Systematic review and meta-analysis of prediction models for diabetic kidney disease progression to ESRDKidney disease prediction models show promise but have major flaws

BackgroundDiabetic kidney disease (DKD) is a major cause of end-stage renal disease (ESRD). Early identification of DKD patients at high risk of progressing to ESRD is essential, yet the overall performance, methodological quality, and translational readiness of prediction models for this transition remain unclear. To our knowledge, we conducted the first systematic review and meta-analysis focused specifically on prediction models for progression from established DKD to ESRD.MethodsWe searched PubMed, Embase, Cochrane Library, Web of Science, China National Knowledge Infrastructure (CNKI), Wanfang, VIP Chinese Journal Service Platform, and the Chinese Biomedical Literature Database (CBM) for English- and Chinese-language studies published through 27 September 2025 that developed or validated DKD to ESRD prediction models. Two reviewers independently screened records and extracted data. Risk of bias was assessed using the updated PROBAST-AI checklist. Reported AUCs were pooled using random-effects meta-analysis (Stata 18.0) with 95% confidence intervals (CIs). We performed sensitivity analyses, assessed publication bias, and conducted subgroup analyses by inclusion of pathological predictors and prediction horizon.ResultsFifteen studies met inclusion criteria. PROBAST-AI judged all 15 studies at high risk of bias, mainly due to retrospective single-center designs, lack of blinded predictor assessment, use of predictors not routinely available in practice, and inadequate calibration and external validation; seven studies were limited to biopsy-proven DKD, limiting their applicability to routine clinical populations. Meta-analysis included training datasets from six studies (seven models) and validation datasets from three studies (three models). Pooled AUCs were 0.896 (95% CI, 0.853–0.940) for training models and 0.863 (95% CI, 0.803–0.923) for validation models. Five prespecified sensitivity analyses yielded broadly similar pooled AUCs, but interpretation remained exploratory because of persistent heterogeneity and universal high risk of bias. Subgroup analyses found no significant differences by pathological predictor inclusion or prediction horizon.ConclusionsPublished DKD to ESRD models show promising discrimination in development and internal validation cohorts. However, pervasive methodological limitations, extreme heterogeneity, and scarce independent external validation severely restrict clinical generalizability. Prespecified sensitivity analyses yielded broadly similar pooled AUCs, but these results remain exploratory. Future prospective multicenter studies with rigorous external validation and calibration are urgently needed.Systematic Review Registrationhttps://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD420251127778, identifier CRD420251127778.