An Interpretable Machine Learning Framework for Predicting Disease Progression in Early-Stage Cardiovascular-Kidney-Metabolic(CKM) Syndrome: A Prospective Cohort Study

<title>Abstract</title> <p> <bold>Objective:</bold> Cardiovascular-kidney-metabolic (CKM) syndrome constitutes a continuum of metabolic, renal, and cardiovascular dysfunction, yet practical tools for stratifying progression risk in its early, modifiable stages remain absent. We aimed to develop an interpretable machine learning (ML) framework to estimate stage-advancement risk among individuals with early-stage CKM syndrome, emphasizing calibration stability and subgroup heterogeneity. <bold>Methods:</bold> This prospective cohort study used data from the China Health and Retirement Longitudinal Study (CHARLS). A total of 3,773 participants aged ≥45 years with baseline (2011) CKM stages 0–2 and complete follow-up through 2015 were included. Progression was defined as an increase of at least one CKM stage. Six ML models were evaluated through a three-stage selection process that sequentially prioritized discrimination, calibration stability (calibration slope and Integrated Calibration Index [ICI]), and clinical utility (Decision Curve Analysis). SHapley Additive exPlanations (SHAP) were applied to interpret model predictions. Subgroup analyses were stratified by baseline CKM stage. <bold>Results:</bold> Over the 4-year period, 1,560 participants (41.3%) progressed. This rate substantially exceeded the 15–34% stage-transition frequencies documented in Western cohorts, reflecting marked dynamic instability of early CKM in Chinese middle-aged and older adults. LASSO retained eight predictors: age, BMI, SBP, HbA1c, eGFR, baseline CKM stage, sex, and education. CatBoost yielded the highest discrimination (test AUC = 0.772) but was poorly calibrated (calibration slope = 1.305; ICI = 0.060). Random Forest (RF) was selected as the optimal model given its calibration stability (slope = 1.056; ICI = 0.026) and highest net benefit. Global SHAP analysis identified baseline CKM stage, HbA1c, and SBP as the leading predictors and uncovered a non-monotonic inverse stage effect. Subgroup analyses by baseline stage revealed a predictive ceiling effect in Stages 0–1 (AUCs ~0.60), attributable to progression rates exceeding 70%, while calibration remained acceptable across all stages. <bold>Conclusions:</bold> We present a well-calibrated, interpretable RF framework for risk stratification. The ceiling effect observed in early CKM stages indicates that individuals in Stages 0–1 may benefit more from immediate preventive care than from further conventional risk stratification. </p>