For the Singapore Adult Metabolism Study (SAMS), ethical approval was obtained from the National Healthcare Group Domain Specific Review Board prior to conduct of the study, and written informed consent was obtained from all participants.

For the independent longitudinal cohort, ethics approval was obtained from two Institutional Review Boards (National University of Singapore and Singapore General Hospital). Written informed consent was obtained before conduct of the study.

Healthy male volunteers, aged 21–41 years and with body mass index (BMI) between 18.5–30 kg/m², were invited to participate in SAMS. Individuals with known hypertension, diabetes mellitus, dyslipidemia, ischemic heart disease, epilepsy or any medication that might affect insulin sensitivity (eg, corticosteroids) were excluded. Also excluded were individuals with recent changes in or attempts to change body weight, bleeding diathesis, inaccessible veins, recent investigational medicine use, or contraindications to magnetic resonance imaging (MRI).

Anthropometric measures included height, weight, waist and hip circumference, measured twice and the average taken, as well as skinfold thicknesses at four sites (biceps, triceps, subscapular and suprailiac), measured in triplicate. The waist-to-hip ratio (WHR) was calculated by dividing waist circumference (in cm) by hip circumference (HC, in cm). Percentage body fat was derived from the skinfolds measured using the following formula [16]: Body fat % = 36.7 * log (sum of 4 skinfolds: triceps, biceps, subscapular and suprailiac) −39.5.

All subjects also underwent a whole body Dual Energy X-ray Absorptiometry (DEXA) scan (a single Hologic Discovery Wi densitometer, Hologic, Inc, Massachusetts, USA) to estimate total fat mass, and magnetic resonance imaging (MRI) of the abdomen (3T Trio, Siemens AG, Medical Solutions, Germany) from lumbar segments T12 to L5, to quantify subcutaneous and visceral fat.

We looked at data from participants from two cross-sectional surveys, the National Health Survey (NHS 1992) [19] and the National Health Survey (NHS 1998) [20], who repeated the health survey between 2004 and 2007. Briefly, both studies were a random sample of from the Singapore population, with disproportionate sampling stratified by ethnicity to increase the number of the minority ethnic groups (Malays and Asian Indians). Details of participant characteristics and biochemical analysis have been published previously [19], [20].

A total of 6,302 subjects who participated in one of the NHS (1992 or 1998), were recontacted between 2004 and 2007 for follow up examination. Of these 4,023 subjects had complete data for anthropometric parameters, biochemical measures and diabetes mellitus (DM) and cardiovascular disease (CVD) status at baseline (NHS 1992 or 1998) and follow-up. Subjects with DM (n = 202) or CVD (n = 71) at baseline were excluded. Thus, a total of 3,750 subjects were included for the final analysis.

Diabetes mellitus was defined using fasting plasma glucose based on ADA criteria, as well as history of diabetes/diabetes medication. Cardiovascular disease in NHS 1992 was defined as history of ischaemic heart disease/angina/stroke. Cardiovascular disease in NHS 1998 was defined as history of angina/stroke.

Cardiovascular disease at follow up was defined as history of ischaemic heart disease/blockage of coronary arteries/angina/stroke. Homeostatic Model of Assessment of Insulin Resistance (HOMA-IR) was calculated as [I₀ (µU/mL) × G₀ (mmol/L)] ÷22.5 [6], I₀ =  fasting insulin, G₀ = fasting glucose.

A total of 253 subjects (96 Chinese, 80 Malay, and 77 Indian) from SAMS with complete clamp, anthropometric and biochemical data were included in the analysis. All variables were checked for normality and log-transformed to improve normality assumptions as necessary. A sub-sample was drawn from the Chinese subjects in the study using the random sampling option in the statistical software to select approximately half of the subjects available. In this sub-sample of individuals, Pearson's correlation coefficients were used to identify significant associations between anthropometric, biochemical and imaging measures with ISI-clamp. Stepwise multiple linear regression was performed to identify predictors of ISI-clamp, and to generate the prediction equation (ISI-cal). Alternative models were evaluated by comparing the total variation explained by the models. Findings were confirmed using robust regression. Bootstrapping was used to generate bias corrected and accelerated confidence intervals for the beta coefficients.

The equation was validated by comparing non-parametric correlation coefficients (Kendall's tau-a) between various estimates of insulin sensitivity and ISI-clamp, and testing for significant differences between the correlation coefficients (using the lincom command in STATA). Lin's concordance coefficient was also calculated to estimate agreement between ISI-clamp and ISI-cal. Receiver operating characteristic (ROC) curves were used to compared the predictive function of different methods in identifying individuals with insulin resistance, as defined by the lowest tertile of insulin sensitivity by ISI-clamp, i.e. ISI ≤6.94 mg min⁻¹ kg lean mass⁻¹ mU insulin⁻¹.

We then examined the ability of the derivative ISI-cal in predicting cardio-metabolic events in an independent longitudinal cohort. ISI-cal and HOMA-IR were calculated for the baseline cohorts (NHS 1992 and NHS 1998). ROC curves were used to compare predictive function between HOMA-IR and ISI-cal in identifying new onset of DM and CVD in the follow – up cohort. Youden's index was used to identify cut-off values with optimum specificity and sensitivity. McNemar's statistic was used to test for differences in sensitivity and specificity between ISI-cal and metabolic syndrome definition in predicting CVD.

The random sample was generated using the random sampling facility in SPSS (Version 19, IBM Statistics, USA). All other statistical analyses were performed using Stata 10.0 for Windows (Stata Corporation, College Station, Texas, USA). Data are presented as means ± SD unless stated otherwise. All statistical tests were two-sided, with any p<0.05 being considered significant.

In the random sample of 45 Chinese subjects, various anthropometric and biochemical measures were associated with ISI-clamp on univariate analysis: BMI (r = −0.45, p = 0.002), waist circumference (r = −0.59, p<0.001), WHR (r = −0.66, p<0.001), skinfold-derived body fat% (r = −0.63, p<0.001), DEXA derived total body fat% (r = −0.55, p<0.001), serum triglycerides (r = −0.46, p = 0.001), fasting insulin (r = −0.59, p<0.001), MRI measured visceral (VAT, r = −0.58, p<0.001) and subcutaneous (SAT, r = −0.56, p<0.001) adipose tissue.

To identify which set of variable(s) best explained the variation in ISI-clamp, we examined several combinations of these variables in separate regression models. A model with WHR (ΔR²  = 0.42), I₀ (ΔR²  = 0.13) and TG (ΔR²  = 0.04) best predicted insulin sensitivity (total adjusted R²  = 0.58), over models with I₀ alone (adjusted R²  = 0.33), and HOMA-IR alone (adjusted R²  = 0.32). Replacing WHR with various MRI-derived fat measures, VAT (adjusted R²  = 0.52), SAT (adjusted R²  = 0.51), VAT/SAT ratio (adjusted R²  = 0.46), VAT over hip circumference (adjusted R²  = 0.55), SAT over HC (adjusted R²  = 0.52), did not improve the adjusted R² over the model based on WHR, I₀ and TG. Neither did replacing fasting insulin with other surrogate measures of insulin resistance, HOMA-IR (adjusted R²  = 0.55) or QUICKI (adjusted R²  = 0.56). There was no multi-collinearity in the final model (VIF <1.5, tolerance 0.75). The relationship between these variables and ISI-clamp was confirmed using robust regression, and confidence intervals for the coefficients obtained by boot-strapping (Table S1). The prediction equation finally derived was given as ISI-cal  =  exp(2.65 – (2.63*ln [WHR]) – (0.39*ln [I₀])- (0.21*ln [TG])) and simplified to ISI-cal  = 14.15* (WHR)^−2.63 * (I₀)^−0.39 * (TG)^−0.21.

The prediction equation, ISI-cal, was validated using the remaining 208 subjects from SAMS. Overall, the correlation between ISI-clamp and ISI-cal (Kendall's tau τ = 0.42, p<0.001) was stronger than the correlation between ISI-clamp and HOMA-IR (τ = −0.37, p<0.001) (p _(comparison)  = 0.045). The correlation coefficients for ISI-cal and HOMA-IR, for Chinese (0.43 versus  = −0.39) and Indians (0.34 versus  = −0.32) were not significant, except for Malays (0.47 versus  = −0.38, p_(comparison)  = 0.03). Lin's concordance coefficient between ISI-clamp and ISI-cal was 0.53 for the whole group, 0.46 for Chinese, 0.6 for Malay and 0.42 for Indian.

The predictive function of these indices in identifying individuals with low insulin sensitivity was tested using ROC curves, with the ISI cutoff set at 6.94 mg min = ⁻¹ kg lean mass⁻¹ mU insulin⁻¹ (the lowest tertile of insulin sensitivity in the group). For this comparison, we used inverse values of HOMA-IR to indicate insulin sensitivity. ISI-cal had a significantly larger area under the ROC curve compared to HOMA-IR (0.82 versus 0.78, p = 0.024) (Figure 1).

In the longitudinal cohort (with 6 to 15 years of follow-up), there were 99 individuals who subsequently developed CVD (70 with ischaemic heart disease, 29 with stroke or transient ischaemic attack, and 7 with both), and 213 who developed DM, among those without DM or CVD at baseline. We compared the ability of ISI-cal (inverse) and HOMA-IR in identifying these individuals using ROCs. ISI-cal and HOMA-IR had similar AUCs in predicting DM (Table 2). However, ISI-cal had a significantly larger AUC compared to HOMA-IR in predicting CVD. These findings were replicated across all three ethnic groups and in both genders (Table 2). ISI-cal showed better discrimination for both ischaemic heart disease (ROC ISI-cal 0.75±0.03 vs HOMA-IR 0.63±0.04, p<0.001) and stroke (ROC ISI-cal 0.73±0.04 vs HOMA-IR 0.58±0.05, p<0.001).

Using Youden's index, the optimal ISI-cal cut-off for predicting DM was 9.3 with sensitivity and specificity of 71%, and for predicting CVD was 9.23 with a sensitivity of 71% and specificity of 70%. The optimal cut-offs for HOMA-IR in this population was 1.99 for DM with sensitivity 70% and specificity 71%, and 1.59 for CVD with sensitivity 64% and specificity 54%. These results did not change materially when we repeated the analysis by cohort of origin (NHS 92, NHS 98) (data not shown).

We also examined whether ISI-cal and metabolic syndrome, as defined by the NCEP ATP III criteria, performed comparably in predicting CVD using McNemar's statistic. We did this to verify whether the gain in predictive accuracy over HOMA-IR was due to improved estimation of the insulin resistance state, or due to the inclusion of triglycerides and WHR, which are independent risk factors for CVD. ISI-cal had a significantly higher sensitivity (70%) compared to metabolic syndrome (25%; p<0.001), but also lower specificity (71% ISI-cal vs 92% metabolic syndrome, p<0.001). We also ran a logistic regression to evaluate if ISI-cal was associated with future CVD events independent of metabolic syndrome, and found that the association between CVD and ISI-cal was highly significant (OR 4.7, 95% CI 2.9–7.4, p<0.001), even when metabolic syndrome was in the model (OR 1.9, 95% CI 1.1–3.1, p = 0.02). There was no collinearity between ISI-cal and metabolic syndrome in the model (VIF 1.2, tolerance 0.8).