The inclusion criteria were as follows: 1) study subjects aged ≥18 years; 2) participants confirmed as Chinese patients with type 2 diabetes mellitus according to diabetes diagnostic guidelines or expert consensus; 3) participants diagnosed with diabetic kidney disease according to the diagnostic guidelines for diabetic kidney disease or expert consensus; 4) studies focused on the construction and validation of risk prediction models for diabetic kidney disease in patients with type 2 diabetes mellitus; 5) study types include cohort studies, cross-sectional studies, and case–control studies; and 6) articles written in Chinese or English.

The exclusion criteria were as follows: 1) studies that only analyzed risk factors without establishing a risk prediction model; 2) studies with incomplete data or where the original texts are inaccessible; 3) reviews, systematic reviews, and meta-analyses; and 4) conference papers.

A systematic search was conducted in PubMed, Web of Science, the Cochrane Library, Embase, China National Knowledge Infrastructure (CNKI), Wanfang Data, Chinese Biomedical Literature Database (CBM), and VIP Chinese Scientific Journals Full-text Database for studies on the construction and validation of risk prediction models for diabetic kidney disease among Chinese patients with type 2 diabetes mellitus. The search was limited to articles published up to 28 December 2023 and included both Chinese and English languages. The search strategy combined the use of MeSH terms and free words, supplemented by manual searches to include references traced back. The search terms included diabetes mellitus, type 2, type 2 diabetes, diabetic nephropathy, diabetic nephropathies, diabetic kidney disease, risk prediction, risk assessment, risk factors, predict, validate, model*, tool*, scale*, and score*. The specific search strategy is shown in Figure 1 .

In our systematic review, we employed the PICOTS framework as recommended by the Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS) (10). This framework assists in defining the review’s objectives, search strategy, and criteria for including and excluding studies:

Two researchers, trained in evidence-based courses, independently screened the literature, eliminating duplicate publications of titles and abstracts. Selection was based on the inclusion and exclusion criteria, followed by cross-checking. In case of disagreement, a third researcher was consulted to make a decision, ultimately determining the included studies.

The CHARMS checklist was first proposed in 2014 by Douglas G. Altman of the University of Oxford and Karel G. M. Moons of the University of Utrecht, among others. It is designed to clarify the issues of systematic reviews and analyze the methodology and quality of original studies (10). Researchers used the CHARMS checklist to formulate a data extraction form for extracting characteristics of the literature. Specific items extracted included the first author, year of publication, type of study, modeling methods, validation methods, calibration methods, predictive performance, predictive factors, etc. (11). Two researchers independently extracted the information from the literature, and in case of disagreement, a third researcher was consulted for resolution.

The Prediction Model Risk of Bias Assessment Tool (PROBAST) was developed in 2019 by Wolff and colleagues as a specialized tool for assessing studies that develop or validate diagnostic or prognostic multifactorial prediction models (12). The risk of bias assessment covers four domains: participants, predictors, outcome, and analysis. Each domain includes two to nine key questions, answered with “yes/probably yes,” “no/probably no,” or “no information,” corresponding to the judgment of “low risk of bias,” “high risk of bias,” or “unclear.” If all questions in a domain are judged as “low risk of bias,” that domain is considered to have a “low risk of bias.” If one or more questions are judged as “high risk of bias” or “unclear,” the domain is considered to have a “high risk of bias” or to be “unclear.” The applicability assessment covers three domains: participants, predictors, and outcomes, excluding the key questions, with evaluation outcomes similar to those of the risk of bias (13, 14). Two researchers used the PROBAST tool to assess the risk of bias and applicability of the included prediction model studies. In cases of disagreement, a third researcher was consulted for a decision.

In this study, STATA 14.0 software was used to perform meta-analysis. Heterogeneity among studies was evaluated using the Higgins I² statistic. An I² value of ≤25% indicates low heterogeneity, 25% < I² ≤ 50% indicates moderate heterogeneity, and I² > 50% indicates high heterogeneity. If P > 0.05 and I² ≤ 50%, the heterogeneity is considered acceptable, and a fixed-effect model is used; otherwise, a random-effects model is employed to combine the effect sizes. Furthermore, the area under the curve (AUC) of the receiver operating characteristic (ROC) curve is used as the measure of effect size, with specific values and 95% confidence intervals (CI) provided. An AUC range of 0.7 to 0.9 indicates moderate predictive accuracy, while an AUC > 0.9 suggests high diagnostic accuracy. A P < 0.05 is considered statistically significant. The Egger’s test was used to identify publication bias. If the P-value is greater than 0.05, it indicates a low likelihood of publication bias. Finally, a funnel plot was created (15). Funnel plot inspection is a commonly used statistical tool in meta-analysis to assess the reliability and consistency of the study results. It visually represents the relationship between effect size and study size, helping researchers identify potential publication bias and other forms of heterogeneity. In a funnel plot, each small dot represents a study or a set of data. The horizontal axis measures the study results (such as effect size or sample size), while the vertical axis represents each study’s weight or confidence (such as standard error or sample variance). Ideally, if all studies are accurate and unbiased, these dots should form a symmetrical funnel shape, with small sample studies scattering around larger sample studies, which typically show smaller effect sizes and tighter clustering. An asymmetrical or skewed funnel plot may indicate the presence of publication bias, where small sample studies (possibly because they are more likely to report extreme effect sizes) are missing on one side of the plot. Alternatively, this asymmetry might be due to other factors like selective reporting or methodological quality differences, which distort the relationship between effect size and study size. When a funnel plot shows asymmetry, researchers need to investigate the possible causes further and consider the impact of these factors on the study conclusions.

The database search yielded a total of 28,809 articles. After removing duplicates using the Note Express software, 18,337 articles remained. Subsequent screening of titles and abstracts reduced the pool to 759 articles. After a full-text review, 32 articles were ultimately included in the study (16–47). Figure 2 shows the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flow diagram, illustrating the comprehensive literature screening process and results.

Among the 32 studies included, 12 were published domestically (16–25), and 20 were published internationally (26–47). The most contributions came from scholars in Jiangsu Province, with a total of seven studies (18, 20, 28, 30, 38, 40, 41). The earliest studies were published in 2017 (35, 40), and there were 23 studies published in the last three years (18, 20, 21, 23–32, 35, 37–39, 41, 43, 45, 46). In terms of study types, there were 10 case–control studies (16, 17, 21, 24, 25, 30, 33, 36, 45, 47), 5 cross-sectional studies (19, 22, 34, 35, 39), and 16 cohort studies (18, 20, 23, 27–29, 31, 32, 37, 38, 40, 41, 43, 44, 46, 47). The specific characteristics of the included studies are detailed in Table 1 .

The sample sizes of the studies included ranged from 102 to 141,516, with the incidence of diabetic nephropathy in patients with type 2 diabetes ranging from 6.0% to 62.3%. Modeling methods included logistic regression analysis, LASSO regression, Cox regression, machine learning, and R language. The range of the area under the curve (AUC) for the models was from 0.713 to 0.949. A total of 31 studies (16–24, 26–47) conducted internal validation. Specific details on the model construction methods and predictive efficacy of the included studies are provided in Table 2 .

Among the 32 studies, 30 reported the presentation form of the prediction models, mainly divided into nomograms, model equations, machine learning models, risk score charts, and neural networks, among others. The most frequently occurring predictive factors were hypertension, urinary protein, creatinine, duration of diabetes, age, and fasting blood glucose. The distribution of predictive factors for each model can be seen in Figure 3 .

This study systematically reviews existing risk prediction models for diabetic nephropathy in patients with type 2 diabetes in China. The aims are to evaluate the overall performance of these models and provide additional information to help select specific screening tools for prediabetic conditions. Most models include similar predictors, but there are differences in the weights assigned to each component. The study found that common predictors include hypertension, urine protein, creatinine, duration of diabetes, age, and fasting blood glucose. These factors are low cost and easily accessible.

Sixteen studies (20, 23, 24, 27, 31–34, 36–38, 40, 42, 44, 45, 47) consider hypertension a predictive factor for the risk of diabetic nephropathy in patients with type 2 diabetes. Sustained high blood pressure increases the burden on the kidneys. Compression of small arteries leads to arteriosclerosis of the renal arteries, thickening of the vascular wall, narrowing of the lumen, and increased glomerular filtration pressure and filtration rate, further causing increased albumin filtration. Renal lesions can also exacerbate hypertension, creating a vicious cycle, making hypertension a significant factor in the development of diabetic nephropathy (48). Advance et al. (49) found that lowering blood pressure can reduce vascular complications of diabetes by 10%. Therefore, for patients with diabetic nephropathy and hypertension, strict blood pressure control is crucial for protecting renal function in diabetic nephropathy patients. Optimizing blood pressure control can reduce the incidence of end-stage renal disease and adverse cardiovascular events. Therefore, patients should monitor their blood pressure under a doctor’s guidance as well as regularly review kidney function, urinary microalbumin, and other relevant indicators. If necessary, they should undergo individualized antihypertensive treatment under a doctor’s guidance. Dietary measures should focus on a low-salt diet and eating smaller, more frequent meals.

Thirteen studies (16, 18, 21, 24–26, 28, 30, 31, 33, 37, 39, 41) consider urine protein a predictive factor for the risk of diabetic nephropathy in patients with type 2 diabetes. When urine protein is positive, it indicates that protein is being lost through the urine, suggesting a disruption in the glomerular barrier function (50). Therefore, urine protein can provide valuable information for the diagnosis of diabetic nephropathy. Morphological studies by Said et al. (51) have shown that moderate increases in proteinuria are associated with glomerular and interstitial tubular damage in some diabetic patients. Therefore, patients should adhere to a low-salt diet, with a daily salt intake less than 6 g, avoiding salty pickles and preserved products and aiming for a bland diet. Moreover, moderate protein intake is a core and critical issue in the dietary management of diabetic nephropathy. Excessive protein intake can increase the glomerular filtration rate, promote thickening of the glomerular basement membrane, and increase kidney damage. Therefore, individuals with diabetic nephropathy should avoid high-protein diets and appropriately increase the proportion of high-quality protein in their diet.

Thirteen studies (16, 23, 24, 26, 28, 30–32, 34, 37, 40, 41, 46) have identified creatinine as a predictive factor for the risk of diabetic nephropathy in patients with type 2 diabetes. Creatinine, which is primarily filtered freely by the glomeruli, serves as a marker of kidney function impairment (26). When renal excretion is compromised, urates accumulate in the kidneys, leading to the buildup of creatinine and subsequent kidney damage (52). Early increases in creatinine levels are reversible; during this period, it is crucial to control blood sugar and reduce damaging factors. Additionally, medication, including traditional Chinese medicine or Western drugs, can help the body excrete excess metabolic products. In the case of treatment during the uremic phase, blood dialysis, peritoneal dialysis, and kidney transplantation are required.

Ten studies (17, 19, 22–25, 27, 44, 45, 47) have identified the duration of diabetes as a predictive factor for the risk of diabetic nephropathy in patients with type 2 diabetes. The duration of diabetes is a major factor in the occurrence of microvascular complications. As the duration of the disease increases, long-term dysregulation of glucose and lipid metabolism can cause irreversible damage to the kidneys. This leads to a continuous increase in urinary microalbumin, resulting in a gradual loss of glomerular function and the occurrence of end-stage renal events (33). Morton et al. (53) indicated that an extended duration of diabetes increases the risk of end-stage renal disease. For T2DM patients with a long history of the disease, it is important to take early measures such as lowering blood pressure, reducing lipid levels, adopting a low-salt diet, and improving kidney function to reduce the risk of developing diabetic nephropathy.

Twelve studies (17–19, 22, 28, 30–32, 36, 40, 45, 47) have identified age as a predictive factor for the risk of developing diabetic nephropathy in patients with type 2 diabetes. With increasing age, the reserve function of the kidneys gradually diminishes. After the age of 40, the glomerular filtration rate decreases by 1% with each additional year of age, meaning that between the ages of 50 and 90, the glomerular filtration rate can decrease by approximately 50%. Qiao et al. (54) found that the average age of the diabetic nephropathy patient group was 57.03 ± 12.425 years, which is approximately 6 years older than the average age of the non-diabetic nephropathy group. Therefore, for diabetic patients, greater attention should be paid to changes in kidney function as they age, including regular health check-ups. Additionally, a balanced diet, regular exercise, and stress relief are important. Many elderly individuals have chronic diseases requiring medication, but many drugs are nephrotoxic, so medication should be taken under the guidance of a professional physician.

Eight studies (21, 23, 32, 34, 35, 38, 42, 47) have identified fasting blood glucose as a predictive factor for the risk of developing diabetic nephropathy in patients with type 2 diabetes. Fasting blood glucose is the most fundamental marker of the progression of type 2 diabetes (34). Persistent hyperglycemia can induce mesangial expansion, leading to thickening of the basement membrane and hardening of the glomerular capillary walls, ultimately resulting in proteinuria and kidney damage. Furthermore, sustained high blood sugar levels cause the glomeruli to be in a state of high perfusion and filtration, leading to arteriosclerosis and changes in vascular permeability. This eventually results in progressive vascular proliferation, hyalinization, and intravascular thrombosis due to the extensive deposition of plasma proteins on the capillary walls (55). Lou et al. (56) found that for every 1 mmol/L increase in fasting blood glucose, the risk of proteinuria increases by 1.15 times. Therefore, medical professionals should make recommendations on the frequency of checks based on the patient’s blood sugar levels to ensure they are maintained within the target range. Dietarily, adherence to a low-sugar, high-fiber diet is advised, including consuming more fruits, non-starchy vegetables, whole grains, and legumes. Intake of saturated fats, processed meats, sweets, and sodium should be limited. Moreover, patients should follow their doctor’s advice on choosing the appropriate oral hypoglycemic drugs or insulin injections.

Generally, the performance metrics included in the models primarily report the sensitivity indicator AUC. Typically, the larger the AUC value, the better the model’s predictive ability. Among the 32 studies included in this research, 18 articles (16–18, 20, 21, 24, 26–30, 32, 35, 38, 42, 43, 45, 47) reported an AUC greater than 0.8. Furthermore, the results of the meta-analysis show a combined AUC value of 0.810 (0.780–0.840), indicating that the overall predictive performance of the existing risk prediction models for diabetic nephropathy in Chinese patients with type 2 diabetes is relatively good. Discrimination, calibration, and the clinical utility of the model are important parameters for evaluating predictive models. This study includes information on model calibration results and more from 22 articles (16–18, 21, 23, 24, 27, 28, 31, 33, 34, 36–42, 44–47). Moreover, only seven articles (21, 33, 36, 37, 39, 41, 44) conducted external validations, with the remaining articles lacking such verification; thus, the generalizability of these findings requires further validation. Conducting external studies on risk prediction models is crucial as it assesses the model’s generalizability and robustness across different populations and settings, thereby enhancing the model’s credibility and practical value. Additionally, external validation can reveal the model’s performance on new datasets, helping to identify potential overfitting and bias issues, thus promoting further optimization and improvement of the model (57). In terms of modeling methods, 21 articles (17–19, 21–25, 27, 30, 33, 34, 36–39, 42, 44–47) utilized logistic regression models, 6 articles (16, 20, 26, 29, 40, 41) used machine learning models, and 7 articles (23, 24, 28, 33, 34, 42, 45) employed LASSO regression models. Among these, 4 articles (31, 32, 35, 43) compared traditional models with machine learning models and found that the predictive performance and application effects of machine learning models were superior to those of traditional models. This indicates that prediction models based on machine learning have a promising future and high application value. Future researchers could develop risk prediction models based on machine learning within an interpretability framework to further improve the sensitivity, accuracy, and credibility of the models.

All 32 predictive models included in the analysis were assessed as having a high risk of bias, which limits their clinical applicability. The bias mainly stems from the type of study subjects; 27 studies (16–18, 21, 23–33, 35–47) were retrospective and deemed to have a high risk of bias. In terms of statistical analysis, 20 studies (16, 17, 20, 21, 23–28, 30, 33, 37–39, 41, 43–46) were conducted as single-center studies or only within the same region, 5 studies (27, 29, 31, 32, 37, 38) did not clearly mention how missing data were handled, and 15 studies (19–21, 23, 24, 29–31, 35, 37, 39–42, 47) used random split validation methods for model internal or external validation. Moreover, 12 studies (16, 17, 19, 21–23, 25, 26, 38, 41, 42, 46) had insufficient sample sizes for model development and validation. For future risk prediction model development and validation studies, it is recommended that researchers refer to the PROBAST standards to reduce bias in the research process and data analysis (58). Overall, although the models demonstrate moderate to good performance, the risk of bias remains high. Future improvements are needed in the type of studies conducted, the time interval between predictor assessment and outcome determination, the number of events, handling of missing values, and methods for selecting predictors.

In prediction model research, there is a direct relationship between the overall sample size and the number of participants with specific outcomes. In model development studies, sample size is often determined by the events per variable (EPV). An EPV of at least 10 to 20 is widely accepted as a standard to minimize overfitting (59). However, this study found issues with small sample sizes in the models. Therefore, it is necessary to use larger sample sizes for future model development and validation. A larger sample size yields more precise results, with smaller standard errors and narrower confidence intervals. It also helps ensure the model’s generalizability, meaning it can accurately predict outcomes in new datasets. Using a sufficiently large sample size allows researchers to capture more variability and potential confounders, leading to more accurate estimates of predictor effects. Additionally, an adequate sample size enhances the model’s stability and reliability. Thus, future research should strive to ensure sufficiently large sample sizes to support model development and validation (60). Furthermore, this study found that some models were developed using single-center studies, which may limit the models’ broad applicability. Future researchers should prioritize multicenter studies. Multicenter studies, especially those conducted across multiple regions, increase sample diversity and representativeness, reduce biases inherent in single-center studies, and improve the generalizability and extrapolability of the findings. They allow researchers to validate the interventions’ effectiveness across different cultural, environmental, and medical settings, promote interdisciplinary and international collaboration, and provide more comprehensive data for policy-making and public health practices. Despite challenges in coordination, data management, and cultural differences, multicenter studies are widely adopted for providing richer and more reliable scientific evidence.

A good prediction model should include not only the actual influencing factors but also statistical indicators that reflect the model’s effectiveness. However, these statistical indicators’ prediction and calibration estimates can vary significantly among different disease populations. When applying a prediction model to different populations, recalibration may be necessary. This means that when using the prediction model on new and diverse populations, we need to adjust the model to ensure its accuracy and reliability. This may involve reassessing the weights of risk factors in the model, updating statistical indicators, and performing necessary calibration or validation steps. Only then can we ensure that the prediction model can accurately predict the risk of diabetic nephropathy in other type 2 diabetes populations.

Due to the lack of established expert consensus or guidelines, there is no standardized method for the systematic evaluation of risk prediction models. The methods used in this study to summarize the reliability and clinical applicability of the included models are primarily based on existing research (61). Systematic evaluation of models is still in the exploratory stage, indicating the need for more prediction model studies. Although some models have undergone external validation, there is high heterogeneity when studying different regions or populations (62). Therefore, caution must be exercised, and thorough evaluation and adjustment must be conducted when applying risk prediction models to clinical settings or populations different from the original model (63). We recommend future studies conduct external validation of models in different centers and countries to ensure their generalizability. Additionally, we suggest adjusting these models based on local environments and study populations to achieve better predictive performance. Further research is needed to determine effective strategies for integrating models into clinical practice and to establish a unified method for systematically evaluating prediction models.

In recent years, risk prediction models have become a hotspot in clinical research. Early screening and identification of high-risk populations help healthcare professionals develop targeted preventive measures, improve the utilization of medical resources, and enhance patient outcomes. This study’s methodological and quality evaluation of risk prediction models found that statistical analysis is a major factor contributing to the overall risk of bias in research. Future researchers could follow the PROBAST tool to standardize the reporting of study items. In terms of model-building methods, employing machine learning algorithms to establish disease risk prediction models has become a research focus in the fields of computing and medicine. Models built with machine learning algorithms outperform those created with logistic regression methods in predictive performance (64). Machine learning offers new possibilities for complex medical predictions, especially when dealing with multidimensional, non-linear, and interrelated data. Therefore, applying machine learning algorithms to improve model performance is an important direction for future research. Further exploration and application of machine learning models may enhance prediction accuracy and provide patients with more personalized risk assessment and management strategies (65). Regarding the presentation of models, converting models into probability calculation equations, simplified scoring system tables, line charts, or online calculators makes them more clinically applicable. Integrating these with electronic medical record systems meets healthcare professionals’ needs for model use. This approach supports the use of complex algorithms while reducing the workload of manual calculations, facilitating practical clinical operations. In terms of study design, risk prediction models should be constructed using prospective studies. Prospective studies can collect data in real time, reduce recall bias, and ensure a clear temporal sequence of data, thereby improving the accuracy and reliability of predictions. Such study designs can better control confounding factors and provide higher-quality evidence, helping to establish more precise causal relationships during model development. This ultimately enhances the model’s effectiveness and generalizability in practical applications (66). Since the models included in the literature have mostly not been validated with large samples externally, their extrapolation ability is somewhat hindered. Therefore, future studies could employ large-scale, multicenter research in China to externally validate models and explore the construction of predictive models best suited to the Chinese population. Moreover, healthcare professionals could leverage data mining technology and artificial intelligence to establish a risk prediction information platform, improving the accuracy of model predictions to enhance clinical work efficiency.

One of the strengths of this study is that it systematically retrieved, screened, and extracted a wide range of data, including general characteristics, predictors, and study outcomes, to compare and summarize the performance of the models. However, in conducting the literature search and selection for this study, due to language restrictions, we only searched for literature in Chinese and English. While evaluating and analyzing the study results, we considered all available literature but did not account for regional biases, which may lead to potential publication bias. Also, considering the differences caused by genetics among different ethnic groups, this study only summarized existing research in China and provided information on predictive models. This may limit the generalizability of the study results to Western populations, and adjustments may be necessary when applying these models to different regions. Moreover, our focus was on the development, validation, and methodological description of the models, without evaluating their actual application effects. This is because application effects are influenced by multiple factors, such as data quality and the actual environment. To ensure the objectivity and reliability of our evaluation, the research focus was placed on the characteristics and performance of the models themselves. Future research could further assess the effectiveness and accuracy of these models in practical applications.