Assessment of risk prediction models for chronic kidney disease: a global perspective

Background Numerous risk prediction models for chronic kidney disease (CKD) had been developed and validated worldwide. It was imperative to offer a comprehensive overview and comparative analysis for their performance. Methods PubMed, Embase, and Web of Science databases were searched to identify studies developing and/or validating CKD risk prediction models. The basic characteristics, discrimination (AUC), and calibration (calibration slope) of these models were summarized and compared. The Prediction Model Risk Of Bias Assessment Tool (PROBAST) was applied to evaluate the Risk of Bias (ROB) of the included prediction models across four dimensions: participant, prediction, outcome, and analysis. Each dimension was classified as high-risk, low-risk, or uncertain risk. Results After an initial screening of 14,149 studies, 26 studies on 18 distinct prediction models were identified. Among them, six models specifically targeted patients with type 2 diabetes, while the other 12 models focused on high-risk or general populations. The main modeling algorithms were logistic regression and Cox proportional hazards regression. There were a total of 43 predictive factors which spanned six categories, encompassing demographic information, comorbidities, lab tests, lifestyle factors, medication use, and specific indicators. The number of predictive factors for each model ranged from 3 to 18. In terms of discrimination for 18 models, the AUCs of internal validation ranged from 0.680 to 0.920, whereas the AUCs of external validation ranged from 0.623 to 0.918. Fourteen models reported the calibration, and 7 of them provided the calibration slope. The calibration slope of internal validation ranged from 0.880 to 1.129, whereas the calibration slope of external validation ranged from 0.841 to 1.842. However, the PROBAT indicated that 15 of the 18 models had high ROB. It was primarily due to overlapping predictors and outcomes, improper handling of missing data, lack of internal validation or reliance on simple random splits, and/or inadequate performance evaluation. Conclusions Although the risk prediction models for CKD had been developed and validated, these models exhibited significant heterogeneity. Future modeling studies should avoid using eGFR as both a predictive factor and outcome determination, consider appropriate missing data processing in the study population, and follow the reporting guidelines. PROSPERO registration number: CRD420251270491.