Interpretable machine learning for multiclass trajectory prediction in cardiovascular-kidney-metabolic syndrome stage: development and external validation

Background Cardiovascular-kidney-metabolic (CKM) syndrome imposes a substantial global burden, yet most studies reduce stage change to a binary outcome that cannot distinguish clinically meaningful trajectories. This study developed and externally validated a multiclass machine learning framework for predicting CKM stage trajectories. Methods A longitudinal cohort of 2,971 adults with baseline CKM stages 0–3 was used for development, with trajectories classified as Improvement, Stable, Mild Progression and Rapid Progression. Six algorithms were compared by Macro-AUC and decision curve analysis. SHAP values identified consensus predictors for a parsimonious model, externally validated in an independent hospital-based cohort (N = 291). Results XGBoost achieved the highest Macro-AUC (0.786, 95% CI 0.757–0.813), with the six algorithms performing comparably. For rapid-progression screening, the primary model reached a sensitivity of 0.767, a specificity of 0.739 and a negative predictive value of 0.949 at a threshold of 0.23, and identified 41.9% of rapid progressors within the highest-risk 10%. Baseline CKM stage was the dominant prognostic determinant, with a stage-only benchmark reaching 0.689. Six consensus predictors were identified: baseline CKM stage, age, fasting glucose, TyG-BMI index, systolic blood pressure and triglycerides. The parsimonious model retained over 98% of full-model discrimination and reached a Macro-AUC of 0.752 externally without retraining. Conclusion This framework distinguishes CKM trajectories and is intended for screening rapid progression rather than assigning individuals to a single trajectory. Baseline CKM stage is the principal prognostic determinant, and the six-predictor model preserves discrimination using routinely available measurements, supporting stratified follow-up pending confirmation in larger cohorts.