CoChat
Sign Up
Cited by 22Chronic Kidney Disease and DiabetesArtificial Intelligence in HealthcareMachine Learning in Healthcare

Machine learning-based risk predictive models for diabetic kidney disease in type 2 diabetes mellitus patients: a systematic review and meta-analysis

Open access
Li, Yihan, Jin, Nan, Zhan, Qiuzhong, Huang, Yue, Sun, Aochuan, Yin, Fen, Li, Zhuangzhuang, Hu, Jiayu +1 more·2025·Frontiers in Endocrinology
DOI 10.3389/fendo.2025.1495306PMC PMC11911190
Open this paper in CoChat — free account

Abstract

Background: Machine learning (ML) models are being increasingly employed to predict the risk of developing and progressing diabetic kidney disease (DKD) in patients with type 2 diabetes mellitus (T2DM). However, the performance of these models still varies, which limits their widespread adoption and practical application. Therefore, we conducted a systematic review and meta-analysis to summarize and evaluate the performance and clinical applicability of these risk predictive models and to identify key research gaps. Methods: We conducted a systematic review and meta-analysis to compare the performance of ML predictive models. We searched PubMed, Embase, the Cochrane Library, and Web of Science for English-language studies using ML algorithms to predict the risk of DKD in patients with T2DM, covering the period from database inception to April 18, 2024. The primary performance metric for the models was the area under the receiver operating characteristic curve (AUC) with a 95% confidence interval (CI). The risk of bias was assessed using the Prediction Model Risk of Bias Assessment Tool (PROBAST) checklist. Results: 26 studies that met the eligibility criteria were included into the meta-analysis. 25 studies performed internal validation, but only 8 studies conducted external validation. A total of 94 ML models were developed, with 81 models evaluated in the internal validation sets and 13 in the external validation sets. The pooled AUC was 0.839 (95% CI 0.787-0.890) in the internal validation and 0.830 (95% CI 0.784-0.877) in the external validation sets. Subgroup analysis based on the type of ML showed that the pooled AUC for traditional regression ML was 0.797 (95% CI 0.777-0.816), for ML was 0.811 (95% CI 0.785-0.836), and for deep learning was 0.863 (95% CI 0.825-0.900). A total of 26 ML models were included, and the AUCs of models that were used three or more times were pooled. Among them, the random forest (RF) models demonstrated the best performance with a pooled AUC of 0.848 (95% CI 0.785-0.911). Conclusion: This meta-analysis demonstrates that ML exhibit high performance in predicting DKD risk in T2DM patients. However, challenges related to data bias during model development and validation still need to be addressed. Future research should focus on enhancing data transparency and standardization, as well as validating these models’ generalizability through multicenter studies. Systematic Review Registration: https://inplasy.com/inplasy-2024-9-0038/ , identifier INPLASY202490038.

AI snapshot

Written from the paper’s full text.

TL;DR

This systematic review and meta-analysis found that machine learning models predict diabetic kidney disease risk in type 2 diabetes with pooled AUCs of 0.839 in internal validation and 0.830 in external validation. Deep learning and random forest models showed the highest pooled performance, but most included studies had high risk of bias and few performed external validation.

Key findings

  • Across 26 included studies and 94 developed ML models, the pooled AUC was 0.839 (95% CI 0.787–0.890; I²=99.8%) in internal validation and 0.830 (95% CI 0.784–0.877; I²=95.6%) in external validation.
  • By ML type, pooled AUC was 0.797 (95% CI 0.777–0.816) for traditional regression ML, 0.811 (95% CI 0.785–0.836) for ML models, and 0.863 (95% CI 0.825–0.900) for deep learning.
  • Among models used three or more times, random forest had the best pooled AUC at 0.848 (95% CI 0.785–0.911), while decision tree had the lowest at 0.748 (95% CI 0.567–0.929); these differences were not statistically significant.
  • PROBAST assessment found only 2 studies at low risk of bias and 1 unclear, with the remainder high risk; 14 studies failed to properly assess discrimination and calibration, and 34.6% relied solely on random data splitting for internal validation.

Methods

  • Systematic review and meta-analysis conducted per Cochrane, PRISMA-DTA, and CHARMS guidance, registered as INPLASY202490038; searched PubMed, Embase, Cochrane Library, and Web of Science Core Collection for English-language studies up to April 18, 2024.
  • Included 26 studies using ML algorithms to predict DKD risk in T2DM patients; extracted AUC and 95% CI as primary performance metrics and assessed risk of bias with PROBAST.
  • Pooled AUCs using a DerSimonian and Laird random-effects model, with subgroup analyses by study design, prediction model type, ML type, and individual algorithms used three or more times; analyses performed in STATA version 18.

Limitations

  • The paper states that limited external validation constrains generalizability; most included studies were retrospective single-center cohort studies; sample sizes varied widely; predictor sources varied; missing data handling was often inadequate; and predictor selection methods differed across studies.
  • Evident from the text: heterogeneity was very high (I²=99.8% internal, 95.6% external), most studies were rated high risk of bias, only 8 of 26 studies performed external validation, and the internal-validation 95% prediction interval (0.56–1.00) included 1, indicating uncertainty in future performance.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3389/fendo.2025.1495306

Cites

85 of 85 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

22 of 31 cited sources checked · 13 supported · 1 flagged · accuracy not yet measured

  • Interpretationown data

    DL models consistently outperform both regression-based and other ML models.

    “Our results indicated that DL models consistently outperform both regression-based and other ML models, aligning with findings from previous studies (71, 72).”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [71] Smith, 2023, Machine learning and deep learning predictive models for long-term prognosis in patients with chronic obstructive pulmonary disease: a systematic review and meta-analysis, Lancet Digit Health source

      The Model performance for deep learning section reports no significant difference in pooled mortality AUC between deep learning and other machine-learning models (p=0.44). The Discussion says evidence that these models outperform regression models is limited: externally validated performance was worse than or equal to regression models. These findings contradict the claim of consistent outperformance. (LLM, conf=0.90)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites [72] Tu, 1996, Advantages and disadvantages of using artificial neural networks versus logistic regression for predicting medical outcomes, J Clin Epidemiol source
  • Premisecitation

    Previous studies have demonstrated the feasibility of employing ML techniques to utilize various data types for identifying and predicting the risk of DKD in T2DM patients.

    “Previous studies have demonstrated the feasibility of employing ML techniques to utilize various date types for identifying and predicting the risk of DKD in T2DM patients (26, 27).”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [26] Sabanayagam, 2023, Prediction of diabetic kidney disease risk using machine learning models: A population-based cohort study of Asian adults, Elife source

      The Introduction describes using imaging, metabolite, and genetic data with ML to predict DKD risk and identify predictors; the Discussion reports improved prediction of incident DKD. However, the study describes its participants as adults with diabetes, without specifying that they all had T2DM, so the claim’s population qualifier is not established by this source. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [27] Nicolucci, 2022, Prediction of complications of type 2 Diabetes: A Machine learning approach, Diabetes Res Clin Pract source
  • Premisecitation

    Conventional risk prediction models for DKD, such as linear regression, face significant limitations in capturing complex nonlinear relationships and interactions among multidimensional risk factors.

    “Conventional risk prediction models for DKD, such as linear regression, estimate risk by calculating a weighted sum of known risk factors (8–10). While these models are effective when a limited number of variables or predictors are involved, they face significant limitations in capturing the complex nonlinear relationships and interactions among multidimensional risk factors (11, 12).”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [11] Chen, 2011, The worldwide epidemiology of type 2 diabetes mellitus–present and future perspectives, Nat Rev Endocrinol source
    • cites [12] Jacobs, 2024, Artificial intelligence and machine learning for improving glycemic control in diabetes: best practices, pitfalls, and opportunities, IEEE Rev BioMed Eng source

      In the Best practice 14 passage (page unavailable), the review says a simple linear regression predictor works well when glucose changes at a constant rate, supporting a limited point about when that model is useful. It does not address DKD risk prediction or say that conventional models fail to capture nonlinear interactions among multidimensional risk factors. (LLM, conf=0.72)

      Experimental — accuracy not yet measured
  • Premisecitation

    Age has been consistently validated in numerous studies as a pivotal predictor of DKD.

    “It is noteworthy that age has been consistently validated in numerous studies as a pivotal predictor of DKD (5).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [5] Koye, 2018, The global epidemiology of diabetes and kidney disease, Adv Chronic Kidney Dis source
  • Premisecitation

    Longer duration of T2DM significantly escalates the risk of DKD.

    “Extensive research indicates that a longer duration of T2DM significantly escalates the risk of DKD, a phenomenon that is intrinsically linked to the chronic renal damage induced by sustained hyperglycemia (83, 84).”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [83] Wang, 2019, The analysis of risk factors for diabetic nephropathy progression and the construction of a prognostic database for chronic kidney diseases, J Transl Med source

      The Background states that diabetic nephropathy (DKD) affects people with T1DM or T2DM and has peak incidence 10–20 years after diabetes onset. This supports a relationship with diabetes duration, but it does not establish that risk rises continuously with longer T2DM duration or report a statistically significant increase. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [84] Siddiqui, 2022, Risk factors of chronic kidney disease among type 2 diabetic patients with longer duration of diabetes, Front Endocrinol (Lausanne) source
  • Classified as probably not a substantial claim. 36
  • May combine two claims. 32

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductcitation

    The systematic review was conducted in accordance with Cochrane Collaboration and PRISMA-DTA guidelines.

    “This systematic review and meta-analysis were conducted in accordance with the guidelines recommended by the Cochrane Collaboration and the Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy Studies (PRISMA-DTA) statement (28).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [28] Page, 2021, The PRISMA 2020 statement: an updated guideline for reporting systematic reviews, Bmj source

      The Scope of the guideline section recommends using PRISMA 2020 alongside the diagnostic-test-accuracy extension when reporting such reviews. But this paper is the PRISMA 2020 statement, not the PRISMA-DTA statement, and it explicitly says PRISMA 2020 is not intended to guide review conduct. The cited passage does not establish that the citing authors followed Cochrane guidance. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductcitation

    The CHARMS checklist and TRIPOD guidelines were used to develop standardized data extraction forms.

    “We employed the CHARMS checklist and the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) guidelines to develop standardized forms for data extraction (29, 32).”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [29] Moons, 2014, Critical appraisal and data extraction for systematic reviews of prediction modelling studies: the CHARMS checklist, PloS Med source
    • cites [32] Collins, 2024, TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods, Bmj source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductcitation

    The CHARMS checklist was employed to define the objectives of the review.

    “The Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS) was employed to define the objectives of this systematic review and meta-analysis (29).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [29] Moons, 2014, Critical appraisal and data extraction for systematic reviews of prediction modelling studies: the CHARMS checklist, PloS Med source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductcitation

    The protocol for this systematic review and meta-analysis has been registered a priori at INPLASY.

    “The protocol for this systematic review and meta-analysis has been registered a priori at the International Platform of Registered Systematic Review and Meta-analysis Protocols (INPLASY) (30).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [30] Canellas, 2023, The international platform of registered systematic review and meta-analysis protocols (INPLASY) at 3 years: an analysis of 4,658 registered protocols on inplasy.com, platform features, and website statistics, Front Res Metr Anal source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationcitation

    These complications serve as hallmarks of systemic microvascular pathology, indicating an increased likelihood of renal impairment.

    “These complications serve as hallmarks of systemic microvascular pathology, indicating an increased likelihood of renal impairment (85).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [85] Jha, 2024, Diabetes and renal complications: an overview on pathophysiology, biomarkers and therapeutic interventions, Biomedicines source

      The Abstract describes diabetic kidney disease as a microvascular complication that leads to kidney dysfunction, supporting the general link between microvascular damage and renal impairment. It does not say that the unspecified “these complications” are hallmarks of systemic microvascular pathology or that their presence indicates an increased likelihood of renal impairment; searches for retinopathy and neuropathy did not establish that link. (LLM, conf=0.76)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    High-dimensional datasets enrich the input variables for complex models, thereby enhancing predictive precision and bolstering the models' generalizability.

    “These high-dimensional datasets enrich the input variables for complex models, thereby enhancing predictive precision and bolstering the models’ generalizability (81).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [81] Alicic, 2017, Diabetic kidney disease: challenges, progress, and possibilities, Clin J Am Soc Nephrol source

      The supplied passages discuss diabetic kidney disease, its risk factors, and treatment, not high-dimensional datasets or complex models. Searches of our copy for the claim’s wording and for predictive-model accuracy and generalizability found no passage addressing whether richer inputs improve either outcome. (LLM, conf=0.75)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Wu et al. reported that neural network algorithms outperformed others in screening for diabetic retinopathy.

    “Similarly, Wu et al. reported that neural network algorithms outperformed others in screening for diabetic retinopathy (80).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [80] Wu, 2021, Performance and limitation of machine learning algorithms for diabetic retinopathy screening: meta-analysis, J Med Internet Res source

  • May combine two claims.

    Interpretationcitation

    The superior performance of RF may be attributed to its unique ability to handle feature importance, strong generalization, avoidance of overfitting, and proficiency in processing high-dimensional data.

    “This superior performance may be attributed to unique ability to its unique ability to handle feature importance, strong generalization, avoidance of overfitting and proficiency in processing high-dimensional data (75, 76).”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [75] Liu, 2024, Self-supervised random forest on transformed distribution for anomaly detection, IEEE Trans Neural Netw Learn Syst source

      The Datasets results report that sForest outperformed baselines on three high-dimensional datasets, and the classifier comparison describes RF as competitive. But the paper attributes the high-dimensional gains partly to dimensionality reduction and does not establish that RF’s feature-importance handling, generalization, or avoidance of overfitting caused its superior performance. Page numbers were not available in our copy. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [76] Breiman, 2002 source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    RF is an ensemble learning algorithm that solves classification and regression problems by constructing multiple DTs.

    “RF is an ensemble learning algorithm that solves classification and regression problems by constructing multiple DTs, enabling the model to effectively assess the relative importance of each feature in prediction (77).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [77] Speiser, 2019, A comparison of random forest variable selection methods for classification prediction modeling, Expert Syst Appl source

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationcitation

    The heterogeneity of these predictive factors reflects the intricate and multifaceted pathophysiological underpinnings of DKD risk.

    “The heterogeneity of these predictive factors is reflective of the intricate and multifaceted pathophysiological underpinnings of DKD risk (81, 82).”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [81] Alicic, 2017, Diabetic kidney disease: challenges, progress, and possibilities, Clin J Am Soc Nephrol source
    • cites [82] Pugliese, 2019, Diabetic kidney disease: New clinical and therapeutic issues. Joint position statement of the Italian Diabetes Society and the Italian Society of Nephrology on "The natural history of diabetic kidney disease and treatment of hyperglycemia in patients with type 2 diabetes and impaired renal function, Nutr Metab Cardiovasc Dis source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The meta-analysis by Feng et al. found that the LR model achieved a higher AUC in predicting stroke onset timing.

    “Unlike our study, the meta-analysis by Feng et al. (79) found that the LR model achieved a higher AUC in predicting stroke onset timing.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [79] Feng, 2023, The value of applying machine learning in predicting the time of symptom onset in stroke patients: systematic review and meta-analysis, J Med Internet Res source

  • May combine two claims.

    Premisecitation

    Diabetic retinopathy and peripheral vascular disease are closely associated with an increased risk of DKD.

    “Furthermore, complications, such as diabetic retinopathy and peripheral vascular disease are also closely associated with an increased risk of DKD.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [85] Jha, 2024, Diabetes and renal complications: an overview on pathophysiology, biomarkers and therapeutic interventions, Biomedicines source

      The Abstract and Section 2 discuss DKD and its renal mechanisms but do not link diabetic retinopathy or peripheral vascular disease to increased DKD risk. Searches of our copy for “retinopathy” and “peripheral vascular disease” found no passage addressing either association; the indexed passages have no page numbers. (LLM, conf=0.76)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    Momenzadeh et al. applied the MissForest imputation method.

    “Momenzadeh et al. (42) applied the MissForest imputation method”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    All included studies used classical methods to assess the discriminative ability of ML models.

    “All studies used classical methods to assess the discriminative ability of ML models.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    There was substantial variation in sample sizes among the included studies, with the largest sample size being 147,664 and the smallest being 133.

    “However, there was substantial variation in sample sizes among the studies, with Nicolucci et al. (27) having the largest sample size of 147,664, and Maniruzzaman et al. (57) having the smallest with 133.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    4 studies employed external validation using datasets independent of the training set.

    “4 studies employed external validation using datasets independent of the training set (27, 39, 44, 45).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Six studies used mean or median imputation.

    “Among them, six studies used mean or median imputation (45, 46, 52, 54, 56, 61).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    3 studies used a temporal split method for external validation, dividing data chronologically into training and external validation sets.

    “3 studies used a temporal split method, dividing data chronologically into training and external validation sets, with earlier data used for training and later data for external validation (46, 52, 58).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Although the vast majority of studies (96.1%) conducted internal validation, 34.6% of these studies relied solely on random data splitting as internal validation, which cannot be considered a proper application of internal validation methods.

    “Although the vast majority of studies (96.1%) conducted internal validation, a portion of these studies (34.6%) relied solely on random data splitting as internal validation, which cannot be considered a proper application of internal validation methods.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    3 studies did not mention their method to handling missing data.

    “There were also 3 studies that did not mention their method to handling missing data (57, 59, 62).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The most commonly applied ML algorithms were logistic regression (n=11), random forest (n=9), decision tree (n=9), support vector machine (n=6), gradient boosted decision trees (n=5), Bayesian networks (n=3), Naive Bayes (n=3), Adaptive Boosting (n=3), and LASSO (n=3).

    “The most commonly applied ML algorithms included logistic regression (LR, n = 11), random forest (RF, n = 9), decision tree (DT, n = 9), support vector machine (SVM, n = 6), gradient boosted decision trees (GBDT, n = 5), Bayesian networks (BN, n = 3), Naive Bayes (NB, n = 3), Adaptive Boosting (AdaBoost, n = 3), and Least Absolute Shrinkage and Selection Operator (LASSO, n = 3).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    One study used all available features as predictors.

    “one used all available features as predictors (42)”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Three studies excluded missing data directly.

    “Three studies excluded missing data directly (39, 44, 48)”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    2 studies used ML algorithms for imputation.

    “Additionally, 2 studies used ML algorithms for imputation (55, 60).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    Another study determined predictors based on existing literature and expert consensus.

    “and another determined predictors based on existing literature and expert consensus (56).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Most studies addressed missing data using various imputation methods.

    “Most studies addressed missing data using various imputation methods.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    6 studies selectively excluded missing data based on a certain proportion.

    “while 6 studies selectively excluded missing data based on a certain proportion (26, 41, 47, 49, 51, 58).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    Liu et al. employed the Multivariate Imputation by Chained Equations (MICE) algorithm.

    “Liu et al. (40) employed the Multivariate Imputation by Chained Equations (MICE) algorithm.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    3 studies only mentioned the use of imputation without specifying the exact methods.

    “However, 3 studies only mentioned the use of imputation without specifying the exact methods employed (27, 50, 53).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Eight studies (30.8%) had unreasonable sample size designs, with 4 studies having an event-per-variable (EPV) ratio of less than 10 and another 4 studies having a validation cohort sample size of fewer than 100 cases.

    “Additionally, 8 studies (30.8%) had unreasonable sample size designs, with 4 studies having an event-per-variable (EPV) ratio of less than 10 during the development of predictive models, and another 4 studies having a validation cohort sample size of fewer than 100 cases.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Of the studies reviewed, two did not mention the method used for predictor selection.

    “Of the studies reviewed, two did not mention the method used for predictor selection (27, 44)”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    A total of 94 ML models were developed, with 81 models in the internal validation set and 13 models in the external validation set.

    “A total of 94 ML models were developed, with 81 models in the internal validation set and 13 models in the external validation set.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The most frequently employed predictor selection method was Recursive Feature Elimination with Cross-Validation (RFECV), followed by LASSO.

    “The most frequently employed method was Recursive Feature Elimination with Cross-Validation (RFECV), followed by the LASSO.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingcitation

    One study conducted external validation across 5 different centers, with reported cohort sizes ranging from 3,912 to 200,007.

    “One study conducted external validation across 5 different centers (27), with reported cohort sizes ranging from 3,912 to 200,007, though specific sample sizes for the other centers were not provided.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    A total of 94 ML models were developed, with 81 evaluated in internal validation sets and 13 in external validation sets.

    “A total of 94 ML models were developed, with 81 models evaluated in the internal validation sets and 13 in the external validation sets.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingreasoning

    The remaining 22 studies detailed their methods for predictor selection.

    “The remaining 22 studies detailed their methods for predictors selection.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    25 studies performed internal validation, and 8 studies conducted external validation.

    “25 studies performed internal validation, but only 8 studies conducted external validation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingcitation

    One prospective study had no missing data and did not involve any imputation methods.

    “Furthermore, one prospective study had no missing data, and thus, did not involve any imputation methods (43).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The pooled AUC for case-control studies was 0.858 (95% CI 0.817-0.900) with significantly reduced heterogeneity (I2 = 23.0%; P = 0.254).

    “Notably, case-control demonstrated a pooled AUC of 0.858 (95% CI 0.817-0.900) with significantly reduced heterogeneity (I² = 23.0%; P = 0.254).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The pooled AUC in the internal validation set was 0.839 (95% CI 0.787-0.890; I2= 99.8%; P = 0.000 for heterogeneity), with a 95% PI ranging from 0.56 to 1.00.

    “The pooled AUC in the internal set was 0.839 (95% CI 0.787-0.890; I²= 99.8%; P = 0.000 for heterogeneity), with a 95% PI ranging from 0.56 to 1.00.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The pooled AUC for traditional regression ML models was 0.797 (95% CI 0.777-0.816) with significant heterogeneity (I2= 96.7%, P = 0.000).

    “The results indicated that the pooled AUC for traditional regression ML models was 0.797 (95% CI 0.777-0.816) with significant heterogeneity (I²= 96.7%, P = 0.000)”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The pooled AUC in the external validation set was 0.830 (95% CI 0.784-0.877; I2= 95.6%; P = 0.000 for heterogeneity), with a 95% PI of 0.64-0.97 that excluded 1.

    “In contrast, the pooled AUC in the external validation set was 0.830 (95% CI 0.784-0.877; I²= 95.6%; P = 0.000 for heterogeneity), with a narrower 95% PI of 0.64-0.97 that excluded 1, indicating more stable performance and better generalizability across external datasets, despite significant heterogeneity.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given

    This study represents the first comprehensive review of the application of ML models for predicting DKD risk in patients with T2DM.

    “This study represents the first comprehensive review of the application of ML models for predicting DKD risk in patients with T2DM.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The results remained consistent with the original pooled AUC (0.84, 95% CI 0.79-0.89), indicating the stability of the findings.

    “The results remained consistent with the original pooled AUC (0.84, 95% CI 0.79-0.89), indicating the stability of the findings.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The pooled AUCs for models used three or more times were: AdaBoost 0.798 (95% CI 0.729-0.868); BN 0.815 (95% CI 0.777-0.854); DT 0.748 (95% CI 0.567-0.929); GBDT 0.821 (95% CI 0.757-0.885); LASSO 0.820 (95% CI 0.816-0.825); LR 0.793 (95% CI 0.764-0.823); NB 0.788 (95% CI 0.737-0.840); RF 0.848 (95% CI 0.785-0.911); SVM 0.825 (95% CI 0.690-0.960); and XGBoost 0.829 (95% CI 0.724-0.934).

    “The results indicated that the pooled AUCs were as follows: AdaBoost, 0.798 (95% CI 0.729-0.868); BN, 0.815 (95% CI 0.777-0.854); DT, 0.748 (95% CI 0.567-0.929); GBDT, 0.821 (95% CI 0.757-0.885); LASSO, 0.820 (95% CI 0.816-0.825); LR, 0.793 (95% CI 0.764-0.823); NB, 0.788 (95% CI 0.737-0.840); RF, 0.848 (95% CI 0.785-0.911); SVM, 0.825 (95% CI 0.690-0.960); and XGBoost, 0.829 (95% CI 0.724-0.934).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The RF model yielded the highest performance, with a pooled AUC of 0.848 (95% CI 0.785-0.911), outperforming the other models.

    “Our results indicated that the RF model yielded the highest performance, with a pooled AUC of 0.848 (95% CI 0.785-0.911), outperforming the other models.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The meta-analysis of the AUC included data from 25 studies for internal validation sets and 8 studies for external validation sets.

    “The meta-analysis of the AUC included data from 25 studies for internal validation sets and 8 studies for external validation sets.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    The studies included in this meta-analysis varied in design, with the majority being retrospective single-center cohort studies.

    “The studies included in this meta-analysis varied in design, with the majority being retrospective single-center cohort studies, which are prone to selection bias and information bias”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductcitation

    Risk of bias in the included prediction models was assessed using the PROBAST checklist.

    “We assessed the risk of bias in the prediction models using the Prediction Model Risk of Bias Assessment Tool (PROBAST) (33), which is specifically designed for studies involving multivariable prediction models for individual prognosis or diagnosis.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [33] Moons, 2019, PROBAST: A tool to assess risk of bias and applicability of prediction model studies: explanation and elaboration, Ann Intern Med source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    The search covered English-language publications up to April 18, 2024.

    “We included all relevant English-language publications up to April 18, 2024.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    ML techniques have the capacity to extract common features from diverse data sources and unveil intricate relationships between DKD and various contributing factors.

    “Specifically, ML techniques have the capacity to extract common features from diverse data sources, such as text, images, behavioral data, and physiological indicators, thereby unveiling the intricate relationships between DKD and various contributing factors (19).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [19] Jordan, 2015, Machine learning: Trends, perspectives, and prospects, Science source

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationnone given

    Challenges related to data bias during model development and validation still need to be addressed.

    “However, challenges related to data bias during model development and validation still need to be addressed.”
    • Also stated in: § Conclusion · plain
      “However, challenges related to data bias during model development and validation still need to be addressed.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Retrospective single-center cohort studies are prone to selection bias and information bias.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductreasoning

    Two independent reviewers performed study screening and data extraction, with disagreements resolved by a third reviewer.

    “Two independent reviewers (Y.H. Li and N. Jin) performed a preliminary screening of the titles and abstracts of the included studies.”
    • Also stated in: § Study selection · plain
      “In cases of disagreement regarding the inclusion of a study during the screening process, consensus was achieved through discussion or by consulting a third independent reviewers (Q.Z. Zhan) for arbitration.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    The search was conducted across PubMed, Embase, Cochrane Library, and Web of Science Core Collection.

    “we conducted a comprehensive search across the following databases: PubMed, Embase, Cochrane Library, and Web of Science Core Collection.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisereasoning

    The performance of ML models for predicting DKD risk in T2DM varies, limiting their widespread adoption and practical application.

    “However, the performance of these models still varies, which limits their widespread adoption and practical application.”
    • Also stated in: § Introduction · hedged
      “Moreover, although ML has garnered significant attention within the medical field, its robustness in clinical practice remains uncertain, and its widespread adoption and application are somewhat constrained.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductnone given

    Two independent reviewers systematically accomplished literature screening, data extraction, and a detailed risk of bias assessment using the PROBAST checklist.

    “Two independent reviewers systematically accomplished literature screening, data extraction, and a detailed risk of bias assessment using the PROBAST checklist, which improved the credibility and reliability of the study results.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Among the models, RF exhibited the best performance, while DT had the lowest performance; however, these differences were not statistically significant.

    “Among the models, the RF exhibited the best performance, while the DT model had the lowest performance; however, these differences were not statistically significant.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    RF's parallelization capability results in high computational efficiency, markedly reducing training time.

    “Moreover, RF’s parallelization capability results in high computational efficiency, markedly reducing training time (78).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [78] Paul, 2018, Improved random forest for classification, IEEE Trans Image Process source

  • May combine two claims.

    Interpretationreasoning

    Studies with smaller sample sizes may experience a reduction in machine learning model performance, which could serve as a potential source of heterogeneity.

    “However, studies with smaller sample sizes may experience a reduction in machine learning model performance, which could also serve as a potential source of heterogeneity.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The external validation set showed more stable performance and better generalizability across external datasets, despite significant heterogeneity.

    “with a narrower 95% PI of 0.64-0.97 that excluded 1, indicating more stable performance and better generalizability across external datasets, despite significant heterogeneity.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The authors extensively searched multiple databases in accordance with PRISMA guidelines.

    “In addition, we extensive searched multiple databases in accordance with PRISMA guidelines.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    Traditional ML algorithms can still achieve high performance with less computational power and smaller datasets.

    “However, traditional ML algorithms can still achieve high performance with less computational power and smaller datasets.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Based on the included studies, several risk factors for the development of DKD in patients with T2DM were identified, including age, gender, race, history of hypertension, cardiovascular disease, cerebrovascular disease, T2DM duration, smoking, alcohol, family history of diabetes, use of antidiabetic and antihypertensive medications, complications such as diabetic retinopathy and diabetic periv

    “Based on these studies, we can identify several risk factors for the development of DKD in patients with T2DM.”
    Quote in the paperSubstantial claim
63 of 63

  • The paper relies on this without support.

    The included studies' populations are representative of the broader T2DM patient population, allowing generalization of the pooled performance.

    “This meta-analysis demonstrates that ML exhibit high performance in predicting DKD risk in T2DM patients.”

  • The paper relies on this without support.

    Retrospective single-center cohort studies are prone to selection bias and information bias.

    “which are prone to selection bias and information bias”

  • The paper relies on this without support.

    Studies with smaller sample sizes may experience greater random error in estimating feature factors, leading to potential overfitting or underfitting of the model.

    “Studies with smaller sample sizes may experience greater random error in estimating feature factors, leading to potential overfitting or underfitting of the model”

  • The paper relies on this without support.

    Diversity of data sources may introduce issues related to data quality and consistency.

    “While the diversity of data sources enriches the model’s input dimensions, it may also introduce issues related to data quality and consistency”

  • The paper relies on this without support.

    Discrepancies in handling missing data may affect the prediction accuracy and stability of the models.

    “These discrepancies may affect the prediction accuracy and stability of the models”

  • The paper relies on this without support.

    The choice of method for selecting predictive factors can influence both the complexity and interpretability of the model.

    “the choice of method can influence both the complexity and interpretability of the model”

  • The paper relies on this without support.

    ML-based prediction of DKD risk is a more selective and cost-effective screening method for diabetic nephropathy.

    “the application of machine learning in predicting diabetic nephropathy risk holds significant clinical value and represents a more selective and cost-effective screening method for diabetic nephropathy.”
7 of 7

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Prognostic-study list omits a study classified as prognostic

    Minor

    The text says eight studies were prognostic but lists only seven citations. Table 1 classifies the omitted Nicolucci study (27) as a prognostic model.

1 of 1

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Li, Y., Jin, N., Zhan, Q., Huang, Y., Sun, A., Yin, F., Li, Z., Hu, J., & Liu, Z.
Affiliations
  • Xiyuan Hospital (CN)
  • Macau University of Science and Technology (MO)
  • Beijing University of Chinese Medicine (CN)
Corresponding author
  • Hu, Jiayu
  • Liu, Zhengtang
ORCID
Venue
Frontiers in Endocrinology
Publisher
Frontiers Media SA
Published
2025-03-03
Volume
16
Pages
1495306-1495306
ISSN
1664-2392 (ISSN-L)
DOI
10.3389/fendo.2025.1495306
PMC
PMC11911190
OpenAlex
W4408096697
Semantic Scholar
8e988f3b84385cb5592dafc2966ca34fc35f9f3c
Cited by
22
References
84
Citations by year
2026: 15 · 2025: 7
FWCI
5.01
Type
journal article
Language
en
Open access
Licence
Locations
  • Frontiers in Endocrinology (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Europe PMC (PubMed Central) (submittedVersion, cc-by)
  • PubMed Central (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
National Key Research and Development Program of China
Topics
  • Health Sciences › Medicine › Nephrology › Chronic Kidney Disease and Diabetes
  • Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
Keywords
diabetic nephropathy, type 2 diabetes, risk prediction model, machine learning, deep learning, random forest, model performance, external validation, risk of bias, PROBAST, clinical applicability
MeSH terms
Machine Learning, Diabetes Mellitus, Type 2, Diabetic Nephropathies, Humans, Risk Factors, Risk Assessment

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-13.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-13.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review