Social determinants of health substantially enhance identification of high-risk CKM syndrome progression: a longitudinal prediction study based on data from CHARLS

Abstract

Cardiovascular-Kidney-Metabolic (CKM) syndrome represents a critical public health challenge, with a large portion of the Chinese population classified as CKM Stage 2. While the impact of Social Determinants of Health (SDoH) on chronic disease is known, their longitudinal role in longitudinal CKM stage progression remains unclear. We aimed to apply a hybrid statistical–machine-learning framework to predict 4-year CKM stage progression, with predictive modeling intentionally focused on individuals with CKM Stage 2 at baseline, as this was the largest subgroup and a clinically important transitional stage for preventive risk stratification, and to quantify the predictive contribution of SDoH features. This longitudinal study utilized data from the 2011 and 2015 waves of the China Health and Retirement Longitudinal Study (CHARLS). The primary analytical cohort consisted of 4,963 participants classified as CKM Stage 2 at baseline; this stage was selected a priori because it represented the largest subgroup and a clinically actionable transition point within the CKM continuum. The outcome was defined as 4-year progression to Stage 3 or 4. A hybrid feature-selection strategy, combining univariable/multivariable logistic regression with SHAP-based ranking, was applied to 95 initial features to derive an optimized 14-feature set. Four standard prediction algorithms (RF, LR, SVM, and GBDT) were then trained and evaluated. Four machine learning models, including Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), and Gradient Boosting Decision Tree (GBDT), were developed using a 10-fold cross-validation pipeline incorporating SMOTEENN for imbalance and GridSearchCV for tuning. Models were evaluated on an independent test set (30% of data). A sensitivity analysis was performed by removing SDoH features to assess their contribution. On the independent test set, all models demonstrated good discriminatory power, with the RF model achieving the highest area under the Receiver Operating Characteristic curve (AUC) (0.76) and the LR model showing the highest Recall (0.78). SHAP analysis identified SDoH (e.g., Ratio_New (Price-to-income ratio for new housing), Terrain, Memory) and demographics (Gender, Age) as key predictors, often with complex interaction effects. The contribution analysis showed that SDoH features substantially improved identification of progression, particularly by increasing Recall despite only modest changes in AUC. Removing SDoH features caused a substantial decline in Recall (e.g., RF: 0.75 to 0.47; LR: 0.78 to 0.58), while overall AUC remained stable (e.g., RF: 0.76 to 0.74). Incorporating SDoH features substantially improves risk stratification performance for 4-year progression from CKM Stage 2 to Stage 3 or 4. In the ablation analysis, removing selected SDoH-related features substantially reduced sensitivity for identifying progression cases within the final model framework, despite similar AUC values. These findings suggest that incorporating selected SDoH-related features may enhance screening-oriented prediction of four-year CKM stage progression from Stage 2 to Stage 3 or 4, particularly when the goal is to identify individuals at higher risk.