VALIDATION AND COMPARISON OF CARDIOVASCULAR RISK PREDICTION MODELS IN CHINESE PATIENTS WITH CHRONIC KIDNEY DISEASE

Kidney Disease: Improving Global Outcomes (KDIGO) guidelines recommended chronic kidney disease (CKD)-specific cardiovascular disease (CVD) risk prediction models and kidney measurement-enhanced CVD models for patients with CKD. However, these models still lack validation and comparison within the Asian populations, which limits their further widespread application in this demographic. 

We validated and compared 10 CVD risk prediction models in two large Chinese CKD cohorts: the Chinese Cohort Study of Chronic Kidney Disease (C-STRIDE) based on CKD patients from tertiary hospitals and the China Kidney Disease Network (CK-NET)-Yinzhou study based on CKD patients from a community-based population. We defined two types of CVD outcomes: algorithm-specific outcomes adhering to each model’s original endpoint definitions, and a unified endpoint including myocardial infarction, stroke, heart failure, fatal coronary heart disease, fatal stroke and other cardiovascular deaths for cross-model comparisons. Model performance was evaluated by discrimination and calibration. Discrimination was evaluated by the concordance index (C-index); calibration was assessed by calibration plot and ratio of observed and expected outcome (O/E ratio). To compare the models in the context of clinical relevance, we assessed the categorization of participants across different models with net reclassification improvement (NRI) at the current recommended threshold for statin treatment in the American college of cardiology and the American Heart Association guideline (7.5%), and the integrated discrimination improvement (IDI), using the Add-uACR PREVENT model as the reference model. Decision curve analysis (DCA) was used to quantify the clinical implications. 

In the C-STRIDE cohort, the C-index ranged from 0.682 to 0.749 for algorithm-specific outcomes, and from 0.670 to 0.753 for the unified outcome. In the CK-NET-Yinzhou cohort, the C-index ranged from 0.691 to 0.738 for algorithm-specific outcomes, and from 0.676 to 0.733 for the unified outcome. In both cohorts, the Add-uACR PREVENT model and SCORE2 CKD Add-on model showed slightly better discrimination. In the C-STRIDE cohort, the C-index values for these two models were 0.746 and 0.749 for algorithm-specific outcomes, and 0.753 and 0.745 for the unified outcome. In the CK-NET-Yinzhou cohort, the C-index values for these two models were 0.733 and 0.738 for algorithm-specific outcomes, and 0.733 and 0.726 for the unified outcome. Significant calibration differences were observed between cohorts. In the C-STRIDE cohort, all models overestimated CVD risk for both algorithm-specific and unified outcomes, with SCORE2 demonstrating the best calibration (O/E ratio: 0.526 for algorithm-specific outcomes; 0.856 for unified outcomes). Recalibration improved calibration across all models (O/E ratios range: 0.956-1.016). In contrast, the CK-NET-Yinzhou cohort showed distinct patterns: under algorithm-specific outcomes, most models overestimated risk (particularly in high-risk patients), with SCORE2 and SCORE2 eGFR Add-on model performing optimally (O/E ratios: 1.130 and 0.959, respectively). When unified outcomes were applied, models generally underestimated risk in low-risk patients while overestimating it in high-risk subgroups, where the Add-uACR PREVENT and CRIC model2 showed superior calibration (O/E ratios: 1.154 and 1.153, respectively). NRI and IDI illustrated that the predictive performance of the Add-uACR PREVENT model is better than other models. The Add-uACR PREVENT model also exhibits higher net benefit than other models in the DCA. 

CVD models recommended by KDIGO guidelines, especially the Add-uACR PREVENT model, exhibit superior predictive performance compared to traditional CVD risk prediction models in Chinese patients with CKD. This study provided critical evidence for implementing aforementioned CVD risk prediction models in Chinese patients with CKD.