.05.
3. Results
3.1. Study population characteristics
A total of 14,745 participants were included in the final analysis (Fig. 1). The study population had a mean age of 51.8 ± 17.6 years, with 7519 (51.0%) female participants. The overall CHD prevalence was 5.7% (n = 840). Participants were randomly allocated to training (n = 10,321, 70%) and validation (n = 4424, 30%) sets.
Table 1 presents baseline characteristics stratified by training and validation sets. No statistically significant differences were observed between groups (all P > .05), confirming successful randomization. HDL-related inflammatory biomarkers demonstrated no significant differences: MHR (0.42 ± 0.25 vs 0.42 ± 0.24, P = .687), LHR (1.55 ± 1.34 vs 1.57 ± 1.36, P = .706), NHR (3.11 ± 2.18 vs 3.14 ± 2.10, P = .765), and PHR (185.8 ± 76.7 vs 184.4 ± 75.9, P = .424). CHD prevalence was similar between sets (5.61% vs 5.90%, P = .511).
Table 1.
Baseline characteristics of study participants in training and validation sets.
| Variable names | Overall | Training set | Validation set | P |
|---|---|---|---|---|
| n | 14,745 | 10,321 | 4424 | |
| Age (yr) | 51.821 ± 17.626 | 51.786 ± 17.553 | 51.903 ± 17.797 | .712 |
| Sex (%) | .360 | |||
| Female | 7519 (50.99) | 5289 (51.25) | 2230 (50.41) | |
| Male | 7226 (49.01) | 5032 (48.75) | 2194 (49.59) | |
| Race (%) | .681 | |||
| Mexican_American | 1861 (12.62) | 1311 (12.70) | 550 (12.43) | |
| Non_Hispanic_Black | 3230 (21.91) | 2242 (21.72) | 988 (22.33) | |
| Non_Hispanic_White | 6192 (41.99) | 4320 (41.86) | 1872 (42.31) | |
| Other_Hispanic | 1525 (10.34) | 1068 (10.35) | 457 (10.33) | |
| Other_race | 1937 (13.14) | 1380 (13.37) | 557 (12.59) | |
| Marital_status (%) | .414 | |||
| Divorced | 1641 (11.13) | 1156 (11.20) | 485 (10.96) | |
| Living_with_partner | 1131 (7.67) | 772 (7.48) | 359 (8.11) | |
| Married | 7637 (51.79) | 5389 (52.21) | 2248 (50.81) | |
| Never_married | 2601 (17.64) | 1806 (17.50) | 795 (17.97) | |
| Separated | 503 (3.41) | 339 (3.28) | 164 (3.71) | |
| Widowed | 1232 (8.36) | 859 (8.32) | 373 (8.43) | |
| Education_level (%) | ||||
| 9_11th_grade | 2058 (13.96) | 1416 (13.72) | 642 (14.51) | .627 |
| College_graduate_or_above | 3433 (23.28) | 2431 (23.55) | 1002 (22.65) | |
| High_school_graduate | 3257 (22.09) | 2283 (22.12) | 974 (22.02) | |
| Less_than_9th_grade | 1471 (9.98) | 1032 (10.00) | 439 (9.92) | |
| Some_college_or_AA_degree | 4526 (30.70) | 3159 (30.61) | 1367 (30.90) | |
| Diabetes (%) | ||||
| No | 11,987 (81.30) | 8395 (81.34) | 3592 (81.19) | .854 |
| Yes | 2758 (18.70) | 1926 (18.66) | 832 (18.81) | |
| Cancer (%) | ||||
| No | 13,156 (89.22) | 9250 (89.62) | 3906 (88.29) | .018 |
| Yes | 1589 (10.78) | 1071 (10.38) | 518 (11.71) | |
| Smoking (%) | ||||
| No | 7987 (54.17) | 5602 (54.28) | 2385 (53.91) | .695 |
| Yes | 6758 (45.83) | 4719 (45.72) | 2039 (46.09) | |
| Hypertension (%) | ||||
| No | 8350 (56.63) | 5825 (56.44) | 2525 (57.08) | .486 |
| Yes | 6395 (43.37) | 4496 (43.56) | 1899 (42.92) | |
| BMI | 29.539 ± 6.968 | 29.559 ± 7 | 29.492 ± 6.895 | .591 |
| PIR | 2.476 ± 1.55 | 2.478 ± 1.553 | 2.471 ± 1.542 | .790 |
| MHR | 0.42 ± 0.25 | 0.42 ± 0.25 | 0.42 ± 0.24 | .687 |
| LHR | 1.56 ± 1.35 | 1.55 ± 1.34 | 1.57 ± 1.36 | .706 |
| NHR | 3.12 ± 2.15 | 3.11 ± 2.18 | 3.14 ± 2.10 | .765 |
| PHR | 185.3 ± 76.4 | 185.8 ± 76.7 | 184.4 ± 75.9 | .424 |
| CHD (%) | ||||
| Without | 13,905 (94.30) | 9742 (94.39) | 4163 (94.10) | .511 |
| With | 840 (5.70) | 579 (5.61) | 261 (5.90) | |
BMI = body mass index, CHD = coronary heart disease, LHR = lymphocyte-to-HDL cholesterol ratio, MHR = monocyte-to-HDL cholesterol ratio, NHR = neutrophil-to-HDL cholesterol ratio, PHR = platelet-to-HDL cholesterol ratio, PIR = poverty income ratio.
Table 2 summarizes baseline characteristics comparing participants with and without CHD in the training set. Participants with CHD were substantially older (69.7 ± 10.0 vs 50.7 ± 17.3 years, P < .001), with male sex more prevalent (64.77% vs 47.80%, P < .001). All HDL-related inflammatory biomarkers showed significant associations with CHD. HDL-C concentrations were lower in participants with CHD (1.25 ± 0.38 vs 1.39 ± 0.41 mmol/L, P < .001), while inflammatory ratios were elevated: MHR (0.54 ± 0.35 vs 0.42 ± 0.23, P < .001), LHR (2.05 ± 3.12 vs 1.55 ± 1.02, P < .001), and NHR (4.06 ± 2.89 vs 3.11 ± 1.77, P < .001).
Table 2.
Comparison of baseline characteristics between participants with and without CHD in training set.
| Variable names | Overall | Without CHD | With CHD | P |
|---|---|---|---|---|
| n | 10,321 | 9742 | 579 | |
| Age (yr) | 51.786 ± 17.553 | 50.723 ± 17.329 | 69.674 ± 10.032 | <.001 |
| Sex (%) | <.001 | |||
| Female | 5289 (51.25) | 5085 (52.20) | 204 (35.23) | |
| Male | 5032 (48.75) | 4657 (47.80) | 375 (64.77) | |
| Race (%) | <.001 | |||
| Mexican_American | 1311 (12.70) | 1264 (12.97) | 47 (8.12) | |
| Non_Hispanic_Black | 2242 (21.72) | 2180 (22.38) | 62 (10.71) | |
| Non_Hispanic_White | 4320 (41.86) | 3944 (40.48) | 376 (64.94) | |
| Other_Hispanic | 1068 (10.35) | 1021 (10.48) | 47 (8.12) | |
| Other_race | 1380 (13.37) | 1333 (13.68) | 47 (8.12) | |
| Marital_status (%) | <.001 | |||
| Divorced | 1156 (11.20) | 1094 (11.23) | 62 (10.71) | |
| Living_with_partner | 772 (7.48) | 747 (7.67) | 25 (4.32) | |
| Married | 5389 (52.21) | 5051 (51.85) | 338 (58.38) | |
| Never_married | 1806 (17.50) | 1778 (18.25) | 28 (4.84) | |
| Separated | 339 (3.28) | 328 (3.37) | 11 (1.90) | |
| Widowed | 859 (8.32) | 744 (7.64) | 115 (19.86) | |
| Education_level (%) | <.001 | |||
| 9_11th_grade | 1416 (13.72) | 1319 (13.54) | 97 (16.75) | |
| College_graduate_or_above | 2431 (23.55) | 2328 (23.90) | 103 (17.79) | |
| High_school_graduate | 2283 (22.12) | 2143 (22.00) | 140 (24.18) | |
| Less_than_9th_grade | 1032 (10.00) | 944 (9.69) | 88 (15.20) | |
| Some_college_or_AA_degree | 3159 (30.61) | 3008 (30.88) | 151 (26.08) | |
| Diabetes (%) | ||||
| No | 8395 (81.34) | 8059 (82.72) | 336 (58.03) | <.001 |
| Yes | 1926 (18.66) | 1683 (17.28) | 243 (41.97) | |
| Cancer (%) | ||||
| No | 9250 (89.62) | 8826 (90.60) | 424 (73.23) | <.001 |
| Yes | 1071 (10.38) | 916 (9.40) | 155 (26.77) | |
| Smoking (%) | ||||
| No | 5602 (54.28) | 5358 (55.00) | 244 (42.14) | <.001 |
| Yes | 4719 (45.72) | 4384 (45.00) | 335 (57.86) | |
| Hypertension (%) | ||||
| No | 5825 (56.44) | 5720 (58.71) | 105 (18.13) | <.001 |
| Yes | 4496 (43.56) | 4022 (41.29) | 474 (81.87) | |
| BMI | 29.559 ± 7 | 29.548 ± 7.043 | 29.755 ± 6.227 | .488 |
| PIR | 2.478 ± 1.553 | 2.481 ± 1.557 | 2.425 ± 1.481 | .397 |
| HDL_C (mmol/L) | 1.38 ± 0.41 | 1.39 ± 0.41 | 1.25 ± 0.38 | <.001 |
| MHR | 0.42 ± 0.25 | 0.42 ± 0.23 | 0.54 ± 0.35 | <.001 |
| LHR | 1.56 ± 1.35 | 1.55 ± 1.02 | 2.05 ± 3.12 | <.001 |
| NHR | 3.12 ± 2.15 | 3.11 ± 1.77 | 4.06 ± 2.89 | <.001 |
| PHR | 185.3 ± 76.4 | 186.2 ± 77.1 | 179.6 ± 68.7 | .049 |
BMI = body mass index, CHD = coronary heart disease, 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, PIR = poverty income ratio.
3.2. Feature selection and model performance
LASSO regression selected 15 features, including age, sex, race/ethnicity, education level, marital status, diabetes, cancer history, smoking status, hypertension, HDL-C, and all 4 HDL-related inflammatory ratios. Figure 2 illustrates the LASSO regularization path and cross-validation curve.
Six machine learning algorithms were evaluated on the independent validation set (Table 3). XGBoost demonstrated superior performance, achieving AUROC of 0.8892 (95% CI: 0.863–0.916), accuracy of 0.9655 (95% CI: 0.960–0.970), precision of 0.8600 (95% CI: 0.795–0.907), recall of 0.4942 (95% CI: 0.434–0.554), F1-score of 0.6274 (95% CI: 0.473–0.781), specificity of 0.9946 (95% CI: 0.992–0.996), and Matthews correlation coefficient of 0.6369 (95% CI: 0.598–0.676). Random forest achieved an AUROC of 0.8375 (95% CI: 0.807–0.868), followed by logistic regression (AUROC = 0.8259, 95% CI: 0.795–0.857).
Table 3.
Performance metrics of different machine learning models on test set.
| Model | AUROC (95% CI) | Accuracy (95% CI) | Precision (95% CI) | Recall (95% CI) | F1-score (95% CI) | Specificity (95% CI) | MCC (95% CI) |
|---|---|---|---|---|---|---|---|
| XGBoost | 0.8892 (0.863–0.916) | 0.9655 (0.960–0.970) | 0.8600 (0.795–0.907) | 0.4942 (0.434–0.554) | 0.6274 (0.473–0.781) | 0.9946 (0.992–0.996) | 0.6369 (0.598–0.676) |
| Random forest | 0.8375 (0.807–0.868) | 0.9600 (0.954–0.965) | 0.7800 (0.710–0.840) | 0.4300 (0.370–0.492) | 0.5500 (0.450–0.650) | 0.9930 (0.990–0.995) | 0.5800 (0.520–0.640) |
| Logistic regression | 0.8259 (0.795–0.857) | 0.9580 (0.952–0.964) | 0.7500 (0.680–0.815) | 0.4000 (0.342–0.461) | 0.5200 (0.420–0.620) | 0.9920 (0.989–0.994) | 0.5500 (0.490–0.610) |
| GBDT | 0.8150 (0.783–0.847) | 0.9560 (0.950–0.962) | 0.7200 (0.650–0.785) | 0.3800 (0.323–0.440) | 0.5000 (0.400–0.600) | 0.9910 (0.988–0.993) | 0.5200 (0.460–0.580) |
| KNN | 0.7800 (0.746–0.814) | 0.9450 (0.938–0.951) | 0.6500 (0.575–0.720) | 0.3200 (0.267–0.378) | 0.4300 (0.340–0.520) | 0.9880 (0.985–0.991) | 0.4500 (0.385–0.515) |
| Decision tree | 0.7400 (0.705–0.775) | 0.9380 (0.931–0.945) | 0.6000 (0.520–0.675) | 0.2800 (0.230–0.335) | 0.3800 (0.290–0.470) | 0.9850 (0.982–0.988) | 0.4000 (0.330–0.470) |
AUROC = area under the receiver operating characteristic curve, CI = confidence interval, GBDT = gradient boosting decision tree, KNN = K-nearest neighbors, MCC = Matthews correlation coefficient, XGBoost = eXtreme gradient boosting.
The combination of high precision (86%) and specificity (99.46%) with moderate recall (49.42%) indicates optimal use as a “rule-in” test. The high precision suggests that 86 out of 100 individuals predicted as having CHD actually have the condition, making the model suitable for identifying high-risk individuals requiring further evaluation or intensive intervention. The high specificity minimizes false positives, reducing unnecessary anxiety and downstream testing in low-risk individuals. However, the moderate recall indicates that approximately half of CHD cases would be identified at the optimal threshold, suggesting utility as a complementary rather than standalone assessment tool.
3.3. Model evaluation and interpretability
Figure 3 presents receiver operating characteristic curves, calibration plots, and DCA. XGBoost demonstrated superior discriminative performance with good calibration. DCA confirmed clinical utility across threshold probabilities of 5% to 30%.
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.
SHAP analysis revealed feature importance rankings (Fig. 4). Age emerged as the most influential predictor, followed by hypertension and diabetes. Among HDL-related inflammatory biomarkers, MHR and NHR ranked as the fourth and sixth most important features, demonstrating a substantial contribution beyond traditional risk factors. Elevated inflammatory ratios consistently contributed to increased CHD probability, while HDL-C demonstrated an inverse association.
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.
4. Discussion
4.1. Principal findings
This cross-sectional study of 14,745 US adults demonstrates that machine learning models incorporating HDL-related inflammatory biomarkers achieved high discrimination for classifying prevalent CHD status (AUROC = 0.8892). Among 6 algorithms evaluated, XGBoost demonstrated superior performance with 86% precision and 96.55% accuracy. SHAP analysis identified age, hypertension, and diabetes as primary predictors, while MHR and NHR ranked among the top 5 features, indicating a substantial contribution beyond traditional risk factors. These findings reveal cross-sectional associations between HDL-related inflammatory ratios – calculable from routine laboratory tests – and prevalent CHD, though the temporal relationship remains to be established.
Our findings extend prior research by simultaneously evaluating all 4 HDL-related ratios using machine learning in a nationally representative sample.[28,29] Previous studies using traditional statistical methods reported AUROCs of 0.60 to 0.75; our higher discrimination (AUROC = 0.8892) likely reflects ensemble algorithm advantages, though differences in study populations and outcome definitions limit direct comparison.[30] Unlike neutrophil-to-lymphocyte ratio and platelet-to-lymphocyte ratio, which reflect inflammatory cell balance alone, HDL-related ratios integrate anti-inflammatory lipid metabolism with pro-inflammatory cellular activation. This may explain their contribution to CHD classification, though mechanistic validation is needed.[31]
4.2. HDL-related inflammatory ratios and established biomarkers
HDL-related inflammatory ratios showed associations with prevalent CHD in this cross-sectional analysis. While hs-CRP remains widely studied, it requires a dedicated immunoassay. HDL-related ratios derive from routine complete blood count and lipid panels with no additional cost, potentially facilitating implementation. However, whether these ratios capture mechanistic pathways or merely correlate with CHD requires further investigation. Direct comparison with hs-CRP was not possible in our dataset.
Whether HDL-related ratios provide incremental value beyond hs-CRP and established lipid biomarkers remains unclear.[32,33] The 2025 American College of Cardiology Scientific Statement emphasizes inflammatory biomarkers in cardiovascular assessment, but prospective validation is required before considering clinical integration.[34] Our cross-sectional design cannot establish whether combining marker types improves incident event prediction.
4.3. Machine learning and model interpretability
XGBoost’s ensemble gradient boosting methodology captures nonlinear relationships and complex feature interactions.[35,36] The 86% precision indicates high reliability when classifying individuals as having prevalent CHD. The moderate recall (49.42%) reflects the challenge of identifying low-prevalence conditions (5.7%) while maintaining high specificity (99.46%). In a prospective setting, this performance profile might suggest utility as a “rule-in” test, though validation with incident CHD outcomes is required.
SHAP analysis provides interpretability for complex machine learning models.[37,38] The prominence of MHR and NHR alongside traditional risk factors demonstrates their contribution beyond established predictors in this cross-sectional analysis. Whether these associations reflect causal relationships or confounding requires prospective investigation. Interpretability is important for potential clinical adoption, as transparent models may enhance clinician trust.[39]
4.4. Clinical implications
Our model’s performance profile has important clinical implications. The high precision (86%, 95% CI: 0.795–0.907) and specificity (99.46%, 95% CI: 0.992–0.996) indicate strong “rule-in” capability: individuals predicted as high risk have an 86% probability of truly having CHD, justifying intensive intervention while minimizing false positives (0.54%). However, the moderate recall (49.42%, 95% CI: 0.434–0.554) suggests that approximately half of CHD cases would be identified, necessitating complementary rather than standalone use. This performance profile supports application for risk stratification in intermediate-risk populations, where additional diagnostic clarity guides treatment decisions.
Prospective validation is required before clinical implementation. HDL-related ratios may be affected by acute illness, which transiently elevates inflammatory cells.[40] If validated prospectively, optimal measurement timing and threshold optimization would need to be established for different clinical contexts. Electronic health record integration could facilitate automated calculation, though clinical utility must be demonstrated first.[41]
If prospective studies demonstrate predictive value for incident CHD, potential applications might include risk stratification in intermediate-risk patients and resource-limited settings where specialized biomarkers are unavailable. The combined cost of routine tests (under $50) may offer accessibility. However, clinical implementation requires: prospective validation with incident outcomes; demonstration of improved clinical outcomes beyond discrimination metrics; and validation across diverse populations.[42]
4.5. Strengths and limitations
Study strengths include a large sample size, nationally representative data, rigorous feature selection methodology, comprehensive machine learning algorithm comparison, and SHAP-based interpretability analysis.
From a methodological perspective, several limitations warrant consideration. First, external validation in independent cohorts was not performed. Future studies should validate these models across diverse populations and healthcare settings.[43] Second, SMOTE oversampling was applied to address class imbalance, which may introduce distribution distortion. We mitigated this by applying SMOTE only within cross-validation folds during training while maintaining the original distribution in the validation set. Third, machine learning analyses used unweighted data rather than incorporating NHANES survey weights. Current machine learning implementations do not readily accommodate complex survey weights when combined with SMOTE. Our focus was on identifying feature relationships for classification rather than population-level prevalence estimation.
The cross-sectional design, however, fundamentally limits causal inference. We report associations between HDL-related inflammatory biomarkers and prevalent CHD, not the prediction of future CHD development.[44] Whether elevated inflammatory ratios precede CHD development, result from CHD and its treatments, or reflect shared risk factors remains unclear. Prospective cohort studies tracking incident cardiovascular events are essential to establish temporal relationships and true predictive value.
Self-reported CHD diagnosis introduces substantial misclassification bias through multiple mechanisms. First, recall bias may occur if participants inaccurately remember or report their diagnosis. Anxiety may lead to overreporting of symptoms as “heart disease,” while denial or lack of awareness may cause underreporting. Second, diagnostic heterogeneity exists as participants may interpret the question “has a doctor told you that you have CHD” differently. Some may include conditions beyond CHD (e.g., arrhythmias, valvular disease, heart failure), leading to false positives. Others with diagnosed CHD may not report it accurately. Third, undiagnosed CHD in the control group represents a critical source of misclassification. Participants with subclinical or asymptomatic CHD would be incorrectly classified as non-CHD, diluting observed differences and attenuating associations between biomarkers and true CHD status. Fourth, treatment effects complicate interpretation. Most diagnosed CHD patients receive cardiovascular medications – statins alter lipid profiles, including HDL-C, antiplatelet agents affect platelet counts, and anti-inflammatory effects of some medications may reduce inflammatory cell counts. These treatments likely reduce differences between groups, making it difficult to assess baseline biomarker levels independent of therapeutic intervention.
Validation studies suggest 70% to 85% concordance between self-report and medical records for major cardiovascular conditions, meaning 15% to 30% misclassification could substantially bias effect estimates toward the null.[45] The direction of bias depends on misclassification patterns: false positives (non-CHD reported as CHD) would dilute case characteristics, while false negatives (undiagnosed CHD in controls) would weaken observed associations. These limitations fundamentally constrain our ability to establish true associations between baseline inflammatory biomarker levels and CHD status. Future studies using objective diagnostic criteria (coronary angiography demonstrating ≥50% stenosis, prior myocardial infarction on electrocardiogram or imaging, or medical record verification with specific international classification of diseases codes) would strengthen validity and enable assessment of biomarker levels before treatment initiation.
HDL-related inflammatory ratios reflect dynamic biological processes influenced by acute illness, medications, diurnal variation, and measurement timing.[46] Single-timepoint measurements may not capture the chronic inflammatory state. Cardiovascular medications (statins, antiplatelet agents, antihypertensives) alter both inflammatory cells and HDL-C, complicating interpretation in CHD patients already receiving treatment.[47] This bidirectional relationship between treatment and biomarkers limits assessment of baseline risk independent of interventions.
The moderate recall (49.42%) indicates that our model classified only half of actual CHD cases at the optimal threshold. This performance profile, combined with the cross-sectional design, means that the model cannot reliably identify individuals who will develop future CHD. External validation in independent cohorts with prospective follow-up for incident CHD is necessary to establish true predictive value.
4.6. Future directions
Future research priorities include prospective cohort studies with adjudicated incident CHD to establish temporal relationships and true predictive value; direct comparison with established risk scores (Pooled Cohort Equations and PREVENT) in the same cohort[48]; serial measurements to assess biomarker stability; and cost-effectiveness analyses if predictive value is established.[49]
If prospective validation demonstrates clinical utility, implementation research should examine workflow integration, algorithmic fairness across demographic groups, and health equity implications.[50]
5. Conclusion
Machine learning models incorporating HDL-related inflammatory biomarkers demonstrated high discrimination (AUROC = 0.8892) for classifying prevalent CHD status in this cross-sectional analysis. MHR and NHR contributed substantially beyond traditional risk factors as accessible biomarkers from routine laboratory tests. However, critical limitations constrain interpretation and potential applications. The cross-sectional design precludes causal inference and cannot establish whether elevated ratios precede CHD development or result from disease and treatment. Self-reported diagnosis introduces substantial misclassification bias (15%–30% discordance). These findings represent associations with prevalent CHD rather than prediction of future cardiovascular events. Prospective cohort studies with adjudicated incident CHD outcomes are essential to establish true predictive value and temporal relationships. Only if validated prospectively could HDL-related inflammatory ratios serve as complementary tools for cardiovascular risk stratification, and such applications would require demonstration of improved clinical outcomes beyond discrimination metrics.
Acknowledgments
The authors thank the National Center for Health Statistics and the Centers for Disease Control and Prevention for designing and conducting the NHANES and making the data publicly available.
Author contributions
Conceptualization: Guoxin Zhang.
Methodology: Guoxin Zhang.
Investigation: Yimei Cai, Guoxin Zhang.
Validation: Yimei Cai.
Abbreviations:
- AUROC
- area under the receiver operating characteristic curve
- CHD
- coronary heart disease
- CI
- confidence interval
- DCA
- decision curve analysis
- HDL
- high-density lipoprotein
- HDL-C
- high-density lipoprotein cholesterol
- hs-CRP
- high-sensitivity C-reactive protein
- LASSO
- least absolute shrinkage and selection operator
- LHR
- lymphocyte-to-HDL cholesterol ratio
- MHR
- monocyte-to-HDL cholesterol ratio
- NHANES
- National Health and Nutrition Examination Survey
- NHR
- neutrophil-to-HDL cholesterol ratio
- PHR
- platelet-to-HDL cholesterol ratio
- SHAP
- SHapley Additive exPlanations
- SMOTE
- Synthetic Minority Oversampling Technique
- XGBoost
- eXtreme gradient boosting
This research was supported by the Nanjing City Health Science and Technology Development Special Fund Project (Grant No.: YKK24233).
The NHANES protocol was approved by the National Center for Health Statistics (NCHS) Research Ethics Review Board, and all participants provided written informed consent. The current study utilized publicly available, deidentified data from NHANES and, therefore, did not require additional institutional review board approval according to federal regulations for research involving human subjects (45 CFR 46.104).
The authors have no conflicts of interest to disclose.
The datasets generated and/or analyzed during the current study are publicly available.
How to cite this article: Cai Y, Zhang G. High-density lipoprotein-related inflammatory ratios and coronary heart disease: A cross-sectional machine learning analysis of NHANES 2009 to 2020. Medicine 2026;105:13(e48214).
References
- [1].World Heart Federation. World Heart Report 2023. World Heart Federation; 2023. [Google Scholar]
- [2].Tsao CW, Aday AW, Almarzooq ZI, et al. Heart disease and stroke statistics-2023 update: a report from the American Heart Association. Circulation. 2023;147:e93–e621. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [3].Roth GA, Mensah GA, Johnson CO, et al. Global burden of cardiovascular diseases and risk factors, 1990–2019: update from the GBD 2019 study. J Am Coll Cardiol. 2020;76:2982–3021. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [4].Khan SS, Matsushita K, Sang Y, et al. Development and validation of the American Heart Association’s PREVENT equations. Circulation. 2024;149:430–49. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [5].Arnett DK, Blumenthal RS, Albert MA, et al. 2019 ACC/AHA guideline on the primary prevention of cardiovascular disease: a report of the American College of Cardiology/American Heart Association Task Force on clinical practice guidelines. Circulation. 2019;140:e596–646. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [6].DeFilippis AP, Young R, Carrubba CJ, et al. An analysis of calibration and discrimination among multiple cardiovascular risk scores in a modern multiethnic cohort. Ann Intern Med. 2015;162:266–75. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [7].Attiq A, Afzal S, Ahmad W, Kandeel M. Hegemony of inflammation in atherosclerosis and coronary artery disease. Eur J Pharmacol. 2024;966:176338. [DOI] [PubMed] [Google Scholar]
- [8].Ridker PM, Everett BM, Thuren T, et al. Antiinflammatory therapy with canakinumab for atherosclerotic disease. N Engl J Med. 2017;377:1119–31. [DOI] [PubMed] [Google Scholar]
- [9].Endo Y, Fujita M, Ikewaki K. HDL functions-current status and future perspectives. Biomolecules. 2023;13:105. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [10].Rosenson RS, Brewer HB, Jr, Ansell BJ, et al. Dysfunctional HDL and atherosclerotic cardiovascular disease. Nat Rev Cardiol. 2016;13:48–60. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [11].Annink ME, Kraaijenhof JM, Stroes ESG, Kroon J. Moving from lipids to leukocytes: inflammation and immune cells in atherosclerosis. Front Cell Dev Biol. 2024;12:1446758. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [12].Ganjali S, Gotto AM, Jr, Ruscica M, et al. Monocyte-to-HDL-cholesterol ratio as a prognostic marker in cardiovascular diseases. J Cell Physiol. 2018;233:9237–46. [DOI] [PubMed] [Google Scholar]
- [13].Canpolat U, Çetin EH, Cetin S, et al. Association of monocyte-to-HDL cholesterol ratio with slow coronary flow is linked to systemic inflammation. Clin Appl Thromb Hemost. 2016;22:476–82. [DOI] [PubMed] [Google Scholar]
- [14].Arabi A, Abdelhamid A, Nasrallah D, et al. Monocyte-to-HDL ratio (MHR) as a novel biomarker: reference ranges and associations with inflammatory diseases and disease-specific mortality. Lipids Health Dis. 2025;24:343. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [15].Juhasz V, Charlier FT, Zhao TX, Tsiantoulas D. Targeting the adaptive immune continuum in atherosclerosis and post-MI injury. Atherosclerosis. 2024;399:118616. [DOI] [PubMed] [Google Scholar]
- [16].Chuang SM, Liu SC, Chien MN, Lee CC, Lee YT, Chien KL. Neutrophil-to-high-density lipoprotein ratio (NHR) and neutrophil-to-lymphocyte ratio (NLR) as prognostic biomarkers for incident cardiovascular disease and all-cause mortality: a comparison study. Am J Prev Cardiol. 2024;20:100869. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [17].Wu W, Jia C, Xu X, et al. Impact of platelet-to-HDL-cholesterol ratio on long-term mortality in coronary artery disease patients with or without type 2 diabetes: insights from a Chinese multicenter cohort. J Inflamm Res. 2024;17:2731–44. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [18].Krittanawong C, Johnson KW, Rosenson RS, et al. Deep learning for cardiovascular medicine: a practical primer. Eur Heart J. 2019;40:2058–73. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [19].Lundberg SM, Lee SI. A unified approach to interpreting model predictions. Adv Neural Inf Process Syst. 2017;30:4766–77. [Google Scholar]
- [20].Akinbami LJ, Chen TC, Davy O, et al. National Health and Nutrition Examination Survey, 2017–March 2020 prepandemic file: sample design, estimation, and analytic guidelines. Vital Health Stat 1. 2022;2:1–36. [PubMed] [Google Scholar]
- [21].Okura Y, Urban LH, Mahoney DW, Jacobsen SJ, Rodeheffer RJ. Agreement between self-report questionnaires and medical record data was substantial for diabetes, hypertension, myocardial infarction and stroke but not for heart failure. J Clin Epidemiol. 2004;57:1096–103. [DOI] [PubMed] [Google Scholar]
- [22].Miller WG, Myers GL, Sakurabayashi I, et al. Seven direct methods for measuring HDL and LDL cholesterol compared with ultracentrifugation reference measurement procedures. Clin Chem. 2020;56:977–86. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [23].Whelton PK, Carey RM, Aronow WS, et al. 2017 ACC/AHA/AAPA/ABC/ACPM/AGS/APhA/ASH/ASPC/NMA/PCNA guideline for the prevention, detection, evaluation, and management of high blood pressure in adults. J Am Coll Cardiol. 2018;71:e127–248. [DOI] [PubMed] [Google Scholar]
- [24].American Diabetes Association Professional Practice Committee. 2. Classification and diagnosis of diabetes: standards of care in diabetes-2024. Diabetes Care. 2024;47:S20–42. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [25].Hastie T, Tibshirani R, Friedman JH. The Elements of Statistical Learning: Data Mining, Inference, and Prediction. 2nd ed. Springer; 2009. [Google Scholar]
- [26].Chen T, Guestrin C. XGBoost: a scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York, NY: ACM; 2016:785–94. [Google Scholar]
- [27].Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Making. 2006;26:565–74. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [28].He L, Chen S, Zhu X, He F. The change of inflammatory markers may predict long-term major adverse cardiovascular events in elderly patients with coronary heart disease: a retrospective cohort study. Front Med (Lausanne). 2025;11:1523581. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [29].Song J, Li W, Zhang M, et al. Association of three composite inflammatory and lipid biomarkers with cardiovascular-kidney-metabolic syndrome: findings from NHANES 2011–2020. Mediators Inflamm. 2025;2025:6691516. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [30].Ogunpola A, Saeed F, Basurra S, Albarrak AM, Qasem SN. Machine learning-based predictive models for detection of cardiovascular diseases. Diagnostics (Basel). 2024;14:144. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [31].Zhu L, Fu Y, Zhu L, Yao Y, Chen L. Identification of routine blood derived hematological and lipid indices in ILD through machine learning: a retrospective case-control study. Front Med (Lausanne). 2025;12:1633713. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [32].Ridker PM. From C-reactive protein to interleukin-6 to interleukin-1: moving upstream to identify novel targets for atheroprotection. Circ Res. 2016;118:145–56. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [33].Dziedzic EA, Gasior JS, Koseska K, et al. The impact of neutrophil-to-high-density lipoprotein ratio and serum 25-hydroxyvitamin D on ischemic heart disease. J Clin Med. 2024;13:6597. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [34].Mensah GA, Arnold N, Prabhu SD, Ridker PM, Welty FK. Inflammation and cardiovascular disease: 2025 ACC scientific statement: a report of the American College of Cardiology. J Am Coll Cardiol. 2025;86:1841–901. [DOI] [PubMed] [Google Scholar]
- [35].Liu T, Krentz A, Lu L, Curcin V. Machine learning based prediction models for cardiovascular disease risk using electronic health records data: systematic review and meta-analysis. Eur Heart J Digit Health. 2025;6:7–22. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [36].Zhu X, Zhang M, Wen Y, Shang D. Machine learning advances the integration of covariates in population pharmacokinetic models: valproic acid as an example. Front Pharmacol. 2022;13:994665. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [37].Ponce-Bobadilla AV, Schmitt V, Maier CS, Mensing S, Stodtmann S. Practical guide to SHAP analysis: explaining supervised machine learning model predictions in drug development. Clin Transl Sci. 2024;17:e70056. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [38].Jin Y, Li X, Han X, et al. SHAP-based interpretable machine learning for Parkinson’s disease severity prediction: integrated analysis of clinical and environmental features. Front Neurol. 2025;16:1678463. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [39].Imrie F, Davis R, van der Schaar M. Multiple stakeholders drive diverse interpretability requirements for machine learning in healthcare. Nat Mach Intell. 2023;5:824–9. [Google Scholar]
- [40].Wei Y, Gao H, Luo Y, et al. Systemic inflammation and oxidative stress markers in patients with unipolar and bipolar depression: a large-scale study. J Affect Disord. 2024;346:154–66. [DOI] [PubMed] [Google Scholar]
- [41].Samir M, Sherief N, Abdelmoez W. Improving bug assignment and developer allocation in software engineering through interpretable machine learning models. Computers. 2023;12:128. [Google Scholar]
- [42].Khan SU, Lone AN, Khan MS, et al. PREVENT equations to estimate 10-year and lifetime cardiovascular disease risk. Circulation. 2024;149:e347–57.38264914 [Google Scholar]
- [43].Xu H, Guan Y, Yang C, et al. Artificial intelligence in the risk prediction models of cardiovascular disease and development of an independent validation screening tool: a systematic review. BMC Med. 2024;22:56. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [44].Conrad N, Verbeke G, Molenberghs G, et al. Autoimmune diseases and cardiovascular risk: a population-based study on 19 autoimmune diseases and 12 cardiovascular diseases in 22 million individuals in the UK. Lancet. 2022;400:733–43. [DOI] [PubMed] [Google Scholar]
- [45].Jurek AM, Greenland S, Maldonado G, Church TR. Proper interpretation of non-differential misclassification effects: expectations vs observations. Int J Epidemiol. 2005;34:680–7. [DOI] [PubMed] [Google Scholar]
- [46].Pearson TA, Mensah GA, Alexander RW, et al. Markers of inflammation and cardiovascular disease: application to clinical and public health practice: a statement for healthcare professionals from the Centers for Disease Control and Prevention and the American Heart Association. Circulation. 2003;107:499–511. [DOI] [PubMed] [Google Scholar]
- [47].Ridker PM, Bhatt DL, Pradhan AD, Glynn RJ, MacFadyen JG, Nissen SE; PROMINENT, REDUCE-IT, and STRENGTH Investigators. Inflammation and cholesterol as predictors of cardiovascular events among patients receiving statin therapy: a collaborative analysis of three randomised trials. Lancet. 2023;401:1293–301. [DOI] [PubMed] [Google Scholar]
- [48].Banach M, Toth PP, Bielecka-Dąbrowa A, Lewek J. Primary and secondary cardiovascular prevention: recent advances. Kardiol Pol. 2024;82:1200–10. [DOI] [PubMed] [Google Scholar]
- [49].Tanguay W, Acar P, Fine B, et al. Assessment of radiology artificial intelligence software: a validation and evaluation framework. Can Assoc Radiol J. 2023;74:326–33. [DOI] [PubMed] [Google Scholar]
- [50].Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. [DOI] [PMC free article] [PubMed] [Google Scholar]