Cluster-based metabolic subphenotypes and projected cross-organ risk in Chinese adults with prediabetes

Abstract

BACKGROUND

Prediabetes is metabolically heterogeneous, but most risk assessment remains glycemia centered. It is unclear whether prediabetes subphenotypes already show coordinated or discordant cardio-kidney-liver risk-score signatures before diabetes. We hypothesized that cluster-defined subphenotypes would differ in projected cardiovascular, kidney, and fibrosis-4 (FIB-4)-defined liver fibrosis-risk profiles.

AIM

To determine whether cluster-defined prediabetes subphenotypes differ in projected cross-organ risk-score profiles.

METHODS

We performed a cross-sectional baseline analysis of a prospective multicenter high-risk diabetes program in China. K-means clustering was applied to 621 adults with prediabetes using age, body mass index, glycated hemoglobin, homeostasis model assessment of insulin resistance, and homeostasis model assessment of β-cell function. Framingham-estimated 10-year cardiovascular disease risk, chronic kidney disease (CKD) prognosis consortium-estimated 5-year CKD risk, and FIB-4-defined fibrosis-risk signal were compared across subphenotypes using Poisson regression with HC3 robust standard errors. Sensitivity analyses evaluated whether the observed patterns were robust to threshold choice, missing-data handling, albumin-to-creatinine ratio availability, reference group, and age loading.

RESULTS

Four subphenotypes were identified: (1) Mild obesity-related dysmetabolism (n = 177); (2) Mild age-related dysmetabolism (MARD) (n = 190); (3) Severe insulin resistance (n = 95); and (4) Severe insulin deficiency (SID) (n = 159). MARD and SID together accounted for most threshold-positive participants across the cardiovascular, kidney, and FIB-4 domains: (1) 75.6%; (2) 86.4%; and (3) 82.9%. Compared with mild obesity-related dysmetabolism, MARD showed the broadest age-loaded projected-risk pattern, with adjusted prevalence ratios of 1.94 (95%CI: 1.28-2.94) for Framingham risk, 8.75 (4.12-18.56) for CKD prognosis consortium risk, and 3.00 (1.76-5.13) for FIB-4. SID showed a nonclassical pattern, remaining enriched for projected CKD risk and FIB-4-defined fibrosis-risk signal despite lower adiposity, with adjusted prevalence ratios of 6.41 and 3.51, respectively (all P < 0.05).

CONCLUSION

Prediabetes subphenotypes show distinct projected cardio-kidney-liver risk patterns: MARD is broadly age-loaded, whereas SID marks a nonclassical kidney/FIB-4 signal requiring organ-outcome validation.

Key Words: Prediabetic state; Metabolic subphenotypes; Cluster analysis; Cardiovascular-kidney-metabolic syndrome; Cardiovascular diseases; Chronic renal insufficiency; Liver fibrosis; Insulin deficiency

Core Tip: Prediabetes is not a uniform metabolic state. In this Chinese multicenter cohort, cluster-defined subphenotypes showed distinct projected cardio–kidney–liver risk-score patterns. Mild age-related dysmetabolism represented a broad, age-loaded projected-risk subtype across Framingham, chronic kidney disease prognosis consortium, and fibrosis-4 domains. Severe insulin deficiency showed the most unexpected signal: Despite lower adiposity and lack of a classical insulin-resistant phenotype, it remained enriched for projected chronic kidney disease risk and fibrosis-4-defined liver fibrosis-risk signal. These findings connect prediabetes subphenotyping with an integrated organ-risk framework and support prospective validation of subtype-informed risk assessment before overt diabetes.

INTRODUCTION

Prediabetes represents a clinically meaningful intermediate metabolic state in which early pathophysiological changes extend beyond dysglycemia alone. Conventional definitions rely primarily on fasting glucose, oral glucose tolerance test (OGTT), or glycated hemoglobin (HbA1c) thresholds and therefore treat prediabetes as a single high-risk state preceding overt diabetes. However, such a glycemia-based framework may inadequately capture interindividual differences in β-cell function[1-3], insulin resistance, adiposity, age-related metabolic decline, and organ vulnerability, differences that may be associated with divergent downstream clinical trajectories[4-7].

Data-driven subclassification has reshaped how metabolic heterogeneity in dysglycemia is understood. Ahlqvist et al[8] first demonstrated that in adult-onset diabetes, cluster-defined metabolic subgroups differ in complication profiles and disease progression. Building on this framework, subsequent studies in prediabetes have shown that such subgroups already diverge in their risks of incident diabetes, vascular complications, and mortality. Rooney et al[6] further reported that prediabetes is associated with adverse clinical outcomes and that this risk is not primarily explained by subsequent progression to overt diabetes. However, most existing studies have primarily focused on diabetes progression or single-system outcomes, leaving it uncertain whether such metabolic heterogeneity[1-3] is already reflected in concordant or discordant multi-organ risk signatures at the prediabetes stage[4-7,9].

The concept of cardiovascular-kidney-metabolic (CKM) syndrome further emphasizes that obesity, dysglycemia, kidney dysfunction, and cardiovascular disease (CVD) are biologically interconnected rather than isolated conditions[10,11]. Within this framework, prediabetes may already lie within the early CKM continuum rather than representing only a glycemia-defined prediabetic stage. In parallel, metabolic dysfunction-associated steatotic liver disease (MASLD) has been increasingly recognized as an integral component of this network, particularly in relation to fibrosis progression and long-term hepatic outcomes[12-15]. Despite these advances, it remains unclear whether CKM-related risk profiles are already differentially distributed across metabolic subphenotypes of prediabetes, and whether liver-related fibrosis-risk signals follow patterns consistent with cardiovascular and renal risk.

However, important gaps remain. First, most studies of prediabetes subclassification have focused primarily on future diabetes risk rather than heterogeneity in cross-organ risk during the prediabetes stage itself. Second, although existing evidence suggests subgroup differences in complication risk, it remains unclear whether cardiovascular, kidney, and liver risk follow concordant or divergent patterns within the same prediabetes cohort under the same subtype framework, as this has not been systematically evaluated. This issue is particularly relevant when risk is assessed using established clinical tools that capture different organ domains, including Framingham-estimated cardiovascular risk, chronic kidney disease prognosis consortium (CKD-PC)-estimated projected incident chronic kidney disease (CKD) risk, and fibrosis-4 (FIB-4)-defined MASLD-related fibrosis-risk signal[16-20].

To address these gaps, we used data from a national multicenter program of adults at high risk of diabetes to identify metabolic subgroups in Chinese adults with prediabetes and to evaluate heterogeneity in cardiovascular, kidney, and liver risk. Specifically, we applied an unsupervised clustering approach to identify reproducible metabolic subphenotypes and compared Framingham-estimated 10-year cardiovascular risk, CKD-PC-estimated projected 5-year incident CKD risk, and FIB-4-defined MASLD-related risk across these subgroups. We hypothesized that metabolic subphenotypes within prediabetes would not share a uniform cross-organ risk projection pattern but instead already exhibit distinct and partially discordant cardio-kidney-liver risk-score signatures, consistent with early divergence in risk-score-derived organ-risk signatures before overt disease onset, thereby supporting precision prevention beyond glycemic thresholds alone.

MATERIALS AND METHODS

Study design and participants

This study was a cross-sectional analysis of baseline data from a prospective multicenter cohort of adults enrolled in the Intervention Study in Individuals at High Risk of Diabetes, a national high-risk diabetes program in China. It was also approved by the Clinical Research Ethics Committee of China-Japan Friendship Hospital (No. 2019-63-k43). Written informed consent was obtained from all participants.

At baseline, 2527 participants underwent standardized screening, including OGTT, HbA1c, insulin measurements, and routine biochemical testing. Individuals with baseline diabetes (fasting blood glucose ≥ 7.0 mmol/L, 2-hour OGTT glucose ≥ 11.1 mmol/L, or HbA1c ≥ 6.5%) were excluded. Prediabetes was defined as a fasting blood glucose concentration of 6.1-7.0 mmol/L and/or a 2-hour OGTT glucose concentration of 7.8-11.1 mmol/L. Clustering was restricted to participants whose fasting and 2-hour glucose measurements were available at baseline, and 621 participants were included in the final clustering sample (Supplementary Figure 1).

Metabolic subtype derivation

Metabolic subtypes were derived by applying the K-means algorithm to the standardized five-dimensional metabolic feature space defined by age, body mass index (BMI), HbA1c, homeostasis model assessment of insulin resistance (HOMA-IR), and homeostasis model assessment of β-cell function (HOMA-β), following the established subtype framework adapted for dysglycemia research[8]. As an unsupervised partitioning method, K-means assigns participants to mutually exclusive clusters by minimizing the within-cluster sum of squared Euclidean distances, without reference to any predefined labels. Right-skewed variables (BMI, HOMA-IR, and HOMA-β) were natural log-transformed to limit the influence of extreme values on the Euclidean distance computation, and all five clustering variables were then standardized to Z scores so that no single variable dominated the distance metric because of its scale. The optimal number of clusters was determined to be four by jointly considering internal validation indices and consensus clustering rather than relying on any single criterion. The optimal number of clusters was selected by jointly considering internal validation indices, clinical interpretability, comparability with the established dysglycemia subtype framework, and downstream stability analyses rather than relying on any single criterion. The elbow plot and Davies-Bouldin index supported K = 4, whereas the silhouette coefficient and Calinski-Harabasz index were interpreted as complementary diagnostics. Therefore, K = 4 was selected as the primary solution (Supplementary Figure 2).

The four clusters were labeled post hoc according to their dominant metabolic profiles to aid clinical interpretation. These labels were used as relative descriptors within the standardized prediabetes clustering space and were not intended to define diagnostic categories. Accordingly, severe insulin deficiency (SID) dysmetabolism denotes the subgroup with the lowest relative β-cell function and insulin-response profile among participants who met the study definition of prediabetes, rather than overt diabetes or absolute insulinopenia. Similarly, severe insulin resistance (SIR) dysmetabolism denotes relative insulin-resistant dysmetabolism within prediabetes, not insulin-resistant diabetes.

Cluster stability and reproducibility were assessed using two complementary approaches. First, bootstrap resampling (80% sampling ratio, 500 replicates) was used to compute the mean Jaccard index of each cluster as a measure of structural consistency under sample perturbation. Second, K-means was re-run 50 times with different random initializations, and the adjusted Rand index (ARI) was used to quantify agreement across runs. A mean Jaccard index > 0.75 together with an ARI approaching 1.0 was prespecified as the criterion for adequate stability. Biological validation of the post hoc subtype labels was performed by comparing OGTT-derived dynamic markers across subtypes, particularly 1-hour and 2-hour insulin concentrations, which were not entered into the clustering algorithm.

For subtype derivation, missingness in core clustering variables was handled using single completed-dataset iterative imputation based on chained-equation principles, which is restricted to variables with low missingness, to preserve the analytic sample for clustering.

Data cleaning and risk score preparation

Before organ-risk estimation, the variables required for risk score calculation were screened for implausible values according to clinically reasonable ranges; invalid observations were recoded as missing, whereas extreme but plausible values were winsorized using the 3 × interquartile range rule. For subtype derivation, missingness in core clustering variables was handled using single completed-dataset iterative imputation based on chained-equation principles; this procedure was restricted to variables with low missingness (< 10%) to preserve the analytic sample for clustering.

The algorithm was allowed to cycle for 10 iterations but generated one completed analytic dataset. The predictors in the imputation models included age, BMI, fasting glucose, 2-hour OGTT glucose, sex, and metabolic subtype; sex was used as a predictor only and was not imputed. Under the final analytic workflow, the main imputed variables were total cholesterol, triglyceride (TG), high-density lipoprotein cholesterol, low-density lipoprotein cholesterol (LDL-C), creatinine (Cr), systolic blood pressure, aspartate aminotransferase (AST), alanine aminotransferase (ALT), and platelet (PLT) count. Variable-level missingness and the number of winsorized observations are summarized in Supplementary Table 1.

Organ-risk estimation and threshold definition

Cardiovascular risk was estimated as Framingham-estimated 10-year CVD risk according to the original published equation[18]. A threshold of ≥ 10% was used as the primary analytic threshold for intermediate/high projected CVD risk; thresholds of ≥ 7.5%, ≥ 15%, and ≥ 20% were examined in sensitivity analyses.

Kidney risk was estimated using the CKD-PC nondiabetic equation for projected 5-year incident CKD risk[16,17]; we treated this as an adapted risk-projection application rather than a formal external validation. Chinese participants were coded as nonblack. Albumin-to-Cr ratio (ACR) was calculated from urinary albumin and urinary Cr when both measurements were available. For participants with positive measured ACR, log10-transformed ACR entered the CKD-PC equation through the ACR deviation term, defined as log10ACR minus expected log10ACR. The expected log10ACR term was used as the reference component required by the CKD-PC equation and was not treated as a generated ACR value for baseline description. For participants without a positive measured ACR, we applied the CKD-PC nondiabetic patch approach by setting the ACR deviation term to zero. Thus, generated ACR values were not combined with measured ACR values in the main analysis. Details of ACR availability and the measured-ACR-only sensitivity analysis are provided in Supplementary Tables. Because no validated clinical cutoff exists for CKD-PC-estimated 5-year risk in prediabetes, we used CKD-PC categories for analytic risk stratification rather than clinical diagnosis. A threshold of 5% was used for the primary intermediate/high-risk category because it provided sufficient events and stable between-subtype comparisons in this cohort; thresholds of ≥ 10%, ≥ P90, and ≥ P95 were examined in sensitivity analyses.

In the liver domain, FIB-4 was used as a noninvasive fibrosis-risk index rather than a diagnostic measure of MASLD burden. We therefore interpreted FIB-4 as a score-defined liver fibrosis-risk signal. We used FIB-4 ≥ 1.30 as the primary analytic threshold and examined FIB-4 ≥ 2.0, an age-aware threshold, and age-restricted analyses in sensitivity analyses[14,15,19,20].

Statistical analysis

Continuous variables are summarized as the mean ± SD or median (interquartile range), and categorical variables are summarized as n (%). For between-subtype comparisons, the Kruskal-Wallis test for skewed continuous variables and the χ² test or Fisher’s exact test for categorical variables were used, as appropriate. Organ-risk analyses included both continuous-risk comparisons and primary analytic threshold-defined intermediate/high-risk subtype comparisons. Threshold-based analyses were organized in three sequential layers: (1) Risk-category composition; (2) Source contribution of intermediate/high-risk individuals; and (3) Adjusted subtype comparisons.

Associations between subtype membership and threshold-defined outcomes were examined using Poisson regression with HC3 robust standard errors to estimate prevalence ratios (PRs) and 95%CI, with mild obesity-related dysmetabolism (MOD) as the reference group in the primary analysis. In addition to unadjusted models, predefined outcome-specific adjusted models were fitted. To limit overadjustment, direct components of the corresponding target risk equation were not reintroduced, and the remaining covariates were selected from complementary metabolic and cross-organ domains. Before finalizing the adjusted models, multicollinearity among candidate covariates was assessed using variance inflation factors. A variance inflation factor greater than 5 was considered indicative of potentially important collinearity; in the primary non-interaction adjusted models, all retained covariates were below this threshold.

Specifically, the CVD model included TG, AST, PLT, eGFR, LDL-C and waist circumference; the CKD model included AST, PLT, TG, LDL-C and waist circumference; and the FIB-4 model included waist circumference, sex, TG, eGFR, LDL-C and systolic blood pressure. Although outcome specific, these covariate sets were constructed under the same cross-organ-informed strategy: Variables already embedded in the corresponding target equation were not reintroduced, whereas complementary markers reflecting shared metabolic burden and adjacent organ-domain information were retained. This approach was intended to test whether the estimated subtype-risk-score patterns were robust beyond the original equation structure rather than to estimate independent causal effects. To examine whether the CKD-PC patch approach influenced the kidney-risk findings, we repeated the CKD-PC threshold analysis after restricting the sample to participants with positive measured ACR.

Sensitivity analyses included alternative risk thresholds, continuous-outcome models, complete-case analyses using observed non-imputed covariates, alternative reference-group analyses, included-versus-excluded comparisons, measured-ACR-only CKD analysis, and age-related robustness analyses.

Age-related robustness analyses retained the original subtype assignments and examined age using median-age stratification at 52 years, subtype-by-age-stratum interaction models, and subtype-by-continuous-age interaction models with age centered at 52 years. For FIB-4, additional age-aware and age-restricted analyses were performed because age is embedded directly in the FIB-4 equation. The measured-ACR-only CKD analysis was performed separately to assess whether the CKD-PC patch approach influenced the kidney-risk findings.

These age-related analyses were designed to characterize age loading and potential effect modification within the original subtype framework, not to eliminate structural age circularity completely.

For binary interaction models, we report the global Wald test for the joint interaction terms. For continuous interaction models, subtype main effects are interpreted as conditional contrasts at age 52 years, whereas interaction terms are interpreted as ratios of age slopes relative to MOD. When sparse or zero-event cells precluded stable estimation, subgroup-specific estimates were treated as exploratory and, where necessary, reported as not estimable. All analyses were performed in Python 3.11, and a two-sided P value < 0.05 was considered to indicate statistical significance.

RESULTS

Study population and outcome-specific analytic samples

Among 621 participants with prediabetes assigned to 4 metabolic subtypes, 177 were classified as having MOD, 95 as having severe insulin-resistant dysmetabolism (SIR), 159 as having severe insulin-deficient dysmetabolism (SID), and 190 as having mild age-related dysmetabolism (MARD). The subtype profiles were differentiated across the five clustering variables, and OGTT-derived measures that were not used for cluster formation also differed across subtypes. The largest between-subtype effects were observed for 1-hour and 2-hour insulin concentrations (ε² = 0.326 and 0.328, respectively; Supplementary Tables 2-4). Across 50 random initializations the mean ARI was 0.972 (range 0.954-1.000), and bootstrap mean Jaccard indices ranged from 0.779 to 0.847 across the four subtypes, indicating overall reproducible cluster assignments with moderate stability for the least stable subtype (Supplementary Tables 5 and 6). Because the required inputs differed across risk equations, the outcome-specific analytic samples were not identical. Framingham-estimated 10-year CVD risk, CKD-PC-estimated projected 5-year CKD risk, and the FIB-4 index were available for 452 participants, 448 participants, and 621 participants, respectively. The corresponding participant-identifier-aligned adjusted-model samples included 452 participants, 448 participants, and 585 participants. In the kidney-risk sample, 426 participants had positive measured ACR values, whereas 22 were handled using the nondiabetic patch approach (Supplementary Table 7). Cluster composition did not differ significantly between participants included in and excluded from any of the three adjusted-model samples, although several individual characteristics differed between the included and excluded groups (Supplementary Table 8 and Supplementary Figure 1).

Across the covariates used in the outcome-specific adjusted models, 216 variable-level values were imputed, including 18 values recoded as missing after implausible-value cleaning, and 15 observed extreme values were winsorized. No missing values remained in these covariates after data processing (Supplementary Table 1). Baseline characteristics of the full clustering sample are presented in Table 1.

Table 1 Baseline characteristics and organ risk measures across cluster-defined metabolic subtypes, mean ± SD/median (interquartile range)/n (%).

Characteristic	Overall	MARD	MOD	SID	SIR	P value
	(n = 621)	(n = 190)	(n = 177)	(n = 159)	(n = 95)
Female sex	319 (51.4)	100 (52.6)	81 (45.8)	91 (57.2)	47 (49.5)	0.197
Age (years)	51.08 ± 10.14	57.85 ± 6.82	44.04 ± 7.84	55.42 ± 7.90	43.39 ± 9.33	< 0.001
BMI (kg/m2)	25.20 (23.40, 27.40)	25.90 (24.49, 27.38)	25.09 (23.70, 26.50)	23.20 (21.40, 24.60)	29.30 (27.38, 31.90)	< 0.001
Waist circumference (cm)	88.92 ± 9.69	91.03 ± 8.29	87.25 ± 7.79	82.50 ± 7.78	98.57 ± 9.39	< 0.001
SBP (mmHg)	124.04 ± 15.53	127.47 ± 14.45	120.68 ± 13.91	121.93 ± 17.85	126.81 ± 14.75	< 0.001
Diastolic blood pressure (mmHg)	77.91 ± 9.95	77.66 ± 9.32	77.83 ± 10.02	75.81 ± 9.80	81.82 ± 10.25	< 0.001
Fasting glucose (mmol/L)	5.96 ± 0.62	6.00 ± 0.56	5.81 ± 0.66	6.07 ± 0.60	5.95 ± 0.65	0.002
2-hour glucose (mmol/L)	8.73 ± 1.38	8.79 ± 1.38	8.61 ± 1.36	8.55 ± 1.45	9.15 ± 1.21	0.013
1-hour glucose (mmol/L)	11.28 ± 2.20	11.54 ± 2.06	10.63 ± 2.16	11.36 ± 2.42	11.83 ± 1.93	< 0.001
HbA1c (%)	5.80 ± 0.34	5.97 ± 0.23	5.51 ± 0.31	5.91 ± 0.27	5.84 ± 0.33	< 0.001
Fasting insulin (0 minute)	10.20 (7.21, 14.43)	12.84 (10.12, 15.72)	9.53 (7.50, 11.44)	6.29 (4.84, 7.44)	21.99 (16.79, 27.87)	< 0.001
Insulin (1 hour)	74.77 (47.34, 116.47)	95.61 (65.53, 135.60)	63.24 (45.91, 89.81)	45.57 (32.63, 67.96)	144.85 (99.99, 180.48)	< 0.001
Insulin (2 hours)	72.96 (46.57, 113.40)	91.33 (61.30, 131.45)	65.01 (46.50, 101.10)	45.31 (31.41, 68.39)	154.40 (101.35, 215.55)	< 0.001
HOMA-IR	2.68 (1.86, 3.89)	3.44 (2.63, 4.32)	2.50 (1.82, 3.06)	1.67 (1.29, 2.08)	5.62 (4.13, 7.56)	< 0.001
HOMA-β cell function	85.71 (58.59, 126.32)	104.34 (83.28, 128.07)	81.12 (64.17, 106.39)	48.66 (37.37, 58.26)	165.29 (134.71, 238.79)	< 0.001
TG (mmol/L)	1.44 (1.04, 2.06)	1.59 (1.19, 2.23)	1.31 (0.96, 2.01)	1.17 (0.90, 1.54)	1.81 (1.32, 2.51)	< 0.001
TC (mmol/L)	4.81 (4.20, 5.51)	4.78 (4.10, 5.45)	4.75 (4.20, 5.33)	4.90 (4.31, 5.74)	4.78 (4.29, 5.47)	0.248
LDL-C (mmol/L)	3.09 (2.54, 3.59)	3.07 (2.54, 3.58)	3.05 (2.49, 3.55)	3.12 (2.56, 3.73)	3.12 (2.67, 3.45)	0.601
AST (U/L)	20.00 (17.00, 24.00)	21.00 (18.00, 25.00)	19.00 (17.00, 23.00)	20.00 (17.00, 22.00)	22.00 (18.00, 28.00)	< 0.001
ALT (U/L)	22.00 (16.00, 30.00)	24.00 (18.00, 32.00)	21.00 (16.00, 29.00)	18.00 (15.00, 23.00)	30.57 (20.00, 44.00)	< 0.001
PLT count (× 109/L)	236.00 (206.00, 273.00)	229.34 (199.25, 260.88)	241.00 (212.00, 281.00)	229.62 (196.50, 258.50)	259.00 (218.50, 296.50)	< 0.001
HDL cholesterol (mmol/L)	1.27 (1.08, 1.50)	1.24 (1.07, 1.42)	1.25 (1.09, 1.50)	1.42 (1.23, 1.62)	1.13 (1.00, 1.29)	< 0.001
10-year CVD risk (%)	6.13 (3.17, 11.26)	10.28 (5.68, 16.55)	3.39 (2.02, 7.61)	6.13 (3.97, 10.16)	5.46 (2.64, 8.23)	< 0.001
FIB-4 index	0.97 (0.69, 1.29)	1.11 (0.90, 1.41)	0.77 (0.59, 1.01)	1.15 (0.96, 1.40)	0.67 (0.47, 0.89)	< 0.001
Cr (μmol/L)	69.00 (59.20, 80.20)	71.00 (59.90, 81.40)	70.00 (59.40, 80.20)	66.10 (58.35, 77.70)	69.20 (59.85, 78.85)	0.290
Measured ACR (mg/g)	5.59 (3.19, 10.57)	6.68 (3.36, 12.63)	4.76 (3.00, 9.24)	5.23 (2.82, 9.81)	7.35 (3.72, 16.95)	0.018
eGFR (mL/minute/1.73 m2)	97.76 (88.75, 105.56)	92.55 (83.14, 98.64)	104.11 (96.86, 111.23)	94.97 (85.42, 102.37)	102.59 (93.61, 112.72)	< 0.001
CKD-PC projected CKD risk (5 years) (%)	2.32 (0.97, 5.26)	4.86 (2.77, 8.43)	0.95 (0.49, 1.89)	3.60 (1.72, 6.15)	1.11 (0.63, 2.69)	< 0.001

P values were derived from the χ² test or Fisher’s exact test for categorical variables and the Kruskal-Wallis test for continuous variables, as appropriate. Score-derived organ-risk measures were calculated in outcome-specific analytic samples: (1) Framingham-estimated 10-year cardiovascular disease risk, n = 452; (2) Chronic kidney disease prognosis consortium-projected 5-year chronic kidney disease risk, n = 448; and (3) Fibrosis-4 index, n = 621. Albumin-to-creatinine ratio was summarized from measured urinary albumin and urinary creatinine values. ACR: Albumin-to-creatinine ratio; AST: Aspartate aminotransferase; ALT: Alanine aminotransferase; CKD: Chronic kidney disease; CKD-PC: Chronic kidney disease prognosis consortium; CVD: Cardiovascular disease; eGFR: Estimated glomerular filtration rate; FIB-4: Fibrosis-4; HDL: High-density lipoprotein; HOMA-IR: Homeostasis model assessment of insulin resistance; HOMA-β: Homeostasis model assessment of β-cell function; LDL-C: Low-density lipoprotein cholesterol; TC: Total cholesterol; TG: Triglyceride; Cr: Creatinine; MARD: Mild age-related dysmetabolism; MOD: Mild obesity-related dysmetabolism; SID: Severe insulin deficiency; SIR: Severe insulin resistance; BMI: Body mass index; SBP: Systolic blood pressure; HbA1c: Glycated hemoglobin; PLT: Platelet.

Continuous organ-risk distributions across metabolic subtypes

Before application of the analytic thresholds, all three continuous risk measures differed across metabolic subtypes. Median Framingham-estimated 10-year CVD risk was highest in MARD (10.28%), followed by SID (6.13%), SIR (5.46%), and MOD (3.39%; Kruskal-Wallis H = 78.56, P < 0.001). Median projected 5-year CKD risk was also highest in MARD [4.86% (interquartile range, 2.77%-8.43%)] and SID [3.60% (1.72%-6.15%)], compared with SIR [1.12% (0.63%-2.69%)] and MOD [0.95% (0.49%-1.89%); H = 161.35, P < 0.001]. Median FIB-4 values were highest in SID (1.15) and MARD (1.11), and lower in MOD (0.77) and SIR (0.67; H = 139.48, P < 0.001). Detailed distributions and pairwise comparisons are provided in Figure 1 and Supplementary Table 9.

Figure 1 Distributions of continuous risk-score-derived organ-risk measures across metabolic subtypes. A: Cardiovascular disease 10-year risk; B: Chronic kidney disease (CKD) risk probability; C: Fibrosis-4 (FIB-4) index. Violin plots show the subtype-specific distributions of Framingham-estimated 10-year cardiovascular disease (CVD) risk, CKD prognosis consortium-estimated projected 5-year CKD risk, and FIB-4 index among adults with prediabetes assigned to four metabolic subtypes. Embedded box plots indicate the interquartile range, whiskers extend to 1.5 × interquartile range, and white circles indicate medians. The continuous distributions differed across subtypes for all three measures by Kruskal-Wallis test (all P < 0.001). Higher values indicate greater risk-score-derived CVD risk, projected CKD risk, or FIB-4-defined liver fibrosis-risk signal. CVD risk was analyzed in 452 participants, projected CKD risk in 448 participants, and FIB-4 in 621 participants. CVD: Cardiovascular disease; CKD: Chronic kidney disease; FIB-4: Fibrosis-4; MARD: Mild age-related dysmetabolism; MOD: Mild obesity-related dysmetabolism; SID: Severe insulin deficiency; SIR: Severe insulin resistance.

At the primary analytic thresholds, 131 of 452 participants (29.0%) had Framingham-estimated CVD risk ≥ 10%, 117 of 448 (26.1%) had projected 5-year CKD risk ≥ 5%, and 152 of 621 (24.5%) had a FIB-4 index ≥ 1.30. Risk-category composition differed across subtypes for all three measures (all P < 0.001; Supplementary Table 10).

The distribution of threshold-positive participants was concentrated in MARD and SID. MARD accounted for 54.2% of participants with Framingham risk ≥ 10%, 55.6% of those with projected CKD risk ≥ 5%, and 44.7% of those with FIB-4 ≥ 1.30. The corresponding contributions from SID were 21.4%, 30.8%, and 38.2%. Together, MARD and SID accounted for 75.6%, 86.4%, and 82.9% of the threshold-defined cardiovascular, kidney, and liver fibrosis-risk burdens, respectively. Within-subtype event rates were highest in MARD for the cardiovascular and kidney measures, whereas MARD and SID had similarly elevated rates for FIB-4 ≥ 1.30 (Figure 2 and Supplementary Table 11).

Figure 2 Within-subtype prevalence and source contribution of threshold-defined intermediate/high-risk status across metabolic subtypes. A: It shows the percentage of participants within each metabolic subtype who met the primary analytic threshold for each risk-score-derived measure; B: It shows the proportion of all threshold-positive participants accounted for by each subtype for the same measure. Primary analytic thresholds were defined as Framingham-estimated 10-year cardiovascular disease risk ≥ 10%, chronic kidney disease (CKD) prognosis consortium-estimated projected 5-year CKD risk ≥ 5%, and fibrosis-4 (FIB-4) index ≥ 1.30. The corresponding threshold-positive counts were 131 of 452 participants for cardiovascular disease risk, 117 of 448 participants for projected CKD risk, and 152 of 621 participants for FIB-4. Percentages are rounded to one decimal place and may not sum exactly to 100.0%. These categories represent score-defined intermediate/high-risk status and FIB-4-defined liver fibrosis-risk signal, not observed incident cardiovascular, kidney, or liver events. CVD: Cardiovascular disease; CKD: Chronic kidney disease; FIB-4: Fibrosis-4; MARD: Mild age-related dysmetabolism; MOD: Mild obesity-related dysmetabolism; SID: Severe insulin deficiency; SIR: Severe insulin resistance.

Adjusted comparisons of threshold-defined intermediate/high-risk outcomes across subtypes

Table 2 presents the unadjusted and outcome-specific adjusted Poisson regression models for the three threshold-defined outcomes. The cardiovascular, kidney, and FIB-4 analyses included 452 participants with 131 threshold-positive cases, 448 participants with 117 threshold-positive cases, and 585 participants with 150 threshold-positive cases in the adjusted models, respectively.

Table 2 Exploratory Poisson regression analyses of threshold-defined intermediate/high-risk outcomes across organ domains, n (%).

Subtype	Model	n	Participants meeting threshold	PR	95%CI (PR)	P value
Framingham-estimated 10-year CVD risk ≥ 10%
MOD (Ref)	Model 1	132	22 (16.7)	Ref	-	-
MARD	Model 1	138	71 (51.4)	3.087	2.040-4.672	< 0.001
SID	Model 1	108	28 (25.9)	1.556	0.946-2.557	0.082
SIR	Model 1	74	10 (13.5)	0.811	0.406-1.618	0.552
MOD (Ref)	Model 2	132	22 (16.7)	Ref	-	-
MARD	Model 2	138	71 (51.4)	1.942	1.283-2.940	0.002
SID	Model 2	108	28 (25.9)	1.805	1.116-2.921	0.016
SIR	Model 2	74	10 (13.5)	0.399	0.206-0.772	0.006
CKD-PC-estimated projected 5-year CKD risk ≥ 5%
MOD (Ref)	Model 1	133	7 (5.3)	Ref	-	-
MARD	Model 1	133	65 (48.9)	9.286	4.423-19.495	< 0.001
SID	Model 1	107	36 (33.6)	6.393	2.964-13.787	< 0.001
SIR	Model 1	75	9 (12.0)	2.280	0.885-5.874	0.088
MOD (Ref)	Model 2	133	7 (5.3)	Ref	-	-
MARD	Model 2	133	65 (48.9)	8.748	4.123-18.560	< 0.001
SID	Model 2	107	36 (33.6)	6.407	2.990-13.728	< 0.001
SIR	Model 2	75	9 (12.0)	2.095	0.790-5.555	0.137
FIB-4 ≥ 1.30
MOD (Ref)	Model 1	177	17 (9.6)	Ref	-	-
MARD	Model 1	190	68 (35.8)	3.726	2.282-6.085	< 0.001
SID	Model 1	159	58 (36.5)	3.798	2.312-6.239	< 0.001
SIR	Model 1	95	9 (9.5)	0.986	0.457-2.127	0.972
MOD (Ref)	Model 2	165	16 (9.7)	Ref	-	-
MARD	Model 2	177	67 (37.9)	3.002	1.758-5.127	< 0.001
SID	Model 2	149	58 (38.9)	3.505	2.100-5.849	< 0.001
SIR	Model 2	94	9 (9.6)	0.912	0.410-2.028	0.821

The n column gives the subtype-specific denominator in the corresponding model sample. Participants meeting threshold are presented as n (%), with percentages calculated using the subtype-specific denominator; these values represent score-defined risk classifications rather than observed clinical outcomes. Poisson regression with HC3 robust standard errors was used to estimate prevalence ratios and 95%CI, with mild obesity-related dysmetabolism as the reference group. Model 1 was unadjusted. Revised model 2 included waist circumference, aspartate aminotransferase, platelet count, estimated glomerular filtration rate, triglyceride (TG), and low-density lipoprotein cholesterol (LDL-C) for the cardiovascular disease analysis; waist circumference, LDL-C, TG, aspartate aminotransferase, and platelet count for the chronic kidney disease analysis; and waist circumference, sex, systolic blood pressure, estimated glomerular filtration rate, TG, and LDL-C for the fibrosis-4 analysis. Covariates were aligned by participant identifier before model fitting. CKD: Chronic kidney disease; CKD-PC: Chronic kidney disease prognosis consortium; CVD: Cardiovascular disease; FIB-4: Fibrosis-4; MARD: Mild age-related dysmetabolism; MOD: Mild obesity-related dysmetabolism; PR: Prevalence ratio; SID: Severe insulin deficiency; SIR: Severe insulin resistance; Ref: Reference.

In the unadjusted models, MARD had a higher prevalence of all three threshold-defined outcomes than MOD. The PRs were 3.09 (95%CI: 2.04-4.67; P < 0.001) for Framingham-estimated 10-year CVD risk ≥ 10%, 9.29 (4.42-19.50; P < 0.001) for CKD-PC-estimated projected 5-year CKD risk ≥ 5%, and 3.73 (2.28-6.09; P < 0.001) for FIB-4 ≥ 1.30. SID was associated with projected kidney risk ≥ 5% (PR = 6.39; 95%CI: 2.96-13.79; P < 0.001) and FIB-4 ≥ 1.30 (PR = 3.80; 95%CI: 2.31-6.24; P < 0.001), whereas its association with Framingham risk ≥ 10% did not reach statistical significance (PR = 1.56; 95%CI: 0.95-2.56; P = 0.082). SIR did not differ significantly from MOD in any of the unadjusted models.

After adjustment for the prespecified outcome-specific covariates, MARD remained associated with all three outcomes. Compared with MOD, the adjusted PRs were 1.94 (95%CI: 1.28-2.94; P = 0.002) for Framingham risk ≥ 10%, 8.75 (4.12-18.56; P < 0.001) for projected CKD risk ≥ 5%, and 3.00 (1.76-5.13; P < 0.001) for FIB-4 ≥ 1.30. SID was also associated with all three outcomes, with adjusted PRs of 1.81 (95%CI: 1.12-2.92; P = 0.016), 6.41 (2.99-13.73; P < 0.001), and 3.51 (2.10-5.85; P < 0.001), respectively. For SIR, the adjusted prevalence of Framingham risk ≥ 10% was lower than that in MOD (PR = 0.40; 95%CI: 0.21-0.77; P = 0.006), whereas no significant differences were observed for projected CKD risk ≥ 5% (PR = 2.10; 95%CI: 0.79-5.56; P = 0.137) or FIB-4 ≥ 1.30 (PR = 0.91; 95%CI: 0.41-2.03; P = 0.821) (Figure 3).

Figure 3 Prevalence ratios for threshold-defined intermediate/high-risk status across metabolic subtypes. Forest plot shows prevalence ratios (PRs) and 95%CIs for the association between metabolic subtype and each threshold-defined risk-score-derived outcome, with mild obesity-related dysmetabolism as the reference group. Model 1 was unadjusted. Model 2 used outcome-specific adjusted covariates that were not direct components of the corresponding target risk equation: (1) Triglyceride (TG), aspartate aminotransferase, platelet count, estimated glomerular filtration rate, low-density lipoprotein cholesterol (LDL-C), and waist circumference for cardiovascular disease risk; (2) Aspartate aminotransferase, platelet count, TG, LDL-C, and waist circumference for projected chronic kidney disease (CKD) risk; and (3) Waist circumference, sex, TG, estimated glomerular filtration rate, LDL-C, and systolic blood pressure for fibrosis-4 (FIB-4). PRs greater than 1 indicate a higher prevalence of threshold-positive status relative to mild obesity-related dysmetabolism. 95%CIs crossing 1 indicate that the comparison did not reach statistical significance at the two-sided 0.05 level. The adjusted-model analytic samples included 452 participants for Framingham-estimated 10-year cardiovascular disease risk ≥ 10%, 448 participants for CKD prognosis consortium-estimated projected 5-year CKD risk ≥ 5%, and 585 participants for FIB-4 ≥ 1.30. Estimates describe score-derived projected risk categories and FIB-4-defined liver fibrosis-risk signal rather than observed incident cardiovascular, kidney, or liver outcomes. CVD: Cardiovascular disease; CKD: Chronic kidney disease; FIB-4: Fibrosis-4; MARD: Mild age-related dysmetabolism; MOD: Mild obesity-related dysmetabolism; SID: Severe insulin deficiency; SIR: Severe insulin resistance; Ref: Reference.

Adjustment attenuated the MARD estimates for cardiovascular and FIB-4-defined risk, whereas the kidney-risk estimates for MARD and SID changed little. For SID, the adjusted cardiovascular estimate was larger than the unadjusted estimate and its confidence interval no longer included 1. All variance inflation factors in the primary adjusted models were below 2.0, with a maximum value of 1.855 (Supplementary Table 12).

Sensitivity analyses

The main subtype pattern was retained across alternative risk thresholds, continuous-outcome models, complete-case analyses using observed non-imputed covariates, and alternative reference-group analyses; estimates at stringent thresholds were less precise because of sparse threshold-positive cases (Supplementary Tables 13-16).

Because age was included in subtype derivation and all three risk tools are age-loaded to different degrees, we performed additional age-related sensitivity analyses while retaining the original subtype assignments.

In CVD analyses, age composition was markedly imbalanced across subtypes: MARD and SID were concentrated in the older stratum, whereas MOD and SIR were concentrated in the younger stratum. In median-age-stratified analyses, the MARD signal was more evident below age 52 and attenuated at older ages, whereas the younger SID stratum contained no threshold-positive cases. In the continuous-age model, the global subtype-by-age interaction was significant (Wald χ² = 23.674, df = 3, P = 2.92 × 10^-5), driven mainly by a negative MARD-by-age interaction (interaction PR per year, 0.908; 95%CI: 0.869-0.948). These findings indicate that the projected CVD pattern was strongly age-dependent and that the overall MARD signal should not be interpreted as an age-independent subtype effect (Supplementary Tables 17-26).

In CKD analyses, 108 of 117 CKD-PC-defined intermediate/high-risk events (92.31%) occurred at age ≥ 52 years. Within the ≥ 52-year stratum, the adjusted PRs were 3.69 (95%CI: 1.47-9.27) for MARD and 3.06 (1.20-7.80) for SID relative to MOD, but both contrasts were attenuated relative to the full-sample analysis. Neither the subtype-by-age-stratum interaction (P = 0.902) nor the subtype-by-continuous-age interaction (P = 0.500) was significant. Thus, the projected CKD pattern was strongly age-loaded, but we found no formal evidence that the relative subtype contrasts differed multiplicatively across age (Supplementary Tables 27-34).

The measured-ACR-only sensitivity analysis yielded a similar pattern. After excluding participants who required the CKD-PC patch approach, 113 of 426 participants met the CKD-PC projected 5-year risk threshold of ≥ 5%. In the adjusted model, MARD and SID remained associated with projected CKD risk ≥ 5% compared with MOD, with PRs of 8.63 (95%CI: 4.07-18.31) and 5.99 (2.78-12.89), respectively. SIR was not statistically significant (PR = 2.11; 95%CI: 0.80-5.62; P = 0.134). These findings indicate that the projected CKD-risk pattern was not driven by the patch approach.

Overall, MARD remained the most consistently age-loaded source of projected intermediate/high-risk burden across the three domains, whereas SID retained an additional projected kidney and liver fibrosis-risk signal that was more robust to the examined age-related sensitivity analyses (Supplementary Tables 35 and 36).

For FIB-4 ≥ 1.30, age-related analyses indicated that part of the MARD signal was age-loaded, whereas the SID signal was less fully explained by age. Descriptively, MARD and SID were older than MOD and SIR and had higher median FIB-4 values, supporting the need for age-aware sensitivity analyses. Applying the age-aware FIB-4 definition reduced threshold-positive counts in MARD and SID, but not in MOD or SIR. In the full-sample age-aware model, MARD and SID remained higher than MOD, with adjusted PRs of 2.36 (95%CI: 1.33-4.19) and 3.10 (1.80-5.32), respectively, whereas SIR remained unassociated with the FIB-4-defined higher-risk stratum. The age-restricted analysis among participants aged 35-64 years showed a similar direction: MARD and SID remained higher than MOD, whereas SIR did not. In median-age-stratified models, the younger stratum was sparse, including a non-estimable adjusted MARD contrast; in the older stratum, SID remained higher than MOD, whereas the MARD estimate was attenuated and less precise. In continuous-age interaction models, the global interaction was not significant for the standard FIB-4 threshold after adjustment (χ² = 5.608, df = 3, P = 0.132), but became significant under the age-aware threshold (χ² = 12.569, df = 3, P = 0.006). Subtype-specific age slopes and representative-age marginal estimates supported a cautious interpretation: Age shaped the FIB-4-defined signal, but the age-aware and age-restricted analyses did not shift the pattern toward SIR (Supplementary Tables 37-46).

DISCUSSION

In this baseline cross-sectional analysis, cluster-defined prediabetes subphenotypes differed in risk-score-derived profiles across cardiovascular, kidney, and FIB-4-defined liver fibrosis domains (Supplementary Figure 3). MARD accounted for the broadest age-loaded intermediate/high projected-risk burden across the three domains. SID, characterized in our cohort by the lowest HOMA-β and fasting and post-load insulin concentrations despite prediabetes-range glycemia, was also enriched for projected CKD risk and the FIB-4-defined liver fibrosis-risk signal. This interpretation is consistent with the Japan Diabetes Prevention Study, which identified an insulin-insufficient prediabetes cluster with reduced β-cell function and fasting and 2-hour insulin concentrations of 6.7 ± 2.0 μIU/mL and 49.1 ± 26.6 μIU/mL, respectively, values close to those observed in our SID subgroup[21]. These findings suggest that glycemic status alone may not fully capture score-derived organ-risk heterogeneity in prediabetes.

Framingham-estimated cardiovascular risk: Age loading in MARD and additional SID enrichment

For the Framingham-estimated 10-year CVD risk threshold, MARD contributed the largest share of the intermediate/high projected-risk group; SID was also enriched, whereas SIR was not.

In our data, MARD had the highest within-subtype proportion above the Framingham 10% threshold (51.4%) and contributed 54.2% of all individuals in this risk stratum. In the outcome-specific adjusted model, MARD remained higher than MOD (PR = 1.94; 95%CI: 1.28-2.94; P = 0.002).

SID was also higher than MOD (PR = 1.81; 95%CI: 1.12-2.92; P = 0.016), whereas SIR showed a lower Framingham-estimated projected risk (PR = 0.40; 95%CI: 0.21-0.77; P = 0.006). The predominance of MARD is most plausibly explained by its age-loaded risk structure because age remains major component of the most cardiovascular risk equations. MARD was the oldest subgroup in our cohort, and this age profile would be expected to shift Framingham-estimated risk upward. This interpretation is biologically plausible: Vascular aging is accompanied by reduced arterial compliance, increased arterial stiffness, and increased susceptibility to dysglycemia-related injury[22]. Age-related decreases in β-cell function may further worsen glycemic exposure and reinforce this pattern[23-25].

More importantly, the SID also remained enriched in the overall adjusted projection model. Given its lower adiposity but greater glycemic exposure, the SID pattern suggests that the Framingham-defined signal was not explained by obesity alone. This pattern is biologically compatible with, but does not directly demonstrate, pathways involving impaired insulin signaling and glycemic stress whereas persistent hyperglycemia can further promote oxidative and glycation-related vascular injury[26,27].

In contrast, SIR did not emerge as a major source of Framingham-defined intermediate/high risk in this analysis. This should not be interpreted as evidence of low vascular vulnerability; rather, it more likely reflects the younger age structure of SIR, shorter cumulative glycemic exposure, and the fact that Framingham risk is a projection tool dominated by conventional short-term risk determinants. In this sense, insulin resistance alone may be insufficient to place a prediabetes subgroup at the top of projected short-term cardiovascular risk. This interpretation is broadly consistent with prior subtype analyses showing that the strongest prediagnostic blood pressure enrichment was not concentrated in insulin-resistant subgroups alone.

Thus, in this cohort, the projected cardiovascular intermediate/high-risk burden appeared to be shaped more by age loading and glycemic exposure than by insulin resistance alone.

Projected CKD risk: Enrichment in MARD and SID

Intermediate/high CKD-PC projected risk was concentrated in MARD and SID rather than in SIR, suggesting that this kidney-risk projection did not simply follow insulin resistance severity. When the ≥ 5% threshold was used, the within-cluster proportion of intermediate/high projected risk was highest in MARD (48.9%), followed by that in SID (33.6%), whereas the corresponding proportions were much lower in SIR (12.0%) and MOD (5.3%). ACR handling is important for interpreting this result. Most participants in the CKD-PC analytic sample had positive measured ACR, and only a small proportion required the nondiabetic patch approach. The measured-ACR-only sensitivity analysis showed the same direction, with MARD and SID remaining the main projected CKD-risk subtypes. Therefore, the kidney-risk signal should be interpreted as a CKD-PC-derived projected risk pattern, not as a finding driven by the ACR patch approach.

The MARD pattern in CKD-PC risk also fits its older age profile. Aging is accompanied by loss of functional nephrons, increased glomerulosclerosis, and worsening interstitial fibrosis, rendering the kidney more vulnerable to metabolic insults[28]. In that setting, dysglycemia and blood pressure may contribute to higher projected kidney-risk profiles. Previous evidence has linked prediabetes with CKD-related progression or comorbidity risk, although reported endpoints and effect sizes differ across studies[29,30].

SID reached the same projected-risk tier, but its profile was different from MARD. Rather than age loading alone, the SID pattern is more consistent with secretory defects accompanied by greater glycemic exposure. This is important because dysglycemia-related renal injury is shaped not only by mean glucose levels but also by broader metabolic and hemodynamic stress[31-33]. Mechanistically, altered insulin signaling may affect tubular homeostasis, and tubular metabolic stress may be involved in kidney disease development; however, these pathways were not tested in this study[31-33]. In addition, prior work has suggested a potentially bidirectional amplification between declining kidney function and β-cell dysfunction, whereby renal metabolic disturbance may further impair insulin secretion and thereby reinforce dysglycemia-related renal vulnerability[34]. Accordingly, the SID signal is better described as projected kidney-risk enrichment linked to glycemic exposure rather than to adiposity or insulin resistance alone.

In contrast, SIR did not emerge as the main source of intermediate/high CKD-PC risk in this analysis. This does not mean that insulin-resistant prediabetes is renally benign. It more likely reflects the CKD-PC equation, which gives substantial weight to conventional renal risk factors such as age, eGFR, and blood pressure. This distinction may also explain why our pattern differs from that of subtype studies on the basis of observed diabetic kidney outcomes rather than external risk projection[35]. In our cohort, SIR participants were younger and had higher baseline eGFR despite greater adiposity, insulin resistance, and ACR burden; these features may have lowered their short-term CKD-PC projection. Thus, the relatively modest SIR signal here is best interpreted as a tool-dependent projection pattern, not as evidence of low renal vulnerability; longer follow-up with observed kidney outcomes is needed. Accordingly, the renal pattern is better interpreted as a risk projection driven by differential loading of conventional kidney risk factors plus glycemic vulnerability rather than a linear extrapolation of any single metabolic abnormality.

FIB-4-defined liver fibrosis-risk signal: A nonclassical SID pattern

The FIB-4-defined liver fibrosis-risk signal did not follow insulin resistance severity. SID and MARD accounted for most individuals above the FIB-4 threshold, whereas SIR was not enriched. In our data, SID and MARD together accounted for 82.9% of FIB-4-defined intermediate/high-risk cases.

This pattern should first be interpreted in light of the tool itself. FIB-4 does not quantify steatosis burden, diagnose MASLD, measure MASH, or establish MASLD prevalence or incidence; in this study, it was used as a noninvasive fibrosis-risk signal[14,15,19,20]. Accordingly, divergence across subtypes should not be read as a direct ranking of fatty liver burden but as differential mapping to the FIB-4-defined higher fibrosis-risk stratum.

SIR had features of a more conventional adiposity- and insulin-resistance-related metabolic phenotype, but it was not enriched for the FIB-4-defined higher-risk stratum. Previous data-driven liver and diabetes-subtype studies have linked metabolic phenotypes to MASLD heterogeneity and liver-related outcomes[36,37]. In contrast, our analysis evaluated a FIB-4-defined fibrosis-risk signal, not steatosis burden. Our baseline data were consistent with this metabolic profile: SIR had higher adiposity, TG, ALT, insulin levels, and HOMA-IR. Higher adiposity and ALT therefore should not be taken as equivalent to a higher FIB-4-defined fibrosis-risk signal.

The SID pattern was less expected. Despite lacking the typical lipotoxic profile of SIR, SID was more strongly represented in the FIB-4-defined higher-risk stratum. In our cohort, SID had lower BMI, waist circumference, TG, ALT, insulin, and HOMA-β levels, but higher fasting glucose and HbA1c, consistent with an insulin-deficient, higher-glycemic-exposure profile. Current evidence links diabetes and hyperglycemia with MASH/MASLD progression, although our study did not directly test this pathway[38]. Thus, the SID finding is better described as a higher-glycemic-exposure FIB-4 pattern rather than a typical steatotic pattern. The MARD signal likely reflects age loading in part because age is embedded directly in the FIB-4 equation. Even so, the persistence of the association after adjustment suggests that age structure alone may not fully account for MARD enrichment.

The main point is not that SID has a greater MASLD burden than SIR, but that SID was more strongly represented in the FIB-4-defined liver fibrosis-risk stratum.

Cross-organ pattern and clinical implications

These subtypes did not distribute intermediate/high-risk burdens evenly across organs; instead, they showed discordant, tool-dependent patterns across cardiovascular, kidney, and FIB-4-defined liver fibrosis-risk domains. This finding is consistent with the CKM continuum and its liver-inclusive extension, in which a shared metabolic background may be expressed differently across organ systems and across the specific tools applied[10-12,39]. The findings do not support subtype-specific treatment. They suggest that subtype information may help prioritize organ-oriented screening after prospective validation. In future validation studies, MARD should be examined as a broad age-loaded projected-risk subtype; SID as a subtype with projected CKD risk and FIB-4-defined fibrosis-risk enrichment despite lower adiposity; and SIR as a subtype in which lower short-term projected risk should not be equated with low long-term organ risk[10-12,39].

Limitations and future directions

This study mapped baseline cross-organ risk projections rather than observed organ outcomes. Accordingly, the subtype differences described here should be interpreted as divergence in projected cardiovascular, kidney, and FIB-4-defined fibrosis-risk signal rather than direct evidence of established organ injury or future clinical events.

Second, applying external tools to subtype-defined phenotypes introduces an important interpretive constraint. The Framingham equation, CKD-PC equation, and FIB-4 were developed in specific populations and within predefined variable structures; therefore, higher projected risk in a given subtype may partly reflect differential loading of conventional risk determinants rather than stronger intrinsic pathogenicity of the subtype itself[16-20]. This is particularly relevant for CKD-PC, which estimates 5-year incident CKD risk rather than existing renal pathology, and for FIB-4, which is a partly age-loaded fibrosis-risk index rather than a measure of steatosis burden, MASLD prevalence, MASLD incidence or MASH.

Age also requires caution because it contributed to subtype derivation and is embedded directly or indirectly in the three risk tools used here. Although the age-stratified, age-interaction, and age-aware FIB-4 analyses helped characterize this issue, they do not remove structural age loading from the original subtype framework. For CKD-PC, a small proportion of participants required the nondiabetic patch approach for ACR handling. The measured-ACR-only sensitivity analysis produced a similar subtype pattern, but validation with observed kidney outcomes remains necessary.

And missing covariates were handled using a single completed-dataset iterative imputation procedure. This preserved the aligned analytic samples but did not propagate imputation uncertainty as multiple imputation with pooled estimates would.

Sample size and sparse subtype-by-risk strata may have affected estimate stability. Although the main heterogeneity pattern remained directionally consistent across sensitivity analyses, some comparisons became less stable at higher thresholds because of limited events. This is particularly relevant for SIR, where lower or nonsignificant projected risk may reflect younger age, limited events, or cutoff-based instability rather than truly low long-term risk. Cluster stability was acceptable overall, but the least stable cluster had only moderate bootstrap stability; subtype labels should therefore be viewed as clinically interpretable summaries of the clustering solution rather than fixed disease categories.

Fourth, selection and generalizability remain limited. The analytic sample was derived from a larger high-risk cohort after restricting to participants with computable clustering variables and organ-risk inputs, which may have affected representativeness. In addition, this was a Chinese cohort, and prior studies suggest that subtype composition and subtype-outcome associations may vary across populations, obesity backgrounds, and clinical settings[1,2,5,9]. Accordingly, these findings are best viewed as evidence of projected cross-organ risk heterogeneity in Chinese adults with prediabetes and require validation in other populations and care settings.

Future studies should validate these patterns using observed organ outcomes and direct organ phenotyping, and test whether subtype-informed screening adds value beyond a uniform glycemia-based approach[4-7,37].

CONCLUSION

In Chinese adults with prediabetes, cluster-defined metabolic subphenotypes showed clear heterogeneity in score-derived cardiovascular, kidney, and FIB-4-defined liver fibrosis-risk profiles. The main burden was not distributed according to insulin resistance alone. MARD represented the broadest age-loaded projected-risk subtype, contributing the largest share of the Framingham-defined cardiovascular, CKD-PC-defined kidney, and FIB-4-defined fibrosis-risk strata. SID showed a distinct nonclassical pattern: Despite lower adiposity and the absence of a typical insulin-resistant phenotype, it remained enriched for projected CKD risk and FIB-4-defined liver fibrosis-risk signal. This SID pattern is the main subtype-specific finding of the study and suggests that insulin-deficient, higher-glycemic-exposure prediabetes may carry organ-risk signals that are not captured by adiposity or insulin resistance alone. More broadly, this study links prediabetes subphenotyping with a cardio-kidney-liver risk-score framework, extending prediabetes classification beyond diabetes progression or single-organ risk. These findings support prospective validation of subtype-informed, organ-oriented risk assessment in prediabetes, while emphasizing that the present results reflect risk-score-derived projections rather than observed cardiovascular, kidney, or liver outcomes.