Potentially modifiable predictors in models for undiagnosed and incident type 2 diabetes: A systematic review

Abstract

BACKGROUND

Type 2 diabetes mellitus (T2DM) prediction models are increasingly used to identify undiagnosed disease and estimate future T2DM risk, but the clinical relevance of potentially modifiable predictors remains unclear.

AIM

To systematically review diagnostic and prognostic T2DM prediction models, focusing on potentially modifiable predictors, model performance, validation, and methodological quality.

METHODS

PubMed, Embase, and Web of Science were searched for studies published from January 2016 to February 2026 that developed or substantially updated multivariable models for prevalent undiagnosed or incident T2DM and included at least 1 potentially modifiable predictor. Study characteristics, modeling methods, predictors, validation, and performance were extracted. Risk of bias and applicability were assessed using the Prediction Model Risk of Bias Assessment Tool. Findings were synthesized descriptively because of substantial heterogeneity.

RESULTS

Eighty-two studies were included. Most were conducted in China or Japan, and 47.6% used prospective cohort designs. Only 28 studies performed external validation. Overall risk of bias was high in 67 studies (81.7%), mainly because of analytical limitations, including data-driven predictor selection and insufficient calibration assessment. Common predictors included age, body mass index, glucose measures, sex, family history, and triglycerides. Potentially modifiable domains included adiposity, smoking, physical inactivity, diet, blood pressure, and glycemic and lipid biomarkers.

CONCLUSION

T2DM prediction models increasingly incorporate potentially modifiable predictors, but clinical translation is limited by high risk of bias, inadequate calibration, and insufficient external validation. More rigorous development and prospective evaluation are needed before routine implementation.

Key Words: Type 2 diabetes mellitus; Prediction models; Potentially modifiable predictors; Risk stratification; Systematic review

Core Tip: This systematic review of 82 studies developing type 2 diabetes mellitus prediction models (2016–2025) shows that while modifiable risk factors (e.g., obesity, smoking, and physical inactivity) are increasingly included, most models have a high risk of bias, use inappropriate variable selection methods, and lack external validation. Future models should prioritize rigorous methodology and clinically actionable predictors to improve early detection and prevention.

INTRODUCTION

Diabetes, particularly type 2 diabetes mellitus (T2DM), has emerged as a significant global public health concern. According to the International Diabetes Federation, the number of individuals affected by diabetes is predicted to reach 643 million by 2030, with an anticipated continued increase in the coming years[1]. Diabetes not only affects patients’ quality of life but also causes multiple complications and increases the healthcare burden. This trend will have a significant economic impact and pose a significant challenge to the health of patients and society as a whole, emphasizing the importance of early detection and intervention.

The objective of developing risk prediction models for diabetes mellitus is to assist healthcare professionals in the early identification of individuals with undiagnosed diabetes mellitus and those at high risk. These models provide a scientific foundation for early diagnosis and timely treatment, with the ultimate goal of reducing the incidence of diabetes[2]. At present, the methodologies employed in the construction of these risk prediction models predominantly encompass traditional statistical techniques in conjunction with artificial intelligence (AI) methodologies. While traditional methods typically rely on basic statistical models such as logistic regression and Cox regression, the field of diabetes risk modeling has undergone significant advancements over the past decade. The advent of big data and medical testing tools has enabled researchers to collect vast quantities of data, including genomic, clinical, and environmental data. These datasets can then be integrated with clinical, pharmacological, and socioeconomic information, facilitating the pursuit of precision medicine[3]. Furthermore, in recent years, the models developed have progressed beyond traditional statistical models to include an increasing number of AI-based machine-learning models. While traditional models have been instrumental in past research and clinical practice, the advent of data science has given rise to a new generation of AI-driven models that offer a promising avenue for more effective identification of high-risk individuals[4]. These models analyze complex, multidimensional data and integrate multiple risk factors and biomarkers associated with the onset of T2DM, thereby improving the accuracy and efficiency of risk prediction and providing individuals with more scientific and effective medical care that can reduce the likelihood of diabetes onset or improve prognosis[5].

Although existing T2DM prediction models have demonstrated varying degrees of predictive ability, it remains uncertain whether their development has been sufficiently aligned with the practical needs of clinicians and patients. Previous systematic reviews have primarily evaluated model development and reporting methods, predictor composition, discrimination and calibration, validation practices, implementation, geographic variation, or the application of machine-learning algorithms[2,6-10]. However, comparatively limited attention has been paid to distinguishing directly actionable predictors from anthropometric measures, treatable clinical conditions, and intermediate metabolic biomarkers or to evaluating their accessibility, potential actionability, and relevance to clinical implementation. This distinction is important because the inclusion of a variable in a prediction model indicates predictive value but does not necessarily establish a causal relationship or demonstrate that modifying the predictor will reduce the incidence of T2DM. Therefore, the present systematic review examined advances in T2DM prediction models over the past decade by summarizing their study characteristics, modeling approaches, validation strategies, predictive performance, and methodological quality, with a particular focus on the classification and clinical relevance of potentially modifiable predictors. By evaluating the accessibility, potential actionability, and predictive contribution of these variables, this review seeks to identify opportunities for developing more practical and clinically applicable risk-stratification tools and to inform future diabetes prevention strategies.

MATERIALS AND METHODS

This systematic review was conducted in accordance with the PRISMA 2020 guidelines[11]. The study protocol was registered in the International Prospective Register of Systematic Reviews (PROSPERO; registration number: CRD420261304786).

Search strategy

Studies developing or substantially updating multivariable models to detect prevalent undiagnosed T2DM or predict incident T2DM were identified. A comprehensive systematic search was performed in PubMed, Embase, and Web of Science for articles published from January 1, 2016 to February 1, 2026. The search strategy combined Medical Subject Headings or Emtree terms with free-text keywords related to T2DM, prediction models, and risk factors or predictors.

The search was restricted to English-language articles published between January 2016 and February 2026. Reference lists of relevant systematic reviews and included articles were manually screened to identify additional eligible studies. The complete database-specific search strategies, including the database platforms, controlled vocabulary and free-text terms, field tags, Boolean operators, filters, date limits, dates of execution, and numbers of records retrieved, are provided in the Supplementary material.

Inclusion and exclusion criteria

Diagnostic prediction models were defined as models intended to identify prevalent but previously undiagnosed T2DM, whereas prognostic prediction models estimated the future risk of incident T2DM among individuals without diabetes at baseline.

Studies were eligible for inclusion if they described the development or significant update of multivariable prediction models—including risk scores and machine-learning algorithms—intended either to identify prevalent undiagnosed T2DM or to estimate the future risk of incident T2DM among adults without diabetes at baseline. Eligible T2DM outcomes were defined according to the criteria used in the original studies, including fasting plasma glucose (FPG), glycated hemoglobin (HbA1c), 2-hour post-load glucose or oral glucose tolerance test, physician diagnosis, use of glucose-lowering medication, or validated registry or administrative definitions. A core inclusion requirement was that the model incorporated at least 1 potentially modifiable predictor, as operationally defined in the “Data synthesis and analysis” section.

Conversely, several exclusion criteria were applied to maintain the focus on modifiable risk prediction in general adult populations. Models that relied exclusively on non-modifiable factors, such as genetic markers (e.g., polygenic risk scores) or demographic data (age, sex, and ethnicity) without lifestyle or clinical components, were excluded. The review excluded studies focusing on pediatric populations (age < 18 years), pregnant women (gestational diabetes), or those specifically targeting type 1 diabetes or secondary forms of diabetes. Studies that focused solely on the external validation of pre-existing models were also excluded, although their reference lists were manually screened to identify original development papers. Additionally, articles were excluded if they lacked sufficient technical details (e.g., model coefficients or scoring systems); were published as reviews, meta-analyses, conference abstracts, or editorials; or were written in languages other than English.

Study selection and data extraction

The selection process was conducted in 3 stages. First, all identified records were imported into Rayyan, and duplicates were removed. Second, 2 independent reviewers screened the titles and abstracts of the remaining records based on the predefined inclusion and exclusion criteria. Any discrepancies were resolved through discussion or consultation with a third senior reviewer. Third, the full texts of potentially eligible studies were retrieved and assessed for final inclusion.

Data extraction was conducted independently by 2 reviewers using a standardized data collection form, with any discrepancies resolved through consensus or consultation with a third senior reviewer. For each eligible study, we systematically extracted information regarding the first author, year of publication, country of origin, study design, target population, and total sample size. Furthermore, details pertaining to model development and characteristics were documented, including the specific modeling methods employed, the key predictors incorporated into the final models, and the clinical format of the model. To assess the predictive performance and robustness of the developed tools, we recorded the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity alongside the methods utilized for internal or external validation.

Quality assessment

The risk of bias and applicability of the included studies were assessed using the Prediction Model Risk of Bias Assessment Tool (PROBAST). PROBAST evaluates risk of bias across 4 domains: Participants, predictors, outcome, and analysis. Each domain was judged as having a low, high, or unclear risk of bias according to the PROBAST signaling questions and guidance. Particular attention was given to participant selection, consistency and availability of predictor assessment, appropriateness of outcome definition, sample size and number of outcome events, handling of continuous predictors and missing data, predictor selection procedures, model overfitting, calibration assessment, and internal or external validation. An overall high risk of bias was assigned when at least 1 domain was judged to be at high risk; an overall low risk was assigned when all domains were judged to be at low risk. Applicability concerns were evaluated separately for the participants, predictors, and outcome domains and were classified as low, high, or unclear. Two researchers independently performed the assessments, and disagreements were resolved through discussion and consensus. Detailed study-level judgments and supporting reasons are provided in Supplementary Table 1.

Data synthesis and analysis

We conducted a descriptive analysis of the included studies and presented the findings in tabular form to synthesize the key characteristics of the risk prediction models. First, we summarized the study design, source population, geographic region, sample size, data source, prediction target, and outcome definition. Second, we extracted and summarized information on model development and validation, including modeling methods (e.g., Cox regression, logistic regression, and machine learning), candidate predictors, the number of outcome events and events per variable, handling of missing data, the final predictors retained in each model, validation methods, and model presentation formats (e.g., risk scores, nomograms, and web-based calculators). We also analyzed the predictors incorporated across the included models, with particular emphasis on potentially modifiable predictors. In this review, potentially modifiable predictors were operationally defined as behavioral, anthropometric, clinical, treatment-related, environmental, or biological variables whose measured values could plausibly change through lifestyle, clinical, pharmacological, or environmental interventions. This classification refers to the potential modifiability of the predictor and does not imply a causal relationship or demonstrate that modifying the predictor would necessarily reduce the incidence of T2DM. Given the substantial heterogeneity across studies in terms of prediction settings, modeling approaches, predictor selection, outcome definitions, follow-up durations, decision thresholds, validation methods, and performance metrics, we performed a structured narrative and descriptive synthesis rather than a quantitative meta-analysis. This approach was pre-specified before completion of the review.

RESULTS

In total, 82 studies were included in this systematic review. The included studies comprised both diagnostic models for prevalent undiagnosed T2DM and prognostic models for incident T2DM. The detailed literature screening process is presented in Figure 1. The database searches identified 8636 records. After the removal of 4073 duplicates, 4563 records were screened by title and abstract, of which 4391 were excluded. The full texts of 172 reports were assessed for eligibility, and 90 were excluded for the reasons shown in Figure 1. Ultimately, 82 studies were included in the systematic review.

Figure 1  Flow diagram of selected studies.

Model-related statistics

Study characteristics:Table 1 summarizes the basic characteristics of the included studies, while Table 2 details the specific parameters of the models utilized for predicting T2DM risk.

Table 1 Summary of study design, modeling techniques, and validation strategies in type 2 diabetes mellitus models.

Characteristic	Category		n	Percentage (%)
Overall	Studies included		82
Country/region	China		30	36.6
	Japan		12	14.6
	South Korea		4	4.9
	Iran		3	3.7
	Mexico		3	3.7
	United States		3	3.7
	Germany		2	2.4
	Qatar		2	2.4
	Singapore		2	2.4
	United Kingdom		2	2.4
	United Arab Emirates		2	2.4
	Other1		17	20.7
Study design	Prospective cohort		39	47.6
	Cross-sectional		20	24.4
	Retrospective cohort		14	17.1
	Mathematical modeling		2	2.4
	Case-control		2	2.4
	Population-based cohort		2	2.4
	Nested case-control		1	1.2
	Multi-cohort prospective		1	1.2
	Meta-analysis + prospective cohort validation		1	1.2
Modeling method	Traditional regression	Logistic regression	42	-2
		Cox proportional hazards model	32	-
	Regularized regression	LASSO/elastic net	17	-
	Machine learning	Support vector machine	5	-
		Decision tree	4	-
		Neural network	2	-
		Random forest	1	-
	Competing risk model	Fine–Gray model	2	-
	Other methods	Meta-analysis–based model	1	-
Model validation	Internal validation		67	-
		Bootstrap	21	-
		Cross-validation	13	-
		Split-sample	32	-
	External validation		28	-
	Both internal and external validation		15	-
	No validation		0	-

¹Other includes Denmark, Finland, Indonesia, Jordan, Malaysia, the Netherlands, Nigeria, Oman, Peru, Portugal, Russia, Saudi Arabia, Spain, Switzerland, Taiwan (China), 12 European countries, and the United Kingdom/Germany.

²Modeling methods and validation categories were not mutually exclusive because some studies applied more than one modeling or validation approach.

LASSO: Least absolute shrinkage and selection operator; SVM: Support vector machine.

Table 2 Methodological characteristics and predictive performance of the included type 2 diabetes mellitus prediction models.

Ref.	Country/region	Study design	Sample size	Modeling method	Key predictors	AUC (95%CI)	Sensitivity (95%CI)	Specificity (95%CI)	Validation	Model format	Target population
Bernabe-Ortiz et al[17], 2016	Peru	Derivation: Cross-sectional national survey (ENINBSC); validation: Prospective cohort (CRONICAS)	Derivation: 2472 (48 events); validation: 2945 (85 events at baseline; 121 incident cases)	Logistic regression with stepwise backward elimination	Age, family history of diabetes, WC	Derivation: 0.73 (0.65–0.78)	Derivation: 69.6%	Derivation: 65.8%	Internal validation: Bootstrap; external validation: CRONICAS cohort	Points-based scoring system	Peruvian adults aged ≥ 35 years
						Validation (cross-sectional): 0.68 (0.62–0.73)	Validation: 70.2%	Validation: 58.9%
						Validation (incident): 0.66 (0.61–0.71)
Liu et al[14], 2016	China	Prospective cohort (Beijing Longitudinal Study of Ageing)	1857 (144 events over 20 years)	Subdistribution hazards model (Fine–Gray) accounting for competing risk of non-diabetes death	Age, BMI, FPG, self-rated health, physical activity	0.76 (0.72–0.80)	0.74	0.65	Internal validation by bootstrapping	Points-based scoring system	Community-dwelling Chinese adults aged ≥ 55 years
Zhang et al[60], 2016	China	Prospective cohort (rural Henan province)	Derivation: 11564 (659 events); validation: 1285 (70 events)	Cox proportional hazards with forward selection	Age, BMI, TG, FPG	Derivation: 0.768 (0.760–0.776)	Derivation: 66.7%	Derivation: 74.0%	Internal validation: Random split (90/10); external validation: Compared with 5 existing models in validation set	Risk score formula	Rural Chinese adults aged ≥ 18 years
						Validation: 0.766 (0.742–0.789)	Validation: 70.0%	Validation: 72.5%
Han et al[49], 2016	China	Derivation: Cross-sectional (HDNNCDS); validation: Prospective cohort (HPHS)	Derivation: 8764 (1096 prevalent cases); validation: 3430 (394 incident cases over 4.2 years)	Multivariable logistic regression with backward selection	Classic model: Age, sex, hypertension, family history of diabetes, alcohol consumption, regular exercise, BMI, WC	Derivation (classic): 0.78 (0.77–0.80)			External validation in independent cohort (HPHS)	Points-based scoring system	Northern Chinese adults aged 20–74 years
						Validation (classic): 0.76 (0.73–0.78)
					Diet model: Above + staple foods, livestock, eggs, potato, fresh fruit, fresh vegetables, dairy products	Derivation (diet): 0.81 (0.79–0.82)	Derivation (diet): 57.6%	Derivation (diet): 84.5%
						Validation (diet): 0.79 (0.77–0.81)	Validation (diet): 49.0%	Validation (diet): 85.8%
Miyakoshi et al[20], 2016	Japan	Retrospective cohort (health examinees)	Total: 4159 (279 events); derivation: 2080 (138 events); validation: 2079 (141 events)	Multivariate Cox proportional hazards with backward stepwise selection	RSFull: Age, family history of diabetes, SBP, sex, smoking, FPG, 2hPG, HbA1c	RSFull: 0.80 (0.76–0.84)	NR	NR	Internal validation (random split 50/50)	Points-based scoring system	Japanese adults undergoing health checkups (baseline non-diabetic)
					RS-2hPG: Age, family history of diabetes, SBP, sex, FPG, HbA1c, TG	RS-2hPG: 0.75 (0.70–0.78)
					RSNI: Age, family history of diabetes, SBP, sex, BMI	RSNI: 0.68 (0.63–0.72)
Zhang et al[55], 2017	China	Prospective cohort (rural Henan, 6-year follow-up)	Total: 15768 (702 events); derivation: 12654; validation: 3114	Cox proportional hazards with forward selection; scores = β × 10 (rounded down)	Drinking tea, BMI, WHtR, TG, FPG	Derivation: 0.791 (0.783–0.799)	69.6% (derivation)	75.6% (derivation)	Internal validation: Random split (80/20)	Points-based scoring system	Rural Chinese adults aged ≥ 18 years
						Validation: 0.731 (0.713–0.748)
Asadollahi et al[18], 2017	Iran	Cross-sectional (populationbased survey)	Derivation: 2158 (254 diabetes, 168 IFG); validation: 404	Multivariate logistic regression with univariate prescreening	Family history of diabetes, place of life, age, hypertension, daily exercise, marital status, sex, WC, smoking, BMI	Derivation: 0.71	Range: 12.9%–30.7% (for diabetes at different score cut-offs)	Range: 72.8%–88% (for diabetes at different score cut-offs)	External validation in independent clinical sample (n = 404)	Points-based scoring system	Iranian adults aged ≥ 25 years
						Validation: 0.872
Hippisley-Cox and Coupland[32], 2017	United Kingdom (England)	Prospective open cohort (QResearch database)	Derivation: 8186705 (178314 events); validation: 2629940 (62326 events)	Cox proportional hazards with fractional polynomials; multiple imputation for missing data	Model A: Age, ethnicity, deprivation, BMI, smoking, family history of diabetes, CVD, treated hypertension, corticosteroids, atypical antipsychotics, statins, schizophrenia/bipolar disorder, learning disability, gestational diabetes, PCOS	Model A: Women: Harrell’s C 0.834 (0.832–0.837); men: 0.814 (0.812–0.816)	Model A (threshold 5.6%): 68.1% (top 20%)	Model A: 81.1% (top 20%)	External validation in separate set of 363 practices (not used in derivation)	Cox regression equations; web calculator (qdiabetes.org)	United Kingdom adults aged 25–84 years without diabetes
					Model B: Model A + FPG	Model B: Women: 0.889 (0.887–0.891); men: 0.866 (0.863–0.868)
					Model C: Model A + HbA1c	Model C: Women: 0.878 (0.875–0.881); men: 0.855 (0.849–0.861)
Chen et al[50], 2017	China	Prospective cohort (Rural Deqing Cohort Study)	Whole cohort: 28251 (387 events over mean 4.2 years); sub-cohort: 3043 (191 events over mean 8.7 years)	Cox proportional hazards; scores based on β-coefficients	Non-invasive model: Age, BMI, family history of diabetes, meat diet, hypertension	Non-invasive: Whole cohort: 0.705; subcohort: 0.638	Noninvasive: 70.4%	Noninvasive: 60.4%	Internal validation: Bootstrap; external validation: Subcohort validation	Points-based scoring system	Rural Chinese adults aged ≥ 18 years
					Plus-FPG model: Age, BMI, family history of diabetes, meat diet, hypertension, IFG	Plus-FPG: Whole cohort: 0.754; sub-cohort: 0.667	Plus-FPG: 63.0%	Plus-FPG: 73.1%
Tian et al[72], 2017	China	Cross-sectional (oil field workers)	1995 (776 dysglycemia)	Multivariable logistic regression	Sex, age, BMI, HHBG, smoking, TG, FPG	0.791 (0.773–0.809)	0.736	0.735	Internal validation; compared with m-FINDRISC)	Pointsbased scoring system	Chinese oil field workers aged 20–64 years
Yatsuya et al[15], 2018	Japan	Prospective cohort (Aichi Workers’ Cohort Study)	3540 (342 events over median 12.2 years)	Cox proportional hazards with backward selection; Framingham point system method	Age, BMI, smoking status, family history of diabetes, TG, FPG	C-index: 0.77	0.66	0.76	Internal validation; cross-validated with Ibaraki model	Points-based scoring system	Japanese male workers aged 35–64 years
Abdullah et al[26], 2017	Malaysia	Casecontrol (nested within Malaysian Cohort Project)	Total: 4077 (cases: 1962; controls: 2115); Malay: 1323, Chinese: 1344, Indian: 1410	Multivariable logistic regression; GRS based on 62 T2DM-associated SNPs	Environmental model: Sex, age, WHR, physical inactivity	Environmental only: Malay: 0.75 (0.72–0.78); Chinese: 0.83 (0.80–0.86); Indian: 0.78 (0.75–0.81); Combined: 0.79 (0.77–0.81)	NR	NR	Internal validation; compared model fit with/without GRS)	GRS combined with environmental risk factors in logistic regression model	Multiethnic Malaysian adults (Malay, Chinese, Indian) aged 35–70 years
					Full model: Environmental + GRS	Environmental + GRS: Malay: 0.77 (0.73–0.80); Chinese: 0.84 (0.82–0.87); Indian: 0.79 (0.76–0.82); Combined: 0.80 (0.78–0.82)
Wang et al[65], 2018	China	Prospective cohort (health checkup attendees)	9735 (1099 events over 3 years)	Cox regression (HR); logistic regression for prediction models (full, stepwise, MFP)	Male: Age, hypertension, 2hPG, FPG, TG	Male (stepwise model): 0.779 (0.759–0.799)	Male: 74%	Male: 67%	Internal validation using 500 bootstrap resamples	Nomogram; points-based scoring system (separate for male and female)	Chinese adults (aged 17–96 years) attending annual health check-ups in Wuhan, China
					Female: Age, FPG, 2hPG, TG	Female (stepwise model): 0.860 (0.836–0.883)	Female: 88%	Female: 69%
Han et al[37], 2018	China	Prospective cohort (Dongfeng-Tongji cohort)	17690 (1390 events over 5 years)	Cox proportional hazards	BMI, FPG, hypertension, hyperlipidemia, current smoking, family history of diabetes	0.751 (0.737–0.764)	0.656	0.729	Internal validation using 90/10 random split (repeated 10 times)	Points-based scoring system	Middle-aged and older Chinese adults (mean age 63.3 years, retired workers)
Sulaiman et al[19], 2018	United Arab Emirates	Cross-sectional (United Arab Emirates National Diabetes and Lifestyle Study)	872 (219 undiagnosed diabetes)	Stepwise forward logistic regression	Age, family history of diabetes, hypertension, BMI, WHR	0.82 (0.78–0.86)	0.754	0.7	External validation in a separate sample of Arab non-UAE citizens residing in the UAE (n = 631)	Pointsbased scoring system	United Arab Emirates citizens aged ≥ 18 years (without known diabetes)
Hu et al[83], 2018	Japan	Prospective cohort (J-ECOH Study)	Total: 46198 (3385 events over 7 years); derivation: 30500 (2216 events); validation: 15698 (1169 events)	Cox proportional hazards with backward selection	Noninvasive: Sex, age, BMI, abdominal obesity, smoking, hypertension	Noninvasive: 0.73 (0.72–0.74)	NR	NR	Internal validation (random split 2/3-1/3)	Cox regression equations; risk calculators and charts	Japanese employees aged 30–59 years
					Invasive: Age, BMI, smoking, hypertension, dyslipidemia, FPG, HbA1c	Invasive: 0.89 (0.89–0.90)
Félix-Martínez and Godínez-Fernández[84], 2018	Mexico	Cross-sectional (National Health and Nutrition Surveys)	Model 1 (2006): 8743 (258 UDM); model 2 (2012): 5103 (63 UDM)	Multivariable logistic regression with backward stepwise selection (AIC-based, 10-fold CV)	Model 1: Age, WC, SBP	Model 1: 0.73 (validation)	Model 1: 74%	Model 1: 62%	external validation in NHS 2000 (n = 28913)	Logistic regression equation; interactive web applications	Mexican adults aged ≥ 20 years
					Model 2: Age, WC, height, family history of diabetes	Model 2: 0.71 (validation)	Model 2: 76%	Model 2: 55%
Kim et al[77], 2019	Korea	Prospective cohort (KoGES)	Clinical model: 7912 (1345 events); integrated model: 6818 (1162 events)	Cox proportional hazards with backward elimination (clinical) + forward selection (SNPs)	Clinical model: Age, sex, FBS, ALT, γ-GTP, HDL, TG, SBP, BMI, parental history of DM	Clinical model (validated C-index): 5-year 0.744, 7-year 0.732, 10-year 0.732	NR	NR	Internal validation (bootstrap 200 iterations)	Cox regression equations; points-based scoring system	Korean adults aged 40–69 years (urban/rural)
					Integrated model: Above + 5-SNP point score	Integrated model (validated C-index): 5-year 0.747, 7-year 0.736, 10-year 0.738
Zhang et al[51], 2019	China	Crosssectional (community-based)	1432 (142 newly diagnosed diabetes)	Multivariable logistic regression	Age, family history of diabetes, hypertension, WC, BMI, smoking, vegetable intake, fruit intake	0.758 (0.714–0.803)	0.831	0.6	Internal validation	Pointsbased scoring system	Chinese communitydwelling adults aged 18–60 years without known diabetes
Wang et al[38], 2019	China	Prospective cohort (3-year followup)	Training: 5557 (595 events); validation: 1870 (206 events)	Multivariable logistic regression with stepwise selection	Female: Age, BMI, FPG, TG	Training: Female 0.863 (0.837–0.888); male 0.751 (0.729–0.774)	Training: Female 82.6%; male 73.8%	Training: Female 79.1%; male 65.7%	Internal validation (500 bootstrap); temporal validation using later cohort	Nomogram	Central Chinese adults attending annual health check-ups
					Male: Age, BMI, FPG, HDL-C, TG	Validation: Female 0.847 (0.801–0.892); male 0.755 (0.717–0.794)	Validation: Female 90.2%; male 63.3%	Validation: Female 66.7%; male 76.0%
Kraege et al[41], 2019	Switzerland	Prospective cohort (CoLaus/PsyCoLaus)	5277 (405 events over mean 10.9 years)	Logistic regression with bivariate prescreening	Age, sex, WC, family history of diabetes, hypertension, physical inactivity	0.772 (0.752–0.792)	0.605	0.771	External validation in 3 cohorts: EPIC-Europe (n = 26270), Tlalpan 2020 (Mexico, n = 1076), Shahedieh (Iran, n = 904)	Points-based scoring system	European adults aged 35–75 years, generalizable to populations with different diabetes prevalence
Xie et al[80], 2019	United States	Cross-sectional (BRFSS 2014)	138146 (20467 T2DM)	Multiple machine-learning methods: SVM, decision tree, logistic regression, random forest, neural network, Naïve Bayes	Age, BMI, income, marital status, race/ethnicity, sleep duration, check-up frequency, physical activity, smoking, etc. (27 variables)	Neural network: 0.795	Decision tree: 51.6% (highest)	Neural network: 90.2% (highest)	Internal validation (random split 2/3 training, 1/3 test)	Machine learning models	United States adults aged ≥ 30 years (Behavioral Risk Factor Surveillance System)
						Logistic regression: 0.793
						Linear SVM: 0.781
						Rbf SVM: 0.779
						Random forest: 0.761
						Naïve Bayes: 0.760
						Polynomial SVM: 0.759
						Decision tree: 0.718
Pei et al[82], 2019	China	Cross-sectional (retrospective, health checkup data)	10436 (1570 with diabetes)	Decision tree (J48 algorithm, C4.5)	Age, BMI, family history of diabetes, work-related stress, physical activity, salty food preference, hypertension, sex, history of CVD/stroke	0.948	0.942	94.2% (accuracy)	Internal validation (random split 70/30)	Decision tree with 19 IF-THEN rules	Chinese adults undergoing annual health check-ups in Shenyang, China
Wang et al[29], 2020	Singapore	Nested case-control within prospective cohort (Singapore Chinese Health Study)	485 cases, 485 controls (matched on age, sex, dialect, date of blood collection)	Weighted biomarker score based on logistic regression β coefficients; 10-fold cross-validation	TG/HDL ratio, ALT, ferritin, adiponectin (all in quartiles)	Base model + biomarker score: 0.83 (0.81–0.86); increment from base model (AUC 0.81) significant (P = 0.002)	NR	NR	Internal validation (10-fold cross-validation)	Weighted biomarker score	Singapore Chinese adults aged 45–74 years
Wu et al[76], 2019	China	Crosssectional (Shanghai Nicheng Cohort)	15309 (2344 prevalent undiagnosed diabetes)	Multivariable logistic regression with backward stepwise selection	Age group, family history of diabetes, hypertension, overweight/obesity (BMI ≥ 25 kg/m2), central obesity (WC ≥ 90/85 cm)	Exploratory: 0.654 (0.637–0.670)	0.655	0.587	Internal validation (random split 50/50)	Pointsbased scoring system	Middleaged and elderly Chinese (45–70 years) in suburban Shanghai
						Validation: 0.669 (0.653–0.686)
Arellano-Campos et al[67], 2019	Mexico	Prospective cohort (metabolic syndrome cohort)	6144 (331 events over median 2.4 years)	Cox proportional hazards; points based on β-coefficient categories	Office-based model: Age, family history of diabetes, WHtR, hypertension, BMI	Office-based model: C-statistic 0.656 (after cross-validation)	NR	NR	Internal validation (k-fold cross-validation, bootstrap)	Points-based scoring system	Middle-aged Mexican adults (≥ 20 years, urban settings)
					Biochemical model: Age, TG, FPG, hypertension, abdominal obesity	Biochemical model: C-statistic 0.741 (after cross-validation)
Cai et al[23], 2020	China	Retrospective cohort (Rich Healthcare database)	Total: 1688 (103 events); derivation: 1273 (81 events); validation: 415 (22 events)	LASSO regression for feature selection; multivariate Cox regression	Age, BMI, FPG, TC	Derivation: AUC 0.878 (0.861–0.895), C-index 0.862 (0.830–0.894)	NR	NR	Internal validation (random split 75/25); 1000 bootstrap resampling	Nomogram (continuous variables; no simplified pointsbased score)	Chinese adults with hypertension (aged ≥ 20 years) without baseline diabetes
						Validation: AUC 0.855 (0.836–0.874), C-index 0.841 (0.817–0.865)
Bahijri et al[85], 2020	Saudi Arabia	Cross-sectional (communitybased, stratified 2-stage cluster sampling)	1403 (328 with dysglycemia: 69 diabetes, 259 prediabetes)	Multivariable logistic regression; points assigned based on β-coefficient categories	Sex, age, WC, history of hyperglycemia, family history of diabetes	0.76 (0.73–0.79)	Approximately 70% (at cutoff 5 or 6)	Approximately 70% (at cutoff 5 or 6)	Internal validation; model performance assessed on development sample	Pointsbased scoring system (range 0–15)	Saudi adults aged ≥ 20 years without known diabetes
Hu et al[86], 2020	China	Prospective cohort (Dongfeng-Tongji cohort)	4833 (171 events over 4.6 years)	Cox proportional hazards; restricted cubic splines; Youden index for optimal cut-offs	Age, sex, smoking, physical activity, BMI, TG, FPG	0.75 (0.70–0.81)	44.9% (BMI ≥ 24 kg/m2)	75.4% (BMI ≥ 24 kg/m2)	Internal validation	Cox regression model (no simplified pointsbased score)	Chinese adults aged ≥ 50 years without major underlying diseases (relatively low-risk population)
							46.3% (TG ≥ 1.1 mmol/L)	67.6% (TG ≥ 1.1 mmol/L)
							52.9% (FPG ≥ 5.89 mmol/L)	81.7% (FPG ≥ 5.89 mmol/L)
Liu et al[58], 2020	China	Prospective cohort (Beijing Health Management Cohort)	43404 (2623 events over median 6.8 years)	Logistic regression with forward selection	Age, education, parental history of diabetes, WHR, BMI, sleep duration, FPG	Comprehensive nomogram: C-index 0.74 (0.71–0.76) in test set	NR	NR	Internal validation (bootstrap); random split (80/20)	Nomogram (points-based scoring system; risk groups defined)	Chinese adults aged 18–80 years from urban Beijing
						Concise nomogram: C-index 0.73 (0.71–0.75)
Wilkinson et al[61], 2020	United States	Prospective cohort (REGARDS, ARIC)	Development: 12043 (1602 events) over approximately 10 years; validation: 9710 (927 events)	Bayesian logistic regression with weakly informative Cauchy priors	Age, sex, race, BMI, TG, HDL, SBP, DBP, blood glucose	Development (REGARDS): 0.79 (0.78–0.80)	NR	NR	External validation in ARIC (independent cohort)	Bayesian logistic regression equation; probability calculation formula provided	United States Black and White adults aged ≥ 45 years (REGARDS) and 45–64 years (ARIC)
						External validation (ARIC): 0.85 (0.83–0.86)
Zhang et al[78], 2020	China	Cross-sectional (Henan Rural Cohort)	36652 (3356 T2DM)	Multiple machine-learning methods: LR, CART, ANN, SVM, RF, GBM; SMOTE for class imbalance	60 variables including sociodemographic, lifestyle, anthropometric, laboratory, and medical history variables. Top 10: Sweet flavor, urine glucose, age, heart rate, creatinine, WC, uric acid, pulse pressure, insulin, hypertension	GBM (with laboratory data): 0.872	GBM: NR	GBM: 81.7%	Internal validation (10-fold cross-validation, repeated 100 times)	Machine learning models	Rural Chinese adults aged 18–79 years
						RF: 0.868	RF: NR
						ANN: 0.858	ANN: Highest sensitivity among all models
						LR: 0.841
						SVM: 0.835
						CART: 0.811
						Non-laboratory models ranged from 0.767 to 0.817
Shao et al[36], 2020	China	Prospective cohort (China Health and Nutrition Survey, 2004-2015)	Total: 6023 (349 events over 10 years); derivation: 4498 (257 events); validation: 1525 (92 events)	Multivariable logistic regression; LASSO for feature selection; nomogram	Model A: Age, sex, ethnicity, hypertension, smoking, alcohol, WC, BMI	Model A: 0.788 (0.761–0.816) derivation; 0.818 (0.775–0.861) validation	Model A: 74.7%	Model A: 71.2%	Internal validation: Bootstrap; external validation: Validation in independent provinces	Nomogram (4 separate models with increasing complexity)	Chinese adults aged 20–80 years
					Model B: Model A + education, soft drinks, tea, physical activity, calories, carbohydrates, fat, protein, triceps skinfold, sleep hours	Model B: 0.807 (0.780–0.834) derivation; 0.823 (0.780–0.865) validation	Model B: 77.5%	Model B: 70.8%
					Model C: Model B + LDL, HDL, TC, TG, insulin, FPG, HbA1c	Model C: 0.905 (0.879–0.932) derivation; 0.915 (0.877–0.953) validation	Model C: 84.1%	Model C: 83.0%
					Model D: LDL, HDL, TC, TG, insulin, FPG, HbA1c	Model D: 0.882 (0.853–0.912) derivation; 0.862 (0.813–0.912) validation	Model D: 73.1%	Model D: 92.3%
Wang et al[87], 2020	Japan	Retrospective cohort (DRYAD database, Murakami Memorial Hospital, 2004–2015)	Total: 11113 (373 events); training: 8296 (275 events); validation: 2817 (98 events)	Multivariable Cox regression; nomogram	Age, WC, smoking, fatty liver, FPG, HbA1c	Training: 0.87 (0.85–0.89)	Training: 82%	Training: 79%	Internal validation (random split 75/25)	Nomogram (points-based scoring system)	Japanese adults attending annual health checkups (healthy at baseline)
						Validation: 0.87 (0.86–0.90)	Validation: 87%	Validation: 73%
Yang et al[88], 2020	United States	Crosssectional (NHANES 2011–2016)	Total: 8057 (1336 T2DM); training: 4246 (664 T2DM); test: 1060 (165 T2DM); validation: 2751 (507 T2DM)	Multiple machine-learning methods: LDA, RF, SVM; easy ensemble methods with 100 base models	Age, sex, WC, sagittal abdominal diameter, relative leg length, 60-second pulse, smoking, alcohol, hypertension, family history, general health condition, weight control or loss	LDA ensemble: 0.849 (validation)	LDA ensemble: 81.9%	LDA ensemble: 70.9%	Internal validation (random split 80/20); external validation using NHANES 2015–2016	Machine-learning models	United States adults aged 18–80 years
						RF ensemble: 0.836 (validation)	RF ensemble: 81.3%	RF ensemble: 71.3%
						SVM ensemble: 0.848 (validation)	SVM ensemble: 82.4%	SVM ensemble: 71.4%
Del Razo-Olvera et al[33], 2020	Mexico	Prospective cohort (derivation: Metabolic syndrome cohort; validation: INPer-IMSS cohort)	Derivation: 3161 (147 events over 2.4 years); Validation: 925 (97 events over 2.8 years)	Cox proportional hazards; stepwise selection; points based on β-coefficients (Framingham method)	History of gestational diabetes, BMI, hypertriglyceridemia, FPG	Derivation: 0.75 (0.73–0.80)	73% (at cut-off ≥ 15)	67% (at cut-off ≥ 15)	External validation in independent cohort (INPer-IMSS)	Points-based scoring system	Mexican women of reproductive age (20–49 years)
						Validation: 0.91 (0.87–0.95)
Katsimpris et al[52], 2021	Germany	Crosssectional (KORA FF4 study)	1591 (139 T2DM)	Elastic net-penalized regression with bootstrap (5000 samples); logistic regression	Potatoes, mushrooms, onions and garlic, whipping cream, cookies, soft drinks, coffee, tomato sauce, mayonnaise, xylitol, galactose	0.79 (0.76–0.83)	10%–90% (range across thresholds)	53%–98% (range across thresholds)	Internal validation (5-fold crossvalidation; bootstrap)	Logistic regression equation (coefficients provided)	German adults aged 39–88 years
Rhee et al[34], 2021	Korea	Prospective cohort (NHIS-HEALS)	Total: 335302 (23420 events over 10.4 years); training: 268241; validation: 67061	Deep learning (RNNLSTM); Cox longitudinal summary model for comparison	Age, sex, BMI, SBP, DBP, FPG, TC, hemoglobin, AST, ALT, GGT, proteinuria, smoking, alcohol, exercise, personal history (hypertension, heart disease, stroke, others), family history (hypertension, heart disease, stroke, DM, others)	DL model 5-year: 0.877 (0.869–0.885)	DL 5-year: 81.6%	DL 5-year: 76.5%	Internal validation (random split 80/20)	Deep learning (RNN-LSTM) algorithm	Korean adults aged ≥ 40 years
						DL model 10-year: 0.827 (0.821–0.833)	DL 10-year: 75.1%	DL 10-year: 74.0%
						Cox model 5-year: 0.842 (0.832–0.852)
						Cox model 10-year: 0.807 (0.801–0.813)
Awad et al[24], 2021	Qatar	Mathematical modeling (simulated population based on epidemiological model)	5000 simulated individuals per time point (2020, 2030, 2050)	Logistic regression on simulated samples	Age, sex, obesity, smoking, physical inactivity	2020: 0.79 (0.77–0.80)	2020: 79.0%	2020: 66.8%	External validation using 2012 Qatar STEPwise survey data	Pointsbased scoring system	Qatari adults aged 15–79 years
						2030: 0.78 (0.76–0.79)	2030: 77.5%	2030: 65.8%
						2050: 0.76 (0.75–0.78)	2050: 74.4%	2050: 64.5%
Cai et al[63], 2021	Japan	Retrospective cohort (NAGALA database)	Total: 12940 (201 events); derivation: 9651 (154 events); validation: 3289 (47 events)	LASSO regression for feature selection; multivariate Cox regression	Age, fatty liver, GGT, TG, HbA1c, FPG	Derivation: C-index 0.906 (0.878–0.934), AUC 0.916 (0.889–0.943)	NR	NR	Internal validation (random split 75/25); 1000 bootstrap resampling	Nomogram (continuous variables)	Japanese nonobese adults (BMI < 25 kg/m2) attending health checkups
						Validation: C-index 0.837 (0.760–0.914), AUC 0.829 (0.753–0.905)
Wang et al[89], 2021	China	Crosssectional (national health examination)	Total: 458153 (38335 T2DM); development: 366523 (30758 T2DM); validation: 91630 (7577 T2DM)	LASSO regression for feature selection; multivariable logistic regression	Age, family history of T2D, WC, TC, TG, BMI, HDL-C, hypertension	Male: Dev 0.864 (0.861–0.866), Val 0.865 (0.859–0.871)	Male: 80.1%	Male: 77.7%	Internal validation (random split 80/20)	Nomogram (sex-specific)	Adults in Western Xinjiang, China (aged ≥ 19 years)
						Female: Dev 0.816 (0.813–0.819), Val 0.815 (0.808–0.821)	Female: 77.9%	Female: 70.7%
Cai et al[21], 2021	Japan	Retrospective cohort (NAGALA database)	Total: 2741 (223 events over 8 years); training: 2058 (157 events); validation: 683 (66 events)	LASSO regression for feature selection; multivariate Cox regression	Smoking status, WC, FPG, HbA1c	Training: C-index 0.839 (0.804–0.874), AUC 0.811	Training: 77.5%	Training: 72.4%	Internal validation (random split 75/25); 1000 bootstrap resampling	Nomogram (continuous variables)	Japanese adults with NAFLD attending health check-ups
						Validation: C-index 0.822 (0.777–0.868), AUC 0.805	Validation: 69.9%	Validation: 75.6%
de Sousa Tavares and Jorge[90], 2021	Portugal (Azores)	Cross-sectional (retrospective, primary care data)	6834 (4661 with T2DM)	Binary logistic regression with interaction terms; backward selection	CVD history, hypertension history, sex, age, BMI, TG, glucose	0.863 (0.854–0.872)	NR	NR	Internal validation (bootstrapping)	Logistic regression equation with interaction terms	Azorean adults aged ≥ 18 years
Mustafina et al[39], 2021	Russia	Prospective cohort (HAPIEE)	7739 (915 events over 14 years)	Cox proportional hazards	Full model: Age, BMI, WC, FPG, TG, HDL-C, BP, family history of T2DM	Full model: NR	Full model: Male 76.0%, female 71.7%	Full model: Male 71.5%, female 69.2%	Internal validation; performance assessed on development sample	Points-based scoring system	Russian adults aged 45–69 years
					Simplified model: WC, hypertension, family history of T2DM	Simplified model: 0.71	Simplified model: 74.7%	Simplified model: 60.0%
Deberneh and Kim[91], 2021	Korea	Retrospective cohort (EHRs)	253395 subjects; after preprocessing: 17131 per class for training, 200 per class for testing	Machine learning (LR, RF, SVM, XGBoost, CIM, stacking, soft voting) with ANOVA, χ2, RFE feature selection	FPG, HbA1c, TG, BMI, γ-GGT, age, uric acid, sex, smoking, drinking, physical activity, family history	Test: 0.71–0.73 (accuracy-based, no AUC reported)	Test: 71%–74% (accuracy)	Test: 71%–74% (accuracy)	Internal validation: 10-fold cross-validation	Machine-learning models	Korean adults attending health check-ups
Oh et al[40], 2021	Korea	Prospective cohort (KoGES)	Derivation: 3973 (NR); validation: 1700 (NR)	Multivariable logistic regression; points based on β coefficients	KDR: Age, living location, WC, hypertension, family history of T2D, smoking	KDR: 0.657 (0.626–0.715)	NR	NR	Internal validation (random split 70/30)	Points-based scoring system (sex-specific risk scorecard with 10-year risk table)	Korean adults aged 40–69 years without diabetes
					KDR + FPG: KDR + FPG	KDR + FPG: 0.690 (0.660–0.720)
					KDR + HbA1c: KDR + HbA1c	KDR + HbA1c: 0.746 (0.717–0.775)
Stiglic et al[12], 2021	12 European countries (Austria, Belgium, Denmark, France, Germany, Greece, The Netherlands, Israel, Italy, Spain, Sweden, Switzerland)	Prospective cohort (SHARE survey)	16363 (1360 events over 10 years)	LASSO logistic regression with inverse probability weighting	Global model: BMI, alcohol consumption, physical activity, CASP score, orientation in time, chronic health problems, hypertension, high cholesterol, self-rated health, country indicators (Spain, Israel, Italy)	0.702 (0.698–0.706)	0.642 (0.633–0.650)	0.648 (0.643–0.652)	Internal validation (repeated 10-fold cross-validation with 10 runs)	LASSO logistic regression equation	European adults aged ≥ 50 years without diabetes at baseline
Al-Lawati et al[92], 2022	Oman	Cross-sectional (WHS 2008 for development; Sur Survey 2006 for validation)	Development: 2720 (315 T2DM); validation: 1355 (189 T2DM)	Multivariable logistic regression with backward stepwise elimination	Age, male sex, no education, marital status, urban residence, hypertension, BMI, current tobacco use	Development: 0.80 (0.78–0.82); Validation: 0.74 (0.70–0.78)	Development: 71.1%; Validation: 54.0%	Development: 74.4%; Validation: 79.0%	External validation in independent cohort (Sur Survey 2006)	Points-based scoring system; risk assessment table provided	Omani adults aged ≥ 20 years
Sulistiowati and Pradono[93], 2022	Indonesia	Prospective cohort (Bogor Cohort Study)	4418 (568 events over 6 years)	Multivariable logistic regression	Age, obesity, central obesity, hypertension, physical activity	0.70 (0.675–0.721)	0.766	0.503	Internal validation (split-sample not specified)	Bogor Diabetes Risk Prediction Chart	Indonesian adults aged 25–60 years without diabetes at baseline
Tan et al[48], 2022	Japan	Prospective cohort (NAGALA)	2012 (incident T2DM 6.81%)	LASSO regression; Cox proportional hazards	WC, HbA1c, FPG, HDL, TG, smoking	Training: 0.819 (0.776–0.858); validation: 0.853 (0.820–0.886)	Training: 74.7%; validation: 77.6%	Training: 77.0%; validation: 88.1%	Internal validation (random split 70/30 with 1000 bootstrap resampling)	Nomogram; online calculator	Japanese adults with abdominal obesity
Wang et al[35], 2022	China	Prospective cohort (Guangzhou Biobank Cohort Study)	15934 (1302 events over mean 4.1 years)	Bayesian network with structural expectation maximization algorithm	Hypertension, impaired fasting glucose, WC, age, sex, education, occupation, family income, smoking, drinking, physical activity, daytime napping, snoring, self-rated health, heart rate, lipid-lowering drugs, coronary heart disease history, family history of diabetes, BMI, WHR, TG, HDL-C, LDL-C	0.746 (0.732–0.760)	0.727 (0.703–0.752)	0.660 (0.652–0.667)	Internal validation (5-fold cross-validation)	Bayesian network with conditional probability table; post-test probability table for risk combinations	Chinese adults aged ≥ 50 years without diabetes at baseline
Nicolaisen et al[28], 2022	Denmark	Population-based cohort (Danish National Health Survey linked to healthcare registries)	26007 with incident HbA1c-defined pre-diabetes (2449 progressed to T2D within 5 years)	Fine–Gray competing risk survival model with LASSO variable selection	Baseline HbA1c, age, sex, BMI, antihypertensive drug use, pancreatic disease, cancer, self-reported unhealthy diet, doctor’s advice to lose weight/change diet, lack of social support, good self-rated health	72.7 (71.2–74.3)	68.3 (63.9–72.7)	66.3 (65.4–67.1)	Internal validation (random split 80/20)	Fine–Gray competing risk regression model	Danish adults aged ≥ 30 years with incident HbA1c-defined pre-diabetes
Zhao et al[44], 2022	China	Prospective cohort (Rural Chinese Cohort Study)	9860 (653 events over median 6.01 years)	Cox proportional hazards	Age, sex, BMI, WC, SBP, family history of diabetes	0.675 (0.666–0.684)	72.6% (at cut-off 24)	53.6% (at cut-off 24)	External validation of existing New Chinese Diabetes Risk Score in independent cohort	New Chinese Diabetes Risk Score: Points-based scoring system	Chinese rural adults aged 20–74 years without diabetes at baseline
Sadek et al[94], 2022	Qatar	Cross-sectional (Qatar Biobank)	Development: 1660 (169 T2DM); Validation: 930 (61 T2DM)	LASSO logistic regression	Age, sex, WHR, history of hypertension, history of hyperlipidemia, education level	Development: 0.870 (0.843–0.896); validation: 0.815 (0.765–0.864)	Development (cut-off 26): 84.6%; Validation: NR	Development: 76.2%; validation: NR	External validation in independent QBB cohort (n = 930)	Qatari Diabetes Mellitus Risk Score: Points-based scoring system	Qatari adults aged 18–89 years
Edlitz and Segal[16], 2022	United Kingdom	Prospective cohort (United Kingdom Biobank) with external validation in Israeli Clalit cohort	Development: 44709 (1.79% incident T2D over mean 7.3 years); validation: 17132 (4.1% T2D)	Logistic regression; Cox regression; scorecard based on β coefficients	Anthropometric model: Age, sex, weight, height, WC, hip circumference, BMI, WHR	Anthropometric: 0.81 (0.77–0.84)	NR	NR	Internal validation (train 56%/val 24%/test 20%); external validation of 4-blood-test model only	Scorecard (points-based); logistic regression; Cox regression	United Kingdom adults aged 40–69 years without diabetes at baseline
					Four-blood-test model: Age, sex, HbA1c, GGT, HDL, TG	Four blood test: 0.87 (0.85–0.90); external validation: 0.75 (0.74–0.75)
Liu et al[22], 2022	China	Retrospective cohort (health screening data, Wuhan)	127031 (8298 incident T2D over 2 years)	LASSO feature selection; machine-learning algorithms: LR, DT, RF, XGBoost	Age, sex, education, marital status, hypertension, exercise, current smoking, current drinking, WC, SBP, FPG, TC, TG, HDL-C, LDL-C, ALT, AST, TBIL, Scr, BUN, uric acid	XGBoost: 0.7805; RF: 0.7772; LR: 0.7601; DT: 0.7280	XGBoost: 0.6452; RF: 0.6428; LR: 0.6320; DT: 0.5821	XGBoost: 0.7577; RF: 0.7524; LR: 0.7636; DT: 0.7633	Internal validation (random split 80/20)	Machine learning models (no simplified scoring system); SHAP for feature importance	Chinese adults aged ≥ 65 years without diabetes at baseline
Martínez-Hervás et al[95], 2022	Spain	Prospective cohort (Di@bet.es Study)	2408 (156 incident T2D over 7.5 years)	CHAID decision tree algorithm	FPG, fasting TG, age, sex, BMI	0.764 (0.722–0.807) for FPG alone; model correctly classified 93.5%	NR	NR	Internal validation (cross-validation)	CHAID decision tree diagram with risk stratification nodes	Spanish adults aged ≥ 18 years without diabetes at baseline
Dong et al[57], 2022	China (Hong Kong)	Cross-sectional (Hong Kong Population Health Survey 2014/15)	1857 (280 newly detected DM/pre-DM)	Logistic regression (AIC stepwise)	Age, BMI, WHR, smoking, sleep duration, vigorous recreational activity time, fruit consumption	LR: 0.812 (0.769–0.853)	LR at 15% threshold: 77.6%	LR: 68.1%	Internal validation (random split 2/3 development, 1/3 validation)	Logistic regression equation; online risk calculator	Chinese adults aged 18–84 years without self-reported DM/pre-DM or major chronic diseases
				XGBoost with isotonic calibration	Age, BMI, WHR, SBP, WC, sleep duration, smoking, vigorous recreational activity time	XGBoost: 0.822 (0.779–0.863)	XGBoost at 15% threshold: 72.4%	XGBoost: 77.9%		XGBoost model with SHAP interpretation; online risk calculator
Tong et al[75], 2022	China	Retrospective cohort (Beijing Health Management Cohort)	44852 (2912 incident T2D over median 3.58 years)	Multivariable Cox regression	Sex, age, BMI, FPG, hypertension, dyslipidemia, sCr	Dynamic nomogram: 0.909 (0.897–0.920) (10-fold CV)	NR	NR	Internal validation (random split 70/30); 10-fold cross-validation	Dynamic nomogram	Urban Chinese current or retired employees aged 16–68 years without diabetes at baseline
Boersma et al[59], 2022	Netherlands	Prospective cohort (Lifelines)	Derivation: 58377 (2146 new diabetes)	Multivariable logistic regression	Model 1: Age group, BMI group, SAF group	Model 1: 0.783 (0.774–0.793)	Model 1: 78	Model 1: 66	Internal validation (bootstrapping); external validation comparison to concise FINDRISC model	Points-based scoring system	Dutch general population adults (aged 25–50 years and their family members)
					Model 2: Age group, BMI group, SAF group, parental history of diabetes	Model 2: 0.792 (0.783–0.802)	Model 2: 74	Model 2: 70
Wang et al[54], 2023	China	Cross-sectional	Pre-SMOTE: 4106 (149 T2DM); Post-SMOTE: 8013 (4023 T2DM)	BP neural network, logistic regression, SVM, CART, C4.5, DNN	Age, DBP, FPG, TC, TG, HDL-C, LDL-C, drinking, consumption frequency of cereals, potatoes, beans, fruits, eggs, milk, poultry, fish	BPNN: 0.977 LR: 0.962 SVM: 0.911 CART: 0.906 C4.5: 0.888 DNN: 0.845	BPNN: 92.8 LR: 86.0 SVM: 89.0 CART: 84.8 C4.5: 84.9 DNN: 82.9	BPNN: 94.6 LR: 93.0 SVM: 93.3 CART: 93.3 C4.5: 92.7 DNN: 86.1	Internal validation (10-fold cross-validation, random split 70/30)	Machine learning models; logistic regression equation provided	Chinese adults aged ≥ 18 years in Dongguan, China
Zheng et al[46], 2023	China	Prospective cohort (Kailuan Study)	Total: 18384 (5697 events); training: 12870; internal validation: 5514; external validation: 10620 + 8513	Cox regression with 2-step selection (univariate + LASSO-Cox)	Age, BMI, WC, DBP, TG, FPG, fatty liver	Men: 0.741 (0.723–0.760)	NR	NR	Internal validation (random split 70/30) + external validation (2 independent cohorts)	Points-based scoring system; Web calculator	Chinese adults with IFG
						Women: 0.762 (0.720–0.802)
Yonel et al[74], 2023	Germany	Population-based cohort (SHIP-TREND-0 for development; SHIP-START-0 for external validation)	Development: 3339 (329 events); external validation: 2381 (403 events)	Logistic regression with backward selection; Heuristic shrinkage	Age, sex, BMI, smoking status, first-degree relative with diabetes, presence of dental prosthesis, presence of mobile teeth, history of periodontal treatment, probing pocket depths ≥ 5 mm	Development: 0.72 (0.69–0.75)	Point score (development): 77.0 (76.8–77.2)	Point score (development): 51.5 (51.4–51.7)	Internal validation (backward selection with multiple imputation, heuristic shrinkage); external validation in independent cohort	Full model (coefficients); Simplified point-based scoring system	German adults aged ≥ 40 years without known diabetes, attending dental settings
						External validation (full model): 0.69 (0.67–0.72)	Point score (external validation): 79.2 (79.0–79.4)	Point score (external validation): 49.9 (49.8–50.0)
						External validation (point score): 0.69 (0.66–0.71)
Hu et al[62], 2023	China	Retrospective cohort	Total: 184188 (19699 events); training: 92177; validation: 92011	LASSO regression (compared with full Cox, stepwise Cox, MFP, XGBoost)	Age, BMI, FPG, TG, SBP, Scr	Training: 0.7341 (0.7290–0.7392)	Training: 68.00	Training: 67.13	Internal validation (random split 50/50)	Nomogram; Risk prediction formula	Chinese adults aged ≥ 20 years without prediabetes at baseline
						Validation: 0.7336 (0.7285–0.7387)	Validation: 65.35	Validation: 69.91
Yang et al[96], 2024	China	Prospective cohort (physical examination center)	Total: 4602 (760 events); training: 3221; test: 1381	Logistic regression with stepwise selection	Age, sex, BMI, FPG, HDL-C, LDL-C, fatty liver, liver dysfunction (ALT > 50 U/L and AST > 40 U/L)	Training: 0.787 (0.765–0.808)	NR	NR	Internal validation (random split 70/30; 1000 bootstrap resampling)	Nomogram; logistic regression equation	Chinese adults aged ≥ 20 years with newly diagnosed prediabetes
						Test: 0.800 (0.770–0.829)
Ojurongbe et al[27], 2024	Nigeria	Case-control (hospital-based)	Total: 444 (172 cases, 272 controls); training: 312; test: 132	Penalized logistic regression with nested cross-validation and sequential backward feature selection	WHR, TG, catalase, atherogenic indices of plasma	Training: 0.99 (0.97–1.00)	Training: 96	Training: 93	Internal validation (10-fold nested cross-validation with 20 repetitions)	Web-based application; logistic regression equation	Nigerian adults aged > 18 years attending a diabetes clinic
						Test: 0.94 (0.89–0.99)	Test: 94	Test: 88
Hafezi et al[68], 2024	Iran	Prospective cohort (MASHAD study)	Total: 6710 (1256 events); training: 5032; test: 1678	Logistic regression; decision tree (CHAID)	Model A: BMI, VAI, LAP, abdominal volume index	Model A (LR): Approximately 0.67	NR	Model A (train): 99%; model A (test): 100%	Internal validation (random split 75/25)	Logistic regression equations; Decision tree with 14 IF-THEN rules	Iranian adults aged 35–65 years without diabetes at baseline
					Model B: Model A + age, sex, physical activity level, anxiety, depression, smoking, dietary composition, energy intake	Model B (LR): Approximately 0.70		Model B (train): 99%; model B (test): 100%
Xu et al[97], 2024	Japan	Prospective cohort (JPHC Study)	Development: 10986 (707 events); external validation: 11345 (673 events)	Logistic regression with backward selection	Model 1 (non-invasive): Sex, BMI, family history of DM, DBP	Model 1: 0.643	NR	NR	Internal validation (bootstrap, 1000 iterations); internal –external cross-validation (leave-one-area-out); external validation (J-ECOH Study)	Logistic regression equations; Excel spreadsheet for risk calculation	Japanese adults aged 46–75 years without diabetes at baseline
					Model 2 (invasive): Family history of DM, HbA1c	Model 2: 0.786
					Model 3 (invasive): Family history of DM, HbA1c, FPG	Model 3: 0.845
Lin et al[70], 2024	China	Retrospective cohort	Total: 2235 (99 events); training: 1564; validation: 671	LASSO logistic regression	Age, sex, BMI, FPG, TG, ALT, ALT/AST, BUN, Hb	Training: 0.824 (0.765–0.883)	NR	NR	Internal validation (random split 70/30; 1000 bootstrap resampling)	Dynamic online nomogram; logistic regression equation	Chinese older adults (≥ 60 years) with normal FBG
						Validation: 0.809 (0.732–0.886)
Kawasoe et al[64], 2025	Japan	Retrospective cohort	Total: 31084 (1548 events); derivation: 15542; validation: 15542	Logistic regression	Model 1 (invasive): Age, sex, BMI, FPG, HbA1c	Model 1: 0.89	Model 1: 81	Model 1: 81	Internal validation (random split 50/50)	Logistic regression equations	Japanese adults aged 30–69 years without diabetes at baseline
					Model 2 (non-invasive): Age, sex, BMI, hypertension	Model 2: 0.71	Model 2: 75	Model 2: 54
Xie et al[13], 2025	United Kingdom/Germany	Prospective cohort (United Kingdom Biobank; ESTHER)	Development: 60362 (2477 events); internal validation: 25870 (1059 events); external validation: 4383 (495 events)	LASSO-Cox regression with bootstrap (1000 resamples)	Clinical CDRS: Age, sex, BMI, smoking, family history of diabetes, antihypertensive medication, steroid medication, HbA1c; plus 11 metabolites (glucose, citrate, lactate, pyruvate, 3-hydroxybutyrate, acetate, glutamine, tyrosine, IDL-CE-pct, M-LDL-TG-pct, LA-pct)	Clinical CDRS: 0.815 (0.800–0.829)	NR	NR	Internal validation (random split 70/30); external validation (ESTHER cohort)	Cox regression equations	European adults (United Kingdom and Germany) aged 50–69 years without diabetes at baseline
						United Kingdom Biobank-DRS (full): 0.834 (0.820–0.847)
						United Kingdom Biobank-DRS (concise): 0.830 (0.815–0.844)
Wang et al[98], 2025	Japan	Prospective cohort (NAGALA)	Total: 15464 (262 events)	Cox proportional hazards regression	Age, WC, ALT, HbA1c, FPG, fatty liver, current smoking, heavy alcohol consumption, HDL-C	Combined model: 0.90 (0.88–0.92)	86.3	82.8	Internal validation (not specified)	Risk score combining 9 factors	Japanese adults without diabetes at baseline (mean age 43.7 years)
Shahraki et al[66], 2025	Iran	Prospective cohort (Isfahan Diabetes Prevention Study)	Total: 1765 (diabetes: 7.4%; prediabetes: 32.6%)	Cox proportional hazards regression with optimal cutoff categorization	Diabetes Model 1: WHR, HbA1c, OGTT-AUC, lipid index	Diabetes model 1: 0.71 (0.66–0.77)	Diabetes model 1: 78	Diabetes model 1: 52	Internal validation (ROC analysis; compared with FINDRISC)	Points-based scoring system	First-degree relatives of Iranian patients with type 2 diabetes, with normal glucose tolerance at baseline
					Diabetes model 2: WHR, HbA1c, FPG, lipid index	Diabetes model 2: 0.69 (0.63–0.74)	Diabetes model 2: 77	Diabetes model 2: 52
Chee et al[73], 2025	Singapore	Cross-sectional	1074 (65 confirmed diabetes, 83 confirmed prediabetes)	Logistic regression with stepwise variable selection	Model IV: BMI ≥ 23 kg/m2, family history of DM, smoking, stage III/IV periodontitis	Model IV: 0.717 (0.689–0.744)	Model IV: 71	Model IV: 62	Internal validation	Clinical risk model based on ORs; risk stratification using 4 simple factors	Asian dental patients (Singapore) aged ≥ 21 years without prior diabetes diagnosis
					Model V: BMI ≥ 23 kg/m2, family history of DM, smoking, severe periodontitis	Model V: 0.721 (0.693–0.748)	Model V: 71	Model V: 60
Munekawa et al[47], 2025	Japan	Retrospective cohort (Panasonic Cohort)	Development: 72124 (5133 events); external validation: 12885 (712 events)	Cox proportional hazards model with interval censoring; natural cubic splines; bootstrap variable selection	Age, sex, BMI, SBP, log-TG, HDL, log-ALT, FPG, weight gain, smoking status	Internal validation (optimism-corrected C-index): 0.877 (0.852–0.882)	NR	NR	Internal validation (2000 bootstrap resamples); external validation (NAGALA cohort)	Cox regression equation	Japanese adults aged ≥ 40 years without diabetes at baseline
						External validation C-index: 0.882
Kiviniemi et al[30], 2025	Finland	Multi-cohort prospective (FINRISK 1997-2012, NFBC1966)	Development: 4926 (96 events); validation: 17385 (397 events)	Logistic regression with penalization (ridge, elastic net, LASSO) and stepwise selection	Baseline model: Age, sex, BMI, WC, systolic/diastolic BP, HDL-C, TC, TG, glucose, current smoking, BP medication, lipid medication, diabetes family history	Baseline model: 0.832–0.913 (validation cohorts)	NR	NR	External validation (4 independent cohorts)	Logistic regression equations	Finnish adults aged 25–74 years without diabetes at baseline
					Extended models: Baseline + 154 NMR metabolites	Best extended model: 0.840–0.915 (validation cohorts)
Alareeki et al[25], 2025	Jordan	Mathematical modeling (simulated population based on epidemiological model)	Derivation: 5000 simulated individuals per time point (2020, 2030, 2050)	Multivariable logistic regression	Age group, sex, obesity, smoking, physical inactivity	2020: 0.79 (0.77–0.80)	2020: 78.7 (77.5–79.8)	2020: 64.2 (62.9–65.6)	Internal validation (compared with survey-based risk score in 2017 data); external validation comparison with regional and international risk scores	Points-based scoring system	Jordanian adults aged 20–79 years without known diabetes
						2030: 0.78 (0.77–0.80)	2030: 76.1 (74.9–77.3)	2030: 68.4 (67.1–69.6)
						2050: 0.77 (0.76–0.79)	2050: 74.2 (73.0–75.4)	2050: 65.6 (64.2–66.9)
Chen et al[99], 2025	China (Taiwan)	Prospective cohort (Taiwan Biobank)	Development: 65112 (2485 events); external validation (United Kingdom Biobank): 86484 (1187 events); external validation (APHS): > 4300000	LASSO logistic regression	HbA1c, blood glucose, PEF, PRS, age, WHR, family history of diabetes, BMI, TG, SGPT, prediabetes, self-reported hyperlipidemia, mixed hyperlipidemia (ICD)	Taiwan Biobank testing: 0.9058 (0.8915–0.9202)	Taiwan Biobank testing: 78.34 (71.65–86.61)	Taiwan Biobank testing: 85.88 (77.08–91.20)	External validation (United Kingdom Biobank, APHS)	Logistic regression coefficients; risk prediction equation	Han Chinese adults (Taiwan) aged ≥ 20 years without diabetes at baseline
						United Kingdom Biobank external validation: 0.8869 (0.8543–0.9112)	United Kingdom Biobank external validation: NR	United Kingdom Biobank external validation: NR
Han et al[31], 2025	China	Meta-analysis + prospective cohort validation	Development: Meta-analysis of 70 studies; validation: 3 cohorts (Hongguang, Bao’an, Kailuan)	Pooled effect estimates from meta-analysis; Framingham method for risk scoring	Sex, age, smoking, central obesity, exercise, BMI, WHtR, hypertension, fatty liver, family history of diabetes, FPG, TG, TyG, HDL-C, uric acid	Hongguang: 0.783	Hongguang: NR	Hongguang: NR	External validation (3 independent Chinese cohorts)	Points-based scoring system (iCARE-DM)	Chinese elderly adults (≥ 50 years) without diabetes at baseline
						Bao’an: 0.766	Bao’an: NR	Bao’an: NR
						Kailuan: 0.741	Kailuan: NR	Kailuan: NR
Zhang et al[53], 2025	China	Prospective cohort (HAELS)	Total: 1127 (107 events); training/validation: Bootstrap resampling	Cox proportional hazards regression with stepwise selection (SLE/SLS = 0.1)	Age, exercise, WC, FPG, TG, SBP, dessert intake	Training: 0.752	NR	NR	Internal validation (bootstrap resampling, 2 iterations)	Nomogram	Chinese older adults (≥ 65 years) with prediabetes (FPG 5.6–7.0 mmol/L)
						Validation: 0.793
Baynouna AlKetbi et al[71], 2025	United Arab Emirates	Retrospective cohort (Weqaya screening program)	Total: 8699; analysis: 6772 (643 events)	Cox proportional hazards regression with stepwise variable selection	HbA1c, random blood glucose, BMI, current smoking, sex, age, mean BP, eGFR percentile, HDL, hypertension at baseline	Full model: 0.837 (0.818–0.856)	75.1	78.1	Internal validation (time-dependent ROC)	Cox regression HRs; risk score based on coefficients	United Arab Emirates adults (Emirati nationals) aged ≥ 18 years without diabetes at baseline
						HBA1C alone: 0.784 (0.761–0.807)
						RBS alone: 0.698 (0.672–0.723)
Satoh et al[45], 2025	Japan	Retrospective cohort (SKDB)	Total: 463248 (78431 events); derivation: 308832 (52152 events); validation: 154416 (26279 events)	Cox proportional hazards regression with stepwise selection	Age, sex, BMI, SBP, TG, HDL-C, LDL-C, AST, γ-GTP, HbA1c, eGFR, urine protein, antihypertensive medication, dyslipidemia medication, heavy drinking, exercise habits	C-index (validation): 0.656 (0.652–0.659)	NR	NR	Internal validation (random split 2:1)	Points-based scoring system	Japanese adults aged ≥ 40 years without diabetes at baseline

2hPG: 2-hour post-load plasma glucose; 95%CI: 95% confidence interval; AIC: Akaike information criterion; ALT: Alanine aminotransferase; ANN: Artificial neural network; ANOVA: Analysis of variance; APHS: Adult Preventive Healthcare Services; ARIC: Atherosclerosis Risk in Communities; AST: Aspartate aminotransferase; AUC: Area under the receiver operating characteristic curve; BMI: Body mass index; BP: Blood pressure; BPNN: Backpropagation neural network; BRFSS: Behavioral Risk Factor Surveillance System; BUN: Blood urea nitrogen; CART: Classification and regression tree; CASP: Control, Autonomy, Self-Realization, and Pleasure; CDRS: Cambridge Diabetes Risk Score; CHAID: χ²d Automatic Interaction Detector; CIM: Confusion matrix-based classifier integration approach; CoLaus: Cohorte Lausannoise; CRONICAS: CRONICAS Cohort Study; CV: Cross-validation; CVD: Cardiovascular disease; DBP: Diastolic blood pressure; DL: Deep learning; DM: Diabetes mellitus; DNN: Deep neural network; DRS: Diabetes risk score; DT: Decision tree; eGFR: Estimated glomerular filtration rate; EHR: Electronic health record; ENINBSC: National Survey of Nutritional and Biochemical Indicators for Noncommunicable Diseases; EPIC-Europe: European Prospective Investigation into Cancer and Nutrition; ESTHER: Epidemiological Study on Potentials for Prevention, Early Detection and Optimized Treatment of Chronic Diseases in the Elderly Population; FBG: Fasting blood glucose; FBS: Fasting blood sugar; FF4: Second follow-up examination of the KORA S4 cohort; FINDRISC: Finnish Diabetes Risk Score; FINRISK: National FINRISK Study; FPG: Fasting plasma glucose; GBM: Gradient boosting machine; GGT: Gamma-glutamyl transferase; GRS: Genetic risk score; HAELS: Health Aging Evaluation Longitudinal Survey; HAPIEE: Health, Alcohol and Psychosocial Factors in Eastern Europe; Hb: Hemoglobin; HbA1c: Glycated hemoglobin; HDL: High-density lipoprotein; HDL-C: High-density lipoprotein cholesterol; HDNNCDS: Harbin Cohort Study on Diet, Nutrition and Chronic Noncommunicable Disease; HHBG: History of high blood glucose; HPHS: Harbin People Health Study; HR: Hazard ratio; iCARE-DM: Integrated Chinese Adapted Risk Evaluation for Diabetes Mellitus; ICD: International Classification of Diseases; IDL-CE-pct: Cholesteryl esters to total lipids in intermediate-density lipoprotein percentage; IFG: Impaired fasting glucose; INPer-IMSS: Instituto Nacional de Perinatología–Instituto Mexicano del Seguro Social; J-ECOH: Japan Epidemiology Collaboration on Occupational Health Study; JPHC: Japan Public Health Center-based Prospective Study; KDR: Korean Diabetes Risk; KoGES: Korean Genome and Epidemiology Study; KORA: Cooperative Health Research in the Region of Augsburg; LA-pct: Linoleic acid to total fatty acids percentage; LAP: Lipid accumulation product; LASSO: Least absolute shrinkage and selection operator; LDA: Linear discriminant analysis; LDL: Low-density lipoprotein; LDL-C: Low-density lipoprotein cholesterol; LR: Logistic regression; m-FINDRISC: Modified Finnish Diabetes Risk Score; M-LDL-TG-pct: Triglycerides to total lipids in medium low-density lipoprotein percentage; MASHAD: Mashhad Stroke and Heart Atherosclerotic Disorder Study; MFP: Multivariable fractional polynomial; NAFLD: Non-alcoholic fatty liver disease; NAGALA: NAfld in Gifu Area, Longitudinal Analysis; NFBC1966: Northern Finland Birth Cohort 1966; NHANES: National Health and Nutrition Examination Survey; NHIS-HEALS: National Health Insurance Service–Health Screening Cohort; NHS: National Health Survey; NMR: Nuclear magnetic resonance; NR: Not reported; OGTT: Oral glucose tolerance test; OGTT-AUC: Oral glucose tolerance test area under the curve; OR: Odds ratio; PCOS: Polycystic ovary syndrome; PEF: Peak expiratory flow; PRS: Polygenic risk score; PsyCoLaus: Psychiatric arm of the CoLaus cohort; QBB: Qatar Biobank; RBS: Random blood sugar; REGARDS: REasons for Geographic And Racial Differences in Stroke; RF: Random forest; RFE: Recursive feature elimination; RNN-LSTM: Recurrent neural network long short-term memory; ROC: Receiver operating characteristic; RS-2hPG: Risk score without 2-hour post-load plasma glucose; RSFull: Full risk score; RSNI: Non-invasive risk score; SAF: Skin autofluorescence; SBP: Systolic blood pressure; Scr: Serum creatinine; SGPT: Serum glutamate pyruvate transaminase; SHAP: SHapley Additive exPlanations; SHARE: Survey of Health, Ageing and Retirement in Europe; SHIP-START-0: Baseline examination of the first Study of Health in Pomerania cohort; SHIP-TREND-0: Baseline examination of the second Study of Health in Pomerania cohort; SKDB: Shizuoka Kokuho Database; SLE: Significance level for entry; SLS: Significance level for stay; SMOTE: Synthetic minority over-sampling technique; SNP: Single nucleotide polymorphism; SVM: Support vector machine; T2D: Type 2 diabetes; T2DM: Type 2 diabetes mellitus; TBIL: Total bilirubin; TC: Total cholesterol; TG: Triglycerides; TyG: Triglyceride–glucose index; UDM: Undiagnosed diabetes mellitus; UKB-DRS: UK Biobank Diabetes Risk Score; VAI: Visceral adiposity index; WC: Waist circumference; WHR: Waist-to-hip ratio; WHS: World Health Survey; WHtR: Waist-to-height ratio; XGBoost: Extreme Gradient Boosting; γ-GGT: Gamma-glutamyl transferase; γ-GTP: Gamma-glutamyl transpeptidase.

The geographical distribution of these 82 studies reveals a concentrated focus on regions characterized by high diabetes prevalence and aging populations. Asia accounted for the majority of the research, with China contributing the largest share (n = 30, 36.6%), followed by Japan (n = 12, 14.6%) and South Korea (n = 4, 4.9%). Europe contributed 13 studies, including 2 each from the United Kingdom and Germany, 1 multinational study spanning 12 European countries[12], and 1 study that utilized cohorts from both the United Kingdom and Germany[13]. Furthermore, 6 studies originated from North America, equally divided between the United States (n = 3) and Mexico (n = 3).

Regarding study design, prospective cohort studies were the most common (n = 39, 47.6%), exemplified by large-scale studies such as the Beijing Longitudinal Study of Ageing[14], the Aichi Workers’ Cohort Study[15], and the United Kingdom Biobank study[16]. Cross-sectional studies accounted for 20 studies (24.4%), including national survey-based studies from Peru[17], Iran[18], and the United Arab Emirates[19]. Retrospective cohort studies represented 17.1% (n = 14), with notable examples from Japan[20,21] and China[22,23]. Other designs included mathematical modeling[24,25], case-control[26,27], population-based cohort[28], nested case-control[29], and multi-cohort prospective[30] studies, as well as a study combining meta-analysis with prospective cohort validation[31]. Sample sizes varied considerably, ranging from 444 participants in a Nigerian case-control study[27] to over 8 million participants in the United Kingdom QResearch database study[32]. Among prospective cohort studies, follow-up durations ranged from 2.4 years[33] to 20 years[14].

Traditional regression methods were the predominant modeling approach. Logistic regression was employed in 42 studies (51.2%), while Cox proportional hazards models were used in 32 studies (39.0%). Regularized regression techniques, such as least absolute shrinkage and selection operator (LASSO) or elastic net, were applied in 17 studies (20.7%) to address overfitting and select predictors. Machine-learning algorithms were utilized in 12 studies, with support vector machines and decision trees being the most common. Deep-learning approaches were employed in a large Korean study[34] using recurrent neural networks. Other methods included competing risk models (Fine–Gray) to account for non-diabetes death[14,28], Bayesian networks[35], and a meta-analysis-based model[31].

Variable selection strategies varied considerably. Stepwise selection methods (forward, backward, or bidirectional) were employed in over half of the studies (n = 48, 58.5%). LASSO regression was used for variable selection in 14 studies. Univariate pre-screening followed by multivariate modeling was also a common practice. Categorization of continuous predictors was frequently reported, often using data-driven cut-offs or clinically defined thresholds.

Internal validation was performed in 67 studies, with split-sample validation being the most common approach (n = 32), followed by bootstrap resampling (n = 21) and cross-validation (n = 13). External validation, using an independent dataset from a different population or setting, was conducted in 28 studies. Notably, 15 studies performed both internal and external validation.

The methodological quality of the included studies was assessed using PROBAST, with detailed study-level judgments presented in Supplementary Table 1. Overall, 67 of the 82 studies (81.7%) were judged to be at high risk of bias, whereas 15 studies (18.3%) were judged to be at low risk. The analysis domain was the principal source of bias, with 64 studies (78.0%) rated as high risk. The main analytical concerns were the use of stepwise or univariate predictor-selection procedures, categorization or dichotomization of continuous predictors, inadequate handling of missing data, and insufficient assessment of model calibration. The participants domain was rated as high risk in 24 studies (29.3%), principally because of selected or non-representative study populations and case-control, retrospective, or cross-sectional sampling designs that were not fully aligned with the intended prediction setting. The outcome domain was rated as high risk in 7 studies (8.5%), mainly because outcomes were based on self-report, a single biochemical measurement, or insufficiently robust diagnostic definitions. All studies were rated as low risk in the predictors domain. Regarding applicability, 75 studies (91.5%) were judged to have a low concern for applicability, whereas 7 studies (8.5%) had a high concern, generally because they were developed in narrowly selected occupational, clinical, or otherwise restricted populations.

Model performance and validation: Discrimination was the most commonly reported aspect of model performance, primarily using the AUC or C-statistic. Internal validation was conducted in 67 studies, whereas only 28 studies used an independent external validation dataset, and 15 performed both internal and external validation. Calibration was less consistently assessed and was frequently absent from model reports, contributing to a high risk of bias in the PROBAST analysis domain. Evidence regarding clinical utility, including decision-curve analysis, implementation feasibility, or prospective impact evaluation, was limited. Therefore, models with high development-set AUCs should not necessarily be regarded as clinically superior to simpler models with external validation and readily obtainable predictors.

Risk factor statistics

To systematically identify the most commonly incorporated predictors across the included risk models, we extracted and synthesized data on all predictor variables from the final models of the 82 included studies. The specific frequencies of each predictor variable are detailed in Table 3, while Supplementary Table 2 provides a comprehensive overview of all predictor variables included in the studies, model presentation, and their intended clinical applications. Our analysis revealed a diverse range of risk factors utilized in model development, spanning non-modifiable characteristics, obesity indices, lifestyle behaviors, and biochemical markers. Among these, age was the most frequently incorporated predictor (n = 83), followed by body mass index (BMI) (n = 68), glucose levels (fasting or random; n = 51), sex (n = 48), and family history of diabetes (n = 41). In addition to these primary factors, several clinical and lifestyle variables showed high utilization rates, including triglycerides (TG) (n = 39), smoking status (n = 36), hypertension (n = 34), and waist circumference (WC) (n = 28). Other notable predictors appearing in multiple models included HbA1c (n = 24), physical activity (n = 23), and high-density lipoprotein cholesterol (HDL-C) (n = 20). Overall, the selection of predictors reflects a multifaceted approach, balancing easily accessible anthropometric data with specialized laboratory parameters to estimate diabetes-related risks. Figure 2 provides an overview of the included studies, the major potentially modifiable predictor domains incorporated into T2DM prediction models, and the principal methodological findings of this review.

Figure 2 Overview of potentially modifiable predictors and key findings of the included type 2 diabetes mellitus prediction studies. BMI: Body mass index; FPG: Fasting plasma glucose; HbA1c: Glycated hemoglobin; HDL-C: High-density lipoprotein cholesterol; LDL-C: Low-density lipoprotein cholesterol; T2DM: Type 2 diabetes mellitus; TC: Total cholesterol.

Table 3 Frequency of predictor variables across the included models.

Category	Predictor	Frequency
Non-modifiable	Age	83
	Sex	48
	Family history of diabetes	41
	Education	7
	Race/ethnicity	5
Obesity	Body mass index	68
	Waist circumference	28
	Waisttohip ratio	12
	Obesity/central obesity	5
	Waist-to-height ratio	3
Lifestyle	Smoking	36
	Physical activity	23
	Dietary factors (e.g., fruit, vegetable, meat, dessert intake, tea)	14
	Alcohol	11
	Sleep duration	5
Glycemic	Glucose (FPG or random glucose)	51
	HbA1c	24
	Post-load glucose measures (2hPG or OGTT-AUC)	4
	Insulin	3
Lipid	Triglycerides	39
	HDL	20
	Total cholesterol	9
	LDL	8
	Hyperlipidemia/dyslipidemia medication use	8
	Lipid accumulation product/lipid indices	3
Blood pressure	Hypertension	34
	SBP	15
	DBP	6
Liver	Liver dysfunction/ALT/AST	8
	Fatty liver	6
Metabolic	Uric acid	4
	Creatinine/eGFR/renal function	3

2hPG: 2-hour post-load glucose; OGTT-AUC: Oral glucose tolerance test area under the curve; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; DBP: Diastolic blood pressure; eGFR: Estimated glomerular filtration rate; FPG: Fasting plasma glucose; HbA1c: Glycated hemoglobin; HDL: High-density lipoprotein; LDL: Low-density lipoprotein; SBP: Systolic blood pressure.

Fixed factors in diabetes risk models: Fixed risk factors are those that cannot be modified by individual behavior or medical intervention. Across the 82 included studies, non-modifiable characteristics—particularly age, sex, and family history of diabetes—consistently formed the backbone of risk prediction models for T2DM.

Age was universally incorporated, appearing in 83 models. Its operationalization varied substantially: Several studies employed predefined age cut-offs, such as ≥ 55 years[17,20] or decade-based categories[14,36], whereas others treated age as a continuous linear or spline-transformed term[16,32].

Sex was included in 48 studies. Among studies reporting directionality, male sex was generally associated with a higher risk of undiagnosed or incident diabetes, particularly in Asian populations[15,37,38]. However, several models developed sex-specific scoring systems, reflecting substantial differences in risk factor profiles between men and women[38-40].

Family history of diabetes, present in 41 studies, was typically defined as a first-degree relative with diabetes and often carried a weight comparable to that of major metabolic predictors, as seen in models from the United Arab Emirates[19] and Switzerland[41].

Other fixed factors appeared less frequently. Education was included in 7 models across 6 studies, serving as a proxy for socioeconomic status, with lower educational attainment consistently associated with increased diabetes risk. Race or ethnicity featured in 5 studies, most notably in the United Kingdom QDiabetes model[32], where South Asian (Indian, Pakistani, Bangladeshi), Black African, and Black Caribbean ethnicities were associated with significantly higher risk compared with White populations. A multi-ethnic Malaysian case-control study[26] reported that the prevalence of T2DM was highest in Malaysian Indians (25%–28%), followed by Malays (17%–19%) and Chinese (9%–14%). After adjusting for age, sex, waist-to-hip ratio (WHR), and physical inactivity, the environmental-only model achieved an AUC of 0.83 [95% confidence interval (95%CI): 0.80–0.86] in Chinese, 0.78 (95%CI: 0.75–0.81) in Indians, and 0.75 (95%CI: 0.72–0.78) in Malays.

Collectively, these fixed factors provide the stable foundation upon which modifiable risk elements are added to achieve context-specific risk stratification.

Modifiable factors in diabetes risk models: In the identified diabetes risk models, the modifiable risk factors predominantly encompass several key domains. These include anthropometric indices of adiposity, such as BMI, WC, WHR, and waist-to-height ratio (WHtR). Lifestyle behaviors constitute another major category, comprising smoking status, alcohol consumption, dietary patterns, sleep duration, and physical activity. Biochemical markers are further classified into glycemic indicators—namely FPG, HbA1c, 2-hour post-load glucose, and insulin—and lipid profiles, including total cholesterol, TG, HDL-C, and low-density lipoprotein cholesterol. Additionally, manageable clinical and metabolic parameters are frequently incorporated, notably blood pressure (BP) markers such as systolic BP (SBP) and diastolic BP (DBP), as well as hepatic indicators involving liver function and fatty liver and renal-metabolic factors including uric acid and creatinine. Because proactive lifestyle modifications or therapeutic interventions targeting these modifiable variables represent a pivotal strategy for mitigating or delaying the onset of T2DM, the present review prioritizes the synthesis of these actionable indicators.

Obesity-related factors: Anthropometric indices of adiposity emerged as the most prevalent modifiable predictors across the included models, reflecting the established role of adipose tissue in driving multi-organ insulin resistance and β-cell dysfunction[42]. BMI served as the cornerstone indicator, integrated into 68 models. While many studies categorized BMI using standardized thresholds for overweight (≥ 24.0 kg/m² or 25.0 kg/m²) and obesity (≥ 28.0 kg/m² or 30.0 kg/m²), particularly in cohorts from China[14,37,43,44] and Japan[20,45], others preserved BMI as a continuous variable to maximize predictive granularity. Furthermore, several models utilized sex-specific BMI thresholds or point-based allocations to account for dimorphic risk profiles[39,46]. Notably, Munekawa et al[47] introduced a longitudinal perspective by incorporating “weight gain of ≥ 10 kg since age 20 years” into their Cox regression analysis. This variable was independently associated with diabetes risk—yielding a hazard ratio (HR) of 1.07—which suggests that temporal weight fluctuations may capture cumulative metabolic risk more effectively than a static BMI measurement[47].

WC was incorporated into 28 studies as a primary surrogate for central adiposity[17]. The clinical application of WC varied; for instance, the Peruvian risk score[19] allocated 5 points for elevated WC (≥ 90 cm in men and ≥ 85 cm in women). The Japanese NAGALA studies[21,48] and the Chinese impaired fasting glucose (IFG) model[46] also utilized WC as either a continuous or categorical predictor. Of particular interest, Tan et al[48] developed a T2DM prediction model specifically for Japanese adults with abdominal obesity, integrating WC alongside HbA1c, FPG, HDL-C, TG, and smoking status. Other adiposity-related metrics included WHR, utilized in 12 studies, and WHtR, featured in 3 studies. Additionally, 5 studies employed a composite category of obesity or central obesity, defined by dichotomous or ethnic-specific criteria. Collectively, these findings underscore that BMI and measures of abdominal adiposity remain robust and consistent predictors of both undiagnosed and incident diabetes across diverse global populations.

Lifestyle-related factors: Lifestyle-related factors represented a cornerstone of modifiable risk assessment across the reviewed models, reflecting a shift toward integrating granular behavioral data into T2DM screening tools. Among these, dietary habits emerged as a diverse and potent category of predictors. In 2016, Han et al[49] showed that adding dietary predictors, including staple foods, livestock, eggs, potatoes, dairy products, fresh fruits and vegetables, significantly improved model discrimination [ΔAUC = 0.03, relative integrated discrimination improvement (IDI) = 24.6%, net reclassification improvement (NRI) = 0.155] in a northern Chinese population, with external validation confirming these gains. Several models incorporated meat consumption patterns; notably, Chen et al[50] identified a “meat diet” (a self-reported preference for meat over vegetables) as a significant independent predictor (adjusted HR = 2.41, 95%CI: 1.87–3.11), contributing 3 points to their total risk score. Conversely, higher intake of plant-based foods demonstrated protective effects. Zhang et al[51] highlighted that daily vegetable intake of < 400 g and fruit intake of < 300 g were assigned 8 and 6 points, respectively, emphasizing the heavy weight of low fiber intake in risk stratification. The predictive power of nutrition was further underscored by a specialized model utilizing solely dietary data, which achieved an AUC of 0.79; this study identified novel predictors such as xylitol (positive correlation) and the protective roles of mushrooms, tomatoes, and galactose[52]. Other dietary elements integrated into models included “sweet flavor” preference, which Zhang et al[53] found to have a risk contribution comparable to physiological markers such as TG and SBP, as well as the consumption of soft drinks, cereals, and poultry[36,54]. Interestingly, frequent tea consumption was identified as a protective factor, receiving a negative score (−2 points) in 1 Chinese screening tool[55].

Smoking emerged as the most prevalent lifestyle determinant, integrated into 36 studies. Both passive and active exposure to cigarette smoke have been linked to an elevated risk of T2DM compared with nonsmokers[56]. For example, a large Japanese study quantified current smoking as an independent risk factor (HR = 1.09, 95%CI: 1.07–1.11) compared with never smoking, while former smoking also showed a modest elevation in risk[45]. Alcohol consumption was featured in 11 models, typically characterized as a binary variable reflecting current drinking status. While often treated as an independent risk factor, its predictive weight was generally lower than that of smoking, appearing less frequently in non-invasive screening tools.

Physical activity and sleep hygiene have gained prominence in recent models as critical behavioral predictors. Physical activity was included in 23 models, almost invariably as a protective factor. Physical inactivity was associated with increased risk across multiple settings: The Iranian model[18] assigned 1.5 points for no daily exercise, while the Swiss CoLaus model[41] defined inactivity as less than 2 sessions of ≥ 20 minutes of leisure-time physical activity per week.

Sleep duration has emerged as a novel chronobiological risk factor incorporated in 5 models. Evidence suggests a U-shaped association in which both insufficient sleep (≤ 6 hours) and excessive sleep (≥ 9 hours) increase the incidence of T2DM. Dong et al[57] confirmed this U-shaped relationship in Chinese adults, reporting the lowest risk at 7–8 hours per day and a significantly higher risk for short sleep (< 7 hours). One innovative nomogram demonstrated that incorporating sleep duration significantly enhanced model performance (IDI = 0.017, NRI = 0.619), with short sleep duration (< 6 hours/day) associated with a 98% increased risk (odds ratio = 1.98)[58].

Finally, some models sought to capture the cumulative biological impact of long-term lifestyle choices through innovative biomarkers. Boersma et al[59] introduced skin autofluorescence as a surrogate marker for advanced glycation endproducts. Skin autofluorescence serves as a non-invasive “metabolic memory” that quantifies the long-term biological effects of modifiable factors such as chronic poor nutrition and smoking, providing a more stable predictive indicator than transient behavioral reports. Evidence regarding incremental predictive value was limited. In individual models, adding dietary variables increased the AUC by 0.03 and improved the IDI and NRI, while inclusion of sleep duration improved the IDI and NRI; however, these findings were derived from study-specific comparisons and were not consistently externally validated.

Glycemic factors: Glycemic markers represented the most direct physiological predictors in the identified models, serving as essential components for refining risk estimates beyond non-invasive screening. Glucose measures (FPG or random glucose) appeared in 51 models, making them the most common laboratory predictor. Many studies used FPG as a continuous variable[60-62], while others applied categorical cut-offs based on clinical thresholds for IFG. For example, the Chinese rural model[50] assigned 4 points for IFG (FPG: 6.1–6.9 mmol/L), and the Japanese NAGALA studies[21,48,63] incorporated FPG as a continuous or categorical predictor in combination with HbA1c. The United Kingdom QDiabetes model[32] included FBG as an optional addition to model B, demonstrating improved discrimination beyond model A alone.

HbA1c has seen increasing adoption as a stable indicator of long-term glycemic control, appearing in 24 models. HbA1c was often utilized to enhance the accuracy of traditional clinical models, particularly when differentiating between prediabetes and undiagnosed T2DM[64]. Furthermore, post-load glucose measures, including 2-hour post-load glucose or oral glucose tolerance test-AUC markers were included in 4 models across 3 studies, capturing dynamic glucose intolerance that fasting measures might overlook[20,65,66].

Lipid-related factors: Lipid profile dysregulation, a hallmark of metabolic syndrome, was extensively integrated into the identified prediction models as a key modifiable risk category. Among lipid parameters, TG emerged as the most prevalent predictor, utilized in 39 models.

Most studies treated TG categorically using standard cut-offs; for instance, both the Mexican biochemical model[67] and the Japanese SKDB cohort study[45] employed a threshold of > 150 mg/dL, while the Chinese IFG model[46] assigned risk points for TG ≥ 2.3 mmol/L.

HDL-C was the second most frequently incorporated marker (n = 20) and was consistently characterized as a protective factor. The Japanese NAGALA study[48] incorporated HDL as a continuous variable in its nomogram, while the Russian model[39] assigned risk points for HDL-C ≤ 0.9 mmol/L in males. The United Kingdom anthropometric scorecard included HDL alongside TG and HbA1c in its 4-blood-test model[16]. Total cholesterol appeared in 9 studies and low-density lipoprotein cholesterol in 8 studies, often as continuous variables or categorized according to local guidelines.

Hyperlipidemia or dyslipidemia medication use was incorporated in 8 studies as a binary predictor. For instance, the Finnish multi-cohort study included lipid medication use as a covariate[30].

Notably, recent modeling efforts have shifted toward composite lipid indices to better capture synergistic metabolic burden. These include the lipid accumulation product and visceral adiposity index utilized in the Iranian model[68] and the triglyceride–glucose index incorporated into the iCARE-DM model for elderly Chinese adults[31]. In the latter, the triglyceride–glucose index contributed up to 14 points to the total risk score, underscoring its potent predictive value as a surrogate for insulin resistance.

BP-related factors: Hypertension and elevated BP metrics were ubiquitous across the reviewed models, serving as critical hemodynamic indicators of metabolic risk. Hypertension (defined as a binary diagnosis, treatment status, or measured BP ≥ 140/90 mmHg) was the most frequently incorporated BP indicator, appearing in 34 models. SBP was included as a separate predictor in 15 models, while DBP appeared in 6 models. Among these, only 2 studies[34,61] simultaneously incorporated both SBP and DBP as continuous variables. The remaining studies included either SBP or DBP alone.

Notably, 1 study[23] specifically developed a prediction model for incident T2DM in Chinese adults with hypertension, a high-risk subgroup in whom the risk of developing diabetes is approximately 2.5 times that of normotensive individuals, and the coexistence of hypertension and diabetes confers worse cardiovascular outcomes[69].

Liver-related factors: Liver-related predictors reflect the central role of the liver in glucose homeostasis and the pathophysiological link between hepatic steatosis and systemic insulin resistance. Liver dysfunction markers [e.g., alanine aminotransferase (ALT), aspartate aminotransferase (AST)] were included in 8 models, while fatty liver (diagnosed by ultrasonography) appeared in 6 models.

Among studies incorporating fatty liver, a Japanese model developed specifically for non-obese individuals (BMI < 25 kg/m²) found that fatty liver was a strong independent predictor of incident T2DM (HR = 1.98, 95%CI: 1.38–2.84), contributing substantially to the nomogram[63]. Another Japanese study focused exclusively on the non-alcoholic fatty liver disease population—a high-risk group with an established link to diabetes—and developed a prediction model using LASSO regression to select smoking status, WC, FPG, and HbA1c, achieving a C-index of 0.822 in validation[21]. In a large Chinese cohort of adults with IFG, fatty liver severity (absent, mild, moderate, or severe) was incorporated into sex-specific points-based scoring systems, with a web-based calculator provided[46].

Beyond individual enzymes, recent research has shifted toward composite liver indices to improve predictive accuracy. Lin et al[70] pioneered the inclusion of the ALT/AST ratio in T2DM risk models. This ratio is considered a more sensitive indicator of hepatic fat infiltration and hepatic insulin resistance than single enzyme levels, providing a more nuanced reflection of the metabolic state of the liver.

Broader metabolic and renal indicators: Recent models have expanded beyond traditional glycemic and lipid panels to incorporate indicators of renal function and systemic metabolic waste clearance, reflecting the intricate crosstalk between the kidneys and glucose metabolism. Uric acid appeared in 4 models. Creatinine, estimated glomerular filtration rate (eGFR), or renal function markers appeared in 3 models. A novel contribution came from Baynouna AlKetbi et al[71], who reported that eGFR percentiles—particularly the 97^th percentile—were significantly associated with diabetes risk (HR = 1.379) in a large Emirati cohort, representing the first use of eGFR percentiles rather than absolute values in T2DM prediction. Lin et al[70] incorporated blood urea nitrogen into a risk model for older Chinese adults with normal fasting glucose, noting that blood urea nitrogen levels are influenced by protein intake, renal function, and hydration status and may serve as an accessible marker of metabolic health. Satoh et al[45] demonstrated that both reduced eGFR and positive urinary protein (proteinuria) serve as independent predictors of T2DM. Their model integrated these renal markers alongside liver enzymes (AST and gamma-glutamyl transferase) to capture a multi-organ profile of metabolic risk.

Other modifiable factors: Beyond the established domains of obesity, lifestyle, glycemia, lipids, BP, and liver function, several studies incorporated less conventional modifiable predictors reflecting occupational, psychosocial, and oral health exposures. Tian et al[72] developed a diabetes risk score specifically for Chinese oil field workers—a population potentially exposed to environmental pollutants such as benzene, which has been linked to increased diabetes risk. This study, conducted in a single occupational setting, highlighted the potential importance of work-related exposures as predictive factors in high-risk subpopulations.

In the diabetes risk models developed for dental patients in 2 studies, periodontal health indicators were included as modifiable predictors. Chee et al[73] incorporated periodontitis (stage III/IV) into a risk model for Asian dental patients, finding an independent association with undiagnosed diabetes (OR = 2.01). This finding underscores the oral–systemic health link, suggesting that controlling inflammation through periodontal therapy may not only improve oral health but also reduce diabetes risk. Similarly, Yonel et al[74] demonstrated that periodontal pocket depth ≥ 5 mm, self-reported mobile teeth, and history of periodontal treatment were independent predictors of undiagnosed non-diabetic hyperglycemia or T2DM in German adults attending dental settings. These periodontal indicators are readily obtainable in dental practice and may be modified through periodontal therapy and daily oral hygiene[74].

Overall, anthropometric and glycemic predictors were the most consistently incorporated potentially modifiable domains across different populations. Lifestyle predictors were less frequently included but provided additional predictive information in several individual models, particularly dietary and sleep-related variables. The relative frequency of predictors differed between regions; however, these differences may reflect variation in study design, data availability, target populations, and modeling objectives rather than true biological differences between Asian and Western populations. Because predictor definitions and modeling methods varied substantially, frequency of inclusion should not be interpreted as a direct measure of causal or predictive importance.

DISCUSSION

This systematic review synthesizes 82 T2DM risk prediction models published between 2016 and 2025, providing a comprehensive overview of their methodological characteristics, predictive performance, and modifiable risk factor composition. Over the past decade, the field has witnessed a substantial geographical shift: Whereas a 2011 review included only 4 Asian models[6], the present analysis reveals that East Asia—particularly China—now dominates model development, accounting for 36.6% of all included studies. This expansion reflects both the rising diabetes burden in the region and increasing local efforts to tailor risk stratification to specific populations, including urban[58,67,75], suburban[76], and rural residents[44,50,55,60,77,78]; retired workers[37,75]; and occupational groups such as oil field workers[72]. Despite this progress, a considerable proportion of risk scores developed in high-income settings remain inadequate for low-resource regions, underscoring the need for context-specific adaptation and external validation.

Traditional regression methods—logistic regression and Cox proportional hazards models—remain the predominant analytical tools, owing to their low computational cost, straightforward implementation, and interpretability. However, over half of the studies employed stepwise selection or univariate pre-screening, practices known to produce biased coefficient estimates and inflated type I error rates. Regularized regression techniques such as LASSO or elastic net were used in 17 studies, despite their advantages in high-dimensional settings and overfitting prevention. Machine-learning algorithms, including support vector machines, decision trees, and neural networks, appeared in 12 studies. While these methods excel at capturing non-linear relationships and handling large-scale, multidimensional data[70,79], their “black-box” nature limits clinical interpretability. Direct comparisons between machine-learning and conventional regression approaches were limited. In the study by Xie et al[80], the neural network achieved an AUC of 0.795, only marginally higher than that of logistic regression (AUC = 0.793) when evaluated using the same dataset. This finding is consistent with broader evidence showing no consistent performance advantage of machine learning over logistic regression in clinical prediction models. Therefore, when predictive performance is comparable, interpretability, model complexity, and the availability of external validation should also be considered when selecting models for clinical application[80,81]. One notable exception is a Chinese decision tree model that generated 19 IF-THEN rules, offering transparent clinical decision support[82]. Regardless of the modeling approach, many studies failed to report key development steps (e.g., handling of missing data, calibration assessment), highlighting persistent gaps in methodological rigor.

Validation practices varied considerably. Internal validation was performed in 67 studies, most commonly using split-sample (32 studies) or bootstrap (21 studies) validation. External validation—using independent datasets from different populations or settings—was conducted in only 28 studies, and just 15 studies performed both internal and external validation. The lack of external validation raises concerns about overfitting and generalizability, particularly for models with modest sample sizes or complex predictor sets. For instance, a Nigerian case-control model reported an AUC of 0.99, likely reflecting overfitting given the small sample (444 participants) and nested cross-validation without independent testing[27]. By contrast, the United Kingdom QDiabetes study[32] and the Japanese Panasonic cohort study[47] exemplify best practices with large derivation samples, internal bootstrap or cross-validation, and robust external validation.

The PROBAST assessment indicated that the principal methodological weakness was located in the analysis domain rather than in predictor measurement itself. Data-driven selection procedures, including univariate pre-screening and stepwise regression, may produce unstable predictor sets and overly optimistic regression coefficients, particularly when the number of outcome events is limited. Categorizing continuous variables can lead to loss of information, reduced statistical efficiency, and thresholds that may not generalize to other populations. Inadequate handling of missing data may further introduce selection bias, while the limited reporting of calibration restricts assessment of whether predicted absolute risks agree with observed risks. Consequently, the discrimination reported in the development datasets, particularly high AUC values without adequate internal or external validation, should be interpreted cautiously.

The main contribution of this review is not the identification of entirely new diabetes risk factors but the structured evaluation of how potentially modifiable predictors are incorporated into contemporary diagnostic and prognostic models and whether these models are sufficiently validated for clinical application. A central focus of this review was the identification of modifiable predictors that can be targeted through lifestyle or pharmacological interventions. Obesity-related measures were the most frequently incorporated modifiable domain, with BMI appearing in 82.9% of models. WC (34.1%), WHR (14.6%), and WHtR (3.6%) were also commonly used. Notably, a Japanese model specifically developed for abdominally obese individuals[48] and another that introduced “weight gain ≥ 10 kg since age 20 years” as a predictor[47] illustrate how capturing long-term weight trajectory may provide additional risk information beyond a single BMI measurement. Simple anthropometric indices such as WC are cost-effective and strongly predictive, making them suitable for primary care screening[41]. The WHtR (> 0.5) demonstrated particularly strong performance in Mexican populations[67]. In resource-limited primary care settings, simple non-laboratory models based on demographic characteristics, anthropometric measures, family history, and lifestyle factors may be useful for initial risk stratification, followed by targeted biochemical testing in individuals identified as being at higher risk. However, local recalibration and evaluation of feasibility, clinical impact, and cost-effectiveness are required before implementation.

Lifestyle factors—smoking, physical activity, dietary habits, alcohol consumption, and sleep duration—were variably included. A German study using only nutrition data achieved an AUC of 0.79 and identified novel dietary predictors such as xylitol (positive) and mushrooms (negative)[52]. In rural China, a meat-based diet was associated with a 2.4-fold increased risk[50], while frequent tea drinking was protective (−2 points)[55]. Sleep duration exhibited a U-shaped association; short sleep (< 6 hours/day) increased risk by 98% and significantly improved model performance (IDI = 0.017, NRI = 0.619)[58]. Physical inactivity was consistently harmful, and regular exercise reduced risk (HR = 0.95)[45]. A novel non-invasive marker, skin autofluorescence, was introduced as a proxy for cumulative advanced glycation end products from diet and smoking[59].

Glycemic, lipid, and BP measures remained core laboratory predictors. FPG and HbA1c were the most frequent, with the latter substantially improving discrimination when added to non-invasive models. The triglyceride–glucose index, a composite of fasting glucose and TG, contributed up to 14 points in the iCARE-DM model[31]. Hypertension and individual BP components (SBP and DBP) were also widely included. Notably, a model specifically targeting hypertensive patients demonstrated that this subgroup has a 2.5-fold higher diabetes risk and benefits from tailored prediction[23,69]. Liver and renal markers appeared in a minority of models but offered incremental value. Fatty liver diagnosed by ultrasonography was a strong predictor in non-obese Japanese individuals[63] and in non-alcoholic fatty liver disease-specific models[21]. The ALT/AST ratio was first used in a model of older Chinese adults[70]. eGFR percentiles[71] and proteinuria[45] were independent predictors, suggesting that renal function markers can identify high-risk individuals even before overt kidney disease. Furthermore, the diabetes screening model for dental patients opens up a new area of modifiable risk factors. The Diabetes Risk Assessment in Dentistry Score model and other 4-factor models (incorporating BMI, smoking, and periodontitis) allow dental professionals to identify high-risk individuals without invasive blood tests[73,74].

Although genetic factors are known to contribute to T2DM susceptibility, we excluded several genetic risk score studies from this review because they did not include any modifiable predictors. Among the few gene–environment studies that met our inclusion criteria, a large multi-ethnic Malaysian case-control study[26] found that adding a 62-single nucleotide polymorphism genetic risk score to environmental factors increased the AUC by only 1%–3% and pseudo-R² by 1%–2%. Importantly, although no significant gene–environment interactions were detected, the genetic risk score effect was larger in leaner individuals: OR = 1.49 for normal weight, 1.38 for overweight, and 1.21 for obesity. This gradient suggests that the relative contribution of genetic risk may be diluted in high-risk environments where modifiable factors (e.g., obesity) dominate.

Several limitations should be acknowledged. First, we did not perform a formal meta-analysis or quantitative synthesis of model performance because of substantial heterogeneity in study designs, outcome definitions, and reporting standards. Second, we did not assess publication bias, which may favor models with higher AUCs. Third, the PROBAST assessment indicated a high overall risk of bias in 67 of the 82 included studies (81.7%), predominantly because of limitations in the analysis domain, including inappropriate predictor-selection procedures, categorization of continuous predictors, inadequate handling of missing data, and limited calibration assessment. These methodological limitations may have resulted in optimistic estimates of predictive performance and reduced the reproducibility and transportability of the models. Fourth, the predominance of Chinese and Japanese studies (together 51.2%) may limit generalizability to other ethnicities, particularly African and Latin American populations. Fifth, the review focused on modifiable factors, but the operationalization of “modifiable” varied across studies (e.g., some considered FPG modifiable, while others did not). Finally, we did not assess the clinical impact or cost-effectiveness of the models, which are crucial for real-world implementation.

CONCLUSION

The principal objective of risk models is to estimate the probability that an individual or population will develop a given disease under specified conditions. This systematic review comprehensively assessed T2DM risk models published over the past decade, evaluating their methodological approaches, performance metrics, validation strategies, and predictor selection. Our findings indicate that these models can support early identification of individuals at elevated diabetes risk and enhance awareness of modifiable risk factors and the benefits of lifestyle modification. The identified predictor domains may inform the design of future risk-stratification and prevention strategies. However, evidence that model-guided interventions improve clinical outcomes or are cost-effective remains limited, and prospective impact studies are required before routine implementation.