Insulin resistance and incident prediabetes in underweight individuals: a 5-year Chinese cohort study

Abstract

Objective

To examine the association between surrogate insulin-resistance (IR) indices and incident prediabetes across BMI categories, with particular attention to underweight individuals.

Methods

This retrospective cohort study included 99,016 Chinese adults with normoglycaemia at baseline. Four IR indices—triglyceride-glucose (TyG) index, TyG-BMI, atherogenic index of plasma (AIP) and metabolic score for IR (METS-IR)—were calculated. Participants were classified as underweight, normal-weight or overweight/obese. The primary outcome was new-onset prediabetes (fasting plasma glucose 5.6–6.9 mmol/L). Variable importance was ranked using the random-forest-based Boruta algorithm, and multivariable Cox proportional hazards models were employed to estimate risk associations.

Results

Over a median follow-up of 3.12 years, 12,152 (12.3%) individuals developed prediabetes. All IR indices were significantly and positively associated with prediabetes risk (all P < 0.0001). The strongest associations were observed in the underweight group; for example, TyG-BMI yielded an HR of 1.28 (95% CI 1.12–1.46), higher than in other BMI categories. Boruta identified the IR indices as top-ranking predictors, and ROC analysis showed that TyG-BMI achieved the best discriminative performance for prediabetes.

Conclusions

IR indices are powerful predictors of incident prediabetes, with the most pronounced associations seen in underweight individuals. These findings challenge the notion that low BMI guarantees metabolic health and underscore the need to screen for insulin resistance across the entire BMI spectrum, including underweight persons, to enable more precise diabetes prevention.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13098-026-02127-x.

Introduction

Prediabetes is a critical intermediate metabolic state between normoglycaemia and type 2 diabetes mellitus (T2DM). It is defined by impaired fasting glucose (IFG) and/or impaired glucose tolerance (IGT) [1]. Prediabetes is a robust predictor of future diabetes and cardiovascular disease. Its global prevalence is rising at an alarming rate [2]. Obesity is a well-established risk factor for insulin resistance (IR) and prediabetes [3]. However, recent evidence indicates that underweight individuals (typically defined as a body mass index [BMI] < 18.5 kg/m²) may also develop metabolic dysregulation, including IR and dyslipidaemia. This can predispose them to disturbances in glucose homeostasis [4, 5].

The traditional paradigm of “obesity-driven IR” has been challenged by descriptions of “metabolically unhealthy normal-weight” or “lean diabetes” phenotypes, underscoring that IR can occur independently of excess adiposity [6]. Underweight individuals, often assumed to be metabolically protected, may harbour latent IR attributable to factors such as low muscle mass (sarcopenia), [7] inadequate nutrition, [8] or genetic predisposition [9]. However, there are lack of large-scale epidemiological data examining the association between IR and incident prediabetes in underweight populations, and whether IR indices retain predictive utility in this subgroup remains unclear.

To facilitate early identification of high-risk individuals in both clinical and research settings, several surrogate indices of IR have been developed. The triglyceride–glucose (TyG) index, derived from fasting triglycerides and glucose, correlates strongly with hyperinsulinaemic–euglycaemic clamp results and predicts incident diabetes [10]. The TyG-BMI index further incorporates BMI, [11] whereas the metabolic score for insulin resistance (METS-IR) integrates lipid and glucose parameters with BMI; [12] both have demonstrated utility across diverse populations. The atherogenic index of plasma (AIP), calculated as log (TG/HDL-C), reflects lipoprotein balance and is associated with IR and cardiovascular risk [13]. These indices are simple, cost-effective, and suitable for large cohort analyses, yet their performance across BMI categories, particularly among underweight individuals, has not been systematically evaluated.

In China, rapid socioeconomic transitions have produced a dual burden of malnutrition, with persistent underweight coexisting alongside escalating obesity rates [14]. Although several Chinese cohort studies have investigated IR and diabetes risk, most have focused on overweight/obese or general populations, leaving underweight individuals under-represented [15]. Compared with Western populations, the prevalence of underweight (BMI < 18.5 kg/m²) among Chinese adults is markedly higher [16]. This enrichment, shaped by a cultural preference for leanness and a carbohydrate-dominant diet, provides a cost-effective setting with sufficient statistical power to quantify the link between underweight, insulin resistance and prediabetes [17]. Recent studies also show ethnic heterogeneity in the insulin resistance–glycaemia pathway: south Asian have the highest prevalence of isolated impaired glucose tolerance despite the lowest mean BMI [18]. These findings highlight a greater propensity for “lean but metabolically obese” phenotypes in people of Asian ancestry, making China-specific evidence essential for accurate risk estimation. Clarifying whether IR contributes to the development of prediabetes in this group is essential for targeted screening and early intervention.

We hypothesised that IR indices are positively associated with prediabetes risk even among underweight individuals, and that these associations may exhibit non-linear patterns distinct from those observed in normal-weight and overweight groups. Using a large, longitudinal Chinese health-checkup cohort, this study aimed to: examine the associations of four IR-related indices (TyG, TyG-BMI, AIP, METS-IR) with incident prediabetes across BMI categories, with particular attention to underweight individuals; explore potential non-linear relationships; and evaluate the discriminative performance of these indices using both machine-learning and traditional statistical regression methods.

Methods

Data sources and ethical statement

Data were extracted from an anonymized electronic health-checkup database established by Rich Healthcare Group, covering consecutive medical records of examinees who attended routine health screenings from 2010 to 2016 [19]. The raw data were obtained freely from the DATADRYAD database (www.datadryad.org) provided by Chen et al. (Chen Y, Zhang XP, Yuan J, et al. [2018], Data from Association of body mass index and age with incident diabetes in Chinese adults: a population-based cohort study, Dryad Dataset, 10.5061/dryad.ft8750v) [19]. All anthropometric measurements, laboratory tests, and self-reported lifestyle and medical histories were collected by trained personnel following standardized protocols. The original study protocol was approved by the Institutional Review Board of Rich Healthcare Group, and written informed consent was obtained from every participant at the time of examination. The present investigation constitutes a secondary analysis of fully de-identified data and involved no additional contact with participants or collection of new information; therefore, no further ethical review was required. All procedures were conducted in accordance with the Declaration of Helsinki.

Study population and inclusion/exclusion criteria

This analysis initially included all examinees aged ≥ 20 years who had undergone at least two health check-ups between 2010 and 2016 at 32 Rich Healthcare Group centres located in 11 cities across China (Shanghai, Beijing, Nanjing, Suzhou, Shenzhen, Changzhou, Chengdu, Guangzhou, Hefei, Wuhan and Nantong). Figure 1 summarises the step-wise exclusions. These exclusions left 211,833 participants for the initial cohort. Next, we applied the following study-specific exclusion criteria (total n = 112,817): 1). Baseline FPG ≥ 5.6 mmol/L (n = 26,247); 2). FPG > 6.9 mmol/L during follow-up before prediabetes could be recorded (n = 1,013); [20] 3). Incident diabetes diagnosed without an intermediate prediabetes record (n = 305); 4). Missing fasting triglyceride value (n = 5,094); 5). Missing HDL-C value (n = 78,699); 6). Outlier values (beyond ± 3 SD) for any of the four IR indices—TyG, TyG-BMI, AIP or METS-IR (n = 1,459). After all exclusions, 99,016 participants were included in the final analysis.

Fig. 1.

Flowchart of the study participants in the Chinese population

Participants were classified into three BMI groups according to the WHO criteria: underweight (< 18.5 kg/m²), normal weight (18.5–24.9 kg/m²), and overweight/obese (≥ 25.0 kg/m²) [21]. 

Exposure variables: IR-related indices

Four IR-related indices were calculated from baseline laboratory and anthropometric data:

1. TyG index: ln [TG × FPG/2] [10].

2. TyG-BMI: TyG index × BMI [11].

3. AIP: log₁₀ [TG/HDL-C] [13].

4. METS-IR: ln [2 × FPG + TG] × BMI/ln [HDL-C] [12].

All lipid and glucose values were converted to mg/dL for calculation consistency using standard conversion factors (glucose: 1 mmol/L = 18.0 mg/dL; TC, HDL-C, and LDL-C: 1 mmol/L = 38.67 mg/dL; TG: 1 mmol/L = 88.57 mg/dL) [22]. These indices were treated as continuous variables in primary analyses and as quartiles in subgroup and sensitivity analyses.

Outcome definition

The primary outcome was incident prediabetes, defined as FPG between 5.6 mmol/L and 6.9 mmol/L at any follow-up visit, in accordance with the American Diabetes Association criteria [20]. Participants with normal FPG (< 5.6 mmol/L) at baseline who later developed FPG ≥ 5.6 mmol/L were considered incident cases.

Analytical methods

Continuous variables are presented as mean ± standard deviation or median (interquartile range) based on distribution, and categorical variables as frequencies (percentages). Group differences were assessed using one-way ANOVA or Kruskal–Wallis tests for continuous variables and Chi-square tests for categorical variables.

To handle the limited amount of missing data, we performed multiple imputation by chained equations under the assumption that data were missing at random. All analytic variables were included in the imputation model. Five imputed datasets were created, and the analyses were conducted on every single imputed dataset. Subsequently, Rubin’s rule was applied to combine the main results of five imputed datasets [23, 24]. Kaplan–Meier curves stratified by IR-index tertiles were plotted and compared with the log-rank test to visualise the cumulative incidence of prediabetes across exposure groups. Cox proportional hazards regression models were used to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for the associations between IR indices and incident prediabetes. Multicollinearity was screened with the variance inflation factor (VIF); any variable whose VIF ≥ 5 was dropped to keep the model stable [25]. In practice, almost all covariates satisfied VIF < 5, but total cholesterol (TC) and LDL-C exceeded the threshold (VIF = 8.2 and 6.8, respectively). Consequently, TC and LDL-C were removed from the final adjusted models (Table S1). In our study, three models were constructed: Model I (unadjusted), Model II (adjusted for age and sex), and Model III (further adjusted for systolic and diastolic blood pressure, ALT, BUN, Scr, smoking status, drinking status, and family history of diabetes).

After adjusting for other covariates in Model III, the potential non-linear relationships were explored using restricted cubic splines with four knots placed at the 5th, 35th, 65th, and 95th percentiles [26]. Inflection points were identified where the association slope changed significantly. Likelihood ratio tests were used to compare linear and non-linear models.

Sensitivity analyses included: (1) excluding participants with < 3 years of follow-up; (2) excluding those aged > 60 years; (3) excluding those with baseline systolic BP ≥ 140 mmHg; and (4) using raw, untransformed IR indices. The discriminative ability of IR indices was evaluated using the area under the receiver operating characteristic curve (AUC).

We employed the Boruta algorithm, a wrapper-based feature selection method using random forest classification, to evaluate the predictive importance of insulin resistance indices for prediabetes risk. This algorithm creates randomized “shadow” variables and compares original feature importance against these shadow features across iterations, retaining only variables significantly exceeding random noise. With 500 iterations, variables were ranked by importance Z-scores, where higher values indicate stronger outcome associations with a lower probability of chance findings. This non-parametric approach is robust to multicollinearity and non-linear relationships, complementing traditional regression models.

Results are reported in accordance with the STROBE statement [28]. Analyses were performed with Empower Stats (X&Y Solutions, Boston, MA), R (R Foundation), and Fengrui 2.3; two-sided P < 0.05 was taken as statistically significant.

Results

Participant characteristics

The final analytical cohort comprised 99,016 participants, with a mean age of 42.9 years and 51.5% being male. According to the WHO BMI criteria, 5.8% (n = 5,709) were underweight, 68.3% (n = 67,665) were normal weight, and 25.9% (n = 25,642) were overweight/obese. The baseline characteristics of the total cohort and across BMI categories are detailed in Table 1. Underweight individuals were significantly younger, had a higher proportion of females, and exhibited more favourable metabolic profiles, including lower blood pressure, lipid levels, and insulin resistance indices (all P < 0.001), compared to their normal-weight and overweight counterparts. Despite their ostensibly healthier metabolic profile, 5.0% of underweight participants developed prediabetes over a median follow-up of 3.12 years, compared with 10.3% of those with normal weight and 18.7% of overweight individuals (P < 0.001). The distribution of participant characteristics across quartiles of each IR index (Tables S2-S5) demonstrated strong gradients in age, BMI, blood pressure, and lipid profiles, with higher quartiles associated with progressively less favourable metabolic parameters (all P for trend < 0.001). As shown in Supplementary Figure S1, prediabetes incidence is presented as percentages by sex and age group (< 30, 30–39, 40–49, 50–59, 60–69, ≥ 70 years). Prevalence increases with age in both sexes, and men exhibit higher rates than women in every stratum.

Table 1.

Participant characteristics across BMI States for the Chinese population

Characteristics	Total (n = 99,016)	Underweight (n = 5709)	Normal weight (n = 67,665)	Overweight (n = 25,642)	P-value
Age, years	42.90 ± 12.46	36.55 ± 10.93	42.38 ± 12.21	45.66 ± 12.75	< 0.001
Gender, n (%)					< 0.001
Male	51,028 (51.54%)	1434 (25.12%)	31,064 (45.91%)	18,530 (72.26%)
Female	47,988 (48.46%)	4275 (74.88%)	36,601 (54.09%)	7112 (27.74%)
BMI, kg/m2	23.01 ± 3.12	17.61 ± 0.69	21.93 ± 1.75	27.08 ± 1.80	< 0.001
SBP, mmHg	117.94 ± 16.03	109.24 ± 13.42	115.89 ± 15.24	125.28 ± 16.02	< 0.001
DBP, mmHg	73.65 ± 10.72	68.99 ± 9.07	72.23 ± 10.11	78.44 ± 11.09	< 0.001
FPG, mmol/L	4.79 ± 0.47	4.66 ± 0.48	4.77 ± 0.47	4.86 ± 0.46	< 0.001
TC, mmol/L	4.74 ± 0.88	4.47 ± 0.82	4.70 ± 0.87	4.92 ± 0.89	< 0.001
TG, mmol/L	1.26 ± 0.77	0.81 ± 0.35	1.14 ± 0.67	1.67 ± 0.90	< 0.001
HDL-C, mmol/L	1.39 ± 0.30	1.55 ± 0.32	1.41 ± 0.30	1.28 ± 0.27	< 0.001
LDL-C, mmol/L	2.74 ± 0.67	2.51 ± 0.61	2.71 ± 0.66	2.88 ± 0.68	< 0.001
ALT, U/L	17.50 (12.70–26.10)	12.80 (10.00–16.30.00.30)	16.00 (12.00–23.00)	25.00 (17.70–37.40)	< 0.001
BUN, mmol/L	4.63 ± 1.16	4.34 ± 1.12	4.58 ± 1.15	4.83 ± 1.16	< 0.001
Scr, µmol/L	69.75 ± 15.72	62.46 ± 12.93	68.35 ± 15.33	75.06 ± 15.85	< 0.001
TyG index	8.32 ± 0.56	7.93 ± 0.40	8.23 ± 0.53	8.64 ± 0.52	< 0.001
TyG-BMI	192.27 ± 33.84	139.61 ± 9.04	180.83 ± 21.34	234.18 ± 21.94	< 0.001
AIP	−0.10 ± 0.28	−0.31 ± 0.19	−0.14 ± 0.26	0.07 ± 0.26	< 0.001
METS-IR	2.22 ± 0.17	2.05 ± 0.12	2.19 ± 0.15	2.34 ± 0.15	< 0.001
Smoking Status, n (%)					< 0.001
Current smoker	15,547 (15.70%)	432 (7.57%)	9252 (13.67%)	5863 (22.86%)
Ever smoker	3195 (3.23%)	57 (1.00%)	1837 (2.71%)	1301 (5.07%)
Never smoker	80,274 (81.07%)	5220 (91.43%)	56,576 (83.61%)	18,478 (72.06%)
Drinking Status, n (%)					< 0.001
Current drinker	1872 (1.89%)	42 (0.74%)	1008 (1.49%)	822 (3.21%)
Ever drinker	13,727 (13.86%)	376 (6.59%)	8192 (12.11%)	5159 (20.12%)
Never drinker	83,417 (84.25%)	5291 (92.68%)	58,465 (86.40%)	19,661 (76.67%)
Family History of Diabetes, n (%)					0.071
No	96,853 (97.82%)	5604 (98.16%)	66,145 (97.75%)	25,104 (97.90%)
Yes	2163 (2.18%)	105 (1.84%)	1520 (2.25%)	538 (2.10%)
incident prediabetes					< 0.001
No	86,984 (87.85%)	5422 (94.97%)	60,720 (89.74%)	20,842 (81.28%)
Yes	12,032 (12.15%)	287 (5.03%)	6945 (10.26%)	4800 (18.72%)
Follow-up (years)	3.12 ± 0.95	3.09 ± 0.95	3.12 ± 0.95	3.12 ± 0.95	0.157

Values are n (%), means or medians (quartiles)

BMI, body-mass index; FPG, fasting plasma glucose; SBP, systolic blood pressure; DBP, diastolic blood pressure; TC, total cholesterol; TG, triglycerides; HDL-C, high-density lipoprotein cholesterol; LDL-C, low-density lipoprotein cholesterol; ALT, alanine aminotransferase; BUN, blood urea nitrogen; Scr, serum creatinine; TyG, triglyceride–glucose index; TyG-BMI, triglyceride–glucose–body-mass index; AIP, atherogenic index of plasma; METS-IR, metabolic score for insulin resistance

Associations between IR indices and incident prediabetes

During a median follow-up of 3.12 years, 12,152 (12.3%) incident prediabetes cases were identified. Kaplan–Meier curves showed a clear stepwise decline in prediabetes-free survival across increasing quartiles of every IR index (log-rank p < 0.001; Supplementary Figure S2). Univariable Cox regression showed that all four IR indices were significant predictors of incident prediabetes (Table S6). Multivariable Cox regression showed that these associations remained essentially unchanged after progressive multivariable adjustment. In the fully adjusted model (Model III), all four IR indices were significantly associated with an increased risk of incident prediabetes in the total population (Table 2). The TyG index (HR per unit increase: 1.73, 95% CI: 1.67–1.79), TyG-BMI (HR per 10-unit increase: 1.10, 95% CI: 1.09–1.10), AIP (HR per unit increase: 1.65, 95% CI: 1.54–1.77), and METS-IR (HR per unit increase: 2.01, 95% CI: 1.80–2.24) were all independent risk factors (all P < 0.0001).

Table 2.

Associations between insulin resistance indices and the incident prediabetes in different model

	Exposure	Model I (HR,95%CI) P	Model II (HR,95%CI) P	Model III (HR,95%CI) P
Group	TyG index	2.22 (2.15, 2.28) <0.0001	1.86 (1.80, 1.92) <0.0001	1.73 (1.67, 1.79) <0.0001
All	TyG-BMI (×10)	1.14 (1.14, 1.15) <0.0001	1.11 (1.11, 1.12) <0.0001	1.10 (1.09, 1.10) <0.0001
(N=99016)	AIP	2.84 (2.67, 3.01) <0.0001	1.93 (1.81, 2.07) <0.0001	1.65 (1.54, 1.77) <0.0001
	METS-IR	4.19 (3.79, 4.62) <0.0001	2.51 (2.25, 2.79) <0.0001	2.01 (1.80, 2.24) <0.0001
	TyG index	2.91 (2.23, 3.80) <0.0001	2.14 (1.62, 2.82) <0.0001	1.85 (1.40, 2.45) <0.0001
Underweight	TyG-BMI (×10)	1.54 (1.36, 1.75) <0.0001	1.37 (1.20, 1.56) <0.0001	1.28 (1.12, 1.46) 0.0002
(N=5709)	AIP	3.84 (2.20, 6.70) <0.0001	2.22 (1.26, 3.91) 0.0059	1.74 (0.97, 3.11) 0.023
	METS-IR	8.36 (3.53, 19.80) <0.0001	5.63 (2.32, 13.61) 0.0001	4.56 (1.82, 11.38) 0.0012
	TyG index	2.22 (2.13, 2.31) <0.0001	1.81 (1.73, 1.90) <0.0001	1.73 (1.66, 1.82) <0.0001
Normal weight	TyG-BMI (×10)	1.22 (1.21, 1.24) <0.0001	1.16 (1.15, 1.18) <0.0001	1.15 (1.13, 1.16) <0.0001
(N=67665)	AIP	2.70 (2.48, 2.94) <0.0001	1.77 (1.62, 1.94) <0.0001	1.62 (1.48, 1.78) <0.0001
	METS-IR	3.81 (3.30, 4.40) <0.0001	2.28 (1.96, 2.66) <0.0001	2.06 (1.77, 2.41) <0.0001
	TyG index	1.61 (1.52, 1.70) <0.0001	1.56 (1.48, 1.65) <0.0001	1.50 (1.42, 1.59) <0.0001
Overweight	TyG-BMI (×10)	1.10 (1.09, 1.12) <0.0001	1.10 (1.09, 1.12) <0.0001	1.09 (1.07, 1.10) <0.0001
(N=25642)	AIP	1.27 (1.14, 1.41) <0.0001	1.24 (1.11, 1.38) 0.0002	1.13 (1.01, 1.26) 0.0404
	METS-IR	0.82 (0.68, 0.98) 0.0270	0.86 (0.71, 1.03) 0.1025	0.78 (0.65, 0.95) 0.0112

Model I: we did not adjust other covariates

Model II: we adjust age, gender

Model III: we adjust variables in Adjust II + SBP, DBP, ALT, BUN, Scr, smoking status, Drinking status, Family history of diabetes. HR Hazard Ratio, CI Confidence Interval

Stratified analyses by BMI category showed significant associations in all subgroups, with the largest effect sizes seen among underweight participants (Table 2). After full adjustment, TyG (HR 1.85, 95% CI 1.40–2.45), TyG-BMI (per 10-unit increase: HR 1.28, 95% CI 1.12–1.46), METS-IR (HR 4.56, 95% CI 1.82–11.38) and AIP (HR 1.74, 95% CI 0.97–3.11, P = 0.023) remained significantly associated with higher prediabetes risk in the underweight group.

The non-linearity addressed by cox proportional hazards regression model with cubic spline functions

Restricted cubic spline analyses revealed significant non-linear dose-response relationships between each IR index and the risk of incident prediabetes (P for non-linearity < 0.001 for all) (Fig. 2). The inflection points (K) and piecewise hazard ratios are detailed in Table 3. For the TyG index, the risk increased sharply below the inflection point (K = 8.73, HR for values < K: 2.00, 95% CI: 1.89–2.12) and plateaued thereafter (HR for values ≥ K: 1.38, 95% CI: 1.28–1.50). A similar J-shaped pattern was observed for TyG-BMI, with a steeper slope below its inflection point (K = 20.02). Conversely, the associations for AIP and METS-IR were characterised by a steep initial risk increase at lower values, followed by a much attenuated or non-significant increase at higher values.The risk trend observed below the METS-IR inflection point suggests a potential scenario: abnormally low METS-IR levels may also indicate high risk. However, this does not imply that ‘lowering METS-IR is harmful’, but rather may signal the existence of another high-risk population.

Fig. 2.

Non-linear relationships between insulin resistance indices and prediabetes risk. Restricted cubic spline plots illustrate the dose–response association of each insulin resistance index with incident prediabetes. The solid red line represents the estimated hazard ratio (HR), and the red dashed shading indicates the 95% confidence interval. A: TyG index; B: TyG-BMI; C: AIP; D: METS-IR. P-values for non-linearity were < 0.001 for all indices

Table 3.

Nonlinear associations between insulin resistance indices and prediabetes risk

IR indices	TyG index	TyG-BMI (×10)	TG/HDL	METS-IR
Model I	1.73 (1.67, 1.79) < 0.0001	1.10 (1.09, 1.10) < 0.0001	1.65 (1.54, 1.77) < 0.0001	2.01 (1.80, 2.24) < 0.0001
Model II
Inflection points	8.73	20.02	0.04	2.16
< K	2.00 (1.89, 2.12) < 0.0001	1.15 (1.14, 1.17) < 0.0001	2.44 (2.15, 2.76) < 0.0001	59.19 (38.86, 90.15) < 0.0001
≥K	1.38 (1.28, 1.50) < 0.0001	1.06 (1.05, 1.07) < 0.0001	1.04 (0.90, 1.20) 0.5956	0.87 (0.75, 1.01) 0.0637
P for log-likelihood ratio test	< 0.001		< 0.001	< 0.0001

Data were presented as HR (95% CI) P value; Model I, linear analysis; Model II, non-linear analysis. Adjusted for age, gender, SBP, DBP, ALT, BUN, Scr, smoking status, Drinking status, Family history of diabetes. K indicates inflection point. P for log-likelihood ratio test < 0.05 indicates that model II is significantly different from Model I

Feature importance and sensitivity analyses

The Boruta machine learning algorithm identified the TyG index, TyG-BMI, and age as the top three most important predictors for incident prediabetes in the overall model, underscoring the prominence of IR-related metrics (Fig. 3).

Fig. 3.

The Boruta algorithm ranks the importance of potential risk factors for censor of prediabetes. Ridge plot (Z-score) distributions graphically depict the dispersion characteristics of normalized values during model computation cycles

Results from multiple sensitivity analyses confirmed the robustness of the primary findings. After excluding participants with less than 3 years of follow-up, those older than 60 years, or those with baseline hypertension (SBP ≥ 140 mmHg), the positive associations between all four IR indices and prediabetes risk remained consistent and statistically significant (Table 4). Supplementary results. In addition, analyses performed on the raw, unimputed dataset yielded results consistent in direction and magnitude with the primary analysis (Table S8).

Table 4.

Associations between insulin resistance indices and the incident prediabetes in sensitive analyses.

Exposure	Model I (HR,95%CI) P	Model II (HR,95%CI) P	Model III (HR,95%CI) P
TyG index	1.71 (1.63, 1.79) < 0.0001	1.80 (1.73, 1.87) < 0.0001	1.79 (1.72, 1.86) < 0.0001
TyG-BMI (×10)	1.09 (1.08, 1.10) < 0.0001	1.10 (1.10, 1.11) < 0.0001	1.10 (1.10, 1.11) < 0.0001
AIP	1.74 (1.59, 1.91) < 0.0001	1.71 (1.58, 1.85) < 0.0001	1.73 (1.60, 1.87) < 0.0001
METS-IR	2.13 (1.85, 2.46) < 0.0001	2.10 (1.85, 2.38) < 0.0001	2.16 (1.91, 2.44) < 0.0001

Model I: participants with < 3 years of follow-up were excluded; covariates included age, gender, SBP, DBP, ALT, BUN, Scr, smoking status, alcohol consumption and family history of diabetes.

Model II: participants aged > 60 years were excluded; the same covariates as in Model I were adjusted.

Model III: participants with baseline SBP ≥ 140 mmHg were excluded; identical covariates were adjusted. HR Hazard Ratio, CI Confidence Interval.

Discriminative ability of IR indices for incident prediabetes across BMI categories

Table S7 summarises the discriminative performance of the four IR-related indices. Overall, TyG-BMI achieved the highest AUC (0.651; 95% CI 0.646–0.656), followed by TyG (0.634; 95% CI 0.628–0.639), METS-IR (0.607; 95% CI 0.602–0.612) and AIP (0.605; 95% CI 0.600–0.610). The optimal cut-off identified by maximising the Youden index was 19.3 × 10 for TyG-BMI, 8.38 for TyG, 2.20 for METS-IR and − 0.09 for AIP; corresponding sensitivities ranged from 60% to 66% and specificities from 50% to 58%.

Discussions

In this large-scale, longitudinal analysis of a Chinese health examination cohort encompassing nearly 100,000 individuals, we provide robust evidence that surrogate indices of IR are significantly associated with an elevated risk of incident prediabetes, and this association holds substantial clinical relevance even among underweight adults (BMI < 18.5 kg/m²). Our findings directly challenge the conventional notion that low BMI confers universal metabolic protection. Specifically, in fully adjusted models, the TyG index, TyG-BMI, and METS-IR exhibited significant, and in some cases stronger, hazard ratios for prediabetes in the underweight group compared to their normal-weight and overweight counterparts. Notably, we identified significant non-linear, threshold-dependent relationships for all indices, and machine learning feature selection ranked IR-related metrics among the most important predictors. These results collectively underscore that metabolic dysregulation, as captured by these easily obtainable indices, is a potent risk factor for deteriorating glucose homeostasis across the entire BMI spectrum, with unique implications for the underweight population.

The cumulative incidence of prediabetes in our overall cohort was 12.15% over a median follow-up of 3.12 years. This incidence appears moderate compared to some other Chinese cohort studies. For instance, a nationwide study reported a baseline prevalence of prediabetes at 35.7% among adults, with a substantial annual conversion rate from normoglycemia [29]. The relatively lower incidence observed in our study may be attributed to several factors. Firstly, our cohort was derived from a population undergoing voluntary health check-ups, which may represent individuals with higher health awareness and potentially more favorable baseline metabolic profiles—a common feature of “health check-up” cohorts that can lead to the “healthy screenee” effect. Secondly, our follow-up duration was shorter than some long-term studies that track transitions over decades. Thirdly, we employed a strict definition of incident prediabetes based on a single FPG measurement at follow-up visits, which, while standardized, might be less sensitive than incorporating OGTT or HbA1c, leading to potential under-ascertainment. Importantly, the incidence among underweight individuals (5.03%) was less than half that of the overweight group (18.72%), visually reinforcing the overall protective effect of low adiposity. However, our core finding lies not in the absolute rate but in the relative risk associated with IR within this seemingly low-risk group, revealing a hidden vulnerability.

The most salient and novel finding of our study is the persistent and strong association between IR indices and prediabetes risk among underweight individuals. This contrasts with much of the existing literature that primarily focuses on or finds the strongest associations in overweight/obese populations [30]. For example, while the TyG index is a well-validated predictor of diabetes in general and obese cohorts, [10] our stratified analysis demonstrates its predictive power remains significant, with an HR of 1.85, in the underweight stratum. The effect size for METS-IR in underweight individuals (HR = 4.56) was particularly striking, exceeding that observed in the normal-weight group. This suggests that the pathophysiological impact of IR-related dysmetabolism may be amplified in the context of low body mass, possibly due to the compounding effect of limited metabolic reserve or concurrent conditions like sarcopenia [7]. These findings align with and extend the emerging concept of “metabolically unhealthy non-obese” phenotypesby explicitly quantifying the risk gradient within the underweight category itself [6]. Our data argue against a blanket assumption of low risk for all underweight individuals and advocate for a more nuanced, metabolism-focused risk stratification within this BMI category.

The biological plausibility of IR driving prediabetes in underweight individuals can be explained by several interlinked mechanisms that operate independently of excess adiposity. First, low muscle mass (sarcopenia), common in underweight individuals, is a major site of glucose disposal; its reduction directly impairs insulin-mediated glucose uptake, leading to peripheral IR [31]. Second, even in the absence of generalized obesity, ectopic fat deposition—particularly in the liver (hepatic steatosis) and skeletal muscle—can occur and is strongly associated with IR [32]. This “normal-weight fatty liver” phenomenon and muscle lipid infiltration can disrupt insulin signaling. Third, inadequate nutrition or specific dietary patterns associated with being underweight (e.g., low protein intake, micronutrient deficiencies) may impair pancreatic beta-cell function and insulin secretion, creating a mismatch with prevailing IR [8]. Finally, genetic predispositions that link low BMI with adverse metabolic traits (the so-called “lean but metabolically unhealthy” genotype) may play a role [9]. The non-linear relationships we observed, especially the steep risk increase at lower index values, may reflect threshold effects in these pathways, where a minimal degree of dysmetabolism triggers disproportionate metabolic dysregulation in a constitutionally vulnerable, low-reserve host.

Our findings carry significant implications for diabetes prevention strategies, which have traditionally focused on overweight and obese populations. They advocate for a paradigm shift towards a “metabolism-first” rather than a purely “BMI-first” approach in risk assessment. The practical value of using simple indices like TyG, TyG-BMI, and METS-IR is underscored. These indices, calculated from routine laboratory measurements (lipid profile, fasting glucose) and BMI, offer a cost-effective and accessible tool for clinicians to identify underweight individuals at high risk for prediabetes who would otherwise be overlooked. We recommend that health check-up protocols and public health guidelines for diabetes screening in East Asian populations consider incorporating these IR indices, especially for individuals with BMI < 18.5 kg/m², to enable earlier detection and intervention. For an underweight individual with an elevated TyG index, clinical attention should extend beyond weight gain advice to include a thorough assessment of body composition (e.g., muscle mass), liver fat, dietary quality, and structured lifestyle interventions focusing on resistance training to build muscle and improve insulin sensitivity, alongside nutritional counseling [33]. This targeted approach could improve the efficiency of prediabetes screening programs and personalize prevention efforts.

This study possesses several notable strengths. First, the large sample size (N = 99,016), with a substantial number of underweight participants (n = 5,709), provided robust statistical power to conduct stratified analyses and detect associations within this specific subgroup—an analysis often underpowered in smaller studies. Second, the retrospective cohort design with standardized annual follow-ups strengthens the temporal sequence between exposure (IR indices) and outcome (incident prediabetes). Third, we employed and compared four different IR indices, offering a comprehensive view of the association and demonstrating consistency across related but distinct metrics. Fourth, the use of advanced analytical methods, including restricted cubic splines to model non-linearity and the Boruta machine learning algorithm for feature importance, provided deeper insights beyond traditional linear models and validated the prominence of IR factors. Fifth, extensive sensitivity analyses confirmed the robustness of our primary findings against various assumptions and potential biases.

Several limitations should be acknowledged. First, prediabetes was defined solely by FPG, without confirmation by OGTT or HbA1c, which may have led to some misclassification. However, FPG is the most commonly used screening test in large epidemiological studies and clinical practice in China. Second, while we adjusted for a wide range of potential confounders, residual confounding from unmeasured factors (e.g., detailed dietary data, physical activity levels, body composition metrics like waist circumference or fat mass) cannot be ruled out. The absence of insulin measurement precluded the use of HOMA-IR as a comparator. Third, our cohort consisted of individuals who sought voluntary health examinations, which may limit the generalizability of the absolute incidence rates to the general Chinese population, though the internal validity of the exposure-outcome relationship is likely preserved. Fourth, the observational nature of the study precludes definitive causal inferences. Fifth, a primary limitation of this study lies in the fact that participants were drawn from a population undergoing voluntary health screening, potentially introducing a ‘healthy screener effect’. Sixthly, the data for this study were derived from a Chinese population cohort; therefore, caution is required when applying the findings to other ethnicities or populations. Finally, the follow-up period, while sufficient to detect a sizable number of incidents, was moderate; a longer follow-up is needed to examine progression to frank diabetes.

Future research should validate these findings in other ethnic populations and cohorts with more detailed phenotyping, including direct measures of body composition (DXA, CT), ectopic fat, and insulin secretion. Mechanistic studies exploring the specific pathways linking IR to prediabetes in underweight individuals are warranted. Intervention trials are needed to determine whether modifying these IR indices (e.g., through targeted exercise or nutritional supplementation) in at-risk underweight individuals can effectively reduce the progression to prediabetes and diabetes.

Conclusion

This large retrospective cohort analysis demonstrates that insulin resistance (IR) indices (TyG, TyG-BMI, METS-IR) are significant predictors of incident prediabetes, with associations being particularly pronounced among underweight individuals. These findings challenge the conventional notion that low BMI equates to low metabolic risk and reveal a metabolically susceptible subgroup within the underweight population. Our study supports the incorporation of these simple IR indices into clinical assessment to help identify high-risk individuals overlooked by traditional BMI-based criteria, offering a new strategy for early diabetes prevention across the BMI spectrum.

Abbreviations

IR

Insulin-resistance

TyG

Triglyceride-glucose

AIP

Atherogenic index of plasma

METS-IR

Metabolic score for IR

T2DM

Type 2 diabetes mellitus

IFG

Impaired fasting glucose

IGT

Impaired glucose tolerance

BMI

Body mass index

HRs

Hazard ratios

CIs

Confidence intervals

VIF

Variance inflation factor

TC

Total cholesterol

AUC

The area under the receiver operating characteristic curve

OGTT

/oral glucose tolerance tests

HR

Hazard ratio

Author contributions

Xin Li supervised the study, Ling Chen provided the data access, Haoren Xu and Qiaofeng Wu designed the study, Zhenhua Huang collected the data, performed the analysis and drafted the manuscript. All authors edited the manuscript and approved the submission of this final version.

Funding

The project was supported by the National Key R&D Program of China Intergovernmental Key Projects (No.2023YFE0114300), the Joint Funds of the Natural Science Foundation of China (No.U24A20652), National Science Foundation of China (No. 82272246), Basic and Applied Basic Research Foundation of Guangdong Province (No.2024A1515012697), High- level Hospital Construction Project of Guangdong Provincial People’s Hospital (No.DFJHBF202104) and the Medical Research Foundation of Guangdong Province (No: A2024429).

Data availability

The dataset titled “Association of body mass index and age with incident diabetes in Chinese adults: a population-based cohort study” [19] is publicly available through the Dryad Digital Repository at 10.5061/dryad.ft8750v.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.