Subtypes of pre-diabetes in rural adolescent girls from India (DERVAN-11) Authors: Suvarna Patil, Netaji Patil, Dnyaneshwar Jadhav, Pallavi Bhat, Asmita Jadhav, Anup Nilawar, Charudatta Joglekar Published: 2026 ABSTRACT AbstractIntroductionPre-diabetes is becoming increasingly prevalent in adolescent populations. We recently reported a high prevalence of pre-diabetes among rural adolescent girls from the DERVAN cohort study established in the predominantly undernourished KONKAN region of India. We aimed to investigate the heterogeneity of pre-diabetes by identification of distinct subtypes.Research design and methodsAnthropometry (body mass index (BMI) and waist circumference) and body composition (adiposity), as well as glycemic parameters (glucose, insulin, and glycated hemoglobin), were measured at the recruitment stage. Indices for insulin resistance (homeostatic model assessment (HOMA)-IR), β-cell function (HOMA-β), insulin sensitivity (HOMA-S), and compensatory β-cell response for prevailing insulin insensitivity were calculated using the homeostasis model. Pre-diabetes was diagnosed using the American Diabetes Association criteria. Unsupervised K-means clustering and a subgrouping algorithm were applied to characterize pre-diabetes. We used BMI, waist circumference, and adiposity for characterization. Glycemic parameters and HOMA indices were compared between clusters, as well as subgroups.ResultsAmong 1382 girls, 809 (58.5%) were thin and 550 (39.8%) were pre-diabetic. We identified five clusters based on phenotypic characteristics: Cluster 1: young lean (YL), Cluster 2: old lean (OL), Cluster 3: low BMI-high waist, Cluster 4: normal BMI-low waist, and Cluster 5: obese. The cluster frequencies were 145 (26.4%), 131 (23.8%), 146 (26.5%), 74 (13.5%), and 54 (9.8%). Clusters YL and OL had the lowest fasting insulin concentrations, HOMA-IR, and HOMA-β but high HOMA-S and a compensatory β cell response for prevailing insulin insensitivity. Subgroup analysis further revealed a distinct lean pre-diabetic phenotype (n=314) with the lowest BMI (<18.5 kg/m2), no central obesity and adiposity, and the lowest fasting insulin concentration, HOMA-IR, and HOMA-β compared with remaining subgroups.ConclusionsWe demonstrated the existence of a distinct lean pre-diabetic phenotype in adolescent girls whose characteristics seem to mirror those of adult type 5 diabetes, which has been recently recognized as a distinct entity. This suggests a need for precise strategies for primordial prevention. WHAT IS ALREADY KNOWN ON THIS TOPIC Adult diabetes and pre-diabetes have been extensively subtyped across the globe. However, subtyping among adolescents is lacking. Adult diabetes and pre-diabetes have been extensively subtyped across the globe. However, subtyping among adolescents is lacking. WHAT THIS STUDY ADDS This is the first report analyzing heterogeneity within pre-diabetes among undernourished rural adolescent girls. This is the first report analyzing heterogeneity within pre-diabetes among undernourished rural adolescent girls. A predominant lean pre-diabetic phenotype was observed using non-invasive anthropometric and body composition parameters. The phenotype showed commonly as well as occasionally reported features resembling those described in recent consensus on adult type 5 diabetes. A predominant lean pre-diabetic phenotype was observed using non-invasive anthropometric and body composition parameters. The phenotype showed commonly as well as occasionally reported features resembling those described in recent consensus on adult type 5 diabetes. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY Our approach suggests rethinking obesity-centric prevention and initiation of screening program strategies to address pre-diabetes in undernourished adolescents. Our approach suggests rethinking obesity-centric prevention and initiation of screening program strategies to address pre-diabetes in undernourished adolescents. The only viable solution is to implement appropriate interventions for primordial prevention. The only viable solution is to implement appropriate interventions for primordial prevention. Introduction India’s reputation as the diabetes capital of the world is very well recognized.1 Diabetes, which was thought to be the disease of the affluent and overnourished urban populace, now seems to be spreading in undernourished segments of the rural population. Two national surveys conducted in rural areas of the country have reported a diabetes prevalence of 5.2% among those aged >20 years2 and 15% among adults aged 45 years and older.3 Over the last decade or so, the complexity of diabetes has grown beyond simple classification as type 1 and type 2, after the identification of various subtypes of type 2 diabetes (T2D) in studies across the globe4 5 and in India.6–8 Recently, a distinct form of diabetes in lean individuals with a history of undernutrition was formally classified as type 5 diabetes (T5D) by the International Diabetes Federation (IDF).9 This distinct identity was hotly debated, though its existence was suggested in earlier studies from Africa10 11 and India.12 13 Historically, this identity was also incorporated in the WHO classification in 1985 as malnutrition-related diabetes mellitus,14 which was retracted later in 1999, citing lack of evidence.15 An international consensus has recently emerged that outlines common traits of T5D, suggests pathogenic mechanisms, and recommends management and prevention strategies.16 Chorus is growing globally in developing screening criteria and treatment strategies for adults with T5D, but the debate over its classification as a distinct identity continues to this day.17–20 This variety of evidence as well as opinions has highlighted the heterogeneous nature of T2D. Pre-diabetes (PD) is a state before T2D in the natural history of hyperglycemia, where the blood glucose levels are higher than normal but not high enough to be diagnosed as T2D. A recent national survey has identified PD as a major public health issue.21 The overall PD prevalence of 10.3% among adults aged >20 years was reported in an ongoing India Diabetes study by the Indian Council of Medical Research (ICMR). The estimates were 11.2% for urban India and 9.9% for rural India.2 Thus, in the same survey, the prevalence of PD in rural areas was higher than that of T2D. Adult Indians with PD have been reported to have a better likelihood of reverting to a normoglycemic state than a normoglycemic progressing to PD.22 Thus, there is a need for prevention strategies focusing on the pre-diabetic stage. Similar to T2D, subtyping of PD has been performed in numerous studies internationally23–26 and in India,8 27 and it has also highlighted its heterogeneity, but again, like T2D, the work is limited to adult populations. There are reports about the prevalence of PD among Indian adolescents. Secondary data analysis from the Comprehensive National Nutrition Survey reported a 16.2% prevalence of PD among Indian adolescents (18.7% in boys and 13.6% in girls).28 This has reinforced the need to develop PD prevention strategies even before adulthood by focusing on the adolescent period. BKL Walawalkar Hospital is located in the rural village of Dervan in the western Indian state of Maharashtra. The area falls in the coastal region of ‘Konkan’. The presence of undernutrition (low body mass index (BMI), stunting, wasting, low birth weight, and small placenta) across the life course has been demonstrated in various hospital and community-based cross-sectional studies.29 30 The DERVAN cohort was established in 2019 to prospectively study the associations of growth and nutrition during adolescence with future risk of diabetes in both the prospective mother and offspring. The detailed protocol and the methodologies have already been published.31–33 We observed a 39% prevalence of PD among cohort girls who were predominantly undernourished (low BMI).34 Characterization of PD has been attempted in adults, but such characterization of the heterogeneity of pre-diabetic adolescents is lacking. Here, we report on the characterization of PD among rural adolescent girls of the DERVAN cohort. Materials and methods The DERVAN cohort study recruited 1520 adolescent girls in the 16–18 years age group. The recruitment took place between June 2019 and February 2023. The institute hosted a group of 5–7 girls for a 2-day residential camp. The venous blood sample was taken the following morning at 08:00 after a 12-hour overnight fast. Anthropometry and body composition Height was measured to the nearest 0.1 cm using a stadiometer (Easycare, India). Weight was measured to the nearest 0.1 kg using an electronic weighing scale by SECA. Waist circumference was measured to the nearest 0.1 cm using a measuring tape (SECA). Body fat was measured using a bioimpedance analyzer (MC-780, TANITA Corporation, Japan). The age of menarche was recorded by recall. Laboratory measurements Blood glucose was measured on ERBA 200 with an intrabatch and interbatch coefficient of variation (CV) of <5%. Fasting insulin and ferritin were measured on Abbott Architect i1000SR with a CV of 2.0% and 2.5%, respectively. Glycated hemoglobin (HbA1c) was measured using high-performance liquid chromatography (Bio-Rad D10; Bio-Rad Laboratories, Hercules, California, USA) calibrated against the National Glycosylated Standardization Program with a CV of 2.8%. Hemoglobin was measured using HORIBA YUMIZEN H500 with a CV less than 4.0%. Calculations and classifications Stunting was defined using WHO criteria.35 BMI was categorized into thin, normal, and overweight/obese using the criteria of the International Obesity Task Force (IOTF).36 These cut-offs are linked back to adult BMI thresholds, and they map the child’s BMI-age curve so that it intersects the exact BMI threshold at age 18. Central obesity was defined using an age-specific SD score cut-off of 0.55 for metabolic risk, which corresponds to the 70th percentile for waist circumference using the standards developed by Khadilkar et al.37 Adiposity was defined using a threshold defined by Khadilkar et al,38 which suggests the 95th percentile of body fat% for adolescents at risk for adiposity. Impaired fasting glucose (IFG) was defined as fasting glucose ≥100 mg/dL. Pre-diabetes was defined using American Diabetes Association (ADA) criteria, that is, IFG or HbA1c ≥5.7%.39 Glycemic indices of insulin resistance (homeostatic model assessment (HOMA)-IR), insulin sensitivity (HOMA-S), and β cell function (HOMA-β) were estimated using the homeostasis model.40 The compensatory β cell response for prevailing insulin insensitivity was calculated by (HOMA-S×HOMA-β)/100.41 Girls without PD are referred to as normoglycemic/controls. Statistical methods Variables with skewed distributions were normalized using the appropriate transformation. Group comparisons of continuous data were made using analysis of variance and the χ2 test for categorical variables. Additional adjustments were made for hemoglobin and ferritin concentrations. Pre-diabetes was subtyped using de novo/unsupervised K-means clustering. Age, BMI, waist circumference, and body fat% were used as clustering variables. There was substantial collinearity between measures of anthropometry (BMI and waist circumference) and body composition (body fat%), which prohibited their simultaneous use as clustering variables. We overcame this obstacle by taking a residual approach. Residuals were calculated by regression of BMI on waist circumference and BMI on body fat%. The residuals obtained gave the portion of waist circumference and body fat% independent of BMI. Thus, age, BMI, residual waist circumference, and residual body fat% were used as clustering variables. All variables were standardized to a mean of 0 and SD of 1.0. The clustering ability of the dataset was assessed using Hopkin’s statistics, with a cut-off value of 0.75. The number of optimal clusters was calculated using silhouette width (online supplemental figure 1A,B). Cluster stability was tested using a Jaccard similarity index cut-off of 0.85 and 1000 resamplings of the dataset. K-means clustering was performed in SPSS (V.25.0) and R (V.4.5.2). We decided on the cluster labels based on the clustering variables. We also devised a subgrouping scheme for the entire sample of 1382 adolescent girls based on various combinations of BMI categories (thin, normal, and overweight/obese), whether centrally obese (yes or no), and adipose (yes or no) (figure 1). Group 1 was labeled as lean, group 2 as normal, group 3 as thin-fat, group 4 as overweight-obese+adipose, and group 5 as obese. Subgrouping scheme for the entire sample of 1382 adolescent girls based on various combinations of BMI categories (thin, normal, and overweight/obese), whether centrally obese (yes or no), and adipose (yes or no). BMI, body mass index. Comparisons of glycemic parameters between those with PD and normal in each subgroup were done by calculating subgroup-specific SD scores using the normoglycemic/control subgroup as a reference. SD score for an observation from a pre-diabetic subgroup=(pre-diabetic observation from a subgroup−mean of normoglycemic/control observations for that subgroup/SD of normoglycemic/control observations for that subgroup). Differences were tested using 95% CIs. Analysis sample Height and weight were measured in 1519, and waist circumference in 1510. Adiposity data collected using a bioimpedance instrument were available in 1408. Fasting glucose and fasting insulin were measured in 1518. There were 1387 girls for whom complete data were available. Five girls diagnosed with T2D were excluded. Thus, the final number for the analysis is 1382. Results Of 1382 rural girls, 809 (58.5%) were thin, 514 (37.2%) normal, 59 (4.3%) overweight/obese, 22 (1.6%) centrally obese, and 78 (5.6%) adipose. Normoglycemia was observed in 832 (60.2%), and 550 (39.8%) were diagnosed with PD. Among pre-diabetics, 357 (64.9%) had IFG with normal HbA1c, 74 (21.6%) had IFG with elevated HbA1c, and 119 (13.5%) had elevated HbA1c with normoglycemia. Clinical and biochemical characteristics of the PD and normoglycemic girls are compared in table 1. Average age was similar. Anthropometric parameters (weight, height, BMI, and waist circumference) and body fat% were similar. Stunting, thinness, normal BMI, overweight/obese, and central obesity, as well as adiposity, were also observed in similar proportions in both groups. As expected, the glycemic parameters (glucose, insulin, and HbA1c) were higher, and the glycemic indices (HOMA-β, HOMA-S, and compensatory β cell response for prevailing insulin insensitivity) were lower in those with PD, but HOMA-IR was highest (p<0.001) for all. Hemoglobin and ferritin concentrations were lower in pre-diabetics (p<0.001) for both. Table 1Clinical, glycemic, and biochemical parameters of adolescent girls with PD and normoglycemia (n=1382)CharacteristicsPD/caseNormoglycemic/controlP valueP value 1n (%)550, 39.8%832, 60.2% Age (years)16.6±0.916.6±0.90.6460.588Weight (kg)42.2±8.341.9±8.10.5690.532Height (cm)151.7±5.4151.9±5.50.4900.508Stunted35165 (30.0%)255 (30.6%)0.797–Body mass index (kg/m2)18.3±3.318.1±3.20.4030.375Thin/non-obese36 (<18.5)314 (57.1%)495 (59.5%)0.187–Normal36 (18.5–25.0)206 (37.5%)308 (37.0%) Overweight/obese36 (≥25.0)30 (5.5%)29 (3.5%) Waist circumference (cm)63.7±7.463.3±7.30.3050.238Centrally obese379 (1.6%)13 (1.6%)0.914–Body fat%23.3±7.123.3±6.90.9930.761Adipose3837 (6.7%)41 (4.9%)0.156–Glycemic parameters Fasting glucose (mg/dL)102.5±8.490.2±7.30.0000.000 HbA1c (%)5.4±0.45.1±0.30.0000.000 HbA1c (mmol/mol)36.0±0.432.0±0.30.000– Fasting insulin (µIU/mL)10.2±4.18.8±3.60.0000.000 HOMA-IR1.4±0.51.2±0.50.0000.000 HOMA-β (%)90.8±26.2106.9±36.00.0000.000 HOMA-S (%)82.3±35.096.6±36.50.0000.000 Compensatory β cell response for prevailing insulin insensitivity69.2±21.095.3±25.70.0000.000 Hemoglobin (g%)12.0±1.712.5±1.40.000– Anemic (<12 g%)223 (40.5%)216 (26.0%)0.000– Ferritin (ng/mL)22.2±22.826.5±24.20.001– Low ferritin (<15 ng/mL)282 (51.3%)332 (39.9%)0.000– Age at menarche (years)13.0±1.113.0±1.10.5840.657Data are mean±SD (SD) otherwise n (%). P value 1 refers to additionally adjusted p values for hemoglobin and ferritin concentrations.HbA1c, glycated hemoglobin; HOMA-IR, homeostatic model assessment for insulin resistance; HOMA-S, homeostatic model assessment for insulin sensitivity; HOMA-β, homeostatic model assessment for β cell function; PD, pre-diabetes. Clinical, glycemic, and biochemical parameters of adolescent girls with PD and normoglycemia (n=1382) Data are mean±SD (SD) otherwise n (%). P value 1 refers to additionally adjusted p values for hemoglobin and ferritin concentrations. HbA1c, glycated hemoglobin; HOMA-IR, homeostatic model assessment for insulin resistance; HOMA-S, homeostatic model assessment for insulin sensitivity; HOMA-β, homeostatic model assessment for β cell function; PD, pre-diabetes. Associations of body composition parameters with glycemic parameters and indices are shown in online supplemental table 1. Waist circumference was positively associated with fasting glucose (p<0.05). Girls with higher body composition parameters had higher fasting insulin, HOMA-IR, and HOMA-β, and lower HOMA-S and compensatory β cell response for prevailing insulin insensitivity (p<0.001 for all). Proportions of those with anemia and low ferritin were higher in pre-diabetics. The cluster analysis results are presented in table 2. We used age, BMI, waist circumference, and body fat% as clustering variables and identified five clusters using silhouette width (online supplemental figure 1A). Jaccard indices indicating the cluster stability were 0.72, 0.70, 0.81, 0.83, and 0.82, respectively. The cluster frequencies were 145 (26.4%), 131 (23.8%), 146 (26.5%), 74 (13.5%), and 54 (9.8%). Cluster 1, labeled as young lean (YL), comprised girls who were younger and had low BMI, waist, and body fat%. We labeled cluster 2 as old lean (OL). It comprised girls who were older and had low BMI, waist, and body fat%. Cluster 3, referred to as low BMI-high waist (LBHW), was characterized by low BMI and high waist circumference. We referred to cluster 4 as normal BMI-low waist (NBLW), characterized by girls with normal BMI and low waist circumference. Cluster 5 was labeled as obese (OB) and was characterized by girls with high BMI, waist circumference, and body fat%. Table 2Clinical, body composition, glycemic, and biochemical characteristics of pre-diabetic clusters (n=550)CharacteristicsPre-diabetesYoung leanOld leanLow BMI-high waistNormal BMI-low waistObeseP valueP value 1n (%)(550,39.8%)145 (26.4%)131 (23.8%)146 (26.5%)74 (13.5%)54 (9.8%)Age (years) 16.6±0.9 15.7±0.5 17.5±0.5 16.5±0.8 16.5±0.7 16.8±0.8 0.000 0.000 Weight (kg)42.2±8.337.9±5.039.8±5.342.1±6.142.7±6.158.9±9.00.0000.000Height (cm)151.7±5.4151.4±5.2151.9±5.0153.2±5.1148.8±5.7152.2±6.00.0000.000Stunted35 (less than −2SD)165 (30.0%)40 (27.6%)44 (33.6%)25 (17.1%)38 (51.4%)18 (33.3%)0.000–BMI (kg/m2) 18.3±3.3 16.5±1.6 17.2±2.0 17.9±2.3 19.3±2.3 25.4±3.2 0.000 0.000 Thin/non-obese36 (<18.5)314 (57.1%)109 (75.2%)95 (72.5%)85 (58.2%)25 (33.8%)00.000–Normal36 (18.5–25.0)206 (37.5%)36 (24.8%)36 (27.5%)61 (48.2%)47 (63.5%)26 (48.1%)Overweight/obese36 (≥25.0)30 (5.5%)0002 (2.7%)28 (51.9%)Waist circumference (cm) 63.7±7.4 59.9±4.2 61.2±4.8 65.4±5.8 61.4±5.3 78.7±6.2 0.000 0.000 Centrally obese379 (1.6%)00009 (16.7%)0.000– Body fat% 23.3±7.1 17.8±3.8 20.2±4.5 25.7±4.9 25.8±5.3 36.1±5.6 0.000 0.000 Adipose3837 (6.7%)005 (3.4%)4 (5.4%)28 (51.9%)0.000–Glycemic parameters Fasting glucose (mg/dL)102.5±8.4101.9±8.5102.9±8.9102.7±8.2102.2±8.2102.8±8.30.8690.888 HbA1c (%)5.4±0.45.5±0.35.4±0.45.3±0.45.5±0.45.4±0.30.2440.370 HbA1c (mmol/mol)36.0±0.437.0±0.336.0±0.434±0.437.0±0.436.0±0.3–– Fasting insulin (µIU/mL)10.2±4.19.3±3.39.4±3.610.1±3.510.2±4.214.4±5.50.0000.000 HOMA-IR1.4±0.51.3±0.41.3±0.51.4±0.51.4±0.62.0±0.70.0000.000 HOMA-β (%)90.8±26.287.0±25.085.2±21.690.4±24.891.6±25.1114.6±32.00.0000.000 HOMA-S (%)82.3±35.089.3±44.488.4±36.380.0±26.180.3±25.857.9±21.70.0000.000 Compensatory β cell response for prevailing insulin insensitivity69.2±21.071.7±21.970.8±26.268.2±17.668.8±16.961.7±15.60.0380.079 Age at menarche (years)13.0±1.113.0±1.013.3±1.113.0±1.112.6±1.212.6±1.20.0000.000Concordance between clusters and subgroups Group 1: Lean314 (57.3%)109 (75.2%)95 (72.5%)85 (58.2%)25 (33.8%)0 Group 2: Normal197 (35.9%)36 (24.8%)36 (27.5%)56 (38.4%)45 (60.8%)24 (46.2%) Group 3: Thin-fat9 (1.6%)005 (3.4%)2 (2.7%)2 (3.8%) Group 4: Overweight-obese/adipose19 (3.5%)0002 (2.7%)17 (32.7%) Group 5: Obese9 (1.6%)00009 (17.3%)Data and variables in bold are those used for clustering. Data are mean±SD (SD) otherwise n (%). P value 1 refers to p values additionally adjuste for hemoglobin and ferritin concentrations.BMI, body mass index; HbA1c, glycated hemoglobin; HOMA-IR, homeostatic model assessment for insulin resistance; HOMA-S, homeostatic model assessment for insulin sensitivity; HOMA-β, homeostatic model assessment for β cell function. Clinical, body composition, glycemic, and biochemical characteristics of pre-diabetic clusters (n=550) Data and variables in bold are those used for clustering. Data are mean±SD (SD) otherwise n (%). P value 1 refers to p values additionally adjuste for hemoglobin and ferritin concentrations. BMI, body mass index; HbA1c, glycated hemoglobin; HOMA-IR, homeostatic model assessment for insulin resistance; HOMA-S, homeostatic model assessment for insulin sensitivity; HOMA-β, homeostatic model assessment for β cell function. Fasting glucose concentration and HbA1c were similar across all five clusters. Fasting insulin concentrations, HOMA-IR, and HOMA-β were lowest, and HOMA-S and the compensatory β cell response for prevailing insulin insensitivity were highest in both the lean clusters (YL and OL). Age at menarche was highest in both of these clusters. Glycemic parameters and indices were similar in LBHW and NBLW clusters. Cluster OB had the highest fasting insulin, HOMA-IR, and HOMA-β, and the lowest HOMA-S and the compensatory β cell response for prevailing insulin insensitivity. Concordance between clusters and subgroups is shown in online supplemental figure 2. The concordance of YL and OL clusters with the lean subgroup was 75.2% and 72.5%, respectively, and decreased as we moved away from the lean to the obese cluster. Results of comparisons of the distributions of fasting glucose and HbA1c between the five subgroups of pre-diabetic girls are shown in online supplemental figure 3. Fasting glucose and HbA1c were similar across all subgroups. Comparisons of insulin and indices are shown in figure 2. The lean subgroup had the significantly lowest fasting insulin concentrations, HOMA-IR, and HOMA-β, but the highest HOMA-S. The compensatory β cell response for prevailing insulin insensitivity was highest in the lean subgroup compared with others, except the last subgroup. Group 2 comprised girls with normal BMI but no central obesity and adiposity. It differed from the lean and other subgroups in insulin and HOMA-IR, which were higher than in the lean group but lower than in the other subgroups; HOMA-S was lower than in the lean subgroup but higher than in the remaining subgroups. Comparisons of insulin and glycemic indices across subgroups of anthropometry and body composition. HOMA-β, homeostatic model assessment for β cell function; HOMA-IR, homeostatic model assessment for insulin resistance. Results of comparisons of glycemic parameters between those with PD and normoglycemic controls in each subgroup using SD scores are presented in figure 3. Except for the overweight-obese-adipose subgroup (subgroup 4), pre-diabetic girls in other subgroups had significantly higher fasting glucose and insulin concentrations, higher HOMA-IR, and lower HOMA-S and the compensatory β cell response for prevailing insulin insensitivity. Except for groups 3 and 5, HbA1c was higher, and HOMA-β was lower in all subgroups. Pre-diabetes (cases) versus normoglycemic (controls) across subgroups of anthropometry and body composition using SD scores (normoglycemic as a reference). HOMA-β, homeostatic model assessment for β cell function; HOMA-IR, homeostatic model assessment for insulin resistance. We also repeated the cluster analysis on pre-diabetics diagnosed by IFG (n=431, 31.2%). The results were similar (online supplemental figure 1B, online supplemental table 2, and online supplemental figure 4). Discussion This is probably the first and also the most exhaustive report characterizing PD among rural impoverished adolescent girls. They were predominantly undernourished/non-obese by BMI criteria, and very few were centrally obese and adipose. An overall comparison of PD and normoglycemic girls in our cohort found no differences in BMI, central obesity, and adiposity. Our PD diagnosis was based on ADA criteria, which include HbA1c.39 There is a growing body of evidence for the spuriousness/misclassification of PD by inflated HbA1c status if there is underlying iron deficiency.42–45 Considering the substantial prevalence of iron deficiency anemia, in our cohort, we did appropriate adjustments in the modeling. Our earlier report on pre-diabetic girls of the same cohort had identified two distinct pathophysiologies of PD: poor insulin secretion and β cell function in undernourished and hyperinsulinemia and insulin resistance in obese,34 suggesting biological heterogeneity. We unraveled the entire PD spectrum of pre-diabetic phenotypes by employing the cluster approach widely used in subtyping of T2D and PD in adults but had never been experimented with in adolescents.8 23–27 We used age and easy-to-measure markers, such as BMI (for obesity), waist circumference (for central obesity), and adiposity measured by body fat percentage as clustering variables. Characteristics of the OB cluster (high insulin, high HOMA-IR, and high HOMA-β) resembled those observed in classical T2D. Surprisingly, the glycemic characteristics of both the lean clusters (YL and OL) (low insulin, low HOMA-IR, and low HOMA-β) and those of the OB cluster (high insulin, high HOMA-IR, and high HOMA-β) in adolescent girls mirrored those observed in insulin-deficient and insulin-resistance clusters of adult Indian pre-diabetics described by Jagganathan et al8 and Mathews et al.27 Such clusters have also been reported in adult Indian patients with T2D.6 7 Distributions of clustering variables (BMI, waist circumference, and body fat%) differed significantly across clusters, but there was the presence of girls with varying degrees of BMI (thin, normal, and overweight/obese) in all the clusters. In addition, the glycemic parameters and HOMA indices of both the lean clusters YL and OL were similar despite differences in clustering variables. In order to have more clarity, we used a simple, innovative subgroup analysis approach used by Patil et al,12 which allowed us to isolate a distinct subgroup 1 (lean) of pre-diabetics submerged in all the clusters except the OB cluster (online supplemental figure 2). It had the lowest insulin concentrations and β cell function but the highest insulin sensitivity. Poor insulin secretion by pancreatic β cells may be a result of both intrauterine and childhood undernutrition. There was also a subgroup 2 (normal) of girls with normal BMI without central obesity and adiposity. The subgroup exhibited statistically significant differences from all other subgroups for glycemic parameters as well as HOMA indices, suggesting another identity beyond lean and overweight/obese. Though small in number, we also found subgroup 3 (thin-fat) of girls who were adipose with normal BMI and had high insulin and HOMA-IR, suggesting the role of adiposity and the thin-fat phenotype. Subgroup 4 (overweight-obese+adipose) and subgroup 5 (obese, centrally obese, and adipose) exhibited characteristics of classical T2D. Subgroups 3, 4, and 5 were handicapped by small numbers. Nevertheless, the subgroup-based approach identified distinct phenotypic variations. The lean subgroup resembled the adult T5D described by recent consensus,16 Lontchi-Yimagou et al17 and Patil et al.12 The subgroup analysis revealed subtle biological heterogeneity more explicitly than observed in the clustering method, highlighting its limitation.46 We also observed high concordance between lean pre-diabetic phenotypes obtained using k-means clustering and the subgroup approach. Each PD subgroup was compared with their respective controls. Every subgroup had lower β cell function and insulin sensitivity when compared with the relevant normoglycemic/control subgroup, resulting in failure of the compensatory β cell response for prevailing insulin insensitivity, making them pre-diabetic. As suggested in the consensus, our lean pre-diabetic phenotype was the most insulin sensitive among all other pre-diabetic phenotypes described. Similarly, the lean control phenotype had the highest insulin sensitivity, suggesting that not every undernourished individual will develop PD unless there is a failure of the compensatory mechanism. The importance of adiposity and the presence of a thin-fat phenotype was underscored by the observation of higher insulin levels in individuals with adiposity but a normal BMI in both cases as well as controls. The adolescent period is referred to as a second window of opportunity after intrauterine life for primordial prevention. Our findings foreshadow the features of adult diabetes (T5D or T2D) in pre-diabetic adolescent girls, suggesting the need for precise prevention strategies. Pre-diabetes has been described as a fragile state;22 hence, benefits of prevention strategies at this stage are likely to outperform those designed in adulthood, with a lasting impact beyond one lifecycle to the next generation in terms of reduction of risks for T2D or T5D. Our subgroup analysis has shown that the current one-size-fits-all approach for preventing PD is not appropriate. Our report has addressed a major gap in pediatric and epidemiological literature by shifting the focus of PD subtyping away from overnourished adult cohorts to undernourished rural adolescent populations. The identification of a predominant lean pre-diabetic phenotype in undernourished adolescent girls is an important public health signal, which will dictate preventive strategies. It motivates rethinking prevention beyond weight-centric, calorie-restriction guidance, emphasizing protein/micronutrient sufficiency and maternal-child health. In an epidemiological research setting, development of precision medicine using a cluster approach is feasible but will need a large sample size. We do understand that grouping continuous variables might lead to dilution of exposures as well as outcomes, but this approach has a clinical utility that can be used in personalized treatment. The foremost novel feature of our study is the subtyping of PD in ostensibly healthy-looking rural adolescent girls who are prospective mothers. Another novel feature is the use of clustering tools, which have not previously been utilized to subtype PD in adolescents. It also unraveled a distinct lean pre-diabetic phenotype. The inclusion of a non-invasive body composition parameter (adiposity by bioimpedance analysis) as a variable for cluster as well as subgroup analysis added significant depth over basic anthropometrics (BMI and waist circumference alone). The subgroup approach fine-tuned the heterogeneity and identified a lean phenotype, which could not be distinctly identified in cluster analysis. The inclusion of a robust normoglycemic control group nested within the cohort allowed meaningful calculations as well as interpretation of SD scores across subgroups. It enabled us to highlight the importance of insulin sensitivity more clearly in the control group. The limitations include a cohort comprising only girls; hence, we were unable to assess the heterogeneity of PD by gender. We did not carry out pubertal staging, which is known to greatly influence insulin resistance in the given age group of 16–18 years. Lack of an oral glucose tolerance test is also another major limitation. Linkage of adolescent BMI to adult BMI thresholds using IOTF standards enabled us to classify BMI, but there are no such universal pediatric standard thresholds available for central obesity and adiposity in Indians. We opted for pediatric central obesity as well as adiposity thresholds developed by Khadilkar et al,37 38 which are based on Indian ethnicity, though these thresholds are based on urban Indian children and adolescents. This might have tempered the strength of our conclusions, especially in the subgroup approach. Against the backdrop of our earlier cross-sectional evidence of an adult lean diabetic phenotype in the region, we can assert that our lean pre-diabetic adolescent girls could be the harbingers of adult T5D. We labeled the clusters based on clustering variables to facilitate meaningful clinical differentiation for early identification. We were able to cover almost the entire pre-diabetic spectrum in subgroup analysis. Our subgroup results indicated the presence of other pre-diabetic phenotypes. Although small numbers were an issue for some subgroups, the differences observed highlighted the need for a large sample size. The question of whether the different pre-diabetic phenotypes observed in our adolescent cohort track through adulthood can only be answered by longitudinal studies. A recent report from India has shown tracking of glycemia in lean rural Indian women beginning in early childhood, early adulthood, pregnancy, and post-delivery,47 suggesting the need for primordial prevention strategies in adolescence. Many epidemiological birth cohort studies on the DOHaD theme are currently ongoing in India in settings that are diverse in geography, ethnicity, diet, and socioeconomics.48–52 The main focus of these cohort studies has been the identification of risk factors associated with adult non-communicable diseases like diabetes and cardiovascular disease. Anthropometry and body composition measurements are routinely done in these cohorts, and data on glucose, insulin, and many other biochemical parameters are also available. A successful national endeavor is urgently needed to harmonize these cohorts; followed by reanalyzing some data might reveal more heterogeneity of PD among adolescents, thus helping to understand the causality of PD. It will also help with generalizability with a national outlook. We would also like to emphasize that the adiposity measure is very crucial rather than looking at just BMI or central obesity alone. Not every clinic, institute, or research center involved in large multicentric national or international studies will have state-of-the-art technologies like CT, MRI, or dual-energy X-ray absorptiometry to measure body composition. Installing a bioimpedance analyzer to measure adiposity is a practical, non-invasive, feasible, and affordable approach. Our characterization using the subgroup approach is methodologically simple and has demonstrated clinical utility. Nevertheless, our data provide strong evidence of heterogeneity of PD in rural adolescent girls, with footprints of not only T2D but also T5D. This could enable the development of precise preventive measures during adolescence to curb the diabetes epidemic in adults. Conclusions Our findings highlight an under-recognized phenotype of PD in young adolescent girls, primarily driven by lack of adiposity and insulin secretion defects associated with undernutrition, rather than by excess fat or obesity associated with hyperinsulinemia. This form of early metabolic dysfunction may eventually progress to what is now being labeled as T5D in later life. The implications are particularly significant for undernourished adolescent girls, as persistent life course undernutrition puts them at risk of developing gestational diabetes, which in turn increases the likelihood of T5D both in themselves and in their offspring, perpetuating a vicious intergenerational cycle. Adolescence, therefore, represents a critical window of opportunity to intervene with targeted and precise strategies for primordial prevention, aiming to disrupt this cycle before it begins. In summary, our study brings much-needed attention to heterogeneity within adolescent PD in a rural, undernourished population—a demographic that is often overlooked in precision prevention research. The findings make a compelling case for adolescent screening that includes undernourished youth, especially the girls who are the future of the next generation. It offers a brief, pragmatic screening algorithm that takes into account anthropometry and body composition and could aid implementation. Our approach suggests initiation of screening programs for PD not only in obese but also in undernourished adolescent girls. Supplementary material