A Community-Based Machine Learning Model for Predicting Major Adverse Cardiovascular Events in Kazakh Hypertensive Patients

Technical Summary: A Community-Based Machine Learning Model for Predicting Major Adverse Cardiovascular Events in Kazakh Hypertensive Patients

Problem Statement
China faces a significant public health burden with approximately 240 million individuals suffering from hypertension, where cardiovascular diseases remain the leading cause of death. The Kazakh ethnic group in Xinjiang exhibits the highest prevalence of hypertension in the region, characterized by distinct lifestyle and dietary habits (e.g., high salt intake, meat-heavy diet, nomadic lifestyle) that contribute to a high frequency of Major Adverse Cardiovascular Events (MACEs). Existing risk prediction models often lack specificity for this ethnic minority, and healthcare resources in rural and pastoral areas of Xinjiang are unevenly distributed, making early prevention and individualized management difficult. There is a critical need for a validated, interpretable prediction model tailored specifically to the Kazakh hypertensive population to facilitate primary prevention strategies.

Methodology
The study utilized a prospective cohort design involving 1,180 Kazakh hypertensive patients registered at the "Hypertension Control Center" of Altay City People's Hospital and three township health centers. Data collection occurred from January to December 2020, with prospective follow-up continuing until December 31, 2024 (median follow-up: 1,176 days).

• Data Collection: Baseline data included demographics, medical history, physical examinations (office and 24-hour ambulatory blood pressure), and biochemical indicators (lipid profile, glucose, homocysteine, electrolytes, renal function). The primary outcome was the first occurrence of MACEs (unstable angina, non-fatal myocardial infarction, heart failure, ischemic stroke, or death), verified by ICD-10 codes.
• Feature Selection: Predictors were selected through a three-phase process: univariate analysis, Least Absolute Shrinkage and Selection Operator (LASSO) regression, and clinical expert validation. Eight final predictors were identified: age, diabetes status, serum potassium, fasting blood glucose, homocysteine (Hcy), total cholesterol (TC), low-density lipoprotein cholesterol (LDL-C), and office systolic blood pressure.
• Model Development: The dataset was randomly split into a training set (70%) and a validation set (30%). Four machine learning and statistical models were constructed and compared: Cox regression, Random Forest, Decision Tree (DT), and Extreme Gradient Boosting (XGBoost).
• Evaluation Metrics: Performance was assessed using Area Under the Curve (AUC), Accuracy (ACC), Positive/Negative Predictive Values (PPV/NPV), F1-score, calibration curves, and Decision Curve Analysis (DCA).
• Interpretability: The optimal model was interpreted using Shapley Additive Explanations (SHAP) to quantify feature importance and visualize risk contributions. A web-based calculator was developed using the R Shiny framework to facilitate clinical application.

Key Results

• Incidence: During the follow-up period, 327 participants (27.7%) developed MACEs.
• Model Performance:
  • The Decision Tree (DT) model demonstrated the most robust and balanced performance. In the validation set, it achieved an AUC of 0.776, Accuracy of 0.797, PPV of 0.444, NPV of 0.933, and an F1-score of 0.550.
  • XGBoost achieved the highest AUC in the validation set (0.829) but suffered from low precision (PPV = 0.242) and a high false-positive rate, resulting in a lower F1-score (0.390).
  • Random Forest and Cox Regression showed inferior generalizability, with Random Forest yielding the lowest AUC (0.577) and Cox regression showing poor accuracy (0.599) despite high PPV.
• Calibration and Utility: The DT model exhibited excellent calibration (Mean Absolute Error ≈ 0.021) and maintained a stable net benefit across threshold probabilities of 0% to 75% in the Decision Curve Analysis, indicating strong clinical utility.
• Feature Importance (SHAP): The SHAP analysis ranked the top four risk factors as:
  1. LDL-C: The most influential predictor; higher levels correlated with increased risk.
  2. Age: Risk increased progressively with advancing age.
  3. Office Systolic Blood Pressure: Higher values were associated with elevated risk.
  4. Serum Potassium: Lower levels (hypokalemia) were associated with increased risk, while higher levels were protective.
     Other significant factors included Hcy, fasting blood glucose, TC, and diabetes status.

Significance and Claims
The paper claims that the developed Decision Tree-based model effectively predicts the 3-year risk of MACEs specifically for Kazakh hypertensive patients in the Altay region. The study highlights that:

1. Ethnic Specificity: The model addresses the gap in existing literature by focusing on the unique risk profile of the Kazakh population, incorporating factors relevant to their lifestyle (e.g., specific dietary impacts on potassium and salt intake).
2. Clinical Applicability: Unlike complex ensemble models that may act as "black boxes," the Decision Tree model offers an intuitive structure suitable for primary care settings with limited resources. Its high Negative Predictive Value (NPV) allows for the effective identification of low-risk patients, potentially reducing unnecessary interventions.
3. Interpretability: The integration of SHAP analysis provides transparency, confirming that the model's predictions align with established clinical knowledge (e.g., the protective effect of normal potassium levels and the risk associated with elevated LDL-C and age).
4. Deployment: To bridge the gap between research and practice, the authors have deployed the model as an accessible web-based calculator, aiming to support individualized prevention and treatment strategies in community and clinical settings.

The study concludes that while machine learning offers flexibility, the Decision Tree model provided the best balance of predictive accuracy, stability, and clinical interpretability for this specific demographic, serving as a scientific basis for targeted cardiovascular risk management.