A Machine Learning-Based Model for Cardiovascular Disease Risk Stratification in Middle-Aged and Older Adults with Diabetes: A Cross-Sectional Study based on CHARLS

Technical Summary: A Machine Learning-Based Model for Cardiovascular Disease Risk Stratification in Middle-Aged and Older Adults with Diabetes

Problem Statement
Cardiovascular disease (CVD) is the leading cause of mortality among patients with diabetes mellitus (DM). While numerous CVD prediction models exist, there is a critical shortage of tools specifically calibrated for middle-aged and older adults with diabetes. Traditional statistical models (e.g., Cox proportional hazards, linear regression) often rely on rigid linear assumptions and population-level averages, failing to capture the high-dimensional, non-linear, and heterogeneous pathophysiological patterns of individual patients. Furthermore, existing machine learning (ML) approaches often suffer from a "black-box" nature that hinders clinical trust, and many lack dedicated tools for this specific high-risk demographic. This study aims to bridge these gaps by developing an interpretable, ML-driven risk assessment model tailored for diabetic patients aged 45 and older.

Methodology

• Data Source: The study utilized a cross-sectional analysis of the 2020 wave of the China Health and Retirement Longitudinal Study (CHARLS), a nationally representative cohort.
• Cohort: From an initial pool of 19,395 participants, the study selected 2,625 individuals diagnosed with diabetes (aged 45–85) after excluding those without DM, those outside the age range, and those with extreme sleep duration reports. The primary outcome was self-reported physician-diagnosed CVD (including myocardial infarction, coronary heart disease, angina, heart failure, or stroke).
• Feature Selection: Starting with 171 variables, the authors applied a dual-method feature selection strategy on the training set to prevent data leakage:
  1. LASSO Regression: Reduced variables to 17 candidates.
  2. Boruta Algorithm: A wrapper method based on Random Forest to identify relevant features.
  3. Intersection: The final model utilized the intersection of both methods, resulting in 8 robust clinical variables: Age, Self-Rated Health (SRH), Activities of Daily Living (ADL), Hypertension, Dyslipidemia, Renal Disease, Hospitalization within the past year, and Chest Pain.
• Model Development: Six machine learning algorithms were trained and compared: Logistic Regression (LR), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest (RF), Light Gradient Boosting Machine (LightGBM), and eXtreme Gradient Boosting (XGBoost). Hyperparameters were optimized via grid search with five-fold cross-validation.
• Interpretability & Deployment: The optimal model was interpreted using SHapley Additive exPlanations (SHAP) to provide global and local feature importance. The final model was deployed as an interactive, web-based visual decision support system using the Streamlit framework.

Key Results

• Model Performance: Among the six algorithms, the Random Forest (RF) model demonstrated the most robust performance on the independent test set.
  • AUC: 0.752 (95% CI: 0.716–0.786).
  • Sensitivity: 72.53% (superior to other models, crucial for early screening).
  • Specificity: 61.64%.
  • Accuracy: 66.12%.
  • The RF model showed strong generalization with minimal overfitting, maintaining an AUC of 0.773 on the training set.
• Feature Importance (SHAP Analysis): The SHAP analysis identified the hierarchy of risk factors influencing CVD prediction:
  1. Hypertension (highest impact)
  2. Dyslipidemia
  3. Self-Rated Health (SRH)
  4. Age
  5. ADL Score
  6. Renal Disease
  7. Hospitalization in the preceding year
  8. Chest Pain
  • The analysis confirmed that conditions like hypertension, dyslipidemia, and renal disease positively correlated with higher CVD risk, while better SRH and younger age correlated with lower risk.
• Clinical Tool: The study successfully translated the model into a user-friendly web application where clinicians can input patient parameters to receive an instant, personalized risk probability accompanied by a visual breakdown of contributing factors (SHAP force plot).

Significance and Claims
The authors claim this study addresses a specific research gap by providing a dedicated, explainable risk stratification tool for middle-aged and older diabetic patients, a group often overlooked by general CVD models. Key contributions include:

1. Methodological Rigor: The use of a dual-feature selection approach (LASSO + Boruta) on a high-quality, nationally representative dataset (CHARLS) to ensure model parsimony and robustness.
2. Interpretability: Moving beyond "black-box" predictions by integrating SHAP analysis, which elucidates the decision-making process and aligns with established clinical knowledge (e.g., the strong link between hypertension/dyslipidemia and CVD).
3. Clinical Utility: The development of a zero-cost, rapid screening tool that relies on non-invasive, easily accessible data (self-reported history, ADL scores) rather than complex laboratory biomarkers. This makes the tool particularly suitable for resource-limited primary care and community-based settings.
4. Personalized Care: The system facilitates precision interventions by generating individualized risk profiles, potentially aiding in early triage and motivating patients toward early intervention.

The authors acknowledge limitations, noting that the cross-sectional design precludes causal inference and that the model's generalizability to non-East Asian populations requires further validation. Additionally, the reliance on self-reported data introduces potential subjective bias. However, the study concludes that the proposed RF-based system offers a practical, accessible approach to improving cardiovascular risk management in the diabetic population.