Metabolic feature-based clustering for subtype identification and longitudinal outcome predictions in the Chinese prediabetic population

Abstract

BACKGROUND

In China, the prevalence of prediabetes is alarmingly high, affecting approximately 35.7% of adults. Although lifestyle interventions can reduce diabetes onset, the manifestations of prediabetes vary across individuals with age, body composition, insulin resistance, and beta-cell function. Therefore, current diagnostic approaches based solely on glucose thresholds are not sufficient, necessitating improved risk stratification through data-driven, unsupervised clustering methods.

AIM

To investigate the metabolic heterogeneity of prediabetes in Chinese individuals and its association with lifestyle intervention outcomes.

METHODS

A prospective, multicenter cohort study was conducted in China with 2527 adults aged 18-70 years, at high risk for diabetes. Centers were assigned to either enhanced lifestyle management or standard health education. Enhanced management included individualized dietary energy prescriptions, wearable activity trackers, a mobile co-management app, and structured follow-up. Unsupervised K-means clustering identified four metabolic subtypes. Longitudinal outcomes, including incident diabetes and reversion to normoglycemia, were analyzed using Cox models adjusted for intervention and demographic covariates.

RESULTS

Four distinct prediabetes subtypes were identified: Mild obesity-related dysmetabolism (MOD, n = 177), mild age-related dysmetabolism (MARD, n = 190), severe insulin resistance (n = 95), and severe insulin deficiency (n = 159). Of 621 participants, 367 (59.1%) contributed longitudinal data; although attrition differed significantly across subtypes (P < 0.001), inverse probability of censoring weighting confirmed the robustness of all estimates. After a median follow-up of 735 days, MOD had a lower risk of diabetes progression than MARD [adjusted hazard ratio (aHR) = 0.52, P = 0.028] and a greater likelihood of reversion to normoglycemia (aHR = 2.08, P = 0.049). Adjusted for subtype and sex, enhanced lifestyle management reduced diabetes progression risk (aHR = 0.52, 95% confidence interval: 0.30-0.89; P = 0.017), but not reversion to normoglycemia (P = 0.159).

CONCLUSION

Data-driven prediabetes subtyping improved risk stratification. The MOD subtype showed more favorable metabolic trajectories than the MARD subtype. Enhanced lifestyle management was associated with reduced diabetes progression risk.

Key Words: Prediabetes; Metabolic subtypes; Longitudinal outcomes; Lifestyle intervention; Risk stratification

Core Tip: Using unsupervised k-means clustering of core metabolic features, we identified four clinically meaningful prediabetes subtypes: Mild obesity-related dysmetabolism, mild age-related dysmetabolism, severe insulin resistance, and severe insulin deficiency. Among them, mild obesity-related dysmetabolism showed a clear metabolic advantage over mild age-related dysmetabolism, with a lower risk of progression to diabetes and a greater chance of reverting to normal glucose tolerance. Although enhanced lifestyle management was associated with reduced progression risk, metabolic subtype membership, not intervention exposure, was the main driver of outcome heterogeneity, highlighting its value for baseline risk stratification and precision prevention.

INTRODUCTION

China is experiencing a rapid increase in the development of diabetes. Large-scale epidemiological surveys reported that in 2013, the prevalence of prediabetes among Chinese adults reached approximately 35.7%[1-7]. Prediabetes represents a critical intermediate stage preceding type 2 diabetes mellitus, characterized by dysregulated glucose metabolism that does not yet meet the diagnostic criteria for diabetes. Accumulating evidence indicates that timely and intensive lifestyle interventions can substantially reduce the progression from prediabetes to overt diabetes[8-11]. For example, the Diabetes Prevention Program in the United States demonstrated that intensive lifestyle modification reduced the incidence of diabetes by nearly 58%[10]. Similarly, the Da Qing Diabetes Prevention Study in China revealed that lifestyle intervention not only delays the onset of diabetes in individuals with impaired glucose tolerance (IGT) but also decreases the long-term risks of complications and all-cause mortality, marking a milestone in diabetes prevention[11]. However, under the “Healthy China 2030” initiative, scaling effective interventions to China’s vast high-risk population remains challenging.

Although lifestyle intervention is widely regarded as the cornerstone of diabetes prevention and has been shown in landmark trials to substantially reduce the incidence of diabetes, prediabetes is not a homogeneous condition. Individuals with prediabetes differ markedly in terms of age, adiposity, insulin resistance, and β-cell function, resulting in pronounced heterogeneity in clinical trajectories, including progression to diabetes, reversion to normoglycemia, or persistence in a prediabetic state[12-14]. Conventional diagnostic frameworks that rely primarily on single-threshold measures such as fasting or postload glucose are therefore limited in their ability to capture the diverse underlying pathophysiological mechanisms. Treating “prediabetes” as a uniformly high-risk group may lead to inefficient allocation of preventive resources, imprecise risk communication, and limited explanatory power for the observed diversity of outcomes.

In recent years, data-driven subphenotyping has emerged as a promising approach to elucidate this heterogeneity. Using clustering analyses, investigators have identified distinct subtypes among patients with diabetes, such as severe insulin-deficient and severe insulin-resistant phenotypes, with differing rates of progression and complication risks[15,16]. Similar strategies have been extended to prediabetes. A European cohort study proposed six prediabetes subtypes and reported substantial differences in the risks of incident diabetes and complications across clusters[17,18]. In addition, analyses of Diabetes Prevention Program (DPP) trial data, revealed two prediabetes subgroups characterized by different degrees of insulin resistance, with divergent diabetes risk and differential responses to preventive interventions, including intensive lifestyle modification and metformin[19]. More recently, a large prospective study in China (the 4C study), also revealed multiple metabolic subtypes among individuals with prediabetes, demonstrating significant differences between subtypes in terms of progression to diabetes and complication risk[20]. Collectively, these findings suggest that the application of unsupervised clustering to identify latent subtypes of prediabetes may enable refined risk stratification and the optimization of individualized preventive strategies.

However, in Chinese populations with prediabetes, local evidence systematically linking mechanistic subphenotypes to longitudinal outcomes following lifestyle intervention remains limited. Critically, there is a gap in understanding how these subtypes respond within the context of large-scale, real-world public health programs. Unlike strictly controlled randomized controlled trials, the present study was embedded in the pragmatic implementation of a national DPP, where intervention resources were allocated at the center level rather than through individual randomization. This design reflects the real-world accessibility and acceptance to lifestyle management support among individuals with prediabetes under routine healthcare and community management conditions. Such a pragmatic, quasi-experimental approach captures the inherent variability in intervention delivery and uptake found in “daily practice”, thus generating evidence that is more directly translatable to public health decision-making. We used a data-driven, unsupervised clustering approach to classify individuals with prediabetes into clinically interpretable metabolic subgroups and then compared longitudinal outcomes across subgroups, including risks of progression or reversion and their associations with lifestyle intervention under real-world program implementation conditions, with the aim of improving the risk-stratified management of patients with prediabetes and precision prevention strategies.

MATERIALS AND METHODS

Study design and participants

This study was nested within the “Intervention Study in Populations at High Risk for Diabetes”, a national program that deployed lifestyle management resources across multiple participating centers. Participants were drawn from the project database supported by the National Key Research and Development Program of China, No. 2018YFC1313902, community-dwelling adults aged 18-70 years were recruited from 14 provinces/municipalities and underwent standardized baseline assessments, including an oral glucose tolerance test (OGTT), glycated hemoglobin (HbA1c) assessment, lipid profiling, liver and renal function tests. A total of 2527 participants were assessed at baseline. Individuals with diabetes at baseline were excluded (fasting blood glucose concentration of ≥ 7.0 mmol/L, 2-hour OGTT glucose concentration of ≥ 11.1 mmol/L, or HbA1c ≥ 6.5%). Prediabetes is defined as a fasting blood glucose concentration of 6.1-7.0 mmol/L and/or a 2-h OGTT glucose concentration of 7.8-11.1 mmol/L according to the World Health Organisation (WHO)/International Expert Committee (IEC) 2006 criteria, which aligns with the Chinese Diabetes Society diagnostic framework and the clinical context of the parent program. Although HbA1c was not used as an inclusion criterion, it was retained as one of the five clustering variables to capture an additional dimension of long-term glycemic exposure relevant to metabolic heterogeneity. Because this definition is narrower than the American Diabetes Association (ADA) criteria, the findings are most directly applicable to WHO-aligned populations. For baseline data, when the missingness of key covariates was < 10%, missing values were handled using an iterative single-imputation procedure informed by the multiple imputation by a chained equations approach to generate a complete dataset for clustering; variables with > 10% missingness were not used for clustering. After exclusions, missing-data handling, and quality control of the core clustering variables, 621 participants composed the final clustering sample (Figure 1), of whom 367 had longitudinal outcome data and complete sex information. All participants provided written informed consent, and the study was approved by local institutional ethics committees.

Figure 1 Study flowchart of participant selection, data preprocessing and cluster derivation. A total of 2527 participants were screened, 1059 met the prediabetes definition, and 621 remained for clustering after exclusion for diabetes, non-prediabetes or missing glucose data, and missing core clustering variables. The clustering workflow included log transformation of body mass index, homeostatic model assessment of insulin resistance, and homeostatic model assessment of β-cell function, standardization of all five core variables, and K-means clustering (k = 4), which identified four subtypes. FBG: Fasting blood glucose; BG2h: 2-hour blood glucose; HbA1c: Glycated hemoglobin A1c; BMI: Body mass index; HOMA-IR: Homeostatic model assessment of insulin resistance; HOMA-β: Homeostatic model assessment of β-cell function; MICE: Multiple imputation by chained equations; SIR: Severe insulin resistant; SID: Severe insulin deficiency; MOD: Mild obesity-related dysmetabolism; MARD: Mild age-related dysmetabolism.

Intervention framework and exposure definition

In the pragmatic implementation of the national program, centers were allocated to deliver either an enhanced lifestyle management package or standard-of-care health education. The centers designated to provide the enhanced package constituted the intervention arm, whereas one center providing standard health education served as the control arm. Because resource allocation occurred at the center level rather than by individual randomization, intervention status was treated as a non-randomized exposure, and all intervention-associated estimates were interpreted as adjusted observational associations rather than causal effects.

Enhanced lifestyle management (intervention group): The intervention comprised four integrated components: (1) Individualized dietary energy prescription: Daily energy intake targets were assigned according to sex and weight status, for women: 1200 kcal/day (obese), 1400 kcal/day (overweight), and 1600 kcal/day (normal weight); for men: 1400 kcal/day (obese), 1600 kcal/day (overweight), and 1800 kcal/day (normal weight). Each main meal was required to include vegetables, high-quality protein sources, and cereal/grain-based staple foods. For participants who were overweight or obese, a structured eating sequence was recommended, beginning with oil-free vegetables, followed by protein, then staple foods, to support appetite control and improve postprandial metabolic responses. All food amounts were recorded as raw edible-portion weights; (2) Structured physical activity guidance: Participants were provided with wearable step-counting wristbands and encouraged to achieve ≥ 8000 steps per day or ≥ 30 minutes of continuous exercise daily, commencing approximately 30 minutes after meals; fasting exercise was discouraged; (3) Digital co-management platform: Participants installed a dedicated mobile application (“Co-Care” APP) through which they uploaded meal photographs for real-time review and feedback by research staff, accessed health education materials, and recorded lifestyle behaviors; (4) Structured follow-up schedule: In-person visits were conducted at 3 months, 6 months, and every 6 months thereafter, during which anthropometric measurements, health education, and lifestyle counseling were provided; and (5) Personnel and implementation notes: All intervention providers (endocrinologists, nutritionists, sports medicine professionals, community physicians) were certified in their respective fields and trained on the national project’s protocol.

Standard care (control group): Participants in the control arm received routine health education and standard lifestyle counseling during scheduled clinic visits at baseline, 1 year, and 2 years. These visits included general health education and routine metabolic assessment but did not involve structured lifestyle monitoring tools or digital management platforms. Consistent with usual community-based care practices. They did not receive the wearable activity trackers, the mobile co-management application, or the individualized dietary prescription materials.

Because individual-level adherence metrics, such as step counts, photo-upload frequency, and app engagement, were not consistently retrievable across centers during routine implementation, exposure was defined according to the program-offered management package at the center level. Potential informative censoring from follow-up loss was addressed using inverse probability of censoring weighting (IPCW) in sensitivity analyses.

Metabolic subtyping by unsupervised clustering

Selection and preprocessing of clustering variables: Following established frameworks for mechanism-based diabetes subphenotyping, we selected five core metabolic features as inputs for clustering: Age, body mass index (BMI), HbA1c, homeostasis model assessment of insulin resistance (HOMA-IR), and homeostasis model assessment of β-cell function (HOMA-β). Together, these variables capture complementary dimensions of dysglycemia, insulin resistance, and insulin secretory capacity, key pathophysiological domains that underpin heterogeneity in individuals with prediabetes. Because BMI, HOMA-IR, and HOMA-β showed pronounced right-skewed distributions, we applied natural log transformation to attenuate the influence of extreme values on Euclidean distance calculations and to improve numerical stability. All five variables were then standardized to z scores (mean = 0, SD = 1) to remove scale dependence. The preprocessed complete dataset was used for clustering analyses.

Model development and stability assessment: We performed unsupervised subtyping using the k-means clustering algorithm on the standardized variables. To determine the optimal number of clusters (k), we triangulated evidence from multiple internal validation metrics and stability analyses, including the elbow point of the within-cluster sum of squares (inertia), the silhouette coefficient, the Davies-Bouldin index, and the Calinski-Harabasz index. In addition, we applied a consensus clustering procedure to evaluate robustness and between-cluster separation across candidate k values. On the basis of the concordant patterns across these criteria, we selected k = 4 as the optimal solution, which yielded four metabolic subgroups.

For clinical interpretability, clusters were labeled according to their relative profiles across the core metabolic variables as follows: Mild obesity-related dysmetabolism (MOD), mild age-related dysmetabolism (MARD), severe insulin resistance (SIR), and severe insulin deficiency (SID), corresponding to the four k-means-derived clusters.

To quantify the stability and reproducibility of the clustering solution, we implemented two complementary approaches. First, we used bootstrap resampling (80% resampling fraction; 500 iterations) to compute the mean Jaccard similarity index for each cluster. Second, we repeated k-means with 50 random initializations and quantified agreement across runs using the adjusted Rand index (ARI). Mean Jaccard indices > 0.75 and ARI values approaching 1.0 were considered indicative of high stability.

To evaluate whether the preprocessing choice (log transformation) materially influenced subgroup assignment, we compared subtype classifications obtained under two preprocessing schemes (with vs without log transformation), aligned the resulting four subtypes using the same labeling principles, and quantified individual-level concordance between the two solutions using the ARI.

Validation of clustering results: To assess the construct validity and biological plausibility of the derived subtypes, we conducted several validation analyses. First, we compared distributions of the five clustering inputs and additional baseline characteristics across subgroups using the Kruskal-Wallis test for overall differences and quantified effect sizes using epsilon-squared (ε²). When global differences were significant, we performed pairwise subgroup comparisons using the Mann-Whitney U test and controlled for multiple testing with the Benjamini-Hochberg false discovery rate (FDR) procedure. Second, to mitigate circular validation (i.e., validating clusters primarily using the same variables that defined them), we performed external comparisons using OGTT-derived dynamic measures not included in the clustering model, such as 1-hour and 2-hour plasma glucose and insulin concentrations. Finally, we examined the correspondence between data-driven subtypes and conventional impaired fasting glucose (IFG)/IGT categories using contingency-table cross-tabulations.

Definitions of outcomes

During the follow-up period after baseline subtyping, participants were categorized into three mutually exclusive outcomes: (1) Progression to diabetes: Participants who met the WHO criteria for diabetes during follow-up, defined as a fasting blood glucose concentration of ≥ 7.0 mmol/L or a 2-hour OGTT glucose concentration of ≥ 11.1 mmol/L; (2) Reversion to normal glucose tolerance: Participants whose fasting blood glucose and 2-hour OGTT glucose levels returned to the normal range during follow-up; and (3) Sustained prediabetes: Participants who neither progressed to diabetes nor reverted to normal glucose tolerance. Outcome determination was based on follow-up OGTT results and community-based follow-up records.

Statistical analysis

We conducted the following statistical analyses to address the study objectives.

Baseline characteristic comparison: Descriptive statistics were used to summarize the demographic and metabolic characteristics of the overall sample and each metabolic subtype. Categorical variables are presented as n (%), whereas continuous variables are presented as the mean ± SD or medians (interquartile range) on the basis of normality assessment. To compare continuous variables between subtypes, we used the Kruskal-Wallis test for nonnormally distributed data (or one-way ANOVA for normally distributed data), followed by pairwise comparisons using the Mann-Whitney U test with FDR correction. Categorical variables were compared using the χ² test.

Longitudinal outcome analysis: Survival analysis was performed to assess differences in follow-up outcomes across subtypes. First, Cox proportional hazards regression models were used to estimate the hazard ratios (HRs) and 95% confidence intervals (CIs) for the risk of progression to diabetes, reversion to normal glucose tolerance, and persistence of prediabetes. The models were adjusted for clustering subtype, intervention group, and sex to evaluate the independent effects of subtypes and intervention on outcomes. The MARD subtype was used as the reference group (baseline risk), the control group (no intensive intervention) served as the reference for intervention status, and female participants served as the reference for sex.

Competing risks analysis and adjusted cumulative incidence were incorporated, given the mutually exclusive nature of the longitudinal outcomes. For the outcome of diabetes progression, reversion to normoglycemia was treated as a competing event, and vice versa. We first estimated the unadjusted cumulative incidence using the non-parametric Aalen-Johansen estimator and compared the overall cumulative incidence across subtypes using the Gray test. Furthermore, to estimate covariate-adjusted relative risks under the competing risks framework, we used Fine-Gray subdistribution hazard models to calculate subdistribution HRs (sHR).

Crucially, to provide absolute risk estimates that are consistent with the relative risks from the cause-specific Cox models and account for confounding by intervention status and sex, we calculated adjusted 2-year cumulative incidence. This was achieved using the aforementioned Fine-Gray models combined with marginal standardization (also termed direct standardization or G-computation). For each subtype, all individuals’ subtype indicator was set to that group while retaining their original values for intervention and sex, and the predicted cumulative incidence function (CIF) at t = 730 days (2 years) was averaged across the entire study population. Bootstrap resampling (500 iterations) was used to derive 95%CIs for the adjusted estimates.

Intervention effect assessment: To evaluate the effect of lifestyle intervention on prediabetes outcomes, we included the intervention group as a variable in the Cox model and calculated the HR for the intervention group vs the control group. We further included an interaction term for “subtype × intervention” to test whether the intervention effect differed across subtypes. A P value < 0.05 for the interaction term indicates that the intervention effect varied significantly between subtypes.

Sensitivity and bias analysis: Because differential attrition across subgroups can introduce selection bias into longitudinal estimates, we pre-specified IPCW as a sensitivity analysis to evaluate the robustness of the primary findings under potential informative censoring. Specifically, we modeled the probability of remaining under observation at each follow-up time point as a function of baseline covariates (metabolic subtype, intervention status, sex, age, and BMI) using pooled logistic regression. Each participant’s contribution to the Cox model was then weighted by the inverse of their estimated probability of not being censored, thereby up-weighting individuals whose baseline profile resembled those who were lost to follow-up. The IPCW-weighted Cox models were compared with the unweighted primary analyses to assess whether differential attrition materially altered the HR estimates. To mitigate measured confounding, all Cox proportional hazards models included intervention status as a covariate alongside metabolic subtype and sex. We emphasize that the primary inferential target of this study is the association between metabolic subtype and longitudinal outcomes, with intervention status serving as an adjustment variable rather than the principal exposure of interest. In this approach, weights were calculated on the basis of baseline characteristics to account for the probability of completing follow-up, and the Cox models were weighted accordingly to reduce potential loss-to-follow-up bias. Additionally, the proportional hazards assumption of the Cox model was assessed using Schoenfeld residuals to evaluate the robustness of the primary conclusions.

All the statistical analyses were two-sided, with a significance level set at P < 0.05. Data analysis was conducted using Python 3.11 and relevant scientific computing libraries.

RESULTS

Study population and metabolic subtyping

A total of 621 individuals with prediabetes were included at baseline for metabolic subtyping analysis, and their key demographic and metabolic characteristics are detailed in Table 1. Using unsupervised K-means clustering based on age, BMI, HbA1c, HOMA-IR, and HOMA-β, we identified four distinct physiological subtypes with significantly different metabolic profiles: MOD, SIR, SID, and MARD. Significant overall differences were observed across all core clustering variables (Kruskal-Wallis test, all P < 0.001), with the greatest effect sizes for HOMA-IR and HOMA-β (η²: HOMA-IR approximately is 0.59, HOMA-β 0.63) (Supplementary Table 1). Post hoc pairwise comparisons revealed that most pairwise differences in core variables remained significant after FDR correction (28/30, 93.3%), further supporting the discriminative power of this clustering approach (Supplementary Table 2). Furthermore, all pairwise differences between subtypes for BMI, HOMA-IR, and HOMA-β were significant, suggesting not only statistical distinction but also biological relevance, with these three dimensions serving as core axes for mechanistic subtyping. In terms of baseline characteristics, the MARD group was the oldest (mean age approximately 57.85 years) and had the highest HbA1c levels (mean approximately 5.97%), reflecting features of “age-related metabolic decline”. The SIR group exhibited a greater obesity (median BMI approximately 29.30 kg/m²) and marked insulin resistance (median HOMA-IR approximately 5.62), with compensatory β-cell hypersecretion (median HOMA-β approximately 165.29). The SID group had the lowest BMI (median approximately 23.20 kg/m²) and the poorest β-cell function (median HOMA-β approximately 48.66), which was consistent with a “secretion defect-dominated” phenotype. Finally, the MOD group displayed milder overall metabolic dysfunction, with values indicators such as HbA1c levels (mean approximately 5.51%) and HOMA-IR (median approximately 2.50) reflecting less severe metabolic disturbance (Table 1).

Table 1 Baseline characteristics of participants by cluster, mean ± SD/median (interquartile range).

Characteristic	Overall	MOD	MARD	SID	SIR	P value1
Female sex, n (%)	319 (51.4)	81 (45.8)	100 (52.6)	91 (57.2)	47 (49.5)	0.197
Age, years	51.08 ± 10.14	44.04 ± 7.84	57.85 ± 6.82	55.42 ± 7.90	43.39 ± 9.33	< 0.001
Body mass index, kg/m2	25.20 (23.40, 27.40)	25.09 (23.70, 26.50)	25.90 (24.49, 27.38)	23.20 (21.40, 24.60)	29.30 (27.38, 31.90)	< 0.001
Waist circumference, cm	88.80 ± 11.76	86.66 ± 10.55	90.93 ± 10.01	82.24 ± 8.27	98.76 ± 13.74	< 0.001
Systolic blood pressure, mmHg	124.04 ± 15.53	120.68 ± 13.91	127.47 ± 14.45	121.93 ± 17.85	126.81 ± 14.75	< 0.001
Diastolic blood pressure, mmHg	77.91 ± 9.95	77.83 ± 10.02	77.66 ± 9.32	75.81 ± 9.80	81.82 ± 10.25	< 0.001
Fasting glucose, mmol/L	5.96 ± 0.62	5.81 ± 0.66	6.00 ± 0.56	6.07 ± 0.60	5.95 ± 0.65	0.002
2-hour glucose, mmol/L	8.73 ± 1.38	8.61 ± 1.36	8.79 ± 1.38	8.55 ± 1.45	9.15 ± 1.21	0.013
1-hour glucose, mmol/L	11.28 ± 2.20	10.63 ± 2.16	11.54 ± 2.06	11.36 ± 2.42	11.83 ± 1.93	< 0.001
Glycated hemoglobin, %	5.80 ± 0.34	5.51 ± 0.31	5.97 ± 0.23	5.91 ± 0.27	5.84 ± 0.33	< 0.001
Fasting insulin (0 minute)	10.20 (7.21, 14.43)	9.53 (7.50, 11.44)	12.84 (10.12, 15.72)	6.29 (4.84, 7.44)	21.99 (16.79, 27.87)	< 0.001
Insulin (1 hour)	74.77 (47.34, 116.47)	63.24 (45.91, 89.81)	95.61 (65.53, 135.60)	45.57 (32.63, 67.96)	144.85 (99.99, 180.48)	< 0.001
Insulin (2 hours)	72.96 (46.57, 113.40)	65.01 (46.50, 101.10)	91.33 (61.30, 131.45)	45.31 (31.41, 68.39)	154.40 (101.35, 215.55)	< 0.001
HOMA-insulin resistance	2.68 (1.86, 3.89)	2.50 (1.82, 3.06)	3.44 (2.63, 4.32)	1.67 (1.29, 2.08)	5.62 (4.13, 7.56)	< 0.001
HOMA-β cell function	85.71 (58.59, 126.32)	81.12 (64.17, 106.39)	104.34 (83.28, 128.07)	48.66 (37.37, 58.26)	165.29 (134.71, 238.79)	< 0.001
Triglycerides, mmol/L	1.44 (1.04, 2.06)	1.31 (0.96, 2.01)	1.59 (1.19, 2.23)	1.17 (0.90, 1.54)	1.81 (1.32, 2.51)	< 0.001
Total cholesterol, mmol/L	4.81 (4.20, 5.51)	4.75 (4.20, 5.33)	4.78 (4.10, 5.45)	4.90 (4.31, 5.74)	4.78 (4.29, 5.47)	0.248

Sex missing, n = 36 (reported in footnote only).

¹P values: χ² test for categorical variables; Kruskal-Wallis test for continuous variables.

MARD: Mild age-related dysmetabolism; MOD: Mild obesity-related dysmetabolism; SID: Severe insulin deficiency; SIR: Severe insulin resistance; HOMA: Homeostasis model assessment.

To evaluate the external biological consistency of the derived metabolic subtypes, we further examined OGTT-derived dynamic measures that were not included in the clustering procedure. Distinct and physiologically coherent differences in postload insulin responses were observed across subtypes. Across all OGTT time points, both glucose and insulin measures differed significantly among subtypes, with global comparisons demonstrating robust between-group differences (overall Kruskal-Wallis tests for OGTT variables, all P < 0.01). These concordant differences in dynamic glycemic and insulinemic responses provide external physiological support for the pathophysiological validity of the subtype classification (Supplementary Tables 3 and 4).

Moreover, the newly derived subtypes were statistically associated with the conventional IFG/IGT categories (χ² = 13.20, P = 0.04; Cramér’s V = 0.103), yet they further disentangled distinct driving mechanisms underlying similar glycemic phenotypes. For example, the MOD subtype had the highest proportion of IGT (58.2%, 103/177), whereas the MARD subtype showed a higher prevalence of the combined IFG + IGT pattern (36.8%, 70/190) (Supplementary Table 5). These findings indicate that traditional glycemic categorization is insufficient to differentiate the mechanistic determinants of “hyperglycemia”, whether primarily driven by insulin resistance, secretory defects, or age-related metabolic decline, whereas data-driven subtyping can capture these distinctions. Regarding robustness, bootstrap resampling yielded Jaccard indices ranging from 0.764 to 0.847 across subtypes, and random-seed sensitivity analyses produced a high mean ARI of 0.972 (range, 0.954-1.000), indicating good clustering stability and reproducibility (Supplementary Table 6).

Risk of outcome progression across metabolic subtypes

After a median follow-up of 735 days, 367 participants provided valid longitudinal outcome data. In a multivariable Cox proportional hazards model, we simultaneously adjusted for the effects of physiological subtype, lifestyle intervention, and sex. In the primary multivariable model (Table 2; Figure 2), the MOD subtype had a significantly lower cause-specific hazard of progression to diabetes compared with MARD [adjusted HR (aHR) 0.52, 95%CI: 0.29-0.93; P = 0.028]. This relative risk reduction was corroborated by the adjusted 2-year cumulative incidence. The standardized diabetes incidence for MOD was 3.38% (95%CI: 1.70-5.95), roughly half that of MARD at 7.28% (95%CI: 4.27-11.10), confirming that the aHR-based contrast translates into a clinically meaningful absolute risk difference of approximately 3.9 percentage points at 2 years. For regression to normoglycemia, MOD showed a two-fold higher cause-specific hazard than MARD (aHR 2.08, 95%CI: 1.00-4.33; P = 0.049), with a correspondingly higher adjusted 2-year reversion incidence (6.49% vs 3.03%). Neither SID nor SIR differed significantly from MARD in the risk of diabetes progression or glucose reversion (Table 2). From a competing risks perspective, the overall Gray test for CIFs (Figure 3) did not reach statistical significance for diabetes progression (P = 0.170), although pairwise comparison between MOD and MARD showed a significant trend for CIF separation (raw P = 0.027; Supplementary Table 7), reinforcing the pronounced heterogeneity in outcomes between these two subtypes.

Figure 2 Forest plot of adjusted hazard ratios (cause-specific Cox models). Panels show the associations of subtype, intervention, and sex with progression to diabetes, regression to normoglycemia, and persistence of prediabetes. Points indicate adjusted hazard ratios and horizontal lines indicate 95% confidence intervals; the dashed vertical line marks the null value (adjusted hazard ratio = 1.0). Models were adjusted for diabetes subtype, intervention status, and sex, with mild age-related dysmetabolism as the reference subtype. ^aP < 0.05. aHR: Adjusted hazard ratio; CI: Confidence interval; MARD: Mild age-related dysmetabolism; MOD: Mild obesity-related dysmetabolism; SID: Severe insulin deficiency; SIR: Severe insulin resistance.

Figure 3 Cumulative incidence functions (Aalen-Johansen) with adjusted 2-year cumulative incidence function. Panels show Aalen-Johansen cumulative incidence curves for progression to diabetes, regression to normoglycemia, and persistence of prediabetes across the four metabolic subtypes. Shaded bands indicate 95% confidence intervals from 500 bootstrap resamples, and the vertical dotted line marks 2 years. Tables below each panel summarize sample size, events, last event time, unadjusted 2-year cumulative incidence, adjusted 2-year cumulative incidence, and 95% confidence intervals. Adjusted 2-year cumulative incidence was estimated with Fine-Gray competing risk models and marginal standardization, adjusting for subtype, intervention, and sex. MARD: Mild age-related dysmetabolism; MOD: Mild obesity-related dysmetabolism; SID: Severe insulin deficiency; SIR: Severe insulin resistance; CI: Confidence interval.

Table 2 Cluster-specific dysglycemia transitions and intervention effects: Adjusted hazard ratios and cumulative incidence.

Variable	n	Events	Unadjusted 2-year CIF (%)	Adjusted 2-year CIF (%)1	95%CI (adjusted CIF)	aHR2	95%CI3 (aHR)	P value
Progression to diabetes
MARD	104	35	5.8	7.28	4.27-11.10	Reference
MOD	99	17	5.1	3.38	1.70-5.95	0.52	0.29-0.93	0.028a
SID	115	28	7.0	5.20	2.66-8.49	0.84	0.51-1.39	0.496
SIR	49	19	4.1	8.32	4.69-12.84	0.88	0.50-1.57	0.673
Intervention (yes vs no)						0.52	0.30-0.89	0.017a
Sex (male vs female)						1.24	0.82-1.85	0.305
Regression to normoglycemia
MARD	104	11	1.9	3.03	1.41-5.17	Reference
MOD	99	21	7.1	6.49	3.58-10.48	2.08	1.00-4.33	0.049a
SID	115	19	5.2	4.97	2.38-8.28	1.82	0.87-3.84	0.113
SIR	49	9	6.1	5.41	2.09-10.34	1.51	0.63-3.66	0.358
Intervention (yes vs no)						0.63	0.33-1.20	0.159
Sex (male vs female)						0.94	0.56-1.59	0.829
Persistence of prediabetes
MARD	104	58	14.4	15.49	10.99-20.46	Reference
MOD	99	61	14.1	17.50	12.53-22.74	1.18	0.82-1.69	0.380
SID	115	68	22.6	17.21	12.40-22.55	1.25	0.88-1.77	0.222
SIR	49	21	8.2	11.02	6.48-16.28	0.67	0.41-1.11	0.120
Intervention (yes vs no)						0.57	0.40-0.81	0.002a
Sex (male vs female)						0.90	0.68-1.19	0.459

¹Adjusted 2-year cumulative incidence function from Fine-Gray competing risk model with marginal standardization.

²Adjusted hazard ratio from cause-specific Cox models (reference: Mild age-related dysmetabolism, adjusted for intervention and sex).

³95% confidence interval from 500 bootstrap resamples.

^aP < 0.05.

aHR: Adjusted hazard ratio; CIF: Cumulative incidence function; MARD: Mild age-related dysmetabolism; MOD: Mild obesity-related dysmetabolism; SID: Severe insulin deficiency; SIR: Severe insulin resistance; CI: Confidence interval.

Effects of lifestyle intervention

After mutual adjustment for physiological subtype and sex (Table 2), receipt of the center-allocated enhanced lifestyle management package was associated with a lower cause-specific hazard of progression to diabetes (aHR 0.52, 95%CI: 0.30-0.89; P = 0.017) and a significantly reduced hazard of sustained prediabetes (aHR 0.57, 95%CI: 0.40-0.81; P = 0.002). The intervention did not reach statistical significance for promoting reversion to normoglycemia in the primary analysis (aHR 0.63, P = 0.159). Under the Fine-Gray competing risks framework, the intervention effect on cumulative diabetes incidence was attenuated and non-significant (sHR 0.84, 95%CI: 0.55-1.29; P = 0.428), whereas a borderline association with glucose reversion emerged (sHR 0.57, 95%CI: 0.32-1.00; P = 0.051; Supplementary Table 8), suggesting that the primary benefit of the intervention may lie in facilitating glucose reversion rather than solely inhibiting disease progression.

Heterogeneity of intervention effects

No statistically significant subtype × intervention interaction was detected (P for interaction = 0.999; Supplementary Table 9). The coefficients for the interaction terms between subtypes and intervention were also nonsignificant (e.g., MARD × intervention: P = 0.933; SID × intervention: P = 0.974; Supplementary Table 10). This suggests that, within the present sample and follow-up duration, relative intervention associations did not differ clearly across subtypes. Accordingly, the primary value of subtyping in this study is baseline risk stratification rather than demonstration of subtype-specific intervention efficacy.

Robustness of the study

We conducted multiple sensitivity analyses to ensure the reliability of the key findings.

Loss to follow-up bias: Of 621 participants subtyped at baseline, 367 (59.1%) contributed longitudinal data. Subtype distribution differed between completers and non-completers (P = 0.0008), confirming non-random attrition. IPCW-weighted HRs were uniformly concordant with the primary estimates, with all absolute differences within ± 3.2% (Supplementary Table 11). For the core contrast, the IPCW-adjusted HR for MARD vs MOD regarding diabetes progression was 1.89 (95%CI: 1.02-3.48, P = 0.043), compared with 1.89 (1.05-3.38, P = 0.033) in the primary model. These results indicate that differential attrition did not materially distort the subtype-outcome associations.

Proportional hazards assumption: The Schoenfeld residuals test indicated that, except the intervention variable for the outcome “sustained prediabetes” (P = 0.003), all the variables in the primary models satisfied the proportional hazards assumption (Supplementary Table 12). These findings suggest that the effect of intervention on the maintenance of prediabetes status may vary over time, warranting further investigation in future studies.

Resampling stability: A resampling analysis with 1000 bootstrap iterations confirmed the stability of key HR estimates, such as the median HR of 0.52 for MOD vs MARD and the median HR of 0.56 for intervention-induced reversion (Supplementary Table 13), reinforcing the reliability of the results.

Cross-sectional validation: A multistate logistic regression model was used to analyze the final follow-up status, and the findings were consistent with the trends observed in the longitudinal Cox model. For instance, the MARD subtype had a significantly greater risk of diabetes progression compared to reversion to normal glucose tolerance (relative risk ratio 4.22, P = 0.0027) (Supplementary Table 14), providing additional internal consistency with the longitudinal findings from an alternative analytical perspective.

DISCUSSION

This multicenter prospective cohort study revealed significant metabolic heterogeneity within the prediabetic population. We observed distinct differences in the progression of diabetes, reversion to normal glucose levels, and maintenance of prediabetes across metabolic subtypes. These findings support the growing emphasis on refined subtyping of diabetes and prediabetes to enhance risk prediction and intervention outcomes, consistent with the observations of Taurbekova et al[21] and Sevilla-Gonzalez[22]. Using K-means clustering, we identified four metabolic subtypes. These subtypes displayed distinct trajectories during lifestyle intervention follow-up. The MOD group had the best prognosis, with the lowest risk of progression to type 2 diabetes and the highest likelihood of reverting to normal glucose levels. In contrast, the MARD group had the highest risk of diabetes progression and the less favorable overall trajectory during follow-up. The SIR and SID groups exhibited intermediate outcomes with distinct metabolic profiles, highlighting the inadequacy of glucose metrics alone for individualized risk stratification, as emphasized by Yudkin[23]. In contrast, Wagner et al[17] and Zheng et al[20] demonstrated that data-driven, unsupervised clustering based on multidimensional metabolic characteristics offers a more refined method for identifying individuals at high risk before diabetes onset, thus supporting more targeted and precise interventions.

Characteristic differences between the MOD and MARD subtypes and intervention response

The most notable differences in this study were observed between the MOD and MARD subtypes. The MOD group was younger, with moderate obesity and mild metabolic dysfunction (e.g., baseline HbA1c of approximately 5.5% and moderately elevated HOMA-IR); their β-cell function was relatively preserved, and glucose homeostasis was maintained through compensatory insulin hypersecretion. This profile suggests a high potential for reversibility, as intensive lifestyle interventions, including weight control and increased physical activity, can markedly improve insulin sensitivity and correct mild hyperglycemia, as recommended by the ADA Professional Practice Committee[24]. Previous studies, such as that of the DPP Research Group[10], have demonstrated that lifestyle intervention reduces the risk of diabetes by nearly 50% in obese and insulin-resistant populations. Our results further showed that, under the observed implementation context, the MOD group had the most favorable overall trajectory, with a lower risk of progression, a higher likelihood of reversion, a nearly 50% reduction in diabetes risk, and the highest rate of glucose normalization. This observation is in line with the findings of Ahlqvist et al[15], who reported that individuals with mild obesity-related hyperglycemia generally have a favorable prognosis. Similarly, Zaharia et al[25] demonstrated in patients with newly diagnosed diabetes that MOD is characterized by slow disease progression, fewer complications, and lower treatment intensity requirements. In our study, the MOD subtype may represent a prediabetic counterpart of the mild obesity-related phenotype, and its favorable outcomes further support this view. As emphasized by Yudkin[23], the main abnormalities in this group are largely associated with reversible lifestyle-related factors, such as overweight and unhealthy diet; therefore, management should focus primarily on lifestyle modification rather than medical intervention. The ADA Professional Practice Committee[24] has recommended adequate physical activity and a 5%-7% weight reduction, which may restore normal glucose levels in a substantial proportion of these individuals. Regular follow-up monitoring remains essential for consolidating intervention effects and preventing metabolic rebound[26-28].

In contrast, the MARD subtype was characterized by older age, with a mean baseline age of approximately 58 years. Despite the absence of marked obesity, individuals in this group tended to exhibit higher HbA1c levels and other metabolic markers than younger individuals at similar glucose levels. Our follow-up findings showed that the MARD group had the highest risk of progression to diabetes, indicating that age-related metabolic decline may play an important role in accelerating the worsening of glucose tolerance. This finding contrasts with the report by Ahlqvist et al[15], in which mild age-related diabetes was described as a relatively mild disease subtype. This discrepancy may be explained by the fact that age itself is an independent risk accelerator during the prediabetic phase. Zhu et al[12] reported that older adults with prediabetes commonly exhibit progressive β-cell dysfunction and multisystem metabolic vulnerability, with β-cell function declining by approximately 0.7%-1% annually from young adulthood. Echouffo-Tcheugui et al[14] further pointed out that age-related deficits in insulin secretion may blunt the benefits of lifestyle intervention; although peripheral insulin sensitivity may improve, inadequate β-cell reserve often limits complete correction of glycemia.

Additionally, Echouffo-Tcheugui et al[14] emphasized that older individuals often have sarcopenia and chronic inflammation, which may impair exercise tolerance and reduce the effectiveness of lifestyle intervention. Although the DPP Research Group[10] showed that intensive lifestyle intervention reduced diabetes risk by 71% among individuals aged ≥ 60 years, even exceeding the benefit observed in younger groups, caution is warranted when extrapolating these findings to real-world populations, as older participants in the DPP may have represented a relatively selected population with better adherence and health screening. In routine clinical practice, the typical MARD population often presents with multiple comorbidities and declining physical function, which may hinder achievement of the intensive intervention targets used in clinical trials. This may explain why the MARD group in our study derived only limited benefit from intervention and continued to show a relatively high incidence of diabetes.

Mechanistically, the MARD subtype was characterized by relatively impaired insulin secretion and lacked the compensatory hyperinsulinemia seen in the SIR subtype. This pattern suggests that, once hyperglycemia develops in individuals with MARD, it may already reflect markedly reduced β-cell reserve, leaving little capacity for further compensation and thereby favoring a more direct progression to diabetes. By contrast, individuals with the SIR subtype may be able to transiently buffer rising glucose levels through increased insulin secretion. In addition, Li et al[29] reported that although older-onset subtypes showed only a moderate risk of diabetes, they had the highest incidence of cardiovascular complications. Taken together, these findings indicate that the MARD subtype may represent an underrecognized high-risk phenotype among older adults with prediabetes. In clinical practice, this suggests that, beyond standard lifestyle management, a more proactive and comprehensive approach may be warranted in this group, including earlier consideration of pharmacotherapy and closer management of comorbidities, to delay progression to diabetes and cardiovascular events, as recommended by ElSayed et al[30] and Sevilla-Gonzalez[22]. This view is also consistent with current public health strategies that emphasize early intervention in older adults with prediabetes to improve long-term outcomes[31].

Intermediate characteristics and mechanisms of the SIR and SID subtypes

The SIR and SID subtypes present distinct metabolic profiles and clinical implications. The SIR subtype, characterized by SIR, is associated with high levels of obesity and insulin resistance. Individuals in this group show hyperinsulinemia at both the fasting and postprandial stages, which helps to maintain glucose homeostasis. In our cohort, the risk of diabetes progression in the SIR group was greater than that of the MOD group but lower than that in the MARD group, suggesting that residual insulin function mitigates the effects of SIR. This group is closely associated with metabolic syndrome and commonly presents with comorbidities such as dyslipidemia, hypertension, and nonalcoholic fatty liver disease, as described by Taurbekova et al[21]. Ahlqvist et al[15] further showed that individuals with SIR had a significantly higher prevalence of diabetic nephropathy and fatty liver already at the time of diabetes diagnosis. These findings, together with the observations of Baranowska-Jurkun et al[32], suggest that individuals with the SIR phenotype may already face an elevated risk of cardiovascular and renal complications during the prediabetic stage[33].

Clinically, managing weight and addressing insulin resistance through lifestyle changes are crucial. Targeted interventions such as increasing physical activity and improving diet to reduce visceral fat are particularly important. Studies suggest that for high-risk individuals with SIR, combining lifestyle modifications with insulin sensitizers such as metformin may provide additional benefits. For example, Stafford et al[19] found that, in the DPP cohort, metformin was associated with a 50% reduction in diabetes risk among obese individuals with high HOMA-IR relative to those without insulin resistance. Thus, the SIR subtype represents an opportunity for precision intervention, in which early identification and targeted treatments aimed at insulin resistance could help prevent diabetes progression. While our study did not find significant interactions between lifestyle interventions and subtypes, suggesting that lifestyle improvements are broadly beneficial, further exploration in larger randomized studies is needed to assess the benefits of personalized interventions for the SIR group.

The SID subtype, characterized by severe insulin secretion deficiency, follows a distinct pathological course. Individuals with this phenotype generally have a relatively low BMI but exhibit significant declines in β-cell function, with a markedly reduced insulin response during the OGTT compared with that in other subtypes. Our findings indicate that the risk of progression to diabetes in the SID group was slightly greater than that in the MOD group but not significantly greater than that in the MARD group. These findings suggest that while insufficient insulin secretion is the primary issue in this subtype, it has not yet reached a “critical point” at the prediabetic stage. However, once individuals in the SID group progress to diabetes, they may experience more rapid glucose dysregulation and complications. Ahlqvist et al[15] reported that individuals with the SIDD phenotype, which corresponds to the SID subtype in our study, tend to follow a more aggressive clinical course and develop microvascular complications, especially retinopathy, within a few years of diagnosis. Our results suggest that, although baseline glucose compensation in the SID group is not profoundly impaired, reduced β-cell function may create a metabolically fragile state that predisposes these individuals to rapid progression to overt diabetes under external stressors such as infection. Li et al[29] further showed that, compared with other subtypes, individuals with the SIDD phenotype are more likely to develop diabetic nephropathy and severe hyperglycemia. Therefore, close monitoring of glycemic changes and clinical symptoms is warranted in individuals with the SID subtype.

For the SID subtype, intervention strategies may need to differ from standard lifestyle management, as the primary issue is insulin secretion. Simple weight loss may not significantly improve blood glucose levels. If individuals with the SID subtype continue to show hyperglycemia or rapid metabolic deterioration despite lifestyle optimization, pharmacological intervention may need to be considered at an earlier stage. Evidence from Holman et al[34] and le Roux et al[35] suggests that early treatment in high-risk individuals may help delay progression and reduce the likelihood of early complications. Although lifestyle modification remains the main recommendation for most individuals with prediabetes according to the ADA Professional Practice Committee[24], those with high-risk features, including an HbA1c level close to 6.5% or substantial weight gain, may require a more proactive preventive strategy.

We suggest that future research and guideline updates incorporate metabolic subtyping into risk assessment criteria. The SID subtype may represent a high-priority group for proactive intervention, particularly for individuals with progressive β-cell decline, who would benefit from early measures to delay β-cell exhaustion and prevent complications.

Clinical value of clustered subtyping: Risk stratification and precision intervention

This study highlights the clinical value of metabolic subtyping in prediabetes. As noted by Yudkin[23], conventional definitions based on glucose markers such as IFG, IGT, or HbA1c fail to capture important individual differences in pathophysiology and disease progression. Florez[36] further emphasized that a one-size-fits-all prevention strategy is inadequate for such a heterogeneous population. By contrast, Taurbekova et al[21] showed that unsupervised clustering of comprehensive phenotypic data can further distinguish subgroups with different metabolic characteristics and risk profiles.

Large cohort studies have demonstrated this approach. For instance, Wagner et al[17] used a clustering model with multiple biomarkers to identify six subtypes among high-risk individuals, each with distinct risks of diabetes progression and complications. In a Chinese cohort including more than 55000 individuals with prediabetes, Zheng et al[20] identified six data-driven subtypes with a stepwise gradient in diabetes risk. Notably, the obesity-insulin resistance subtype, although associated with only a moderate risk of diabetes, showed increased risks of kidney dysfunction and mortality. These findings are consistent with our results and further indicate that metabolic subtyping may improve risk stratification in individuals with prediabetes by distinguishing those at relatively high or low risk. In contrast, the European TUEF/TULIP cohort used WHO-aligned criteria similar to ours, which may partly explain the closer correspondence in subtype characteristics between our findings and those of Wagner et al[17]. The DPP used a lower fasting glucose entry threshold (≥ 5.3 mmol/L) but required the concurrent presence of IGT[19], resulting in a cohort enriched for postload glucose intolerance. These threshold-related differences in cohort composition should be taken into account when interpreting cross-study comparisons of subtype prevalence, metabolic profiles, and outcome trajectories. Despite these differences in inclusion criteria, the overarching finding that prediabetes is metabolically heterogeneous and amenable to data-driven subclassification appears robust across diagnostic frameworks.

Applying clustering-based subtyping in clinical practice may improve risk stratification at the prediabetes stage. As proposed by Sevilla-Gonzalez[22], this approach could allow high-risk subgroups to receive more intensive interventions, whereas low-risk groups may be managed primarily through lifestyle maintenance. This strategy is in line with the precision medicine framework outlined by Florez[36], which emphasizes tailoring prevention and treatment to individual characteristics in order to optimize cost-effectiveness[37]. Taurbekova et al[21] further highlighted the growing international consensus on incorporating diabetes heterogeneity into clinical practice, including efforts such as the ADA/European Association for the Study of Diabetes Precision Medicine in Diabetes initiative and National Institute of Diabetes and Digestive and Kidney Diseases working groups, which support the development of subgroup-based intervention strategies beyond conventional classification models. With further advances in this field, metabolic subtyping is likely to become increasingly important for risk assessment and clinical decision-making in prediabetes.

Personalized intervention recommendations: Optimizing preventive strategies for different subtypes

Our findings suggest that individuals with prediabetes from different metabolic subtypes should receive targeted interventions to achieve precision prevention, addressing the specific needs of each subgroup. For the MOD subtype, which is characterized by younger age and mild metabolic dysfunction, lifestyle management should remain the cornerstone of intervention, as emphasized by ElSayed et al[30]. This group typically benefits from dietary improvements and increased physical activity, which can significantly reduce the risk of progression without the need for extensive pharmacological treatment. Emphasis should be placed on regular moderate-intensity aerobic exercise combined with resistance training to improve insulin sensitivity, as well as controlling total caloric intake and ensuring balanced nutrition to achieve a 5%-10% weight reduction. These interventions often yield substantial benefits for MOD individuals, who are also the most likely to experience normalization of blood glucose levels[24].

Clinicians should closely monitor changes in weight and metabolic indicators during the intervention process and provide comprehensive behavioral guidance and psychological support to reinforce healthy lifestyle habits. Current evidence suggests that preventive pharmacotherapy may not be necessary for individuals in the MOD group to achieve favorable outcomes; accordingly, Yudkin[23] cautioned against overmedicalization and unnecessary intervention in this low-risk population. If these individuals maintain normal blood glucose levels and healthy body weight after several years of follow-up, they can be considered to have successfully “graduated” from the high-risk category.

In contrast, for the MARD subtype (older individuals), a more proactive and comprehensive intervention approach is necessary. First, lifestyle intervention remains the foundation, including balanced nutrition, moderate exercise, and weight management. However, given the common decline in physical capacity and the presence of multiple comorbidities in older adults, personalized strategies should be introduced. For example, Echouffo-Tcheugui et al[14] noted that nutritional management in older adults should be tailored to individual metabolic needs to achieve glycemic control while avoiding malnutrition. They also emphasized that exercise programs should prioritize aerobic activity according to functional capacity, together with resistance training to preserve lean mass and muscle strength, thereby promoting glucose utilization and improving quality of life.

Additionally, for individuals with the MARD subtype, early initiation of pharmacological interventions as an adjunct to lifestyle modifications may be considered. This is especially relevant for individuals with higher baseline glucose levels (e.g., HbA1c levels approaching or exceeding 6.0%) and multiple cardiovascular risk factors, where the use of insulin sensitizers or low-dose insulin secretagogues can be introduced alongside lifestyle adjustments for preventive therapy. For instance, studies have shown that metformin is safe and has a modest effect on delaying diabetes onset in older adults with IGT, and international guidelines recommend its use in prediabetic individuals with higher BMIs and HbA1c levels approaching the diabetes threshold[24]. Our findings suggest that the MARD group may indeed represent such “ultrahigh-risk” individuals that require early, multidimensional, and combined intervention strategies to address their progressive β-cell dysfunction and clustering of multiple risk factors.

Moreover, given the cardiovascular risk profile of the MARD subtype, management of comorbidities such as hypertension and dyslipidemia should be emphasized. Huang et al[38] and Welsh et al[39] reported that prediabetes is associated with increased cardiovascular risk, underscoring the importance of comprehensive risk-factor control in this population. Accordingly, the use of medications such as statins and angiotensin-converting enzyme inhibitors may be considered to reduce the likelihood of cardiovascular events. This is particularly relevant because both our analysis and previous evidence indicate that cardiovascular complications may be more common in the MARD subtype than in other subtypes. Therefore, interventions for this group should not be limited to preventing diabetes but should also include comprehensive prevention of cardiovascular events.

In summary, the MARD subtype represents a high-risk group that requires increased healthcare system resources and more intensive management. Through personalized interventions, it is possible to reduce the rate of diabetes progression and improve long-term outcomes.

Targeted preventive strategies can be applied to both the SIR and SID subtypes. With respect to the SIR subtype, which is characterized by significant insulin resistance, interventions should focus on improving insulin sensitivity. Weight loss, which is achieved through dietary control and physical activity, is essential for reducing BMI and alleviating insulin resistance, thus easing the burden on β-cells. In clinical practice, more intensive lifestyle interventions, such as specialized weight management or metabolic surgery for individuals with severe obesity, may be necessary. If lifestyle modification remains insufficient after 6 months, pharmacological therapy, such as metformin or thiazolidinediones, may be considered to reduce hepatic glucose production and improve insulin sensitivity. Stafford et al[19] reported that individuals with the SIR phenotype showed a better glycemic response to metformin than those without insulin resistance.

Additionally, associated conditions such as hypertriglyceridemia and low high-density lipoprotein cholesterol should be actively managed, possibly with statins or omega-3 supplements, to reduce cardiovascular risks. This comprehensive approach aims to improve metabolic control and overall cardiovascular health in the SIR group.

For the SID subtype, where β-cell dysfunction is predominant, the focus should be on preserving residual insulin function. In addition to lifestyle management, interventions to increase insulin secretion or reduce β-cell apoptosis should be considered. Low-dose α-glucosidase inhibitors or glucagon-like peptide-1 receptor agonists could help manage postprandial glucose fluctuations and provide β-cell rest, although large-scale trials confirming the long-term benefits of these medications in individuals with prediabetes are lacking. Given the substantial burden of complications reported in the SIDD subtype at the time of diabetes onset by Ahlqvist et al[15] and Zaharia et al[25], earlier pharmacological intervention may be warranted in individuals with the SID subtype.

Individuals with the SID subtype should also undergo regular screening for diabetic retinopathy, renal function impairment, and other microvascular complications, as Ahlqvist et al[15] and Li et al[29] showed that this phenotype may already carry an increased risk of such complications at the time of diabetes diagnosis.

In summary, our findings, together with those of Stafford et al[19] and Taurbekova et al[21], support the use of clustering-based subtyping to guide personalized prevention strategies in individuals with prediabetes. Future prospective trials are needed to test the effectiveness of this approach, such as designing different intervention programs for each subtype and comparing their outcomes, to provide more direct evidence for clinical practice.

Prospective complication risks across subtypes

Metabolic subtyping is valuable not only for predicting diabetes risk but also for understanding the potential trajectory of long-term complications. Although our study’s short follow-up period and limited complication events preclude firm conclusions, the literature suggests that different subtypes are likely to follow distinct complication profiles.

Ahlqvist et al[15] reported that the SIR subtype, characterized by pronounced insulin resistance, is closely associated with atherosclerosis and hypertension and may already present with subclinical vascular changes before the onset of diabetes. They further showed that SIR is strongly linked to diabetic nephropathy and nonalcoholic fatty liver disease. In addition, Zheng et al[20] found in a Chinese cohort that individuals with high insulin resistance had a significantly increased risk of developing chronic kidney disease later in life. Therefore, early protective measures, such as blood pressure control and regular monitoring of renal and liver function, are critical for the SIR phenotype.

The SID subtype, marked by insufficient insulin secretion, is more prone to microvascular complications because of prolonged hyperglycemia. Studies by Ahlqvist et al[15] and Li et al[29] showed that the SIDD subtype is associated with an increased incidence of diabetic retinopathy and nephropathy in the early stage after diabetes onset. Early screening and strict glycemic control in this group are essential to delay the onset of microvascular complications. The MARD subtype, driven by age-related factors, warrants attention for macrovascular complications. Consistent with the findings of Huang et al[38], our study suggests that this subtype may be associated with the highest risk of cardiovascular events. In an East Asian population, Li et al[29] found that older individuals with prediabetes were at significantly increased risk of large-vessel events, particularly stroke. While hyperglycemia in MARD individuals may not be the most severe, atherosclerosis is a long-term process, and age is a strong risk factor. Thus, early cardiovascular prevention, including blood pressure and lipid management, smoking cessation, and preventive medications such as statins and aspirin, should begin in the prediabetes phase. Comprehensive management plans should also be developed early for this group. These insights underscore the importance of recognizing varying complication risks across metabolic subtypes. Although statistical significance was not reached in our study, probably because of the short follow-up duration, previous studies by Zheng et al[20] and Zaharia et al[25] have demonstrated clear differences in long-term risk across subtypes. Future risk models for prediabetic populations should integrate subtype information to better predict complications and guide targeted interventions and monitoring strategies.

Limitations and future directions

This study’s principal strength lies in its use of a data-driven clustering approach to reveal significant metabolic heterogeneity within a real-world cohort of Chinese individuals with prediabetes, linking these subtypes to clinically meaningful longitudinal outcomes. By operating within a pragmatic implementation framework, our findings offer enhanced external validity and direct relevance to public health decision-making. However, the study also has several important limitations that frame the interpretation of its results.

First, the non-randomized, center-level allocation of the intervention is a critical limitation for interpreting intervention-associated findings. This design means that unmeasured center-level confounders (e.g., regional socioeconomic factors, local healthcare capacity) may have influenced both intervention assignment and outcomes. Although we adjusted for intervention status in all models and our IPCW analysis provided consistent estimates, residual confounding variables cannot be fully excluded. Consequently, all intervention-associated effect estimates should be interpreted as adjusted observational associations, not as causal efficacy effects. Importantly, as the subtype × intervention interaction was non-significant, the study’s primary contribution, prognostic characterization of metabolic subtypes, remains robust and is not contingent on a causal interpretation of the intervention.

Second, substantial and non-random loss to follow-up (40.9%) occurred in this study, a common challenge in pragmatic, community-based research. If attrition was systematically related to outcomes, our estimates could be biased. To address this, our pre-specified IPCW sensitivity analysis demonstrated that HR estimates remained highly consistent after weighting for baseline predictors of censoring (all absolute differences < 3.2%), providing strong evidence that differential attrition did not materially distort the core subtype–outcome associations. However, that several intervention-associated P values shifted from nominally significant to borderline after IPCW weighting (e.g., normoglycemia reversion: P = 0.0391 → 0.0505), which is expected given the additional variance introduced by the estimated censoring weights and the modest sample size. This observation does not alter the qualitative interpretation but underscores the need for replication in larger cohorts.

Third, the diagnostic criteria for prediabetes of the WHO/IEC 2006 are narrower than those of the ADA. This choice aligned with the parent program’s clinical context but limits the direct generalizability of our findings to populations defined by the broader ADA criteria. Future research should validate this subtype structure across different diagnostic frameworks.

Fourth, the modest sample size and two-year follow-up duration limited our statistical power, particularly for detecting differences among all subtypes in the competing risks model and for analyzing rare complication events. This may explain why significant outcome differences were primarily observed between the most distinct MOD and MARD subtypes. Validation in larger cohorts with extended follow-up is necessary.

Finally, the K-means clustering methodology is sensitive to the selection and scaling of input variables, this could lead to variations in subtype composition across studies. For example, in a Chinese cohort of patients with newly diagnosed diabetes, Li et al[29] identified a comparable subtype referred to as mild insulin-deficient diabetes. The resulting subtype structure requires external validation in diverse populations to ensure its stability and generalizability.

Despite these limitations, this study has significant implications for precision medicine in individuals with prediabetes and highlights several future research directions. First, large-scale prospective studies are needed to validate risk differences across metabolic subtypes. For example, Sevilla-Gonzalez[22] proposed that clustering algorithms could be applied in national cohorts or electronic health record databases to investigate subtype incidence, outcomes, and healthcare disparities.

Second, integrating biological data into subtyping models is crucial. Current clustering methods rely primarily on clinical markers, but incorporating genetic, multiomic, and physiological data could capture deeper heterogeneity. For instance, genetic risk score analysis by Li et al[40] identified six genetic subtypes of prediabetes, among which two were associated with a significantly increased risk of progression to diabetes. Genetic data could help identify high-risk populations early and guide personalized prevention. Additionally, blood metabolites and epigenetic markers are emerging as tools for more accurate subtyping, which may improve clinical predictions and stability.

Also, there is a critical need for randomized controlled trials designed to explicitly test subtype-specific intervention strategies. In such trials, comparing the effectiveness of tailored vs uniform interventions would provide definitive evidence on the clinical utility of metabolic subtyping. Finally, developing and embedding cost-effective, reliable subtyping algorithms into clinical decision support systems will be essential for translating this precision prevention approach into routine practice, offering a new paradigm to combat the growing diabetes epidemic.

CONCLUSION

In conclusion, pathological subtyping based on core metabolic indicators effectively identified subgroups within the Chinese prediabetic population defined by the WHO/IEC criteria that differed in their risk of disease progression. Among these subgroups, the MOD subtype exhibited more favorable metabolic trajectories than the MARD subtype did. Within this pragmatic, non-randomized implementation framework, receipt of the enhanced lifestyle management package was associated with reduced persistence of prediabetes after covariate adjustment, with suggestive evidence of promoting glucose reversion; however, these associations require confirmation in individually randomized designs. The non-significant subtype × intervention interaction suggests that the benefits of enhanced management, if causal, may extend across metabolic subtypes, although the study did not have sufficient power to detect subtype-specific intervention effects. Future research should combine absolute risk with clinical thresholds to evaluate whether differential intervention intensities across subtypes lead to greater clinical net benefits, ultimately advancing the practice of precision prevention in individuals with prediabetes.