This retrospective cross-sectional study examined data from patients with obesity (BMI 25–45 kg/m², aged 20–75 years) enrolled in a structured one-year online lifestyle intervention program at Freedom from Diabetes, Pune, India, between June 2020 and August 2023. A total of 514 age- and sex-matched case-control pairs (n=1, 028) were identified, comprising individuals with and without T2D.

T2D diagnosis was based on at least one of the following criteria, consistent with American Diabetes Association (ADA) 2023 guidelines (14): (1) fasting plasma glucose ≥126 mg/dL (7.0 mmol/L) on two separate occasions; (2) HbA1c ≥6.5% (48 mmol/mol) confirmed on repeat testing; or (3) a documented diagnosis of T2D in clinical records by a qualified physician, including patients currently on antidiabetic pharmacotherapy.

Cases were individuals meeting the T2D diagnostic criteria with BMI 25–45 kg/m² and complete first-consultation data. Controls were individuals with no diabetes diagnosis, matched 1:1 to cases by age (± 5 years) and sex. Individuals were excluded if they had BMI <25 kg/m², a diagnosis other than T2D (type 1 diabetes mellitus, maturity-onset diabetes of the young, latent autoimmune diabetes in adults, or gestational diabetes), were on external insulin therapy, or had incomplete baseline data. The patient selection flowchart is presented in Figure 1.

Of 13, 066 patients screened, 6, 459 (49.4%) were excluded due to missing values at first consultation, predominantly fasting insulin and hsCRP, which are specialized investigations not universally completed at initial presentation, unlike standard parameters (HbA1c, fasting glucose, lipid profile, BMI) which are collected routinely at enrolment. A further subset had these values recorded at subsequent consultations; only first-consultation values were used to capture pre-intervention baseline metabolic status, as later values would be influenced by the lifestyle intervention.

This study was approved by the Institutional Ethics Committee (Ref. No. FFDRF/IEC/2024/7) and registered with the Clinical trials registry of India (Ref. No. CTRI/2024/03/064596). The requirement for informed consent was waived by the ethics committee owing to the retrospective nature of the study. This study followed the ethical principles outlined in the Declaration of Helsinki.

Data extracted at first consultation included age, sex, anthropometric measurements (height, weight), biochemical parameters (HbA1c, fasting blood glucose, fasting insulin, lipid profile), blood pressure (systolic and diastolic), and medical history (diabetes duration, comorbidities, medication use). Hypothyroidism was identified by current levothyroxine prescription at first consultation. BMI was Classified using WHO Asian-specific cut-offs (15): Obesity Class I (BMI 25.0–29.9 kg/m²) and Obesity Class II (BMI ≥30.0 kg/m²). These cut-offs follow WHO Asian-specific recommendations, where BMI ≥25.0 kg/m² is considered obesity a threshold lower than the global WHO cut-off of ≥30.0 kg/m² (15). Insulin resistance and beta-cell function were assessed using the HOMA2 calculator (16). HOMA2IR ≥2 indicated insulin resistance (17) and HOMA2%B ≤50 indicated poor beta-cell function (18). Low HDL-C was defined as <40 mg/dL in males and <50 mg/dL in females (19). Systemic inflammation was defined as hsCRP ≥3 mg/L (20).

Temporal internal validation was performed using an independent time-cohort of patients (n=966) with obesity (BMI 25–45 kg/m²) enrolled between September 2023 and January 2025, adhering to the same eligibility criteria but without age-sex matching. No patients from the development cohort were included in the validation dataset.

Matching was not applied to the validation cohort as the purpose of temporal internal validation is to assess the discriminative stability of identified markers in an independent time-cohort, rather than to control confounding as in the development phase. The unmatched design better reflects real-world clinical populations and provides a more stringent test of marker stability. Structural differences between cohorts (age, sex distribution, diabetes duration) are acknowledged and systematically documented in Supplementary Table 1.

Post-hoc power was calculated using the method of Kelsey et al. (21) for 1:1 matched case-control designs (discordant pair approach), implemented in Python v3.8 (scipy.stats library), with α=0.05 (two-tailed) and n=514 matched pairs. Power exceeded 85% for five of seven predictors at observed prevalences. Hypothyroidism (power=66.5%) and Obesity Class I (power=4.4% bivariate) were lower, the latter attributable to near-identical Obesity Class I prevalence in both groups consistent with the obesity paradox finding discussed later in discussion. The retrospective design determined sample size by data availability; the post-hoc calculation confirms adequate power for all five primary significant predictors.

Statistical analyses were performed using IBM SPSS v21, R, and Python v3.8. Categorical variables are presented as frequencies and percentages; continuous variables as median (interquartile range) based on skewed data distribution. The McNemar test assessed associations between categorical variables, and the Wilcoxon signed-rank test compared continuous variables between matched pairs. Significant factors from univariate analysis (inflammation, insulin resistance, beta cell function, HDL-C, antihypertensive medication, hypothyroidism, and BMI) were included in conditional logistic regression to estimate adjusted odds ratios (ORs) accounting for the matched-pair structure. Binary logistic regression was subsequently applied to generate absolute predicted probabilities required for discrimination, calibration, decision curve analysis, and nomogram construction, as conditional logistic regression cannot generate predicted probabilities due to the absence of an intercept term (22).

Given the near-complete separation observed for HOMA2%B where poor beta-cell function was present in 285 of 514 diabetic patients (55.4%) but only 19 of 514 non-diabetic patients (3.7%), Firth penalised logistic regression was additionally performed as a confirmatory sensitivity analysis. Firth’s method applies a Jeffreys invariant prior penalty to the likelihood function, producing bias-corrected coefficient estimates that are stable under separation conditions. As the firthlogist package is incompatible with Python 3.12, Firth regression was implemented using the Jeffreys prior penalty using logistf package, R v4.5.3. Results were compared against standard and conditional logistic regression estimates to confirm consistency of findings.

Variance Inflation Factors (VIF) confirmed no multicollinearity among predictors (all VIF <1.1: HOMA2%B=1.06, HOMA2IR = 1.03, HDL-C=1.02, hsCRP=1.01, antihypertensive=1.03, hypothyroidism=1.02, BMI = 1.01).

Binary logistic regression was repeated on the validation dataset (n=966) using the same markers identified in Model 1 to assess discriminative strength in the independent time-cohort. Results are presented as OR with 95% confidence intervals (CI); statistical significance was set at p<0.05.

Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC ROC), interpreted as: poor (0.5–0.7), acceptable (0.7–0.8), good (0.8–0.9), and excellent (≥0.9) (23). Calibration was evaluated using the Hosmer–Lemeshow goodness-of-fit test (p>0.05 indicating good fit) (24), the Brier score (lower values indicating better calibration), and calibration plots (Python v3.8). Isotonic regression was applied as a smoothing technique to reduce bin-level noise in the calibration curve visualization. Clinical utility was assessed using decision curve analysis (DCA), estimating net benefit across threshold probabilities (Python v3.8) (24). Internal validation was performed using bootstrap resampling (1, 000 resamples, SPSS v21). Blood pressure data were available for 761 of 1, 028 patients (74.0%); missing data analysis confirmed a Missing at Random (MAR) pattern, justifying multiple imputation by chained equations (MICE; 20 imputed datasets, Rubin’s Rules, Python v3.8). Model fit was additionally assessed using -2 Log Likelihood (-2LL), Cox & Snell R², and Nagelkerke R² (SPSS v21).

The same discrimination, calibration, and model fit methods were applied to the validation cohort to assess temporal stability of the identified markers.

A clinical scoring nomogram was constructed using Standard LR coefficients from the complete dataset (n=1, 028) using Python. Each marker was assigned points proportional to its regression coefficient, with total scores mapped to four risk categories. Hypothyroidism was retained in the regression model but excluded from the nomogram as it did not reach significance (OR = 0.715, p=0.092).

Of the 1, 028 participants (514 matched pairs), 33% were male. Median age and BMI were 46 (IQR 12) years and 29.7 (IQR 5.5) kg/m², respectively. Obesity Class I (BMI 25.0–29.9 kg/m²) was present in 52% of participants and Obesity Class II (BMI ≥30.0 kg/m²) in 48%. Among diabetic patients, median diabetes duration was 8.5 (IQR 6.6) years.

The diabetes group showed significantly higher prevalence of insulin resistance (23% vs 12%), systemic inflammation (hsCRP ≥3 mg/L: 62% vs 50%), antihypertensive medication use (33% vs 25%), and dyslipidemia medication use (48% vs 20%) compared to the no-diabetes group (McNemar test, all p<0.001) (Table 1). Poor beta-cell function was significantly more prevalent in the diabetes group, while good beta-cell function was more common in the no-diabetes group (p<0.001). Low HDL-C was significantly more prevalent in the diabetes group (72.8% vs 55.1%; OR = 2.2, 95% CI: 1.7–2.8, p=0.001). Hypothyroidism was significantly higher in the no-diabetes group (p=0.001). Regarding BMI category, Obesity Class I was more prevalent in the diabetes group while Obesity Class II was more prevalent in the no-diabetes group (p<0.001).

Subgroup analysis within the no-diabetes group revealed that Obesity Class I non-diabetic patients had significantly worse beta-cell function and greater insulin resistance than their Obesity Class II counterparts (HOMA2%B and HOMA2IR, both p<0.001), despite lower absolute BMI. Conversely, systemic inflammation (hsCRP) was higher in Obesity Class II non-diabetic patients (p<0.001).

Significant differences in diastolic blood pressure and lipid profile were also observed between the diabetes and no-diabetes groups (Wilcoxon test, p=0.013).

Baseline characteristics of the development (n=1, 028) and validation (n=966) cohorts are compared in Supplementary Table 1. The cohorts differed significantly in age, sex distribution, diabetes duration, and antihypertensive medication use (p ≤ 0.001), while core metabolic markers (HOMA2%B, HOMA2IR, HDL-C, hsCRP, and BMI) were comparable between cohorts (all p>0.05), confirming metabolic homogeneity despite demographic differences.

Conditional logistic regression (n=1, 028; 514 matched pairs) identified low HDL-C, systemic inflammation, insulin resistance, Obesity Class I, poor beta-cell function, and antihypertensive medication use as significant markers of prevalent T2D. Poor beta-cell function showed the strongest association, with 47.9 times higher odds of T2D compared to good beta-cell function (Figure 2A).

Standard binary logistic regression was performed to generate predicted probabilities for model performance assessment. The same markers identified in conditional logistic regression were confirmed as significant. Poor beta-cell function again showed the strongest association (OR = 40.9), with all other markers retaining significance in the same direction (Figure 2B).

To assess robustness of findings and address potential circularity of HOMA2%B and HOMA2IR, pre-specified sensitivity analyses were performed excluding insulin secretagogue users (Sensitivity A, n=233 excluded), patients with diabetes duration >10 years (Sensitivity B, n=198 excluded), and both groups simultaneously (Sensitivity C). Complete sensitivity analysis results are presented in Supplementary Table 2. HOMA2%B and HOMA2IR retained significance across all sensitivity subgroups (OR range 47.4–78.6 and 5.2–6.4 respectively, all p<0.001), confirming robustness of these associations independent of treatment effects or disease duration. Inflammation (hsCRP) demonstrated strengthened associations when secretagogue users and long-standing diabetes were excluded (OR range 2.5–3.7), while HDL-C, antihypertensive medication, and Obesity Class I attenuated to non-significance in sensitivity subgroups, consistent with these associations reflecting advanced disease burden rather than independent biological markers. Additionally, antihypertensive medication retained significance after incorporating systolic blood pressure via MICE imputation (OR = 1.756, 95% CI: 1.045–2.951, p=0.034), confirming the association is not solely explained by underlying blood pressure elevation.

Conditional logistic regression, standard binary logistic regression, and Firth penalized logistic regression were all performed on the complete dataset (n=1, 028) as described in the Statistical Analysis section. All three methods demonstrated consistent direction and significance across all seven markers, confirming robustness of findings regardless of analytical approach (Supplementary Table 3).

The markers identified in Model 1 were assessed in the independent validation cohort (n=966). Poor beta-cell function retained the strongest association (OR = 48.8), followed by Obesity Class I, antihypertensive medication use, and insulin resistance (Figure 3). HDL-C and hsCRP did not reach significance in Model 2, likely reflecting greater within-group heterogeneity in the older, unmatched validation cohort (see Discussion).

Bootstrap internal validation (1, 000 resamples, Python v3.8) yielded a mean optimism of 0.000 and an optimism-corrected AUC of 0.860, confirming no overfitting (events-per-predictor ratio=73.4). Model 1 achieved AUC ROC of 0.860 (95% CI: 0.837–0.883, p<0.001), indicating good discrimination (Figure 4A). At the optimal cut-off of 0.28 (Youden index=0.49), sensitivity was 86.6% (95% CI: 83.4–89.3%) and specificity 62.5% (95% CI: 58.5–66.5%). Model 2 achieved AUC ROC of 0.865 (95% CI: 0.842–0.888, p<0.001), with sensitivity 85.5% (95% CI: 82.1–88.3%) and specificity 64.4% (95% CI: 60.1–68.6%) at cut-off 0.30 (Youden index=0.49) (Figure 4B).

The Hosmer–Lemeshow test confirmed good calibration for both models (Model 1: p=0.562; Model 2: p=0.415). Brier scores were 0.14 (Model 1) and 0.162 (Model 2), indicating acceptable calibration. Calibration plots showed good agreement between predicted and observed probabilities across most risk strata, with minor instability in the low-probability range (0.10–0.20), likely attributable to the bimodal distribution of predicted probabilities inherent to the matched-pair design after isotonic adjustment (Figure 5).

Model fit assessed by -2LL was higher in Model 2 (927.045) than Model 1 (615.454), reflecting the larger unmatched validation sample. Model 1 explained 38.4% (Cox & Snell R²) and 51.2% (Nagelkerke R²) of variance; Model 2 explained 39.6% and 52.7% respectively, confirming retained explanatory strength in the validation cohort.

Decision curve analysis demonstrated consistent clinical utility across both development and validation cohorts. In Model 1, net benefit of 0.359 was observed at the clinical cut-off of 0.28, outperforming both ‘treat-all’ and ‘treat-none’ strategies across threshold probabilities 0.10–0.85. Model 2 similarly demonstrated net benefit of 0.347 at a threshold of 0.30 (Figure 6).

Using Standard LR coefficients from the complete dataset (n=1, 028), each marker was assigned points proportional to its regression coefficient (Figure 7): poor beta-cell function (HOMA2%B ≤50) = 100.0 points; insulin resistance (HOMA2IR ≥2) = 40.5 points; low HDL-C = 20.5 points; Obesity Class I (BMI 25.0–29.9 kg/m²) = 19.4 points; elevated hsCRP (≥3 mg/L) = 17.4 points; antihypertensive medication = 13.8 points. Hypothyroidism was retained in the regression model but excluded from the clinical scoring nomogram as it did not reach significance (OR = 0.715, p=0.092). Total scores range from 0 to 212 points, mapped to four risk categories: Low risk (<40 points, <28% probability), Moderate risk (40–79 points, 28–64%), High risk (80–129 points, 64–92%), and Very High risk (≥130 points, >92%).

As a potential future application, contingent on external validation and prospective evaluation, the following risk-stratified actions could be considered based on nomogram total points and corresponding predicted probability of prevalent T2D:

Low risk (<40 points, <28% probability): routine lifestyle counselling, dietary optimisation, and annual metabolic monitoring including fasting glucose and HbA1c.

Moderate risk (40–79 points, 28–64% probability): enhanced structured lifestyle intervention, six-monthly HbA1c and fasting glucose monitoring, targeted dietary modification, and supervised physical activity programme.

High risk (80–129 points, 64–92% probability): comprehensive metabolic assessment including oral glucose tolerance test where indicated, clinical review by an endocrinologist or diabetologist, three-monthly monitoring, and consideration of pharmacological diabetes prevention strategies in accordance with national and international guidelines (25, 26).

Very High risk (≥130 points, >92% probability): urgent clinical evaluation to confirm or exclude a Type 2 diabetes diagnosis, comprehensive workup including HbA1c, fasting and postprandial glucose, and fasting insulin, and immediate initiation of appropriate management if diagnosis is confirmed. These are presented as potential future applications rather than recommendations directly supported by the current study. Given the cross-sectional, retrospective, single-center design, external validation across independent populations and prospective evaluation are required before this tool can be considered for routine clinical practice or before these management actions can be recommended with confidence.

To illustrate clinical application, consider two patients both with Obesity Class I (BMI 25–29.9 kg/m²: +19.4 pts), antihypertensive medication use (+13.8 pts), and low HDL-C (+20.5 pts), with no hsCRP elevation (0 pts). Patient A, with poor beta-cell function (+100.0 pts) and insulin resistance (+40.5 pts), accumulates 194.2 points, corresponding to a predicted probability of 99.2% (Very High Risk, ≥130 points), warranting immediate clinical evaluation. Patient B, without HOMA2 abnormalities (0 pts), accumulates 53.7 points, corresponding to a predicted probability of 39.7% (Moderate Risk, 40–79 points), indicating closer metabolic monitoring. This demonstrates that the tool appropriately stratifies risk across the clinical spectrum and does not default to high risk for all patients.