Lipoprotein Insulin Resistance Index: A Simple, Accurate Method for Assessing Insulin Resistance in South Asians

Abstract

Context

Identification of insulin resistance (IR) in South Asians, who are at a higher risk for type 2 diabetes, is important. Lack of standardization of insulin assays limits the clinical use of insulin-based surrogate indices. The lipoprotein insulin resistance index (LP-IR), a metabolomic marker, reflects the lipoprotein abnormalities observed in IR. The reliability of the LP-IR index in South Asians is unknown.

Objective

We evaluated the predictive accuracy of LP-IR compared with other IR surrogate indices in South Asians.

Methods

In a cross-sectional study (n = 55), we used calibration model analysis to assess the ability of the LP-IR score and other simple surrogate indices (Homeostatic Model Assessment of Insulin Resistance, Quantitative insulin sensitivity check index, Adipose insulin resistance index, and Matsuda Index) to predict insulin sensitivity (S_I) derived from the reference frequently sampled intravenous glucose tolerance test. LP-IR index was derived from lipoprotein particle concentrations and sizes measured by nuclear magnetic resonance spectroscopy. Predictive accuracy was determined by root mean squared error (RMSE) of prediction and leave-one-out cross-validation type RMSE of prediction (CVPE). The optimal cut-off of the LP-IR index was determined by the area under the receiver operating characteristic curve (AUROC) and the Youden index.

Results

The simple surrogate indices showed moderate correlations with S_I (r = 0.53-0.69, P < .0001). CVPE and RMSE were not different in any of the surrogate indices when compared with LP-IR. The AUROC was 0.77 (95% CI 0.64-0.89). The optimal cut-off for IR in South Asians was LP-IR >48 (sensitivity: 75%, specificity: 70%).

Conclusion

The LP-IR index is a simple, accurate, and clinically useful test to assess IR in South Asians.

insulin resistance, south Asian, lipoproteins, and diabetes

Type 2 diabetes mellitus (T2DM) is a significant public health problem among South Asians [1]. The prevalence estimates of T2DM in the South Asian countries of Bangladesh, Bhutan, India, Nepal, Pakistan, and Sri Lanka range from 7.8% in India to 9.8% in Pakistan, with an estimated 87.6 million living with T2DM in the South Asian region [2]. Overall, South Asians have an approximately 2- to 6-fold increased risk of developing T2DM, and this risk increases when South Asians immigrate to Western countries, such as Canada and the United States [3-5].

In the United States, a recent study based on the National Health and Nutrition Examination Surveys (2011-2016) revealed that T2DM prevalence in South Asians was 23% compared with 12% in non-Hispanic White people [6]. Likewise, in the cross-sectional Mediators of Atherosclerosis in South Asians Living in America (MASALA) Study, South Asians had a significantly higher age-adjusted prevalence of T2DM (23%) and prediabetes (24%) than other ethnic groups [1]. South Asians also experience earlier onset impairment of glucose regulation, prediabetes, and T2DM compared with other ethnic groups, contributing to a higher risk of developing premature cardiovascular disease [7, 8]. Findings from the UK Biobank population-based cohort study also confirm higher cardiovascular mortality in this population [9-11]. Existing risk assessment tools, such as FINDRISC and Framingham models, underestimate the risk for these conditions in both South Asian men and women, despite data that suggest a much higher disease burden [12]. Therefore, to enhance metabolic and cardiovascular risk assessment in South Asians, other approaches to assessing insulin resistance (IR) are needed.

IR is one of multiple risk factors, such as genetic predisposition, increased calorie intake, propensity for visceral adiposity, and reduced β-cell function, that may explain the higher prevalence of T2DM in South Asians [1, 2, 7, 13-17]. IR is defined as the decreased responsiveness or sensitivity to the metabolic actions of insulin, which mainly includes glucose disposal [18]. Multiple studies have further identified South Asians as more insulin resistant than other ethnic groups [7, 17, 19, 20]. The gold standard test for insulin sensitivity measurement is the hyperinsulinemic–euglycemic clamp (EHC) [18]. However, this technique is time-consuming, tedious, and requires expertise, thus limiting its use to research settings. Although less complex, the minimal model analysis of a frequently sampled intravenous glucose tolerance test (FSIVGTT) is another indirect method to assess IR that suffers from the same limitations as the clamp study and precludes its use in clinical care settings [18].

Simple surrogate indices derived from plasma, insulin, and glucose concentrations under fasting conditions (eg, the Homeostatic Model Assessment of Insulin Resistance, HOMA-IR, and the quantitative insulin sensitivity check index, QUICKI) or after an oral glucose load (Matsuda index derived from insulin and glucose following an oral glucose tolerance test) are widely used in epidemiological studies [18, 21-24]. These indices are simple, inexpensive, and usually derived from a single fasting blood sample. However, the insulin assay is not well standardized, precluding developing cutoffs for identifying individuals with IR [25]. Consequently, surrogate indices based on fasting or postprandial insulin levels are suboptimal, emphasizing the need for simple, reliable, robust, and universal biomarkers in identifying IR.

Metabolomics-based identification of various metabolites as biomarkers of IR is a powerful new approach [26]. Alterations in lipid and lipoprotein metabolism, characterized by elevations in triglycerides and reductions in high-density lipoprotein cholesterol, are among the earliest and most frequent manifestations of IR [27]. The lipoprotein insulin resistance score (LP-IR) is a novel metabolomic biomarker based on nuclear magnetic resonance (NMR) quantification of lipoprotein levels and sizes. This surrogate index is derived from the weighted score of six lipoproteins (very–low-density lipoprotein [VLDL], low-density lipoprotein [LDL], and high-density lipoprotein [HDL] sizes and concentrations) more strongly related to IR than each subclass [28]. In this study, we examined the predictive accuracy of NMR-based LP-IR compared with other IR-surrogate indices in South Asians.

Materials and Methods

Study Subjects

The study protocol was conducted at the Clinical Research Center, National Institutes of Health in Bethesda, Maryland, after approval from the Institutional Review Board of the National Institute of Diabetes, Digestive and Kidney Diseases. Written informed consent was attained from all subjects. Healthy South Asian volunteers (n = 55), 18 years and older, were recruited as part of the Obesity Phenotyping Study (ClinicalTrials.gov identifier NCT00428987). Participants were excluded based on the diagnosis of diabetes mellitus, pregnancy, liver disease, renal insufficiency, polycystic ovarian syndrome, and medications that affect insulin sensitivity or glucose metabolism. Participants were admitted for a 3-day visit to the NIH Metabolic Research Unit, where they underwent a screening examination comprising physical examination and routine blood and urine chemistries.

Study Procedures

Measurement of body composition

A digital balance (Scale-Tronix 5702; Scale-Tronix, Carol Stream, IL) was used to measure body weight. Dual-energy X-ray absorptiometry was used to assess body composition with a Lunar iDXA scanner (GE Healthcare, Madison, WI).

Frequently Sampled Intravenous Glucose Tolerance Test

Following a 12-hour overnight fast, an intravenous bolus of glucose (20% dextrose, 0.3 g/kg) was administered over 2 minutes. At time 20 minutes, subjects were administered an intravenous injection of insulin (0.03 U/kg). Blood samples were collected to measure plasma glucose, insulin, and free fatty acids (FFAs) at the following time points relative to glucose administration: −10, −1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 10, 12, 14, 16, 20, 22, 23, 24, 25, 27, 30, 40, 50, 60, 70, 80, 90, 100, 120, 150, and 180 minutes. Insulin sensitivity (S_I) was calculated by Minimal Model analysis of glucose and insulin data from the FSIVGTT using the MINMOD software as previously described [29] (version 6.02; MINMOD Millennium, Los Angeles, CA).

Mixed Meal Tolerance Test

Following a 12-hour overnight fast, subjects ingested an 8 fluid ounce (237 mL) liquid meal at time 0 minutes. The approximate total nutrient content was 360 kcal (50% carbohydrates, 34% fat, and 15% protein). Blood samples were collected to measure plasma glucose, serum insulin, C-peptide, and FFAs at the following time points relative to liquid meal ingestion: −10, 0, 15, 30, 60, 90, 120, 180, and 240 minutes.

Research samples were collected in EDTA tubes containing dipeptidyl peptidase-4 inhibitor and protease inhibitor (Millipore-Sigma) and were subsequently frozen at −80 °C until analyzed.

Laboratory Analyses

Routine assays for serum glucose, insulin and C-peptide, HbA1c, and serum lipids were analyzed in the Department of Laboratory Medicine Clinical Center, NIH.

Lipid Profile NMR Spectroscopy and LP-IR Score Calculation

Lipoprotein particle concentration and mean size of LDL, HDL and VLDL were determined by NMR in the Clinical Mass Spectrometry Core, NIH, on fasting blood samples. NMR spectroscopy on thawed plasma samples was conducted using a 400-MHz proton Vantera Clinical analyzer (Liposcience, Inc., Morrisville, NC). As previously reported, proton NMR spectroscopy quantifies lipoprotein subspecies via an advanced algorithm (LP4) [30]. The analysis provides information about particle size and concentration for all lipoprotein classes [low density lipoprotein (LDLPs), and high density lipoprotein (HDLPs) and triglyceride-rich lipoprotein (TRLPs)]. Additionally, concentrations of lipoprotein subclasses by size (eg, small, medium, and large) were measured. This method also provides information for derived measurements of apolipoprotein A1 and apolipoprotein B (apoB).

A combination of 6 NMR-measured lipoprotein variables were used to calculate the LP-IR score: concentrations of small LDL, large HDL, and large VLDL particles, and mean sizes of LDL, HDL, and VLDL. These 6 variables combined are more strongly correlated to IR than any individual variable. Each variable was assigned a numerical weighted score reflecting the strength of their association with HOMA-IR as reported in the Multi-Ethnic Study of Atherosclerosis (MESA) population [28]. The numerical weights of the LP-IR variables are as follows: mean VLDL particle size (32), large VLDL particles concentrations (22), mean HDL particle size (20), large HDL particles concentrations (12), small LDL particles concentration (8) and mean LDL particle size (6) [31]. The LP-IR score ranges from 0, the most insulin sensitive, to 100, the most insulin resistant.

Surrogate Indices of Insulin Sensitivity

HOMA-IR

The HOMA-IR is a surrogate index approximating the homeostasis model assessment of glucose and insulin dynamics. HOMA-IR was calculated as [[fasting insulin (μU/mL)] × [fasting glucose (mmol/L)]]/22.5 [32].

QUICKI

QUICKI is a transformation of fasting blood glucose and plasma insulin concentrations developed to serve as a surrogate index of insulin sensitivity. As described by Katz et al, QUICKI = 1/[log (fasting insulin, μU/mL) + log (fasting glucose, mg/dL)] [23].

Adipo-IR

The adipose tissue insulin resistance index (Adipo-IR) is a surrogate index of IR in adipose tissue. Adipo-IR is calculated from a single measurement of FFAs and insulin concentration, where Adipo-IR is equal to fasting FFA (mmol/L) × fasting insulin (μU/mL) [33].

Matsuda index

The Matsuda index is derived from dynamic mixed meal test data. Matsuda = 10 000/sqrt ([Glucose_fasting × Insulin_fasting] × [Glucose_mean × Insulin_mean]), where fasting glucose and insulin are the average of the −10 and 0 timepoints and mean glucose and insulin are the mean over the entire mixed meal test [24].

Statistical Analysis

Variables were tested for normality using D’Agostino and Pearson omnibus normality test and presented as means ± SD. A calibration model was used to assess the capability of surrogate indices in predicting S_I in which the surrogate index (eg, HOMA-IR, QUICKI, LP-IR, etc.) is regressed on an accurate measurement (eg, S_I) and the new x* (S_I derived) is predicted for a given y* (index derived). Two predicted residuals were used: (1) the difference between Minimal Model derived S_I (x_i for the ith subject) and fitted S_Iderived (⁠${\hat{x}}_i=\alpha +\beta {\hat{y}}_i$⁠), where all subjects were included in the model parameter α and β; (2) cross-validation type predicted residual $e_{(i)}=x_i-{\hat{x}}_{(i)}$ in which x_i is derived S_I from the Minimal Model, but ${\hat{x}}_{(i)}$ is the predicted S_I from the calibration model in which the ith subject is excluded.

Root Mean Squared Error and Leave-One-Out Cross-Validation Type Root Mean Squared Error of Prediction

Two measures of predictive accuracy were calculated from residuals: root mean squared error (RMSE) of prediction and leave-one-out cross-validation type root mean squared error of prediction (CVPE). CVPE is a more robust measure due to its exclusion of the ith subject when predicting the ith subject's results. In addition, the bootstrapping technique was performed with 60 000 replications for each comparison. The bootstrapping technique served as a robust method for cross-validation. P-values obtained from comparisons of RMSE and CVPE were used to compare groups.

Receiver Operating Characteristic Curve and Cut-off Point Analysis

Participants were deemed insulin resistant and assigned a classification variable of “1” if they had an S_I ≤ 2.8. Insulin sensitive subjects (S_I > 2.8) were assigned a classification variable of “0”. Classification variables were compared with the corresponding calculated data point for each surrogate index. Higher values were classified as high risk for HOMA-IR, Adipo-IR, and LP-IR, but not QUICKI and Matsuda. The DeLong et al methodology was used to calculate the SE of the area under the curve (AUC) and CI calculations [34]. Nonparametric curve fitting was applied. The significance level determined for analysis was type 1 error = 0.05. Optimal cut-offs were calculated using the Youden method. Costs of false positives and false negatives were assigned a value of 1. Receiver operating characteristic (ROC) curve analysis was generated using GraphPad Prism. The curve and cut-off analysis were also performed using EasyROC web software (Version 1.3.1, www. biosoft.hacettepe.edu.tr/easyROC/) [35]. Cut-offs were calculated using OptimalCutpoints software (Version 1.1-4) within the EasyROC software.

Results

Clinical and Metabolic Characteristics of the Cohort

The baseline clinical and metabolic parameters of South Asians (24 females and 31 males) are summarized in Table 1. The mean age of participants was 36 ± 7.9 years. The average body mass index of the group was 26.7 ± 5.2 kg/m². Seventy-seven percent of the cohort was obese (n = 36) or overweight (n = 7). Among 55 participants, 8 presented with impaired fasting glucose (>100 mg/dL), and 4 presented with a higher than normal fasting insulin (>25 µU/mL). While the average hemoglobin A1c (HbA1c) was normal (5.26 ± 0.42%), 9 subjects had an HbA1c within the prediabetic range (5.9-6.4%). None of the participants presented fasting metabolic parameters that met the criteria for diabetes. As determined by FSIVGTT and minimal model analysis, mean insulin sensitivity (S_I) in our cohort was 3.88 ± 2.98 ×10⁻⁴ (μU/mL)⁻¹ minute⁻¹. Based on prior studies, S_I was dichotomized using the cut-off point for IR of <2.85 ×10⁻⁴ (μU/mL)⁻¹ minute⁻¹ [36, 37]. The average S_I of the participants with IR, 1.64 ± 0.73 ×10⁻⁴ (µU/mL)⁻¹ minute⁻¹ (n = 28) was lower than individuals with non-IR (6.19 ± 2.63 ×10⁻⁴ (µU/mL)⁻¹ minute⁻¹, n = 27, P < .001). Similarly, HOMA-IR and Adipo-IR were higher in individuals with IR (HOMA-IR: 3.70 ± 2.32 vs 1.74 ± 0.87, P = .0002; Adipo-IR: 9.24 ± 7.57 vs 4.54 ± 2.33, P = .001). Corresponding insulin sensitivity indices, QUICKI and Matsuda index, were lower in subjects with IR (QUICKI: 0.327 ± 0.037 vs 0.359 ± 0.027, P = .0002; Matsuda index: 4.09 ± 3.87 vs 6.20 ± 2.96, P = .0005). Nearly 51% of our cohort, 38% of the women (n = 9) and 61% of the men (n = 19), were insulin resistant.

Table 1.

Subject characteristics.

	South Asians (n = 55)
Age (years)	36.3 ± 7.9
Sex (% female)	44
Weight (kg)	73.9 ± 16.5
Body mass index (kg/m2)	26.7 ± 5.2
Body fat (%)	34.7 ± 7.4
Fat free mass (kg)	47.5 ± 9.5
A/G ratio	1.15 ± 0.22
Systolic blood pressure (mmHg)	113 ± 10
Diastolic blood pressure (mmHg)	70 ± 9
HbA1c (%)	5.26 ± 0.42
Creatinine	0.81 ± 0.16
Blood urea nitrogen (mg/dL)	11.3 ± 3.8
Alanine transaminase (U/L)	26.0 ± 19.6
Aspartate aminotransferase (U/L)	21.1 ± 6.9
Lipid parameters
Total cholesterol (mg/dL)	187 ± 34.2
HDL cholesterol (mg/dL)	50 ± 14
LDL cholesterol (mg/dL)	113 ± 33
Triglycerides (mg/dL)	121 ± 61
Metabolic parameters
Fasting glucose (mg/dL)	91.4 ± 8.6
Fasting insulin (mU/mL)	11.9 ± 8.4
Fasting C-peptide (ng/mL)	2.50 ± 0.98
Fasting free fatty acids (mEq/L)	0.586 ± 0.182
Insulin sensitivity/resistance
Insulin sensitivity ×10−4 (mU/mL)−1 (minute)−1	3.88 ± 2.98
HOMA-IR	2.74 ± 2.01
QUICKI	0.343 ± 0.036
Matsuda index	5.12 ± 3.6
Adipo-IR	6.93 ± 6.07
LP-IR index	47.61 ± 23.5

Data are presented as arithmetic mean ± SD; n = number of subjects.

Abbreviations: HOMA-IR, homeostatic model assessment of IR; QUICKI, quantitative insulin sensitivity check; ADIPO-IR, adipose tissue insulin resistance index; LP-IR, lipoprotein insulin resistance.

NMR Lipoprotein Subclass Profile and Insulin Sensitivity

We examined the relationships between insulin sensitivity (S_I) and lipoprotein subclass concentration and size. Relative weighing of these 6 NMR-derived lipoprotein variables is used to generate LP-IR. In the entire cohort, increasing S_I was negatively correlated with LDL (r = −0.46, P = .004), small LDL (r = −0.38, P = .005), and large VLDL (r = −0.52, P = .0002) particle concentrations, and positively related to large HDL particle concentration (r = 0.54, P < .0001). LDL-C (r = 0.37, P = .005) and HDL cholesterol (r = 0.56, P < .0001) particle diameters were positively related, while VLDL-C (r = −0.37, P = .005) particle size was negatively related to S_I. HDL cholesterol concentration was lower, but total cholesterol, LDL cholesterol, and triglyceride levels were higher in participants with IR (Table 2). Similarly, large LDL particle concentrations were lower, but levels of smaller LDL and large VLDL particles were higher in individuals with IR (Table 2).

Table 2.

Conventional lipid panel and NMR lipoprotein profile

	All (n = 55)	IR (n = 28)	Non-IR (n = 27)	P value
Total cholesterol (mg/dL)	187 ± 34	195 ± 34	179 ± 33	.06
LDL cholesterol (mg/dL)	113 ± 33	123 ± 33	103 ± 29	.03
HDL cholesterol (mg/dL)	50 ± 14	43 ± 11	57 ± 14	.0003
Triglycerides (mg/dL)	121 ± 61	144 ± 67	98 ± 46	.008
NMR-derived lipoprotein concentration
Small LDL (nmol/L)	928 ± 703	1157 ± 777	690 ± 532	.03
Large HDL (µmol/L)	1.33 ± 1.44	0.66 ± 0.69	2.03 ± 1.67	.001
Large VLDL (nmol/L)	3.01 ± 4.65	4.84 ± 5.88	1.09 ± 1.26	.014
Lipoprotein size
LDL (nm)	20.64 ± 0.57	20.49 ± 0.65	20.79 ± 0.42	.10
HDL (nm)	8.69 ± 0.41	8.51 ± 0.25	8.87 ± 0.46	.006
VLDL (nm)	45.36 ± 5.82	46.52 ± 5.92	44.16 ± 5.56	.18

Data presented as arithmetic mean ± SD; n = number of subjects.

Abbreviations: IR, insulin resistance; HDL, high-density lipoprotein; LDL, low-density lipoprotein; NMR, nuclear magnetic resonance; VLDL, very–low-density lipoprotein.

Correlations Between S_I and Simple Surrogate Indices of Insulin Sensitivity/Resistance

As determined by simple linear regression, all surrogate indices show similar, but modest, correlations with S_I (r ≈ 0.53-0.69, P < .0001) (Table 3). The strength of correlations between S_I and LP-IR, HOMA-IR, QUICKI, Matsuda index, or ADIPO-IR was comparable and not significantly different among various indices (P > .05). All surrogate indices examined provided the correct rank order of insulin sensitivity as defined by S_I quartiles and were different between individuals with and without IR. LP-IR was significantly associated with S_I even after adjustment for age, sex, and percent body fat.

Table 3.

Simple correlations between surrogate indices and insulin sensitivity

	r	P-value
HOMA-IR	−0.69	<.0001
QUICKI	0.53	<.0001
Matsuda index	0.62	<.0001
ADIPO-IR	−0.66	<.0001
LP-IR index	−0.54	<.0001

Correlations were calculated from simple linear regression of the natural log of each surrogate index and S_I described in the research methods.

Abbreviations: HOMA-IR, homeostatic model assessment of IR; QUICKI, quantitative insulin sensitivity check; ADIPO-IR, adipose tissue insulin resistance index; LP-IR, lipoprotein insulin resistance.

Predictive Accuracy of Surrogate Indices of Insulin Sensitivity/Resistance Using Calibration Analysis

Further, calibration model analysis was used to evaluate the predictive ability of each surrogate index to predict S_I. Using leave-one-out cross-validation analysis, we generated plots comparing predicted S_I values derived from each surrogate index with actual S_I derived from the FSIVGTT. If a surrogate index perfectly predicted S_I, the slope of the best-fit line of the calibration model would be 1 and a y-intercept of 0. Visual inspection of the plots suggests LP-IR showed a moderate predictive ability comparable to the other surrogate indices (Fig. 1A-1E). RMSE and CVPE were utilized to quantitatively compare the prediction errors between LP-IR and each simple surrogate indices of IR/insulin sensitivity. The RMSE and CVPE analyses reveal that there were no significant differences among RMSE and CVPE from any of the surrogate indices when compared with LP-IR (Table 4). These data suggest that LP-IR predicts insulin sensitivity/resistance to a similar degree as other simple surrogate indices.

Figure 1.

Comparison between measured S_I (derived from FSIVGTT) and predicted S_I derived from surrogate indices of insulin resistance/sensitivity: (A) LP-IR index, (B) Matsuda index, (C) HOMA-IR, (D) QUICKI, and (E) ADIPO-IR. The solid line with a slope of 1 and intercept of 0 represents ideal prediction accuracy. The solid line indicates the least linear squares fit for each predicted vs measured comparison. Correlation coefficients (r) and P values are shown in each panel.

Table 4.

RMSE and CVPE from calibration model analysis of surrogate insulin resistance indices

	RMSE-SI	CVPE-SI	RMSE	CVPE
			(P value)	(P value)
HOMA-IR	0.67	0.71	0.52	.48
QUICKI	0.69	0.74	0.52	.44
Matsuda index	0.66	0.68	0.54	.56
ADIPO-IR	0.64	0.65	0.48	.49
LP-IR index	0.71	0.73	—–	—–

Root mean squared error (RMSE) of prediction as well as leave-one-out cross-validation type RMSE of prediction (CVPE) were calculated from calibration model analysis of HOMA-IR, QUICKI, Matsuda index, ADIPO-IR, and LP-IR as described in the research methods. P-values correspond to comparisons between each surrogate and LP-IR index.

Abbreviations: HOMA-IR, homeostatic model assessment of IR; QUICKI, quantitative insulin sensitivity check; ADIPO-IR, adipose tissue insulin resistance index; LP-IR, lipoprotein insulin resistance.

ROC and Bayes Analyses

We used the ROC AUC analysis to determine the sensitivity and specificity for various surrogate indices to detect IR in South Asians. The presence of IR was defined as an S_I value <2.85 ×10⁻⁴ (mU/mL)⁻¹ minute⁻¹. AUC and 95% CI of each surrogate index are presented in Table 5. The ideal test with 100% sensitivity and specificity would have AUC of 1.0. The LP-IR AUC suggests that LP-IR strongly characterizes IR in South Asian subjects, and all surrogate indices have comparable AUC with overlapping 95% CI (Table 5). ROC analysis curves for the various indices are shown in Fig. 2. Optimal cut-offs in insulin resistant South Asian individuals for surrogate indices were determined as HOMA-IR, >2.29; QUICKI, <0.336; Matsuda, <4.28; Adipo-IR, >6.3. The optimal cut-off in insulin resistant South Asians for LP-IR was >48 (Table 5). Youden index (sensitivity + specificity −1) equals 1 in a perfect test since sensitivity and specificity are 1. Thus, a test with a higher Youden index has a better diagnostic performance. LP-IR has a comparable Youden index, specificity, sensitivity, positive predictive value (PPV), and likelihood ratio to the other surrogate indices (Table 5).

Figure 2.

Receiver operating characteristics area under the curve (ROC AUC) were calculated for each simple surrogate index of insulin sensitivity/resistance: (A) HOMA-IR, (B) QUICKI, (C) ADIPO-IR, (D) LP-IR, (E) Matsuda index, as described in the research methods. The solid line represents insulin resistance in the corresponding index. The dashed line represents the reference line.

Table 5.

ROC AUC and Bayes analyses for simple surrogate indices of insulin sensitivity/resistance

	LP-IR index	HOMA-IR	QUICKI	Matsuda index	Adipo-IR
Cut-off	>48	>2.29	<0.336	<4.28	>6.3
AUROC	0.77	0.79	0.79	0.77	0.76
P value	.0007	.0002	.0002	.0005	.001
95% CI	0.64-0.89	0.67-0.91	0.67-0.91	0.65-0.89	0.63-0.88
Youden index	0.45	0.52	0.52	0.49	0.46
Sensitivity	0.75	0.82	0.82	0.75	0.61
Specificity	0.70	0.70	0.70	0.74	0.85
Positive predictive value	0.72	0.74	0.74	0.75	0.81
Likelihood ratio	2.53	2.77	2.77	2.89	4.09

Quantitative description of AUC of ROC and cut-off analysis for each simple surrogate index of insulin resistance/sensitivity. Cut-off values were determined using the Youden method.

Abbreviations: AUC, area under the curve; AUROC, area under the receiver operating characteristic curve; HOMA-IR, homeostatic model assessment of IR; QUICKI, quantitative insulin sensitivity check; ADIPO-IR, adipose tissue insulin resistance index; LP-IR, lipoprotein insulin resistance; ROC, receiver operating characteristic.

Discussion

South Asians have a high prevalence of type 2 diabetes and premature atherosclerotic disease, conditions characterized by IR, a major risk factor. Identifying and treating IR early in this population is urgent and clinically important. Despite the important role of IR in the pathogenesis of diabetes and cardiovascular disease, we lack a readily available and reliable diagnostic test to identify individuals with IR. Surrogate indices of IR/insulin sensitivity often used in large epidemiological studies are based on fasting or stimulated (eg, oral glucose tolerance test) insulin levels. However, due to the absence of a standardized insulin assay, it is essential to robustly validate and thoroughly compare various insulin-based surrogate indices of insulin sensitivity/resistance and NMR-based LP-IR with the S_I derived from FSIVGTT to determine whether LP-IR may be used to detect IR in South Asians. The present study represents the first comprehensive validation study of LP-IR index in South Asians. We conclude that NMR-derived LP-IR may be an appropriate index to assess IR (above cutoff >48) in South Asians.

The EHC, insulin suppression test, and FSIVGTT are frequently used reference tests to assess insulin sensitivity [18, 38]. Defining cut-off point from these tests to discriminate IR and insulin sensitive subjects is clinically important but also challenging. Some investigators have used a fixed cut-off value, but others use a threshold value that is percentile specific, such as the lower tertile, or greater than the 75th percentile, as a cut point [38-43]. However, other investigators have used a statistical approach of selecting a single threshold value that simultaneously provides high sensitivity and specificity [39, 40]. A single value defining IR may be advantageous over cut-offs based on the population distribution of IR. No prior studies in South Asians suggest a cut-off for IR based on FSIVGTTs. Based on previous studies, albeit predominantly in Europeans, we chose a cut-off point of SI <2.85 ×10⁻⁴ (mU/mL)⁻¹ minute⁻¹ to define IR [36, 37].

Nearly two-thirds of apparently healthy Asian Indian men were insulin resistant as defined by a glucose disposal rate of less than 30 μmol/kg/minute as measured by EHC [17]. Similarly, using a cut-off steady-state plasma glucose of ≥150 mg/dL [20], approximately 44% of healthy women were insulin resistant [44]. In young, healthy South Asians, approximately 37% of the men and 25% of the women were insulin resistant (defined as a value >75th percentile HOMA-IR) [45]. In the Southall And Brent REvisited (SABRE) study, the HOMA-IR distribution in South Asians without diabetes was skewed right with a heavier tail, suggesting a greater prevalence of IR in this population (∼46%) [19]. Using a cut point of S_I <2.85 ×10⁻⁴ (mU/mL)⁻¹ minute⁻¹ [36, 37], nearly 51% of our cohort, were insulin resistant. This prevalence is similar to reported rates in South Asian studies, using dynamic tests to assess IR [17, 20].

It is well known that IR associates with dyslipoproteinemia, which typically precedes glycemic abnormalities [46]. The LP-IR score correlates more strongly with IR (as determined by HOMA-IR) than each of the 6 individual parameters including particle diameter (VLDL-P, HDL-P, LDL-P) and concentration (VLDL-P, large HDL-P, small LDL-P). In our study, HOMA-IR was related to LP-IR (r = 0.57, P < .0001) similarly to that observed in the original study that developed the LP-IR score (r = 0.51, P < .001) [28]. Likewise, S_I from FSIVGTT, a measure of hepatic and peripheral insulin sensitivity, correlates with LP-IR (r = −0.54, P < .0001). The strength of this relationship in our study was similar to the association observed between glucose disposal rate determined during EHC and LP-IR (r = −0.53, P < .001) in the original study [28].

In inulin-resistant South Asians, concentrations of small LDL and large VLDL are higher, and the HDL size and levels of large HDL are lower than in noninsulin resistant individuals. It is worth noting that variables are weighted differentially when a composite LP-IR is calculated. Indeed, VLDL/HDL particle size and large VLDL/HDL particle concentrations primarily determine the magnitude of the LP-IR score.

Hyperinsulinemia and adipocyte/hepatic IR drive increased hepatic delivery of FFAs and increased VLDL production leading to higher LDL levels [46]. Furthermore, higher activity of hepatic lipase and cholesteryl ester transfer protein (CETP) results in smaller LDL and HDL particles. CETP transfers triglycerides from VLDL to LDL/HDL particles in exchange for cholesteryl esters. Hepatic lipase hydrolyzes triglycerides in LDL and HDL, which results in the generation of smaller size lipoprotein particles. Smaller HDL particles are prone to increased clearance leading to lower HDL levels. Higher hepatic steatosis, IR, and CETP activity may together explain the lipoprotein profile seen in insulin resistant South Asians [47, 48]. This suggests that there is a biological basis for the observed relationship between LP-IR and insulin sensitivity. Consistent with this, multiple studies show that LP-IR predicts T2DM [28, 31, 49-51].

Assessing absolute predictive accuracy is essential for rigorous validation of LP-IR. In this study, predictive accuracy was assessed by 2 error function criteria, RMSE and CVPE. Lower RMSE and CVPE indicate less error. LP-IR and other insulin-based surrogate indices had comparable RMSE and CVPE. ROC analysis suggests that the LP-IR index, similarly to other indices, strongly distinguishes individuals with and without IR in South Asians. However, among the indices, the lower sensitivity precludes ADIPO-IR from being a reliable test to screen for IR. Since we have determined the strong relationship between the LP-IR index and its accuracy in predicting IR, we identify that the ROC-defined cutoff of >48 can distinguish South Asians with IR. The mean LP-IR score in our study sample was 47.61 ± 23.5, meaning that on average, our study sample was highly insulin resistant, which is consistent with findings using other simple surrogate indices. Overall, these data suggest that the LP-IR index is valid in South Asians and can be used as a more accessible method of identifying IR in this group.

Among the limitations, the small sample size precluded our ability to detect sex-dependent differences in the predictive accuracy of LP-IR. Although not the objective of this study, the findings are from a cross-section study and do not address whether LP-IR predicts incident T2DM in South Asians. EHC is the gold standard technique to measure IR, and in future studies, LP-IR should be validated against adipose and hepatic IR measures in South Asians. Lastly, LP-IR should be validated in other Asian populations characterized by IR with normal body mass index. The ¹H-NMR assay is robust and cheaper and widely used in large epidemiologic studies. However, standardization of sample processing and transport, harmonization of downstream quantification algorithms from raw NMR spectra, interspectrometer variation, within-spectrometer drift over time, and costs related to instrument acquisition may limit the use of NMR-derived LP-IR in countries constrained by resources.

Ideally, a screening test for IR should be simple, safe, affordable, acceptable to patients and health care providers, accurate, reliable, with sufficient sensitivity and specificity, and easy to administer. LP-IR predicts T2DM and cardiovascular events even in patients without overt glycemia. Furthermore, improvement in IR following nonpharmacological and pharmacological interventions is sufficiently captured by changes in lipoprotein profile. LP-IR score can be estimated by NMR analysis of a single blood sample and is currently offered by commercial labs at a reasonable price. In addition, LP-IR is reasonably accurate with sufficient discrimination power. NMR analysis also provides additional parameters, such as GlycA, a marker of inflammation, that may be useful for cardiovascular risk assessment. Assuming a 40% prevalence of IR in apparently healthy South Asians and given the positive and negative likelihood ratios of LP-IR are 2.53 and 0.35, respectively, the post-test probability of a positive test is 62.5% and a negative test is 18.9%. In contrast, simple fasting estimates of IR/insulin sensitivity (eg, HOMA-IR and QUICKI) are dependent on the fasting insulin level. In addition to the lack of standardization of insulin assay, there is considerable variability (intra-assay and interassay variability). Moreover, to reduce the variability arising out of the pulsatile nature of insulin secretion, pooling of at least 2 samples over a 10- to 15-minute period is recommended [52]. Lastly, we have shown that fasting insulin–based surrogate indices are limited in their predictive ability in South Asians [53].

In conclusion, LP-IR is simple, affordable, accurate, and the preferred method for assessing IR in South Asians in the clinical setting.

Funding

This work was supported in part by the Intramural Research Program of the National Institutes of Diabetes and Digestive and Kidney Diseases (NIDDK), Washington, DC.

Disclosures

The authors have nothing to disclose.

Data Availability

Some or all datasets generated during and/or analyzed during the current study are not publicly available but are available from the corresponding author on reasonable request.

Clinical Trial Information

ClinicalTrials.gov Identifier: NCT00428987 (January 30, 2007)

Abbreviations

 
• Adipo-IR

  adipose tissue insulin resistance index

 
• AUROC

  area under the receiver operating characteristic curve

 
• CETP

  cholesteryl ester transfer protein

 
• CVPE

  leave-one-out cross-validation type root mean squared error of prediction

 
• EHC

  hyperinsulinemic–euglycemic clamp

 
• FFA

  free fatty acid

 
• FSIVGTT

  frequently sampled intravenous glucose tolerance test

 
• HDL

  high-density lipoprotein

 
• HOMA-IR

  Homeostatic Model Assessment of Insulin Resistance

 
• IR

  insulin resistance

 
• LDL

  low-density lipoprotein

 
• LP-IR

  lipoprotein insulin resistance index

 
• NMR

  nuclear magnetic resonance

 
• QUICKI

  quantitative insulin sensitivity check index

 
• RMSE

  root mean squared error

 
• S_I

  insulin sensitivity

 
• T2DM

  type 2 diabetes

 
• VLDL

  very–low-density lipoprotein