Abstract Objective This systematic review aimed to systematically evaluate the methodological quality and predictive performance of existing prognostic models for diabetic kidney disease (DKD) outcomes. Methods A systematic search of PubMed, Web of Science, Cochrane Library, Ovid Embase, China National Knowledge Infrastructure (CNKI), Wanfang, Vip and SinoMed was conducted from database inception to 30 April 2026. Two reviewers independently screened studies, extracted data using the CHARMS checklist and assessed risk of bias with PROBAST. Results A total of 27 studies were included, comprising 80 prediction models. Among them, two were prospective studies and 25 were retrospective studies. The primary outcomes included renal‐related events, all‐cause mortality and cardiovascular endpoints. Twenty‐two studies used traditional regression methods, five applied machine learning algorithms and four combined both approaches for model development. The most frequently used predictors were age, estimated glomerular filtration rate (eGFR), urine–albumin–creatinine ratio (UACR), urinary total protein (UTP), body mass index (BMI), haemoglobin and systolic blood pressure (SBP). Only eight studies performed external validation, and most models showed acceptable discrimination, with area under the curve (AUC)/C‐statistic ranging from 0.70 to 0.90. All studies were judged to have a high risk of bias, and 21 were considered to have high concerns regarding applicability, mainly due to single‐centre design and restrictive inclusion criteria. Conclusions Current DKD prognostic models demonstrate acceptable discrimination but lack adequate calibration and external validation, highlighting the need for more rigorous study designs to enhance reliability and clinical applicability.
Yang et al. (2026) studied this question.