This work is licensed under a Abstract
Background
Type 2 diabetes mellitus (T2DM) poses a significant global public health burden, where early detection of at-risk populations is imperative for implementing targeted preventive strategies. This systematic review and meta-analysis aimed to evaluate the methodological quality and predictive performance of existing T2DM risk prediction models in screening contexts.
Methods
Following the TRIPOD-SRMA statement, eligible studies were selected through searching seven databases (CNKI, WanFang Database, VIP, PubMed, Embase, Web of Science, and the Cochrane Library) from database inception through December 2024. Methodological quality was assessed using the PROBAST tool. Random-effects models synthesized discrimination (AUC). Subgroup analyses explored geographic, modeling, and validation-related heterogeneity. Funnel plots and Egger’s regression test assessed small-study effects.
Results
A total of 65 studies (encompassing 97 distinct prediction models) were included in the analysis. Among 97 models, logistic regression dominated (97.9% of models), achieving moderate discrimination (AUC: 0.628–0.916), while machine learning (ML) models showed marginally higher AUCs (up to 0.998). Geographic and cohort disparities emerged, with USA-based models outperforming others (USA AUC: 0.97 vs China AUC: 0.79) and poor performance in prediabetic cohorts (AUC: 0.72 vs 0.80 in normoglycemic). External validation remained limited (21 models), though spatial/temporal validation cohorts demonstrated stable performance. High risk of bias and application (>80% of models) stemmed from inadequate statistical reporting and external verification definitions.
Conclusion
ML has favorable diagnostic accuracy for the progression of T2DM. This provides evidence for the development of predictive tools with broader applicability. Future research should prioritize external validation to enhance precision.
Keywords: type 2 diabetes mellitus, risk prediction models, meta-analysis, machine learning, screening
Introduction
The global burden of type 2 diabetes mellitus (T2DM) has reached epidemic proportions, with an estimated 537 million adults affected worldwide in 2021, a figure projected to rise to 783 million by 2045 (1). First-line pharmacological interventions demonstrate limited long-term efficacy due to progressive β-cell dysfunction and acquired drug resistance (2). This therapeutic challenge underscores the critical importance of preventive strategies and early detection in mitigating the devastating microvascular and macrovascular complications associated with undiagnosed hyperglycemia (3). Advances in predictive medicine have generated multiple risk stratification tools integrating variables spanning anthropometric indices and biochemical markers (4). However, the selection of optimal diabetes prediction models poses significant challenges due to methodological heterogeneity among existing tools, such as analytical frameworks (machine learning (ML) algorithms versus conventional regression analyses), predictor selection criteria (clinically accessible parameters versus specialized biomarkers), and validation protocols (internal versus external cohort verification) (5, 6). The inherent methodological heterogeneity introduces critical trade-offs among model complexity, generalizability, and clinical feasibility, with each computational approach presenting distinct strengths and limitations (7). Therefore, there is a need to refine systematic reviews of diabetes prediction models to facilitate the selection of the optimal predictive model.
Existing systematic reviews of T2DM prediction models require updating to address fundamental limitations in existing evidence. While conventional regression-based methodologies have been extensively examined in prior analyses (8), contemporary ML developments, particularly deep neural networks and ensemble techniques increasingly employed in clinical investigations, lack comprehensive evaluation (9). Moreover, emerging evidence underscores significant disparities in model performance across ethnic populations, healthcare settings, novel predictors, and social determinants of health (10, 11, 12). However, current meta-analytic approaches remain insufficient in analyzing this multidimensional heterogeneity spanning geographical regions and demographic subgroups. The prediction model risk of bias assessment tool (PROBAST) framework enables a more comprehensive methodological appraisal of prediction models to identify superior modeling approaches (5), which have also not been comprehensively appraised in existing reviews. These gaps compromise the translational relevance of current evidence, necessitating an updated synthesis aligned with the PRISMA statement (13).
Therefore, this study conducts a systematic review to identify, critically evaluate, and synthesize the methodological quality and predictive performance of existing T2DM prediction models, aiming to provide evidence-based recommendations for optimizing the development and clinical application of future risk stratification tools.
Methods
This study adhered to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis – Systematic Review and Meta-analysis (TRIPOD-SRMA) statement (14). The review protocol was registered on the International Prospective Register of Systematic Reviews (PROSPERO) with registration ID CRD42024569731.
Study selection
The following eligibility criteria were used for considering studies for this review. The inclusion criteria were as follows: i) population: adults aged ≥18 years with T2DM; ii) outcomes: development or validation of T2DM risk prediction models; and iii) study design: cohort studies (prospective or retrospective). The exclusion criteria were i) studies investigating T2DM risk factors without model development, ii) non-Chinese/English publications, and iii) cross-sectional studies, case–control studies, or studies with incomplete data/full-text unavailability.
Search strategy
Electronic searches were conducted in the following databases from inception to December 2024: i) Chinese literature databases: CNKI, WanFang Database, and VIP Database; and ii) English databases: PubMed, Embase, Web of Science, and the Cochrane Library. Search terms combined medical subject headings (MeSH) and free-text keywords adapted to database-specific syntax. Example English terms included: diabete*; type 2 diabetes; T2DM; prediction model; prognostic model; risk prediction; risk score; and risk assessment.
The following PubMed search terms were an example of the search strategy:
#1: search (‘type 2 diabetes’[mesh]) OR (‘diabetes mellitus type II’ or ‘diabetes mellitus type 2’ or ‘type 2 diabetes’ or ‘type 2 diabetes mellitus’ or ‘diabetes type 2’ or ‘T2DM’ or ‘diabete*’).
#2: search (‘predictive model’[mesh]) OR (‘prediction model’ or ‘clinical risk score’ or ‘clinical prediction model’ or ‘clinical predictive model’ or ‘clinical scoring system’ or ‘risk assessment model’ or ‘risk prediction model’ or ‘risk predictive model’).
#3: #1 AND #2.
Study screening and data extraction
Study screening and data extraction were independently performed by two investigators (Meichen Li and Aowei Tan) with cross-verification according to the eligibility criteria. Discrepancies were resolved through consensus discussions or adjudication by a third reviewer (Li Zhou). Data extraction followed the Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS) checklist (15), with results tabulated to capture: i) study characteristics, including first author, country, and study design (development/validation); ii) methodological parameters: sample size, handling of missing data, model-building techniques (e.g., regression algorithms), and variable selection methods (stepwise/LASSO); and iii) model performance metrics: discrimination (area under the curve statistics), calibration (observed/expected ratios), and presentation format (nomograms/risk scores).
Risk of bias and applicability assessment
Two investigators (Meichen Li and Aowei Tan) independently evaluated the methodological quality of included studies using the PROBAST tool (4), assessing four domains: i) participant selection: evaluates whether the study population appropriately represents the target clinical population, including inclusion/exclusion criteria clarity and potential selection bias (e.g., inappropriate exclusions or incomplete enrollment). ii) Predictor variables: examines the definition, measurement, and timing of predictors, including whether predictors were clearly defined and measured before outcome assessment (blinding of predictor assessment). iii) Outcome definition: assesses the clinical relevance and objectivity of the outcome, including the handling of missing data (e.g., exclusion, imputation, or transparency in reporting). iv) Statistical analysis: critically appraises model development and validation methods, covering the handling of overfitting, model calibration/discrimination metrics, and internal/external validation procedures. Each domain received three-tiered risk classifications: low risk, unclear risk, and high risk.
In addition, overall applicability (clinical relevance) was graded across the first three domains to evaluate whether the study’s design and execution aligned with real-world clinical practice: i) participant selection: whether the study population matched the intended clinical setting. ii) Predictor variables: feasibility of obtaining predictors in routine care. (iii) Outcome definition: clinical utility of the outcome measure. Each domain was classified as low, unclear, or high concern for bias risk and real-world implementability. Discrepancies between reviewers were resolved through iterative discussion until consensus was reached, with a third senior researcher (Zhou Li) arbitrating unresolved cases.
Data synthesis and analysis
The methodological framework for evaluating predictive model performance was established in accordance with the TRIPOD-SRMA statement for systematic reviews and meta-analyses of prediction models (14). Data synthesis was performed by pooling discrimination metrics (AUC-ROC values) across studies using Hartung-Knapp-Sidik-Jonkman (HKSJ) random-effects models, with between-study variance estimated via restricted maximum likelihood (REML) (16). For the primary metric, in addition to reporting point estimates, 95% confidence intervals (CI) were reported. Heterogeneity assessment was conducted using Cochran’s Q-test and quantified with the I2 statistic, where I2 > 50% indicated substantial heterogeneity. Predefined subgroup analyses were performed to explore heterogeneity sources, including country (geographic origin of study populations), model type (e.g., logistic regression (LR) vs ML), number of predictors (<10 vs ≥ 10), study timeframe (data collection period), and cohort size stratification (<1,000 vs ≥ 1,000 participants). Sensitivity analyses included leave-one-out cross-validation to evaluate the robustness of pooled estimates and iterative exclusion of studies. Reporting bias was assessed using funnel plots complemented by Egger’s regression test for small-study effects when ≥10 studies were included in a meta-analysis. Sensitivity analysis was performed by sequentially omitting each study to assess the robustness of the pooled effect size. For clinical interpretability, decision curve analysis (DCA) was implemented for externally validated models to evaluate net benefit across threshold probabilities. Net benefit was calculated as the proportion of true positives minus the proportion of false positives weighted by the odds of the selected risk threshold. Moreover, co-occurrence network analysis visualized predictors using force-directed algorithms to identify clusters of covariate associations. All statistical analyses were conducted in R v4.1.0 (metafor, dmetar, rmda, and igraph packages).
Results
Study identification
The preliminary search identified a total of 12,340 publications across six databases: CNKI (n = 1,771), WanFang (n = 3,100), CBM (n = 417), PubMed (n = 887), Embase (n = 2,144), and the Cochrane Library (n = 4,021). Following database searching, records underwent two-stage screening. Title/abstract screening excluded 10,052 records. The remaining records underwent full-text assessment. Full-text review excluded 283 records. Following a systematic screening process, 65 studies (encompassing 97 distinct prediction models) were identified for inclusion in the analysis. The literature screening flowchart and detailed results are depicted in Fig. 1.
Figure 1.
Search results and study selection. TC: total cholesterol; FPG: fasting plasma glucose; TG: triglycerides; HDL: high-density lipoprotein; BMI: body mass index; WC: waist circumference; BP: blood pressure; WHR: waist-to-hip ratio; CVD: cardiovascular disease; PDM: pre-diabetes mellitus; OGTT: oral glucose tolerance test; Glu: glucose; HbA1c: glycated hemoglobin; SCR: serum creatinine; ALT: alanine transaminase; LDL: low-density lipoprotein; SBP: systolic blood pressure; DBP: diastolic blood pressure.
Study characteristics
The 65 included studies (97 models) encompassed diverse geographic regions, predominantly China (47.4%), followed by Japan (13.4%), South Korea (9.3%), the USA (7.2%), India (6.2%), Iran (2.06%), England (2.06%), and European countries (2.06%). Study populations comprised non-diabetic individuals (91.8%) and prediabetic participants (8.2%), with follow-up durations ranging from 1 to 20 years, most commonly 5 years (35.1%). Age groups varied widely, including adults aged ≥18 years to elderly populations (≥60 years). Sample sizes spanned from 487 to 2,540,753 participants, with outcome events (e.g., diabetes incidence) reported across cohorts. Of the 97 models, 15 (15.5%) were derived from retrospective cohort studies, and 82 (84.5%) from prospective cohort studies. The characteristics of the included studies are presented in Table 1.
Table 1.
Characteristics of model development studies.
| Study | Country | Study population | Follow-up | Age | Sample size | Outcome events | Predictor variables content | Missing data |
|---|---|---|---|---|---|---|---|---|
| Han 2017 (23) | China | Non-DM | 5-year | ≥50 | 17,690 | 1,309 | BMI; WC; FPG; HBP; lipid disorders; smoke; family history of DM | Deletion |
| Sun 2013 (1) (21) | China | Not-DM | 5-year | 35–70 | T: 12,392 | T: 361 | Age; family history of DM; height; WC; FPG; TG; HDL-C; TC | – |
| EV: 4,215 | EV: 154 | |||||||
| Sun 2013 (2) (21) | China | Not-DM | 5-year | 35–70 | T: 22,936 | T: 220 | Age; family history of DM; height; WC; FPG; TG; HDL-C; TC | – |
| IV: 3,229 | IV: 99 | |||||||
| Sun 2013 (3) (21) | China | Not-DM | 5-year | 35–70 | T: 12,392 | T: 361 | Age; family history of DM; height; WC; FPG; TG; HDL-C; TC | – |
| EV: 3,344 | EV: 139 | |||||||
| Cai 2020 (24) | China | Not-DM | 5-year | ≥20 | T: 22,936 | T: 220 | FPG; TC; TG; HDL-C; LDL-C; smoke; alcohol; family history of DM | No missing data |
| IV: 9,830 | IV: 190 | |||||||
| Hippisley-Cox 2009 (1) (18) | England & Wales | Not-DM | 10-year | 25–79 | Male | Male | Race; age; sex; BMI; smoke; family history of DM; tmvnsend; HBP; CVD; medication history | Multiple imputation |
| T: 2,540,753 | T: 78,081 | |||||||
| EV: 1,232,832 | EV: 3,7535 | |||||||
| Hippisley-Cox 2009 (2) (18) | England & Wales | Not-DM | 10-year | 25–79 | Female | Female | Race; age; sex; BMI; smoke; family history of DM; tmvnsend; HBP; CVD; medication history | Multiple imputation |
| T: 2,540,753 | T: 78,081 | |||||||
| EV: 1,232,832 | EV: 37,535 | |||||||
| Wen 2016 (1) (25) | China | Not-DM | 6-year | ≥30 | T: 2,754 | T: 144 | Age; BMI; WC; family history of DM | Expectation-maximization |
| IV: 1,378 | IV: 74 | |||||||
| Wen 2016 (2) (25) | China | Not-DM | 6-year | ≥30 | T: 2,754 | T: 144 | Age; BMI; WC; family history of DM; FPG; TG | Expectation-maximization |
| IV: 1,378 | IV: 74 | |||||||
| Wang 2018 (26) | China | Not-DM | 6-year | ≥18 | 5,706 | 194 | Age; diet; BMI; FPG; TG | No missing data |
| Bai 2018 (27) | China | Not-DM | 5-year | 18–90 | T: 8,238 | T: 1,197 | Sex; age; BMI; smoke; alcohol; HBP; family history of HBP/DM; SBP; DBP; FPG; TC; TG; fatty liver | – |
| IV: 3,531 | IV: 496 | |||||||
| Lei 2023 (28) | China | Not-DM | 5-year | ≥60 | T: 17,363 | T: 1,072 | Age; FPG; BMI; SBP; TG; ALT | Deletion |
| IV: 7,441 | IV: 468 | |||||||
| Fu 2023 (29) | China | Not-DM | 7-year | ≥60 | 712 | 33 | Sex; age; BMI; ALT; TG; HbA1c; FPG | Deletion |
| Su 2017 (1) (30) | China | Not-DM | 3-year | 20–70 | Male | Male | Age; BMI; FPG; TG; ALT; WBC | – |
| 18,963 | 1,044 | |||||||
| Su 2017 (2) (30) | China | Not-DM | 3-year | 20–70 | Female | Female | Age; FPG; TG; HDL-C; ALT | – |
| 14,482 | 580 | |||||||
| Zhang 2022 (31) | China | Not-DM | 3-year | >60 | 1,512 | 173 | Age; BMI; WC; BP; smoke; alcohol; family history of DM; Glu; insulin; HDL-C; TG | – |
| Yang 2016 (32) | China | Not-DM | 5-year | 20–80 | 16,715 | 858 | Age; BMI; FPG; TG; HBP; WBC | – |
| Zhang 2016 (1) (33) | China | Not-DM | 6-year | ≥18 | Non-invasive | Non-invasive | Sex; age; education; smoke; sleep; BMI; WC; HBP; TC; TG; HDL-C; FPG | – |
| T: 6,143 | T: 376 | |||||||
| IV: 6,142 | IV: 403 | |||||||
| Zhang 2016 (2) (33) | China | Not-DM | 6-year | ≥18 | Invasive | Invasive | HBP; TC; TG; HDL-C; FPG | – |
| T: 6,143 | T: 376 | |||||||
| IV: 6,142 | IV: 403 | |||||||
| Zhu 2023 (1) (34) | China | Not-DM | 3-year | ≥18 | T: 1,133 | 425 | Age; sex; education; smoke; alcohol; medication history; SBP; DBP; TC; LDL; HDL-C; TG; ALT; UA; WC; WHR; BMI; lipid disorders,ect | Deletion |
| IV: 284 | ||||||||
| Zhu 2023 (2) (34) | China | PDM | 3-year | ≥18 | T: 796 | 360 | Age; sex; education; medication history; smoke; alcohol; SBP; DBP; TC; LDL; HDL-C; TG; ALT; UA; WC; WHR; BMI; lipid disorders,ect | Deletion |
| IV: 199 | ||||||||
| Chen 2018 (35) | China | PDM | 1-year | ≥18 | T: 3,803 | T: 99 | Age; BMI; HBP; diet | – |
| IV: 1,853 | IV: 57 | |||||||
| Ma 2020 (36) | China | Not-DM | 2-year | ≥30 | 3,127 | 187 | Smoke; medication history; OGTT; FPG; BMI; family history of DM | Deletion |
| Liu 2020 (37) | China | Not-DM | 6-year | ≥18 | 7,289 | 562 | Age; sex; family history of DM; BMI; SBP; LDL-C; FPG | Multiple imputation |
| Lee 2024 (1) (38) | South Korea | Not-DM | 8-year | 40–69 | T: 5,459 | T: 460 | Age; BMI,; HTN; HDL-C; TG; FPG; HbA1c; γ-GTP | Multiple imputation |
| IV: 2,318 | IV: 180 | |||||||
| EV: 2,043 | EV: 368 | |||||||
| EV: 4,091 | EV: 138 | |||||||
| Lee 2024 (2) (38) | South Korea | Not-DM | 8-year | 40–69 | T: 5,459 | T: 460 | Age; BMI,; HTN; HDL-C; TG; FPG; HbA1c; γ-GTP | Multiple imputation |
| IV: 2,318 | IV: 180 | |||||||
| EV: 2,043 | EV: 368 | |||||||
| EV: 4,091 | EV: 138 | |||||||
| Alssema 2011 (39) | Europe | Not-DM | 5-year | ≥45 | 18,301 | 844 | Age; BMI; WC; medication history; sex; history of GMD; smoke; family history of DM | Multiple imputation |
| Li 2019 (40) | China | Not-DM | 5-year | ≥40 | 17,173 | 1,159 | Age; sex,; FPG; BMI,; WC; TC; TG; BP; family history of DM; ALT | Deletion |
| Gao 2009 (1) (41) | India | Not-DM | 10-year | 20–65 | Male | Male | Obesity; BMI; WC; BP; TG; family history of DM; FPG | Deletion |
| T: 1,141 | T: 259 | |||||||
| IV: 1,182 | IV: 259 | |||||||
| Gao 2009 (2) (41) | India | Not-DM | 10-year | 20–65 | Female | Female | Obesity; BMI; WC; BP; TG; family history of DM; FPG | Deletion |
| T: 1,953 | T: 252 | |||||||
| IV: 1,550 | IV: 245 | |||||||
| Edlitz 2022 (1) (20) | USA | Not-DM | 5-year | 40–69 | IV: 7,948 | IV: 64 | Visit interval; height; HBP; exercise; smoke; diet; family history of DM | Iterative imputation method |
| EV: 10,064 | EV: 201 | |||||||
| Edlitz 2022 (2) (20) | USA | PDM | 5-year | 40–69 | IV: 1,006 | IV: 94 | Visit interval; height; HBP; exercise; smoke; diet; family history of DM | Iterative imputation method |
| EV: 7,059 | EV: 501 | |||||||
| Katsimpris 2021 (42) | Germany | Not-DM | 5-year | 40–69 | 1,591 | 139 | Age; sex; BMI; diet | – |
| Liu 2016 (43) | China | Not-DM | 20-year | ≥55 | 1,857 | 144 | Sex; age; BMI; FPG; self-assessment of health; exercise; disability; marital status; lipid disorders; education; diet | Deletion |
| Hafezi 2024 (44) | Iran | Not-DM | 10-year | ≥55 | T: 5,033 | T: 942 | LAP; BMI; VAI | – |
| IV: 1,677 | IV: 314 | |||||||
| Hu 2020 (45) | China | Not-DM | 5-year | 40–69 | 4,833 | 171 | Age; sex; smoke; exercise; BMI; TG,FPG | – |
| Xu 2014 (46) | China | Not-DM | 4-year | ≥50 | T: 8,000 | 1,063 | Family history of DM; BMI; BP; TC; FPG; HDL | Inverse probability weighting |
| IV: 8,043 | ||||||||
| Rathmann 2010 (47) | Germany | Not-DM | 7-year | 55–74 | 887 | 93 | Age; sex; BMI; family history of DM; smoke; HBP; FPG; HbA1c; UA; OGTT | – |
| Liu 2022 (48) | China | Not-DM | 2-year | ≥65 | T: 10,1625 | 8,298 | Age; gender; education; marital status; BP; exercise; smoke; alcohol; WC; SBP; FPG; TC; TG; HDL-C; LDL-C; ALT; AST; TBIL; Scr; BUN; SUA | – |
| IV: 25,406 | ||||||||
| Zhang 2020 (49) | Australia | Not-DM | 10-year | ≥45 | T: 165,609 | T: 10,019 | Age; sex; BMI; education; family history of DM; residence; income; BP; CVD; lipid disorders; smoke; exercise; diet | – |
| IV: 70,975 | IV: 4,294 | |||||||
| Alghamdi 2017 (50) | USA | Not-DM | 5-year | ≥18 | 32,555 | 5,099 | Age; HR; metabolic equivalent; SBP; DBP; exercise; race; BMI; lipid disorders; medication history; CVD | – |
| Wang 2018 (1) (51) | China | Not-DM | 3-year | ≥18 | Male | Male | Metabolic syndrome; BMI; IGR; FPG; OGTT; BP; lipid disorders; TG; HDL-C | Deletion |
| 5,087 | 801 | |||||||
| Wang 2018 (2) (51) | China | Not-DM | 3-year | ≥18 | Female | Female | Metabolic syndrome; BMI; IGR; FPG; OGTT; BP; lipid disorders; TG; HDL-C | Deletion |
| 3,549 | 298 | |||||||
| Sun 2023 (52) | China | Not-DM | 7-year | 35–60 | 2,755 | 172 | Age; BMl; WC; lipid disorders; HBP; FPG; family history of DM; RvD1; RvD2 | – |
| Stern 2002 (53) | USA | Not-DM | 7-year | 25–64 | 2,903 | 269 | Age; sex; race; FPG; SBP; HDL-CTC; family history of DM; BMI; OGTT | – |
| Tran Quang 2022 (54) | Vietnam | Not-DM | 5-year | 40–64 | 1,448 | 140 | Region; sleep; SBP; WC; FPG | – |
| Sun 2016 (55) | China | Not-DM | 6-year | 50–70 | 2,103 | 507 | Age; sex; region; smoke; alcohol; exercise; family history of DM; BMI; FPG; HbA1c; SBP | – |
| Heianza 2012 (1) (56) | Japan | Not-DM | 5-year | 40–75 | T: 7,654 | T: 289 | NLA (Age; sex; family history of DM; smoke; BMI) | Deletion |
| EV: 1,976 | EV: 57 | |||||||
| Heianza 2012 (2) (56) | Japan | Not-DM | 5-year | 40–75 | T: 7,654 | T: 289 | NLA; FPG; HbA1c | Deletion |
| EV: 1,976 | EV: 57 | |||||||
| Heianza 2012 (3) (56) | Japan | Not-DM | 5-year | 40–75 | T: 7,654 | T: 289 | NLA; FPG | Deletion |
| EV: 1,976 | EV: 57 | |||||||
| Heianza 2012 (4) (56) | Japan | Not-DM | 5-year | 40–75 | T: 7,654 | T: 289 | NLA; HbA1c | Deletion |
| EV: 1,976 | EV: 57 | |||||||
| Oh 2021 (1) (57) | South Korea | Not-DM | 10-year | 40–69 | T: 3,973 | T: 910 | KDR (age; region; smoke; family history of DM; WC) | – |
| IV: 1,700 | IV: 381 | |||||||
| Oh 2021 (2) (57) | South Korea | Not-DM | 10-year | 40–69 | T: 3,973 | T: 910 | KDR; FPG | – |
| IV: 1,700 | IV: 381 | |||||||
| Oh 2021 (3) (57) | South Korea | Not-DM | 10-year | 40–69 | T: 3,973 | T: 910 | KDR; HbA1c | – |
| IV: 1,700 | IV: 381 | |||||||
| Nicolaisen 2022 (58) | Denmark | PDM | 5-year | ≥30 | T: 20,806 | T: 2,449 | HbA1c; age; sex; BMI; BP; diet; self-assessment of physical health | Deletion |
| IV: 5,201 | IV: – | |||||||
| Hu 2018 (1) (59) | Japan | Not-DM | 7-year | 30–59 | T: 30,500 | T: 2216 | Age; sex; abdominal obesity; BMI; smoke; BP; lipid disorders; HbA1C; FPG | Deletion |
| IV: 13,349 | IV: 1,169 | |||||||
| Hu 2018 (2) (59) | Japan | Not-DM | 7-year | 30–59 | T: 30,500 | T: 2,216 | Age; sex; abdominal obesity; BMI; smoke; HBP | Deletion |
| IV: 13,349 | IV: 1,169 | |||||||
| Xu 2024 (1) (60) | Japan | Not-DM | 5-year | 46–75 | T: 10,986 | T: 707 | Sex; BMI; family history of DM; DBP | Multiple imputation |
| EV: 11,345 | EV: 673 | |||||||
| Xu 2024 (2) (60) | Japan | Not-DM | 5-year | 46–75 | T: 10,986 | T: 707 | HbA1c; family history of DM | Multiple imputation |
| EV: 11,345 | EV: 673 | |||||||
| Xu 2024 (3) (60) | Japan | Not-DM | 5-year | 46–75 | T: 10,986 | T: 707 | FPG; HbA1c; family history of DM | Multiple imputation |
| EV: 11,345 | EV: 673 | |||||||
| Arellano-Campos 2019 (61) | Mexico | Not-DM | 3-year | ≥20 | 6,144 | 331 | Age; T; IFG; HBP; abdominal obesity | – |
| Shao 2020 (1) (62) | China | Not-DM | 10-year | ≥20 | T: 4,498 | T: 257 | Age; sex; race; HBP; smoke; alcohol; WC; BMI | Deletion |
| IV: 1,525 | IV: 92 | |||||||
| Shao 2020 (2) (62) | China | Not-DM | 10-year | ≥20 | T: 4,498 | T: 257 | Age; sex; race; BP; smoke; alcohol; WC; BMI; education; exercise; diet; triceps skinfold thickness; sleep | Deletion |
| IV: 1,525 | IV: 92 | |||||||
| Shao 2020 (3) (62) | China | Not-DM | 10-year | ≥20 | T: 4,498 | T: 257 | Age; sex; race; HBP; smoke; alcohol; WC; BMI; education; exercise; diet; triceps skinfold thickness; sleep; LDL; HDL; TC; TG; insulin; FPG; HbA1c | Deletion |
| IV: 1,525 | IV: 92 | |||||||
| Shao 2020 (4) (62) | China | Not-DM | 10-year | ≥20 | T: 4,498 | T: 257 | LDL; HDL; TC; TG; insulin; FPG; HbA1c | Deletion |
| IV: 1,525 | IV: 92 | |||||||
| Rhee 2021 (63) | South Korea | Not-DM | 10-year | ≥40 | T: 268,241 | T: 23,420 | FPG; age; sex; ALT; BMI; GGT; SBP; TC; AST; alcohol | Multiple imputation |
| IV: 67,061 | IV: 5,736 | |||||||
| Cai 2021 (64) | Japan | Not-DM | 5-year | Mean 43.64 | T: 9,651 | T: 154 | Age; fatty liver; GGT; TG; HbA1c; FPG | – |
| IV: 3,289 | IV: 47 | |||||||
| Olivera 2017 (65) | Brazil | Not-DM | 10-year | 35–74 | T: 3,709 | 1,359 | Exercise; alcohol; education; age; BMI; medication history; WHR; income; CVD; diet; family history of HBP; family history of DM; sex | Deletion |
| IV: 8,738 | ||||||||
| Liu 2024 (66) | USA | Not-DM | 5-year | ≥35 | T: 32,372 | T: 411 | Age; sex; race; BMI; SBP; HDL-C; TC; family history of DM | Deletion |
| IV: 13,875 | IV: 205 | |||||||
| Jagannathan 2020 (1) (67) | India | PDM | 3-year | Mean 44.6 | 548 | 167 | Basic model (FPG; age; sex; family history of DM; BMI; SBP; TG; HDL-C) | Deletion |
| Jagannathan 2020 (2) (67) | India | PDM | 3-year | Mean 44.6 | 548 | 167 | Basic model + IGT | Deletion |
| Jagannathan 2020 (3) (67) | India | PDM | 3-year | Mean 44.6 | 548 | 167 | Basic model; 30 min -PG > 182 mg/dl | Deletion |
| Jagannathan 2020 (4) (67) | India | PDM | 3-year | Mean 44.6 | 548 | 167 | Basic model; FPG; 30 min-PG > 182 mg/dl; I11T | Deletion |
| Savolainen 2017 (68) | Sweden | Not-DM | 5-year | 64 | 629 | 69 | Non-invasive (WC; alcohol; smoke; SBP; family history of DM) | – |
| Sun 2009 (69) | Taiwan, China | Not-DM | 5-year | 35–74 | T: 10,294 | T: 483 | Sex; education; age; smoke; BMI; WC | Deletion |
| IV: 10,257 | IV: 441 | |||||||
| Sun 2009 (69) | Taiwan, China | Not-DM | 5-year | 35–74 | T: 10,294 | T: 483 | Sex; education; age; smoke; BMI; WC; FPG | Deletion |
| IV: 10,257 | IV: 441 | |||||||
| Lim 2012 (1) (70) | South Korea | Not-DM | 4-year | 40–69 | 6,342 | 436 | Age; family history of DM; smoke; BMI; HBP | Deletion |
| Lim 2012 (2) (70) | South Korea | Not-DM | 4-year | 40–69 | 6,342 | 436 | FPG; HDL-C; TG; age; family history of DM; smoke; BMI; HBP | Deletion |
| Lim 2012 (3) (70) | South Korea | Not-DM | 4-year | 40–69 | 6,342 | 436 | HbA1c; FPG; HDL-C; TG; age; family history of DM; smoke; BMI; HBP | Deletion |
| Wang 2020 (71) | Japan | Not-DM | 5-year | 20–75 | T: 8,296 | T: 275 | Age; WC; smoke; fatty liver; FPG; HbA1c | Deletion |
| IV: 2,817 | IV: 98 | |||||||
| Tan 2022 (72) | Japan | Not-DM | 6-year | 30–60 | T: 1,518 | T: 10 | WC; HbA1c; FPG; HDL-C; TG; smoke | Deletion |
| IV: 494 | IV: 4 | |||||||
| Yatsuya 2018 (73) | Japan | Not-DM | 10-year | 35–64 | 2,960 | 548 | Age; FPG; BMI; TG; WBC; HDL-C; TC | Deletion |
| Wang 2022 (1) (74) | China | Not-DM | 4-year | Mean 61 | 15,934 | 1,302 | BP; IFG; WC | Deletion |
| Wang 2022 (2) (74) | China | Not-DM | 4-year | Mean 61 | Male | Male | BP; IFG; WC | Deletion |
| 4,324 | 341 | |||||||
| Wang 2022 (3) (74) | China | Not-DM | 4-year | Mean 61 | Female | Female | BP; IFG; WC | Deletion |
| 11,610 | 961 | |||||||
| Wang 2022 (4) (74) | China | Not-DM | 4-year | Mean 61 | 15,934 | 1,302 | Sex; age; family history of DM; BMI; HBP; HDL-C; TG; FPG; WC | Deletion |
| Wilson 2007 (75) | USA | Not-DM | 8-year | Mean 54 | 3,140 | 160 | Age, sex, family history of DM, WC, BMI, BP, HDL-C, FPG, HOMA-IR | – |
| Yan 2020 (19) | China | Not-DM | 5-year | 55 ± 13 | T: 810 | T: 110 | Age; SBP; WC; family history of DM | Deletion |
| EV: 792 | EV: 93 | |||||||
| Liu 2012 (1) (76) | China | Not-DM | 5-year | 25–75 | Female | Female | Age; BMI; WC; SBP; FPG; family history of DM; smoke | – |
| 596 | 70 | |||||||
| Liu 2012 (2) (76) | China | Not-DM | 5-year | 25–75 | Male | Male | Age; BMI; WC; SBP; FPG; family history of DM; smoke | – |
| 487 | 51 | |||||||
| Woo 2016 (77) | Hong Kong, China | Not-DM | 9-year | Mean 50.3 | 1,380 | 123 | Age; family history of DM; smoke; HBP; BMI; FPG | – |
| Jiang 2023 (78) | China | Not-DM | 7-year | 50–75 | 25,2176 | 63,423 | BMI; age; SBP; DBP; diet; exercise | Deletion |
| Asgari 2021 (79) | Iran | Not-DM | 9-year | ≥20 | T: 5,291 | T: 214 | Age; WC; height; SBP; TG; HDL-C; FPG; FHDM | Multiple imputation |
| EV: 3,147 | EV: 54 | |||||||
| Stiglic 2021 (80) | Europe | Not-DM | 10-year | ≥50 | 16,363 | 6,192 | Age; SEX; BMI; alcohol; exercise | – |
| Lindström 2003 (81) | Finland | Not-DM | 10-year | 35–64 | T: 4,746 | T: 196 | Age; BMI; WC; medication history; exercise; Glu; diet | Deletion |
| EV: 4,615 | EV: 67 | |||||||
| Schmidt 2005 (82) | USA | Not-DM | 9-year | 45–64 | T: 3,958 | T: 646 | WC; height; HBP; SBP; DBP; family history of DM; race; age; BMI; FPG; HDL-C; TC; TG | Deletion |
| IV: 3,957 | IV: 646 | |||||||
| Chen 2010 (83) | Australia | Not-DM | 5-year | >25 | T: 6,060 | T: 362 | Age; sex; race; family history of DM; Glu; medication history; smoke; exercise; C | Multiple imputation |
| EV: 2,393 | EV: 115 | |||||||
| EV: 6,034 | EV: 320 |
T, training; IV, internal validation; EV, external validation; DM, diabetes mellitus; BMI, body mass index; WC, waist circumference; FPG, fasting plasma glucose; HBP, high blood pressure; DBP, diastolic blood pressure; SBP, systolic blood pressure; TG, triglycerides; HDL-C, high-density lipoprotein cholesterol; TC, total cholesterol; LDL-C, low-density lipoprotein cholesterol; ALT, alanine aminotransferase; AST, aspartate aminotransferase; BUN, blood urea nitrogen; Scr, serum creatinine; SUA, serum uric acid; OGTT, oral glucose tolerance test; HbA1c, hemoglobin A1c; WBC, white blood cell count; HOMA-IR, homeostatic model assessment of insulin resistance; QUICKI, quantitative insulin sensitivity check index; GGT, gamma-glutamyl transferase; UA, uric acid; WHR, waist-to-hip ratio; LAP, lipid accumulation product; VAI, visceral adiposity index; CVD, cardiovascular disease; IGR, impaired glucose regulation; IFG, impaired fasting glucose; RvD1/RvD2, resolvin D1/D2; γ-GTP, gamma-glutamyl transpeptidase.
Predictive model development
The included studies predominantly employed LR as the primary modeling method (97.9% of models). A subset of models (11.3%) utilized advanced ML techniques, such as random forest (RF) and XGBoost, predominantly in larger cohorts or when modeling complex variable interactions. The top ten predictors ranked by incorporation frequency were as follows: noninvasive predictors: age (69.1%), body mass index (BMI) (64.9%), family history of diabetes (44.3%), waist circumference (WC) (39.2%), smoking status (36.1%), diastolic blood pressure (DBP) (36.1%), sex (34.0%), systolic blood pressure (SBP) (21.6%), exercise (16.5%), diet (14.4), and alcohol consumption (14.4%); noninvasive predictors/invasive predictors (e.g., laboratory biomarkers): age (69.1%), BMI (64.9%), fasting plasma glucose (FPG) (58.8%), family history of diabetes (44.3%), WC (39.2%), triglycerides (TG) (38.1%), smoking status (36.1%), DBP (36.1%), sex (34.0%), and high-density lipoprotein (HDL) (29.9%).
Missing data handling methods, though inconsistently reported, included deletion (e.g., complete-case analysis), multiple imputation, or inverse probability weighting. However, many studies (34.02%) omitted explicit descriptions of data preprocessing. Model development frequently involved partitioning datasets into training and validation subsets (e.g., random split, bootstrap), while a minority applied cross-validation (e.g., five-fold, ten-fold) to optimize generalizability. The predictive model development of the included studies is presented in Table 1.
Network relationships of diabetes predictors
The co-occurrence network analysis of diabetes predictors, derived from hierarchical clustering of the binary dataset, revealed a multidimensional interplay among variables central to diabetes risk assessment. Core metabolic markers, including FPG, glucose levels (Glu), and TG, formed tightly interconnected clusters, emphasizing their established roles in glucose and lipid metabolism dysregulation, a hallmark of diabetes pathophysiology. HDL and waist-to-hip ratio (WHR) co-occurred within this metabolic cluster, suggesting their combined utility in identifying insulin resistance and adiposity-related risks (Figs. 2 and 3).
Figure 2.
Predictor co-occurrence network. Node size reflects the frequency of each predictor in the cohort; line thickness indicates how often two predictors co-occur within the same participant. TC: total cholesterol; FPG: fasting plasma glucose; TG: triglycerides; HDL: high-density lipoprotein; BMI: body mass index; WC: waist circumference; BP: blood pressure; WHR: waist-to-hip ratio; CVD: cardiovascular disease; PDM: pre-diabetes mellitus; OGTT: oral glucose tolerance test; Glu: glucose; HbA1c: glycated hemoglobin; SCR: serum creatinine; ALT: alanine transaminase; LDL: low-density lipoprotein; SBP: systolic blood pressure; DBP: diastolic blood pressure.
Figure 3.
Hierarchically clustered heatmap. Heatmap of standardized predictor levels (columns) for all participants (rows). Ward linkage on Euclidean distance produced four predictor clusters (top dendrogram).
Anthropometric measures such as skinfold thickness, weight, and BMI exhibited strong associations with cardiovascular indicators (SBP, DBP, and family history of HBP), underscoring the synergistic contribution of obesity and hypertension to diabetes progression. Sociodemographic factors (education, marital status, and race) and lifestyle variables (alcohol, exercise, and sleep) clustered distinctly, reflecting their collective influence on behavioral and environmental risk stratification. The centrality of family history of diabetes mellitus (DM) and prediabetes mellitus (PDM) within the network highlights the critical role of genetic predisposition and early metabolic dysregulation in diabetes prediction (Figs 2 and 3).
This interconnectedness underscores the necessity of integrative models that account for biological, behavioral, and hereditary factors in diabetes prediction. The absence of isolated nodes (e.g., income, residence) further supports the hypothesis that socioeconomic and geographic determinants indirectly modulate diabetes risk through intermediate variables such as diet and healthcare access (17).
Predictive model performance and validation
A total of 70.1% prediction models underwent training, 93.8% received internal validation, and 21.6% were externally validated (Supplementary Tables S1, S2 and S3). For training models, LR remained the dominant approach (82.4% of models), while ML algorithms, including XGBoost, RF, and Neural Network (NN), were applied in 16.2% of models. For internal validation, LR remained the dominant approach (73.6% of models), while ML algorithms, including XGBoost, RF, naïve Bayes (NB), decision trees (DT), and NN, were applied in 26.4% of models. In contrast, all externally validated studies exclusively relied on LR (85.7%) and Cox (14.3%), with no ML-based models undergoing external validation. Internal validation techniques (n = 91, 93.8% of models), including random data splitting (n = 38, 39.2% of models), cross-validation (n = 21, 21.6%), and bootstrapping (n = 18, 18.6% of models), were employed more frequently than external validation methods (n = 21, 21.6% of models), such as spatial validation (n = 12, 12.4% of models) or temporal validation (n = 8, 8.2%).
Evaluation of model discrimination using AUC-ROC or concordance index (C-index) revealed consistently moderate to high performance, with values spanning 0.628–0.998 across training, internal validation, and external validation. Calibration, assessed via Hosmer-Lemeshow (HL) tests or Brier scores in 57.7% of models, indicated adequate alignment between predicted and observed risks in most cases. Model presentation formats included risk score formulas (28.9%), nomograms (16.5%), and scoring sheets (13.4%), with limited adoption of dynamic or interactive tools. The predictive model performance and validation of the included studies are presented in Table 1.
Risk of bias assessment using PROBAST tool
The PROBAST-based evaluation of included studies (n = 65, encompassing 98 distinct prediction models) assessed the risk of bias across four domains: study subjects, predictor factors, research results, and statistical analysis, while also assessing the risk of applicability across three domains: study subjects, predictor factors, and research results (Table 2, Figs 4 and 5). For risk of bias, all studies demonstrated low risk in study subjects (100%), and most studies demonstrated low risk in predictor factors (84.5%), reflecting robust participant selection and clinically relevant predictor inclusion (e.g., age, BMI, FPG). However, 87.6% exhibited high risk in statistical analysis, primarily due to deletion of missing data or lack of external validation. Although external validation was performed, 5.2% of studies were rated as having unclear risk due to insufficient description of missing data handling. Overall, 91.8% of models were classified as high risk for bias. As for risk of applicability, the assessments of study subjects revealed low generalizability, with 15.5% rated low risk. In addition, 22.7% of studies were rated as high risk due to age-restricted cohorts. The evidence downgrading rationale for specific predictor factors corresponded to risk-of-bias assessments. Evidence quality was downgraded owing to either inappropriate handling of missing data (46.39%) or absence of documentation describing missing value processing procedures (34.02%).
Table 2.
Risk of bias assessment.
| Included studies | Study subjects | Predictor factors | Research results | Statistical analysis | Study subjects | Predictor factors | Research results | Risk of bias | Applicability |
|---|---|---|---|---|---|---|---|---|---|
| Han 2017 (23) | Low | Low | Low | High | High | Low | High | High | High |
| Sun 2013 (1) (21) | Low | Low | Low | Unclear | Low | Low | Unclear | High | High |
| Sun 2013 (2) (21) | Low | Low | Low | Unclear | Low | Low | Unclear | High | High |
| Sun 2013 (3) (21) | Low | Low | Low | Unclear | Low | Low | Unclear | High | High |
| Cai 2020 (24) | Low | High | Low | High | Low | High | Low | High | High |
| Hippisley-Cox 2009 (1) (18) | Low | Low | Low | Low | Low | Low | Low | Low | Low |
| Hippisley-Cox 2009 (2) (18) | Low | Low | Low | Low | Low | Low | Low | Low | Low |
| Wen 2016 (1) (25) | Low | Low | Low | High | Low | Low | Low | High | Low |
| Wen 2016 (2) (25) | Low | Low | Low | High | Low | Low | Low | High | Low |
| Wang 2018 (26) | Low | Low | Low | High | Low | Low | Low | High | Low |
| Bai 2018 (27) | Low | High | Low | High | Low | High | Unclear | High | High |
| Lei 2023 (28) | Low | High | Low | High | High | High | High | High | High |
| Fu 2023 (29) | Low | High | Low | High | High | High | High | High | High |
| Su 2017 (1) (30) | Low | High | Low | High | Low | High | Unclear | High | High |
| Su 2017 (2) (30) | Low | High | Low | High | Low | High | Unclear | High | High |
| Zhang 2022 (31) | Low | Low | Low | High | High | Low | Unclear | High | High |
| Yang 2016 (32) | Low | High | Low | High | Low | High | Unclear | High | High |
| Zhang 2016 (1) (33) | Low | Low | Low | High | Low | Low | Unclear | High | High |
| Zhang 2016 (2) (33) | Low | Low | Low | High | Low | Low | Unclear | High | High |
| Zhu 2023 (1) (34) | Low | Low | Low | High | Low | Low | High | High | High |
| Zhu 2023 (2) (34) | Low | Low | Low | High | Low | Low | High | High | High |
| Chen 2018 (35) | Low | Low | Low | High | Low | Low | Unclear | High | High |
| Ma 2020 (36) | Low | Low | Low | High | Low | Low | High | High | High |
| Liu 2020 (37) | Low | Low | Low | High | Low | Low | Low | High | Low |
| Lee 2024 (1) (38) | Low | Low | Low | Unclear | Low | Low | Unclear | High | High |
| Lee 2024 (2) (38) | Low | Low | Low | Unclear | Low | Low | Unclear | High | High |
| Alssema 2011 (39) | Low | Low | Low | High | High | Low | Unclear | High | High |
| Li 2019 (40) | Low | Low | Low | High | High | Low | High | High | High |
| Gao 2009 (1) (41) | Low | Low | Low | High | Low | Low | High | High | High |
| Gao 2009 (2) (41) | Low | Low | Low | High | Low | Low | High | High | High |
| Edlitz 2022 (1) (20) | Low | Low | Low | High | Low | Low | Low | High | Low |
| Edlitz 2022 (2) (20) | Low | Low | Low | High | Low | Low | Low | High | Low |
| Katsimpris 2021 (42) | Low | High | Low | High | Low | High | Unclear | High | High |
| Liu 2016 (43) | Low | Low | Low | High | High | Low | High | High | High |
| Hafezi 2024 (44) | Low | Low | Low | High | High | Low | Unclear | High | High |
| Hu 2020 (45) | Low | Low | Low | High | High | Low | Unclear | High | High |
| Xu 2014 (46) | Low | Low | Low | High | High | Low | Low | High | High |
| Rathmann 2010 (47) | Low | Low | Low | High | High | Low | Unclear | High | High |
| Liu 2022 (48) | Low | High | Low | High | High | High | Unclear | High | High |
| Zhang 2020 (49) | Low | High | Low | High | High | High | Unclear | High | High |
| Alghamdi 2017 (50) | Low | High | Low | High | Low | High | Unclear | High | High |
| Wang 2018 (1) (51) | Low | Low | Low | High | Low | Low | High | High | High |
| Wang 2018 (2) (51) | Low | Low | Low | High | Low | Low | High | High | High |
| Sun 2023 (52) | Low | Low | Low | High | Low | Low | Unclear | High | High |
| Stern 2002 (53) | Low | Low | Low | High | Low | Low | Unclear | High | High |
| Tran Quang 2022 (54) | Low | Low | Low | High | Low | Low | Unclear | High | High |
| Sun 2016 (55) | Low | Low | Low | High | High | Low | Unclear | High | High |
| Heianza 2012 (1) (56) | Low | Low | Low | High | Low | Low | High | High | High |
| Heianza 2012 (2) (56) | Low | Low | Low | High | Low | Low | High | High | High |
| Heianza 2012 (3) (56) | Low | Low | Low | High | Low | Low | High | High | High |
| Heianza 2012 (4) (56) | Low | Low | Low | High | Low | Low | High | High | High |
| Oh 2021 (1) (57) | Low | Low | Low | High | Low | Low | Unclear | High | High |
| Oh 2021 (2) (57) | Low | Low | Low | High | Low | Low | Unclear | High | High |
| Oh 2021 (3) (57) | Low | Low | Low | High | Low | Low | Unclear | High | High |
| Nicolaisen 2022 (58) | Low | Low | Low | High | Low | Low | High | High | High |
| Hu 2018 (1) (59) | Low | Low | Low | High | Low | Low | High | High | High |
| Hu 2018 (2) (59) | Low | Low | Low | High | Low | Low | High | High | High |
| Xu 2024 (1) (60) | Low | Low | Low | Low | Low | Low | Low | Low | Low |
| Xu 2024 (2) (60) | Low | Low | Low | Low | Low | Low | Low | Low | Low |
| Xu 2024 (3) (60) | Low | Low | Low | Low | Low | Low | Low | Low | Low |
| Arellano-Campos 2019 (61) | Low | Low | Low | High | Low | Low | Unclear | High | High |
| Shao 2020 (1) (62) | Low | Low | Low | High | Low | Low | High | High | High |
| Shao 2020 (2) (62) | Low | Low | Low | High | Low | Low | High | High | High |
| Shao 2020 (3) (62) | Low | Low | Low | High | Low | Low | High | High | High |
| Shao 2020 (4) (62) | Low | Low | Low | High | Low | Low | High | High | High |
| Rhee 2021 (63) | Low | Low | Low | High | Low | Low | Low | High | Low |
| Cai 2021 (64) | Low | Low | Low | High | Low | Low | Unclear | High | High |
| Olivera 2017 (65) | Low | Low | Low | High | Low | Low | High | High | High |
| Liu 2024 (66) | Low | Low | Low | High | Low | Low | High | High | High |
| Jagannathan 2020 (1) (67) | Low | Low | Low | High | Low | Low | High | High | High |
| Jagannathan 2020 (2) (67) | Low | Low | Low | High | Low | Low | High | High | High |
| Jagannathan 2020 (3) (67) | Low | Low | Low | High | Low | Low | High | High | High |
| Jagannathan 2020 (4) (67) | Low | Low | Low | High | Low | Low | High | High | High |
| Savolainen 2017 (68) | Low | Low | Low | High | Low | Low | Unclear | High | High |
| Sun 2009 (1) (69) | Low | Low | Low | High | Low | Low | High | High | High |
| Sun 2009 (2) (69) | Low | Low | Low | High | Low | Low | High | High | High |
| Lim 2012 (1) (70) | Low | Low | Low | High | Low | Low | High | High | High |
| Lim 2012 (2) (70) | Low | Low | Low | High | Low | Low | High | High | High |
| Lim 2012 (3) (70) | Low | Low | Low | High | Low | Low | High | High | High |
| Wang 2020 (71) | Low | High | Low | High | Low | High | High | High | High |
| Tan 2022 (72) | Low | High | Low | High | Low | High | High | High | High |
| Yatsuya 2018 (73) | Low | High | Low | High | Low | High | High | High | High |
| Wang 2022 (1) (74) | Low | Low | Low | High | High | Low | High | High | High |
| Wang 2022 (2) (74) | Low | Low | Low | High | High | Low | High | High | High |
| Wang 2022 (3) (74) | Low | Low | Low | High | High | Low | High | High | High |
| Wang 2022 (4) (74) | Low | Low | Low | High | High | Low | High | High | High |
| Wilson 2007 (75) | Low | Low | Low | High | High | Low | Unclear | High | High |
| Yan 2020 (19) | Low | Low | Low | High | Low | Low | High | Low | Low |
| Liu 2012 (1) (76) | Low | Low | Low | High | Low | Low | High | High | High |
| Liu 2012 (2) (76) | Low | Low | Low | High | Low | Low | High | High | High |
| Woo 2016 (77) | Low | Low | Low | High | High | Low | Unclear | High | High |
| Jiang 2023 (78) | Low | Low | Low | High | High | Low | High | High | High |
| Asgari 2021 (79) | Low | Low | Low | Low | Low | Low | Low | Low | Low |
| Stiglic 2021 (80) | Low | High | Low | High | High | High | Unclear | High | High |
| Lindström 2003 (81) | Low | Low | Low | High | Low | Low | High | High | High |
| Schmidt 2005 (82) | Low | Low | Low | High | Low | Low | Unclear | High | High |
| Chen 2010 (83) | Low | Low | Low | Low | Low | Low | Low | Low | Low |
Low, low risk; High, high risk; Unclear, unclear risk.
Figure 4.
Risk of bias assessment for methodology.
Figure 5.
Risk of bias assessment for applicability.
Subgroup analysis
Modeling
The subgroup analysis revealed variations in model performance across categories. Geographically, models from China (n = 38; AUC = 0.79, 95% CI: 0.77–0.81) and Japan (n = 12; AUC = 0.82, 95% CI: 0.78–0.86) demonstrated moderate discrimination, while USA-based models (n = 3; AUC = 0.97, 95% CI: 0.91–1.02) showed higher performance. By modeling approach, LR dominated (n = 56; AUC = 0.78, 95% CI: 0.76–0.80), whereas ML methods such as XGBoost (n = 3; AUC = 0.87, 95% CI: 0.76–0.99) and RF (n = 3; AUC = 0.86, 95% CI: 0.72–1.00) achieved higher AUCs than LR. Population-specific analysis indicated stronger performance in non-diabetic cohorts (n = 58; AUC = 0.80, 95% CI: 0.78–0.82) compared to prediabetic groups (n = 10; AUC = 0.74, 95% CI: 0.70–0.77). Follow-up duration (≥5 years: AUC = 0.81, 0.78–0.83), predictors (≥10: AUC = 0.80, 95% CI: 0.77–0.83), and sample size (≥1,000: AUC = 0.80, 95% CI: 0.77–0.82) showed better AUCs (Table 3).
Table 3.
Subgroups of training.
| Study | Numbers of predictive models | AUC (95% CI) | I2 (%) |
|---|---|---|---|
| Total | 68 | 0.79 (0.77–0.81) | 99.5 |
| Country | |||
| China | 38 | 0.79 (0.77–0.81) | 97.4 |
| South Korea | 4 | 0.75 (0.68–0.82) | 97.1 |
| Iran | 3 | 0.77 (0.69–0.86) | 98.0 |
| USA | 3 | 0.97 (0.91–1.02) | 99.8 |
| Japan | 12 | 0.82 (0.78–0.86) | 98.3 |
| Sweden | 1 | 0.64 (0.56–0.71) | – |
| Australia | 1 | 0.78 (0.76–0.80) | – |
| Brazil | 2 | 0.75 (0.74–0.76) | 0 |
| India | 4 | 0.68 (0.65–0.72) | 45.6 |
| Machine | |||
| LR | 56 | 0.78 (0.76–0.80) | 99.2 |
| Cox | 1 | 0.85 (0.82–0.88) | – |
| XGBoost | 3 | 0.87 (0.76–0.99) | 99.3 |
| RF | 3 | 0.86 (0.72–1.00) | 99.6 |
| NN | 5 | 0.79 (0.77–0.82) | 90.8 |
| Population | |||
| Non-diabetic | 58 | 0.80 (0.78–0.82) | 99.5 |
| Prediabetic | 10 | 0.74 (0.70–0.77) | 87.5 |
| Follow up (years) | |||
| ≥5 | 47 | 0.81 (0.78–0.83) | 99.6 |
| <5 | 21 | 0.76 (0.73–0.79) | 93.3 |
| Numbers | |||
| ≥1,000 | 54 | 0.80 (0.77–0.82) | 99.5 |
| <1,000 | 14 | 0.78 (0.72–0.83) | 99.3 |
| Number of predictors | |||
| ≥10 | 51 | 0.80 (0.77–0.83) | 99.7 |
| <10 | 17 | 0.78 (0.75–0.81) | 96.1 |
LR, logistic regression; XGBoost, extreme gradient boosting; RF, random forest.
The results of sensitivity analysis demonstrated minimal variation in the effect estimates, with the 95% CIs remaining stable across all iterations (Supplementary Fig. 1). The overall results remained reliable, supporting the validity of the primary findings (Figs 6 and 7).
Figure 6.
Meta-analysis of training.
Figure 7.
Funnel plots of training.
Internal validation
The subgroup analysis revealed distinct performance patterns across categories. Geographically, models from Germany (n = 2; AUC = 0.86, 95% CI: 0.81–0.91) and Japan (n = 8; AUC = 0.81, 95% CI: 0.75–0.87) demonstrated the highest discrimination, while Indian models (n = 3; AUC = 0.69, 95% CI: 0.66–0.72) showed the lowest. By modeling approach, LR dominated (n = 67; AUC = 0.78, 95% CI: 0.76–0.80), with RF (n = 8; AUC = 0.80, 95% CI: 0.74–0.86) achieving marginally higher performance. Population-specific analysis indicated stronger discrimination in non-diabetic cohorts (n = 79; AUC = 0.78, 95% CI: 0.76–0.80) compared to prediabetic groups (n = 9; AUC = 0.72, 95% CI: 0.71–0.73). Follow-up duration (≥5 years: AUC = 0.79, 95% CI: 0.77–0.82), (<10: AUC = 0.79, 95% CI: 0.77–0.81), and sample size (<1,000: AUC = 0.80, 95% CI: 0.76–0.84) showed a razor-thin differences (Table 4).
Table 4.
Subgroups of internal validation.
| Study | Numbers of predictive models | AUC (95% CI) | I2 (%) |
|---|---|---|---|
| Total | 91 | 0.78 (0.76–0.79) | 99.9 |
| Country | |||
| China | 49 | 0.79 (0.77–0.81) | 99.9 |
| South Korea | 7 | 0.76 (0.71–0.81) | 96.5 |
| Europe | 2 | 0.73 (0.67–0.79) | 97.6 |
| USA | 12 | 0.77 (0.72–0.82) | 99.9 |
| Germany | 2 | 0.86 (0.81–0.91) | 86.2 |
| Iran | 1 | 0.70 (0.68–0.71) | – |
| Australia | 2 | 0.72 (0.66–0.77) | 99% |
| Vietnam | 1 | 0.71 (0.67–0.76) | – |
| Denmark | 1 | 0.73 (0.71–0.74) | – |
| Japan | 8 | 0.81 (0.75–0.87) | 99.5 |
| Mexico | 1 | 0.75 (0.72–0.78) | – |
| Brazil | 2 | 0.74 (0.74–0.76) | 0 |
| India | 3 | 0.69 (0.66–0.72) | 0 |
| Machine | |||
| LR | 67 | 0.78 (0.76–0.80) | 99.5 |
| Cox | 1 | 0.83 (0.82–0.83) | – |
| XGBoost | 5 | 0.77 (0.71–0.83) | 99.9 |
| RF | 8 | 0.80 (0.74–0.86) | 99.9 |
| NB | 1 | 0.68 (0.67–0.68) | – |
| DT | 2 | 0.68 (0.59–0.77) | 99.8 |
| NN | 7 | 0.77 (0.73–0.80) | 98.4 |
| Population | |||
| Non-diabetic | 82 | 0.78 (0.77–0.80) | 99.9 |
| Prediabetic | 9 | 0.72 (0.71–0.73) | 30.1 |
| Follow up (years) | |||
| ≥5 | 44 | 0.79 (0.77–0.82) | 99.9 |
| <5 | 47 | 0.76 (0.75–0.78) | 99.1 |
| Numbers | |||
| ≥1,000 | 75 | 0.77 (0.76–0.79) | 99.9 |
| <1,000 | 16 | 0.80 (0.76–0.84) | 94.6 |
| Number of predictors | |||
| ≥10 | 32 | 0.76 (0.73–0.79) | 99.4 |
| <10 | 59 | 0.79 (0.77–0.81) | 99.9 |
LR, logistic regression; XGBoost, extreme gradient boosting; RF, random forest; NB, naïve Bayes; DT, decision tree.
The results of sensitivity analysis indicated minimal fluctuations in the effect estimates, with 95% CIs consistently overlapping across iterations (Supplementary Fig. 2). Funnel plot asymmetry was assessed using Egger’s regression test, which showed no significant evidence of publication bias (P = 0.0795). The symmetrical distribution of studies in the funnel plot further supported the absence of substantial bias (Figs 8 and 9).
Figure 8.
Meta-analysis of internal validation.
Figure 9.
Full plots of internal validation.
External validation
Geographically, models from China and South Korea showed comparable discrimination. Japanese models (n = 7; AUC = 0.83, 95% CI: 0.76–0.89) displayed slightly lower performance. In addition, England models (n = 2; AUC = 0.84, 95% CI: 0.82–0.86) and the single Finnish model achieved higher AUC (0.87, 95% CI: 0.86–0.88). By sample size, models with ≥1,000 participants (n = 20; AUC = 0.81, 95% CI: 0.77–0.84) outperformed smaller cohorts (n = 1; AUC = 0.69, 95% CI: 0.65–0.74) (Table 5).
Table 5.
Subgroups of external validation.
| Study | Numbers of predictive models | AUC (95% CI) | I2 (%) |
|---|---|---|---|
| Total | 21 | 0.80 (0.73–0.87) | 96.1 |
| Country | |||
| China | 4 | 0.83 (0.77–0.89) | 98.6 |
| South Korea | 2 | 0.83 (0.81–0.85) | 0 |
| England | 2 | 0.84 (0.82–0.86) | 98.9 |
| Japan | 7 | 0.83 (0.76–0.89) | 99.5 |
| Finland | 1 | 0.87 (0.86–0.88) | – |
| Iran | 1 | 0.80 (0.79–0.81) | – |
| USA | 2 | 0.68 (0.67–0.69) | 15.4 |
| Australia | 2 | 0.72 (0.60–0.85) | 87.8 |
| Numbers | |||
| ≥1,000 | 20 | 0.81 (0.77–0.84) | 99.3 |
| <1,000 | 1 | 0.69 (0.63–0.74) | – |
The results of sensitivity analysis demonstrated consistent effect sizes across iterations, with 95% CIs consistently overlapping across iterations (Supplementary Fig. 3). Egger’s regression test for funnel plot asymmetry revealed no statistically significant evidence of publication bias (P = 0.5701). The symmetrical distribution of studies in the funnel plot further supported the absence of substantial bias (Figs 10 and 11).
Figure 10.
Meta-analysis of external validation.
Figure 11.
Funnel plots of external validation.
Decision curve analysis
DCA of 21 externally validated prediction models (Fig. 12) demonstrated substantial variation in clinical utility across threshold probabilities. Model performance exhibited threshold-dependent characteristics strongly associated with methodological and population factors. Models derived from larger cohorts (Hippisley-Cox et al. 2009 (1–2) (18): n = 1,232,832) maintained superior net benefit (>0.04 across 0.15–0.45 thresholds), while smaller studies (Yan et al. 2020 (19): n = 792; Chen et al. 2010 (83): n = 2,393–6,034) showed >30% variability at thresholds >0.4. Multi-predictor models (Heianza et al. 2012 (2) (56): seven predictors, AUC = 0.913; Heianza et al. 2012 (4) (56): six predictors, AUC = 0.879) achieved sustained net benefit (>0.03) over broader threshold ranges (0.2–0.4), contrasting with simpler constructs (Xu et al. 2024 (1) (60): two predictors, AUC = 0.692) that demonstrated threshold-specific limitations. East Asian cohorts (Xu et al. 2024 (3) (60): Japan, AUC = 0.874) peaked at 0.25–0.35 thresholds (net benefit >0.05). Western populations (Edlitz & Segal 2022 (1–2) (20): USA, AUC: 0.68–0.69) showed optimal performance at higher thresholds (>0.35). The English cohort (Hippisley-Cox et al. 2009 (2) (18): AUC = 0.853) demonstrated exceptional generalizability. The highest-AUC model (Heianza et al. 2012 (2) (56): AUC = 0.913) showed threshold-restricted utility (0.2–0.3), while moderate-AUC models with larger samples (Sun et al. 2013 (21), AUC: 0.819–0.85, n = 12,392) maintained >0.035 net benefit across 0.15–0.4 thresholds.
Figure 12.
DCA. Net benefit curves derived from 19 pooled prediction models. The ‘treat-all’ (gray) and ‘treat-none’ (black) strategies serve as references.
Discussion
The escalating global prevalence of T2DM underscores the urgent need for robust risk prediction models to guide early intervention and precision prevention strategies (22, 23). This systematic review and meta-analysis critically appraise 65 multivariable prediction studies (encompassing 97 distinct prediction models) published between 2002 and 2024. The pooled discrimination metrics (AUC: 0.628–0.998) align with prior systematic reviews, which reported similar ranges but highlighted inconsistent validation practices (7). Notably, ML models achieved marginally higher AUCs than LR (e.g., training, XGBoost AUC: 0.87 vs LR: 0.78; internal validation, RF AUC: 0.80 vs LR AUC: 0.78), yet their real-world implementation is hampered by limited interpretability and lack of external validation (21.6% of externally validated models used LR) (4). The challenges identified in the PROBAST framework, where 91.8% of models exhibited high risks of bias due to inadequate statistical reporting and lack of external verification, and 84.5% of models exhibited high risks of applicability due to inadequate statistical reporting (e.g., missing data deletion without justification) and insufficient adjustment for confounders (5). We urge journals to mandate adherence to the TRIPOD-AI statement (22) for transparent reporting of prediction models, especially for ML-based studies.
Predictor network analysis revealed metabolic markers (FPG, TG, HDL) and anthropometric measures (BMI, WC) as interconnected hubs, while socioeconomic factors (education, income) acted through indirect pathways (e.g., diet access). This supports developing integrated models that combine clinical, behavioral, and social determinants.
Geographic disparities (USA AUC: 0.97 vs China AUC: 0.79) and poor performance in pre-diabetic cohorts (AUC: 0.72 vs 0.80 in normoglycemic) highlight the need for context-specific adaptations. To address these gaps, we recommend prioritizing external validation across diverse populations, particularly for ML models, of which only 21.6% have undergone any form of external validation. Furthermore, we propose the establishment of multinational consortia to facilitate the sharing of harmonized data for the purposes of model training and testing. In addition, we advocate for the development of dynamic prediction tools that integrate time-varying predictors, such as lifestyle changes, with a particular focus on the progression of prediabetes.
The study has some limitations, including language bias (only Chinese/English studies) and outcome definition variability. The inclusion of studies published in Chinese may limit the ability of the international scientific community to independently assess the full details of these studies, potentially affecting the reproducibility and critical appraisal of those specific models by non-Chinese readers. We advocate that researchers, especially in regions producing high volumes of research, register study protocols and share model specifications (e.g., coefficients, nomograms, code) on international platforms (e.g., ClinicalTrials.gov, DRYAD, Figshare), which would significantly enhance global accessibility and reproducibility, regardless of the primary publication language. Second, although we requested calibration intercept and slope from authors and searched appendices, these metrics were unavailable for the vast majority of models. We, therefore, present a qualitative summary of the calibration plots. This limitation reinforces our recommendation that future prediction-model studies adopt the TRIPOD-AI statement, which mandates calibration intercept, slope, and full calibration plots.
In conclusion, while T2DM prediction models show promise, their clinical utility is compromised by methodological shortcomings and limited generalizability. By implementing structured reporting standards, fostering cross-border collaboration, and focusing on high-risk subgroups, future models can advance precision prevention and mitigate the global diabetes burden.
Supplementary materials
Declaration of interest
The authors declare that there is no conflict of interest that could be perceived as prejudicing the impartiality of the work reported.
Funding
This research was funded by the State Key Laboratory of Dampness Syndrome of Chinese Medicine (No. 2022KT1315, 2020KT1074), the National Key Technology R&D Program for the 13th Five-year Plan of the Ministry of Science and Technology, China (No. 2018YFC1707407), and the Guangdong Provincial Hospital of Chinese Medicine Special Project for Science and Technology Research in Traditional Chinese Medicine (Grant No. YN2020QN28).
Author contribution statement
Meichen Li designed the study. Linjie Song and Xingying Qiu conducted the study. Meichen Li and Aowei Tan performed the data analysis. Meichen Li and Aowei Tan interpreted the data. All authors drafted the manuscript. Zhou Li and Zeihuai Wen reviewed and edited the manuscript. Zehuai Wen acquired the funding. All authors approved the final version of the manuscript.
References
- 1.GBD 2021 Causes of Death Collaborators . Global burden of 288 causes of death and life expectancy decomposition in 204 countries and territories and 811 subnational locations, 1990–2021: a systematic analysis for the global burden of disease study 2021. Lancet 2024. 403 2100–2132. ( 10.1016/s0140-6736(24)00367-2) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.DeFronzo RA, Ferrannini E, Groop L, et al. Type 2 diabetes mellitus. Nat Rev Dis Primers 2015. 1 15019. ( 10.1038/nrdp.2015.19) [DOI] [PubMed] [Google Scholar]
- 3.Huynh K. Diabetes: diabetes-induced microvascular destabilization. Nat Rev Cardiol 2017. 14 128–129. ( 10.1038/nrcardio.2017.9) [DOI] [PubMed] [Google Scholar]
- 4.Noble D, Mathur R, Dent T, et al. Risk models and scores for type 2 diabetes: systematic review. BMJ 2011. 343 d7163. ( 10.1136/bmj.d7163) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Moons K, Wolff RF, Riley RD, et al. PROBAST: a tool to assess risk of bias and applicability of prediction model studies: explanation and elaboration. Ann Intern Med 2019. 170 W1–W33. ( 10.7326/m18-1377) [DOI] [PubMed] [Google Scholar]
- 6.Omar R. Clinical prediction models: a practical approach to development, validation and updating by Steyerberg, E. W. Biometrics 2010. 66 661–662. ( 10.1111/j.1541-0420.2010.01431.x) [DOI] [Google Scholar]
- 7.Collins GS, Reitsma JB, Altman DG, et al. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. Ann Intern Med 2015. 162 55–63. ( 10.7326/m14-0697) [DOI] [PubMed] [Google Scholar]
- 8.Maiorino MI, Signoriello S, Maio A, et al. Effects of continuous glucose monitoring on metrics of glycemic control in diabetes: a systematic review with meta-analysis of randomized controlled trials. Diabetes Care 2020. 43 1146–1156. ( 10.2337/dc19-1459) [DOI] [PubMed] [Google Scholar]
- 9.Zakizadeh M & Goundar S. Advancements in artificial intelligence algorithms for precise diabetes prediction and analysis in the healthcare landscape: a systematic and analytical investigation. In 2024 20th CSI International Symposium on Artificial Intelligence and Signal Processing (AISP), pp 1–8. Piscataway, NJ, USA: IEEE, 2024. ( 10.1109/AISP61396.2024.10475207) [DOI] [Google Scholar]
- 10.Walker RJ, Smalls BL, Campbell JA, et al. Impact of social determinants of health on outcomes for type 2 diabetes: a systematic review. Endocrine 2014. 47 29–48. ( 10.1007/s12020-014-0195-0) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Zhang Y, Tong M, Wang B, et al. Geographic, gender, and seasonal variation of diabetes: a nationwide study with 1.4 million participants. J Clin Endocrinol Metab 2021. 106 e4981–e4992. ( 10.1210/clinem/dgab543) [DOI] [PubMed] [Google Scholar]
- 12.Witka BZ, Oktaviani DJ, Marcellino M, et al. Type 2 diabetes-associated genetic polymorphisms as potential disease predictors. Diabetes Metab Syndr Obes 2019. 12 2689–2706. ( 10.2147/dmso.s230061) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 2021. 372 n71. ( 10.1136/bmj.n71) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Snell K, Levis B, Damen J, et al. Transparent reporting of multivariable prediction models for individual prognosis or diagnosis: checklist for systematic reviews and meta-analyses (TRIPOD-SRMA). BMJ 2023. 381 e73538. ( 10.1136/bmj-2022-073538) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Fernandez-Felix BM, López-Alcalde J, Roqué M, et al. CHARMS and PROBAST at your fingertips: a template for data extraction and risk of bias assessment in systematic reviews of predictive models. BMC Med Res Methodol 2023. 23 44. ( 10.1186/s12874-023-01849-0) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Riley RD, Higgins JP & Deeks JJ. Interpretation of random effects meta-analyses. BMJ 2011. 342 d549. ( 10.1136/bmj.d549) [DOI] [PubMed] [Google Scholar]
- 17.Hill-Briggs F, Adler NE, Berkowitz SA, et al. Social determinants of health and diabetes: a scientific review. Diabetes Care 2020. 44 258–279. ( 10.2337/dci20-0053) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Hippisley-Cox J, Coupland C, Robson J, et al. Predicting risk of type 2 diabetes in England and Wales: prospective derivation and validation of QDScore. BMJ 2009. 338 b880. ( 10.1136/bmj.b880) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Yan Q, Huang X, Zhong M, et al. Effectiveness of a diabetes-risk-score-based screening for type 2 diabetes mellitus in Pudong New Area, Shanghai. Chin J Diabetes 2020. 12 504–508. ( 10.3760/cma.j.cn115791-20190926-00342) [DOI] [Google Scholar]
- 20.Edlitz Y & Segal E. Prediction of type 2 diabetes mellitus onset using logistic regression-based scorecards. Elife 2022. 11 e71862. ( 10.7554/elife.71862) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Sun F, Tao QM, Tao QS, et al. Type 2 diabetes prevalence and estimated risks on developing type 2 diabetes within 5 years for adults aged 35–74 based on Chinese Taiwan MJ health-checkup database. Chin J Dis Control Prev 2013. 17 369–373. [Google Scholar]
- 22.Collins GS, Moons K, Dhiman P, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 2024. 385 e78378. ( 10.1136/bmj-2023-078378) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Han X. Identification of Risk Factors for Type 2 Diabetes Mellitus and Development of a Prediction Model for Incident Risk in Middle-Aged and Elderly Populations. In [Masters Thesis]. Beijing, China: Peking Union Medical College, 2017. [Google Scholar]
- 24.Cai X, Zhang D, Jing H, et al. Nomogram for predicting the 5-year risk of type 2 diabetes mellitus in a Chinese population. J Hainan Med Univ 2020. 26 1179–1184. ( 10.13210/j.cnki.jhmu.20200403.003) [DOI] [Google Scholar]
- 25.Wen J. Screening of Biomarkers and Establishment of a Prediction Model for Incident Type 2 Diabetes Mellitus in Rural Adults. In [PhD Thesis]. Beijing, China: Chinese PLA General Hospital/PLA Medical School, 2016. [Google Scholar]
- 26.Wang B. A Cohort Study on The Relationship Between Different Obesity and Metabolic Phenotypes and Type 2 Diabetes Mellitus in Rural Adults. In [PhD Thesis]. Zhengzhou, China: Zhengzhou University, 2018. [Google Scholar]
- 27.Bai JL, Zhang CY, Li W, et al. Diabetes prediction models based on health check-up population in a hospital. Pract Prev Med 2018. 25 116–119. ( 10.3969/j.issn.1006-3110.2018.01.035) [DOI] [Google Scholar]
- 28.Lei Q, Li JH, Huo YF, et al. Nomogram prediction of the risk of type 2 diabetes mellitus in senior citizens in China. Sci Technol Eng 2023. 23 5493–5499. ( 10.3969/j.issn.1006-3110.2018.01.035) [DOI] [Google Scholar]
- 29.Fu X, Liu H, Gu Z, et al. Nomogram for risk prediction of type 2 diabetes in the elderly. Chin J Mult Organ Dis Elder 2023. 22 274–279. ( 10.11915/j.issn.1671-5403.2023.04.056) [DOI] [Google Scholar]
- 30.Su P, Yang Y, Yang Y, et al. A risk prediction model for incident type 2 diabetes mellitus in a health-management population. J Shandong Univ 2017. 55 82–86. ( 10.6040/j.issn.1671-7554.0.2017.347) [DOI] [Google Scholar]
- 31.Zhang Y, He Y, Liu F, et al. Analysis of a longitudinal health-examination-data-based risk prediction model for incident diabetes mellitus in the elderly. Geriatr Health Care 2022. 28 866–870. ( 10.3969/j.issn.1008-8296.2022.04.034) [DOI] [Google Scholar]
- 32.Yang Y, Zhang G, Zhang C, et al. A risk prediction model for type 2 diabetes mellitus based on a physical-examination cohort. J Shandong Univ 2016. 54 69–72. ( 10.6040/j.issn.1671-7554.0.2016.074) [DOI] [Google Scholar]
- 33.Zhang H. Development of a Risk-Score Model for Type 2 Diabetes Mellitus In a Rural Population of Henan Province. In [Masters Thesis]. Zhengzhou, China: Zhengzhou University, 2016. [Google Scholar]
- 34.Zhu X. Association of Obesity with Prediabetes and Diabetes Risk and Construction of a Prediction Model. In [PhD Thesis]. Nanjing, China: Southeast University, 2023. [Google Scholar]
- 35.Chen Y, Ye L, Xiao S, et al. The risk predication of type 2 diabetes among rural residents with impaired glucose regulation. Chin J Prev Control Chron Dis 2018. 26 19–24. [Google Scholar]
- 36.Ma Y, Qianzi C, Zheng Q, et al. Common evaluation metrics for type 2 diabetes mellitus risk-prediction models. Chin J Prev Control Chronic Dis 2020. 28 94–100. ( 10.16386/j.cjpccd.issn.1004-6194.2020.02.004) [DOI] [Google Scholar]
- 37.Liu S. Incidence of Prediabetes and Diabetes Mellitus and Their Influencing Factors In a Cohort From 10 Provinces and Cities. In [Masters Thesis]. Beijing, China: Chinese Center for Disease Control and Prevention, 2020. [Google Scholar]
- 38.Lee HA, Park H & Hong YS. Validation of the Framingham diabetes risk model using community-based KoGES data. J Korean Med Sci 2024. 39 e47. ( 10.3346/jkms.2024.39.e47) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Alssema M, Vistisen D, Heymans MW, et al. The evaluation of screening and early detection strategies for type 2 diabetes and impaired glucose tolerance (DETECT-2) update of the finnish diabetes risk score for prediction of incident type 2 diabetes. Diabetologia 2011. 54 1004–1012. ( 10.1007/s00125-010-1990-7) [DOI] [PubMed] [Google Scholar]
- 40.Li Y, Wang J, Han X, et al. Serum alanine transaminase levels predict type 2 diabetes risk among a middle-aged and elderly Chinese population. Ann Hepatol 2019. 18 298–303. ( 10.1016/j.aohep.2017.02.001) [DOI] [PubMed] [Google Scholar]
- 41.Gao WG, Qiao Q, Pitkäniemi J, et al. Risk prediction models for the development of diabetes in Mauritian Indians. Diabetic Med 2009. 26 996–1002. ( 10.1111/j.1464-5491.2009.02810.x) [DOI] [PubMed] [Google Scholar]
- 42.Katsimpris A, Brahim A, Rathmann W, et al. Prediction of type 2 diabetes mellitus based on nutrition data. J Nutr Sci 2021. 10 e46. ( 10.1017/jns.2021.36) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 43.Liu X, Fine JP, Chen Z, et al. Prediction of the 20-year incidence of diabetes in older Chinese: application of the competing risk method in a longitudinal study. Medicine 2016. 95 e5057. ( 10.1097/md.0000000000005057) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 44.Hafezi SG, Saberi-Karimian M, Ghasemi M, et al. Prediction of the 10-year incidence of type 2 diabetes mellitus based on advanced anthropometric indices using machine learning methods in the Iranian population. Diabetes Res Clin Pr 2024. 214 111755. ( 10.1016/j.diabres.2024.111755) [DOI] [PubMed] [Google Scholar]
- 45.Hu H, Wang J, Han X, et al. Prediction of 5-year risk of diabetes mellitus in relatively low risk middle-aged and elderly adults. Acta Diabetol 2020. 57 63–70. ( 10.1007/s00592-019-01375-w) [DOI] [PubMed] [Google Scholar]
- 46.Xu L, Jiang CQ, Schooling CM, et al. Prediction of 4-year incident diabetes in older Chinese: recalibration of the Framingham diabetes score on Guangzhou Biobank Cohort Study. Prev Med 2014. 69 63–68. ( 10.1016/j.ypmed.2014.09.004) [DOI] [PubMed] [Google Scholar]
- 47.Rathmann W, Kowall B, Heier M, et al. Prediction models for incident type 2 diabetes mellitus in the older population: KORA S4/F4 cohort study. Diabetic Med 2010. 27 1116–1123. ( 10.1111/j.1464-5491.2010.03065.x) [DOI] [PubMed] [Google Scholar]
- 48.Liu Q, Zhang M, He Y, et al. Predicting the risk of incident type 2 diabetes mellitus in Chinese elderly using machine learning techniques. J Personalized Med 2022. 12 905. ( 10.3390/jpm12060905) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 49.Zhang L, Shang X, Sreedharan S, et al. Predicting the development of type 2 diabetes in a large Australian cohort using machine-learning techniques: longitudinal survey study. JMIR Med Inform 2020. 8 e16850. ( 10.2196/16850) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 50.Alghamdi M, Al-Mallah M, Keteyian S, et al. Predicting diabetes mellitus using SMOTE and ensemble machine learning approach: the henry ford exercise testing (FIT) project. PLoS One 2017. 12 e0179805. ( 10.1371/journal.pone.0179805) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 51.Wang K, Yang QF, Chen XL, et al. Metabolic syndrome and its components predict the risk of type 2 diabetes mellitus in the Mainland Chinese: a 3-year cohort study. Int J Endocrinol 2018. 2018 1–9. ( 10.1155/2018/9376179) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 52.Sun Q, Wang J, Jing Y, et al. Influences of resolvin D1 and D2 on the risk of type 2 diabetes mellitus: a Chinese community-based cohort study. Front Immunol 2023. 14 1143456. ( 10.3389/fimmu.2023.1143456) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 53.Stern MP, Williams K & Haffner SM. Identification of persons at high risk for type 2 diabetes mellitus: do we need the oral glucose tolerance test? Ann Intern Med 2002. 136 575–581. ( 10.7326/0003-4819-136-8-200204160-00006) [DOI] [PubMed] [Google Scholar]
- 54.Tran Quang B, Pham Tran P, Nguyen Thanh C, et al. High incidence of type 2 diabetes in a population with normal range body mass index and individual prediction nomogram in Vietnam. Diabetic Med 2022. 39 e14680. ( 10.1111/dme.14680) [DOI] [PubMed] [Google Scholar]
- 55.Sun L, Liang L, Gao X, et al. Early prediction of developing type 2 diabetes by plasma acylcarnitines: a population-based study. Diabetes Care 2016. 39 1563–1570. ( 10.2337/dc16-0232) [DOI] [PubMed] [Google Scholar]
- 56.Heianza Y, Arase Y, Hsieh SD, et al. Development of a new scoring system for predicting the 5 year incidence of type 2 diabetes in Japan: the Toranomon Hospital Health Management Center Study 6 (TOPICS 6). Diabetologia 2012. 55 3213–3223. ( 10.1007/s00125-012-2712-0) [DOI] [PubMed] [Google Scholar]
- 57.Oh TJ, Moon JH, Choi SH, et al. Development of a clinical risk score for incident diabetes: a 10-year prospective cohort study. J Diabetes Invest 2021. 12 610–618. ( 10.1111/jdi.13382) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 58.Nicolaisen SK, Thomsen RW, Lau CJ, et al. Development of a 5-year risk prediction model for type 2 diabetes in individuals with incident HbA1c-defined pre-diabetes in Denmark. BMJ Open Diabetes Res Care 2022. 10 e002946. ( 10.1136/bmjdrc-2022-002946) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 59.Hu H, Nakagawa T, Yamamoto S, et al. Development and validation of risk models to predict the 7-year risk of type 2 diabetes: the Japan Epidemiology Collaboration on Occupational Health Study. J Diabetes Invest 2018. 9 1052–1059. ( 10.1111/jdi.12809) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 60.Xu J, Goto A, Konishi M, et al. Development and validation of prediction models for the 5-year risk of type 2 diabetes in a Japanese population: japan public health center-based prospective (JPHC) diabetes study. J Epidemiol 2024. 34 170–179. ( 10.2188/jea.je20220329) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 61.Arellano-Campos O, Gómez-Velasco DV, Bello-Chavolla OY, et al. Development and validation of a predictive model for incident type 2 diabetes in middle-aged Mexican adults: the metabolic syndrome cohort. BMC Endocr Disord 2019. 19 41. ( 10.1186/s12902-019-0361-8) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 62.Shao X, Wang Y, Huang S, et al. Development and validation of a prediction model estimating the 10-year risk for type 2 diabetes in China. PLoS One 2020. 15 e0237936. ( 10.1371/journal.pone.0237936) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 63.Rhee SY, Sung JM, Kim S, et al. Development and validation of a deep learning based diabetes prediction system using a nationwide population-based cohort. Diabetes Metab J 2021. 45 515–525. ( 10.4093/dmj.2020.0081) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 64.Cai X, Ji L, Liu S, et al. Derivation and validation of a prediction model for predicting the 5-year incidence of type 2 diabetes in non-obese adults: a population-based cohort study. Diabetes Metab Syndr Obes 2021. 14 2087–2101. ( 10.2147/dmso.s304994) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 65.Olivera AR, Roesler V, Iochpe C, et al. Comparison of machine-learning algorithms to build a predictive model for detecting undiagnosed diabetes – ELSA-Brasil: accuracy study. Sao Paulo Med J 2017. 135 234–246. ( 10.1590/1516-3180.2016.0309010217) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 66.Liu H, Dong S, Yang H, et al. Comparing the accuracy of four machine learning models in predicting type 2 diabetes onset within the Chinese population: a retrospective study. J Int Med Res 2024. 52 03000605241253786. ( 10.1177/03000605241253786) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 67.Jagannathan R, Weber MB, Anjana RM, et al. Clinical utility of 30-min plasma glucose for prediction of type 2 diabetes among people with prediabetes: ancillary analysis of the diabetes community lifestyle improvement program. Diabetes Res Clin Pract 2020. 161 108075. ( 10.1016/j.diabres.2020.108075) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 68.Savolainen O, Fagerberg B, Vendelbo Lind M, et al. Biomarkers for predicting type 2 diabetes development – can metabolomics improve on existing biomarkers? PLoS One 2017. 12 e0177738. ( 10.1371/journal.pone.0177738) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 69.Sun F, Tao Q & Zhan S. An accurate risk score for estimation 5-year risk of type 2 diabetes based on a health screening population in Taiwan. Diabetes Res Clin Pr 2009. 85 228–234. ( 10.1016/j.diabres.2009.05.005) [DOI] [PubMed] [Google Scholar]
- 70.Lim NK, Park SH, Choi SJ, et al. A risk score for predicting the incidence of type 2 diabetes in a middle-aged Korean cohort: the Korean genome and epidemiology study. Circ J 2012. 76 1904–1910. ( 10.1253/circj.cj-11-1236) [DOI] [PubMed] [Google Scholar]
- 71.Wang H, Zheng X, Bai ZH, et al. A retrospective population study to develop a predictive model of prediabetes and incident type 2 diabetes mellitus from a hospital database in Japan between 2004 and 2015. Med Sci Monitor 2020. 26 e920880. ( 10.12659/msm.920880) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 72.Tan C, Li B, Xiao L, et al. A prediction model of the incidence of type 2 diabetes in individuals with abdominal obesity: insights from the general population. Diabetes Metab Syndr Obes 2022. 15 3555–3564. ( 10.2147/dmso.s386687) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 73.Yatsuya H, Li Y, Hirakawa Y, et al. A point system for predicting 10-year risk of developing type 2 diabetes mellitus in Japanese men: Aichi workers’ cohort Study. J Epidemiol 2018. 28 347–352. ( 10.2188/jea.je20170048) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 74.Wang Y, Zhang WS, Hao YT, et al. A Bayesian network model of new-onset diabetes in older Chinese: the Guangzhou biobank cohort study. Front Endocrinol 2022. 13 916851. ( 10.3389/fendo.2022.916851) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 75.Wilson PW, Meigs JB, Sullivan L, et al. Prediction of incident diabetes mellitus in middle-aged adults: the Framingham offspring study. Arch Intern Med 2007. 167 1068–1074. ( 10.1001/archinte.167.10.1068) [DOI] [PubMed] [Google Scholar]
- 76.Liu H, Hu D, Song L, et al. Methods for assessing the risk of incident diabetes mellitus in health-examination populations. Chin J Health Management 2012. 6 178–181. ( 10.3760/cma.j.issn.1674-0815.2012.03.010) [DOI] [Google Scholar]
- 77.Woo YC, Lee CH, Fong CHY, et al. Serum fibroblast growth factor 21 is a superior biomarker to other adipokines in predicting incident diabetes. Clinical Endocrinology 2017. 86 37–43. ( 10.1111/cen.13229) [DOI] [PubMed] [Google Scholar]
- 78.Jiang L, Xia Z, Zhu R, et al. Diabetes risk prediction model based on community follow-up data using machine learning. Prev Med Rep 2023. 35 102358. ( 10.1016/j.pmedr.2023.102358) [DOI] [PMC free article] [PubMed] [Google Scholar]
- 79.Asgari S, Khalili D, Zayeri F, et al. Dynamic prediction models improved the risk classification of type 2 diabetes compared with classical static models. J Clin Epidemiol 2021. 140 33–43. ( 10.1016/j.jclinepi.2021.08.026) [DOI] [PubMed] [Google Scholar]
- 80.Stiglic G, Wang F, Sheikh A, et al. Development and validation of the type 2 diabetes mellitus 10-year risk score prediction models from survey data. Prim Care Diabetes 2021. 15 699–705. ( 10.1016/j.pcd.2021.04.008) [DOI] [PubMed] [Google Scholar]
- 81.Lindström J & Tuomilehto J. The diabetes risk score: a practical tool to predict type 2 diabetes risk. Diabetes Care 2003. 26 725–731. ( 10.2337/diacare.26.3.725) [DOI] [PubMed] [Google Scholar]
- 82.Schmidt MI, Duncan BB, Bang H, et al. Identifying individuals at high risk for diabetes: the atherosclerosis risk in communities study. Diabetes Care 2005. 28 2013–2018. ( 10.2337/diacare.28.8.2013) [DOI] [PubMed] [Google Scholar]
- 83.Chen L, Magliano DJ, Balkau B, et al. AUSDRISK: an Australian type 2 diabetes risk assessment tool based on demographic, lifestyle and simple anthropometric measures. Med J Aust 2010. 192 197–202. ( 10.5694/j.1326-5377.2010.tb03507.x) [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.