Variability, Mean, and Baseline Values of Metabolic Parameters in Predicting Risk of Type 2 Diabetes

Abstract

Context

The effect of baseline (B) and alteration of metabolic parameters (MPs), including plasma glucose (PG) testing, insulin resistance surrogates, and lipid profile and their mutual interactions on the development of type 2 diabetes mellitus (T2DM), has not been investigated systematically.

Objective

To access the association of the past variability (V), past mean (M), and B values of various MPs and their mutual interaction with the risk of T2DM.

Methods

A community-based, longitudinal analysis was conducted using the Korean Genome and Epidemiology Study comprising 3829 nondiabetic participants with completed MPs measurements during 3 biannually visits who were followed over the next 10 years. Outcomes included the incidence of T2DM during follow-up.

Results

Among predictors, PG concentrations measured during the oral glucose tolerance test were the most prominent T2DM determinants, in which the M of the average value of fasting PG (FPG), 1-hour, and 2-hour PGs had the strongest discriminative power (hazard ratios and 95% CI for an increment of SD: 3.00 (2.5-3.26), AUC: 0.82). The M values of MPs were superior to their B and V values in predicting T2DM, especially among postload PGs. Various mutual interactions between indices and among MPs were found. The most consistent interactants were the M values of high-density lipoprotein cholesterol and the M and V values of FPG. The findings were similar in normal glucose tolerance participants and were confirmed by sensitivity analyses.

Conclusion

Postload PG, past alteration of measurements, and mutual interactions among indices of MPs are important risk factors for T2DM development.

Long-term dysfunction in metabolic response to glucose intake is attributed to the pathogenesis of type 2 diabetes mellitus (T2DM). A decrease in insulin sensitivity and secretion persists silently for many years before the occurrence of T2DM (1). Various metabolic traits, including plasma glucose (PG) testing, insulin resistance surrogates, and lipid profile have been used to predict the occurrence of T2DM. Fasting plasma glucose (FPG) is a traditional marker used for T2DM detection (2) and for determining the risk of future T2DM (3). It has been reported that postload PG concentrations, including 1-hour and 2-hour PG during the oral glucose tolerance test (OGTT) could be considered as better predictors of T2DM and prediabetes (4, 5). In a previous work, we found that the average FPG and postload PG levels are superior to each index in association with T2DM development (6). Insulin sensitivity and secretion status estimated using OGTT information represents a potential marker of T2DM (7). According to existing evidence, the dysregulation of the lipid profile, including total cholesterol (TC), triglycerides (TG), high-density lipoprotein cholesterol (HDL), and their ratios, is associated with a higher risk of T2DM (8, 9). A combination of FPG and lipid profiles may result in additive discrimination power for each marker in predicting T2DM (10). However, a systematic analysis of the interaction between metabolic markers in association with diabetes has not yet been performed.

In recent decades, there has been an emerging interest in the additional effect of the visit-to-visit alteration of risk factors on metabolic-related disorders, especially T2DM. However, the evidence is still controversial. Laspa et al (11) reported that the variability of risk factors, including FPG, lipid profiles, and liver function tests, was not associated with glucose impairment and T2DM. However, in this study, the variability of metabolic indices was calculated using regression that may not reflect the real alteration of the values. Several studies employing widely used indices of visit-to-visit variability, such as SD, coefficient of variation (CV), variability independent of the mean (VIM), and average real variability (ARV), provided consistent findings. The high fluctuation of body mass index (BMI) was a potential risk factor for T2DM in Japanese adults (12). An 8-year follow-up study reported that an SD increment of FPG variability was associated with a 24% higher risk of T2DM development (13). Also, according to existing evidence, the risk of T2DM was substantially higher in those with high fluctuations in their lipid profiles, including HDL (14) and TC (15). However, to date, there has been no study using the past average of risk factors as a proxy indicator of T2DM.

T2DM develops as a consequence of a long-term process that involves several pathogenic pathways with their complex interactions. These interactions may occur between the baseline (B) and the variability within each factor, as well as between factors. Each pathway might play a different role and magnitude in the disease. Analyses for a given metabolic parameter (MP) provide a small slice of the whole picture. In several studies, the interaction of 2 parameters was analyzed by simply dichotomizing continuous parameters into the high and the low and classifying them into 4 groups (14, 15). However, this method could not clarify the direction of the interaction. To the best of our knowledge, there has been no previous study that examined the B, past average, and variability of risk factors for T2DM and their interactions simultaneously. Therefore, this study aims to compare the magnitude of the effect of past variability, past average, and B values of various risk factors and their complex interactions on the development of T2DM.

Material and Methods

Participants

The Korea Genome and Epidemiology Study (KoGES) is an ongoing population-based cohort study designed to investigate the genetic and environmental etiology of common complex diseases in Korean adults (16). A total of 10 030 Korean adults between age 40 and 69 years were first enrolled in 2001 to 2002 in Ansan (urban area) and Ansung (rural area) and were followed up biennially. We allocated the first 6-year follow-up (first, second, and third visits) for variability calculation (Time_var), and the third visit was considered as B. The 10-year follow-up time (Time_fol) ran from B to the date of T2DM incidence or the last visit before the occurrence of T2DM. A total of 4586 participants completed all measurements of 11 MPs, including body weight (BW), BMI, lipid profile (plasma TC, TG, and HDL levels), and OGTT responses (PG and plasma insulin [PI] concentrations during fasting, at 1 hour [1-h], and 2 hours [2-h]) within the Time_var were selected. We excluded those having T2DM during the first 3 visits (n = 700), those with interval-censored (T2DM occurred after the third visit but the time of event could not be determined because of missing data) (n = 51), and those with missing covariate data at the third visit (n = 6). Finally, a total of 3829 nondiabetic participants were included in the main analysis (Fig. 1). The data sets used for sensitivity analyses (SAs) are described in Supplementary Tables S1 and S2 (17). All participants gave their informed consent, and the study protocol was approved by the institutional review board of the corresponding study centers.

Measurement and Definition

A 75-g OGTT was performed during each visit in a fasting state to measure PGs (FPG, 1-h, and 2-h PGs) and corresponding PIs using a radioimmunoassay (INS-IRMA Kit, BioSource). The average of PGs and PIs measured during fasting, at 1-h, and at 2-h during OGTT in each visit were calculated (PG_ave and PI_ave). Insulin resistance was examined using the homeostasis model assessment of insulin resistance (HOMA-IR) and the Matsuda index (18, 19). Glycated hemoglobin A_1c (HbA1c) was measured using high-performance liquid chromatography (Variant II; Bio-Rad Laboratories). Plasma TC, TG, and HDL were measured by using a Hitachi 747 chemistry analyzer (Hitachi Ltd). Blood pressure was measured using a sphygmomanometer (CK-101; Chin Kou Medical Instrument Co Ltd). A self-report questionnaire was completed to survey the history of diabetes diagnosis; hypertension (HTN) diagnosis; medication usage for diabetes, dyslipidemia, and HTN; family history of diabetes; drinking and smoking habits (current user); and physical activity (habitual exercise to work up a sweat). BMI was calculated as BW in kilograms divided by height in meters squared.

At each visit, the presence of DM was determined using the World Health Organization criteria (diagnosed with DM or antihyperglycemic medication or FPG ≥ 126 mg/dL or 2-h PG ≥ 200 mg/dL or HbA_1c ≥ 6.5%) (20). Individuals with prediabetes (Pre-DM) were nondiabetic participants having an FPG between 125 and 100 mg/dL or 2-h PG between 140 and 199 mg/dL or HbA_1c between 5.7% and 6.4% (20). The normal glucose tolerance (NGT) participants were nondiabetics and non–Pre-DM. HTN was defined as having a previous HTN diagnosis or the use of at least one antihypertensive agent or systolic blood pressure greater than or equal to 140 mm Hg or diastolic blood pressure greater than or equal to 90 mm Hg.

Baseline, Mean, and Variability of Metabolic Parameters

The variability of each parameter was defined as the intraindividual alteration during Time_var. To date there has been no consensus on a gold-standard index for visit-to-visit variability, and previous studies have reported no statistically significant difference between variability indices (13, 14). We used the ARV index because it is the most convenient calculation for clinical application. $ARV=(Δ1+Δ2)/2$in which $Δ1$and $Δ2$refer to the absolute difference in measurements between visit 1 and visit 2 and between visit 2 and visit 3, respectively. The mean value of measurements during Time_var (M) was calculated, and the measurement during visit 3 was considered as B for the next 10-year follow-up. The B, M, and ARV of 15 parameters, including BW, BMI, TC, TG, HDL, FPG, 1-h PG, 2-h PG, PG_ave, FPI, 1-h PI, 2-h PI, PI_ave, Matsuda index, and HOMA-IR (45 predictors) were used in the analysis.

Statistical Analysis

Data were analyzed using R version 4.0.4, and the threshold for statistical significance was set at P less than .05. All nonnormally distributed continuous variables were log-transformed before analysis. The distribution of the variables before and after log-transformation is presented in Supplementary Fig S1 (17). The B characteristics of those who had incident DM vs those who did not were compared using the t test and the chi-square test for continuous and categorical variables, respectively. To compare the magnitude of the effect between predictors, each predictor was converted to a sex-specific z score before performing the survival analysis. To examine the effects of each parameter in predicting incident T2DM, we used the Cox proportional hazards model adjusted for covariates in which a given MP was the independent variable. Covariates were age, sex, BMI, drinking (current drinker versus non) and smoking status (current smoker vs nonsmoker), physical exercise (undertaking habitual exercise to sweat or not), family history of DM (yes/no), history of dyslipidemia medication (yes/no), and HTN (yes/no) at B. The covariate BMI was excluded in the model that examines the effect of M, B, and ARV of BMI as a predictor.

The discriminative ability of MPs and their indices in predicting the incidence of T2DM during a 10-year follow-up were compared using the area under the curve (AUC) of the receiver operating characteristic curve. The optimal cutoff for each parameter was estimated using the Youden method.

The interaction terms in log-hazard (logH) between M, B, and ARV of a parameter compared with that of the other parameters in a given pair were calculated to examine the synergistic effect of 2 predictors (X₁ and X₂) in the model. We used the Cox proportional hazards model adjusted for covariates in which 2 MPs in pairs and their interaction term were independent variables. The same covariates were added. A positive logH means there is an additive effect between X₁ and X₂, whereas a negative logH means an increase in X₁ attenuates the effect of an increase in X₂, and vice versa (21). Considering the loss of data due to the high exclusion rate, the SA for the effect of each parameter, their discrimination power, and for the interaction term between B, M, and ARV within each parameter was performed on SA_data1, whereas the SA for the interaction term between indices of parameters in pairs was performed on SA_data2 (see Supplementary Tables S1 and S2) (17).

Results

Characteristics of the Study Participants

The characteristics of 3829 nondiabetic participants at B (visit 3) are presented in Table 1. Those who developed T2DM (progressors) during the 10-year follow-up were older and had higher B and M values of BW, BMI, TC, TG, PGs, PIs, HOMA-IR; lower B and M values of HDL and Matsuda index; and higher ARVs of TC, TG, PGs, PIs, Matsuda index, and HOMA-IR compared with those who remained nondiabetic (nonprogressors). These differences were attenuated in NGT participants but remained consistent for the M values of BW, BMI, TG, HDL, FPG, 1-h PG, 2-h PG, PG_ave, 2-h PI, PI_ave, Matsuda index, and HOMA-IR.

Risk of Type 2 Diabetes Mellitus for Each Index of Metabolic Parameters

The median follow-up time was 8 years both for nondiabetic and NGT participants. A total of 593 nondiabetic (15.5%) and 145 NGT (6.5%) participants developed T2DM during follow-up. The multivariable Cox proportional hazard model revealed that PGs were the most prominent predictors of T2DM. An increase in the SD of M or B values of PGs was associated with a relatively high elevation of the risk of developing T2DM (95% CI of adjusted hazard ratio range (HR_range) = 1.70-3.26 in nondiabetic and 1.08-2.87 in NGT participants, in which M_{PG_ave} had the highest effect), whereas the statistically significant effect of ARV of PGs on T2DM was lower (HR_range = 1.05-1.47 for all PGs in nondiabetics and 1.13-1.60 for 2-h PG in NGT participants). BW had no effect on the development of T2DM, whereas M and B values of the BMI were statistically significantly associated with the development of T2DM. Among lipid profile parameters, TG level was the most prominent predictor of T2DM (HR_range = 1.29-1.50 and 1.22-1.64 in nondiabetic and NGT individuals, respectively). Meanwhile, the effect of M and B values of HDL and ARV of TC was statistically significant in nondiabetic individuals and vanished in NGT participants. Among insulin-related parameters, only M and ARV of 2-h PI and the Matsuda index showed statistically significant effects on T2DM both in nondiabetic and NGT participants (HR_range = 1.11–1.42 for 2-h PI and 0.59-0.83 for Matsuda index in nondiabetic individuals and 1.04-1.48 for 2-h PI and 0.63-0.94 for Matsuda index in NGT participants) (Fig. 2 and Supplementary Table S3) (17). The SA to investigate the effect of each parameter confirmed the trend in the main analysis (Supplementary Fig. S2 and Supplementary Table S4) (17).

Discrimination Power of Metabolic Parameters

In general, among the 3 indices, the discriminative capacity of M was highest, then that of B, and that of ARV was lowest. Table 2 shows that M and B of PGs had the highest discriminative ability (AUC range = 0.70-0.82 in nondiabetic and 0.60-0.75 in NGT participants), in which M_{PG_ave} and M_{1-h PG} were the strongest predictors. Meanwhile, the ARVs of PGs were relatively weak predictors (AUC range = 0.51-0.59). The M and B of TG could be considered the second strongest predictors with higher AUCs compared to that of the other parameters, especially in NGT participants (AUC range = 0.61-0.64), whereas HDL and Matsuda index had the lowest discriminative ability with a high possibility of false positive and false negative (AUC range = 0.34-0.50). This pattern appeared similarly in SA (Supplementary Table S5) (17).

Interaction Term

Fig. 3 shows the matrix of interaction term of predictors in pairs for nondiabetic (upper triangle) and NGT individuals (lower triangle). In terms of interaction within indices of each parameter, among nondiabetic participants, several negative interactions were found for BMI (ARV vs M and B), TC (ARV vs M and B), 1-h PG (B vs M and ARV), 2-h PG (B vs ARV), PG_ave (B vs ARV), FPI (M vs ARV), and HOMA-IR (B vs M and ARV, and M vs ARV), whereas positive interaction was found for 2-h PG (B vs M), and PI_ave (B vs ARV). However, these interactions were eliminated in NGT participants, except that between M_HOMA-IR and ARV_HOMA-IR.

In terms of the interaction of parameters in pairs, a consistent negative interaction both in nondiabetic and NGT participants was found between M_HDL vs B_BMI, M_BMI, B_BW, M_FPI, and M_HOMA-IR (95% CI logHR range [logHR_range] = –0.008 to –0.175), between ARV_{1-h PG} and B_{PG_ave}, and between M_HOMA-IR and ARV_HOMA-IR and M_FPI. A consistent positive interaction in both nondiabetic and NGT participants was found between ARV_FPG vs PI-related indices (B_{1-h PI}, M_{1-h PI}, M_{2-h PI}, M_{PI_ave}, and ARV_{PI_ave} with logHR_range = 0.004-0.283) and ARV_TG (logHR_range = 0.035-0.195), and between B_{Matsuda index} and ARV_{PI_ave} (logHR_range = 0.013-0.167) (Supplementary Table S6) (17). The SA for interaction term confirmed the same results in the synergy effect between HDL vs BMI, BW, and M_FPI, and between ARV_FPG vs PI-related indices (Supplementary Fig. S3 and Supplementary Table S7) (17).

Discussion

The effects of the B, M, and ARV values of 15 MPs on T2DM and their interaction terms were systematically examined. The M values were stronger than the B and ARV values in predicting future T2DM, whereas PGs were superior to other MPs, of which M_{PG_ave} was the strongest predictor. The consistent additive effect of predictors in pairs occurred among various predictors, especially between M_HDL and ARV_FPG vs other parameters.

The association between metabolic-related factors such as obesity, high PGs and PIs, and poor lipid profiles and the development of T2DM has been thoroughly investigated (2-10). However, this evidence mostly relied on B information and the individual alteration of the factor in the past was not considered properly. Recently, there have been several studies that aimed to investigate the role of visit-to-visit variability of MPs in the risk of diabetes. Statistically significant effects were reported for BMI (12), TC (15), HDL (14), and FPG (13). However, it is unclear whether the variability of parameters is clinically meaningful as an alternative for the B measurement. Our study is the first attempt to compare the magnitude of the association of B, M, and ARV with T2DM in various MPs. We found that although the ARV was a statistically significant risk factor of T2DM development for TC, TG, PGs, 2-h PI, the Matsuda index, and HOMA-IR among nondiabetic individuals, only the associations for TG, 2-h PG, 2-h PI, and the Matsuda index remained statistically significant among NGT participants. Furthermore, the role of ARV and its discriminative capacity was significantly lower than that of the B and M values, especially among PGs. Meanwhile, the M value, which includes the information of B as well as the past variability, was superior both to the B and the ARV values in predicting future T2DM. In line with that, a previous study reported a stronger association between M_HDL and T2DM risk compared to those of VIM_HDL and CV_HDL (14), whereas there was no statistically significant difference in the magnitude of the effects of variability indices, including VIM, CV, and SD (13-15). Taken together, these findings suggest that, while using the information of visit-to-visit measurements, the M value appears to be more useful than variability indices.

Evidence suggests that glycemic variability is a proxy of β-cell dysfunction (22), which is associated with the excessive accumulation of oxidative stress (23) and increased risk of insulin resistance (24). In our previous work, we reported that B_{PG_ave} was superior to the B values of FPG, 1-h, and 2-h PG in predicting future T2DM. Interestingly, among 45 screened predictors, M_{PG_ave} was substantially better than B_{PG_ave} and others both in nondiabetic individuals and NGT participants. The present study reconfirmed the important role of postload PGs in predicting future T2DM and the superiority of the averaging method in addressing time-dependent measurements. Clinically, the calculation of the M value is convenient and more applicable than any variability estimation.

The effect of the variability of the lipid profile on the development of T2DM has been debated. Previous studies reported a statistically significant effect of the variability of HDL (14) and TC (15), whereas no relationship with TG was found (25). However, in these studies, analyses were performed by dichotomizing continuous explanatory variables, thereby increasing the risk of a false-positive result (26). Our data revealed only the consistent effect of ARV_TG both in nondiabetic and NGT participants, and the roles of the M, B, and ARV values were comparable. More studies should be conducted to further investigate these findings.

From a complex point of view, this is the first study that systematically examined whether the MPs and their 3 indices interact with each other in predicting T2DM. Among nondiabetic individuals, previous studies reported no interaction between M_FPG and FPG variability (13). We found that no interaction exists between the M, B, and ARV values of FPG, whereas the B value interacted inversely with ARV for 1-h, 2-h PGs, and PG_ave, and this pattern was confirmed in SAs. The clinical application is that the high risk of T2DM development due to a high B value of postload PGs becomes higher if the past measurements of postload PGs do not alternate greatly, and vice versa. The synergistic effect of small variability among high B levels was also seen for BMI, TC, HOMA-IR, and FPI. Thus, future estimations of T2DM risk using MPs should integrate the past alteration of the measurements for better results.

The interaction between parameters was screened. It was recognized that dyslipidemia plays an important role in the progression of glucose metabolic dysfunction, in which the contribution of HDL is substantial (27). The available evidence demonstrated that HDL may directly increase insulin sensitivity and secretion via anti-inflammation and the cholesterol efflux pathway (28, 29), and the infusion of HDL may improve glucose metabolism (30). We found a consistent interaction between M_HDL vs body size parameters (BMI and BW), FPG, and PI-related parameters. Because a low HDL refers to a high risk of T2DM in general, a negative interaction of HDL suggested that the risk increases when low HDL is accompanied by other concomitant factors, including high BMI, high FPG, or high PIs. There is the same interpretation of a negative interaction between M_FPG vs M and B of postload PGs. Clinically, the high risk of T2DM among those having high postload PGs may be more serious if they have low M_FPG. Although ARV_FPG and postload PIs interact, the effects of these 2 factors were modest. The FPG level is a proxy indicator of hepatic insulin resistance, whereas abnormal elevation of postload PGs is related to muscle insulin resistance and β-cell dysfunction (1). Our findings suggest that T2DM prediction using postload measurements should be interpreted within the context of FPG and its variability.

One of the strengths of the present study was the systematic screening of various widely used MPs, considering the B and past alteration information and mutual interactions using a nationwide scale cohort. This approach provides a broader and more complex view of the association between MPs and T2DM development. To validate the result, SAs were thoroughly performed. However, the study has several limitations. First, because the cohort was designed to collect data biannually, the interval between visits was relatively high. Therefore, information about previous measurements was not completely included. Second, because B was set up at the third visit, the participant’s age at B was between 46 and 75 years. Therefore, the results should not be extended to the population with ages out of this range. Third, there are several methods of calculating variabilities such as CV, SD, VIM, and ARV. We reported the results based only on ARV, which may cause criticism. However, in clinical practice, the calculation of ARV is most convenient and its effect was comparable with other variability indices (Supplementary Fig. S4) (17).

Our study revealed that among various MPs, the most prominent predictor of future T2DM was M_{PG_ave}. The alteration of measurements in the past provided additional information for T2DM risk calculation, in which the effect of the M value was highest. There have been statistically significant interactions between parameters, and the most consistent interactants of T2DM risk calculation were M_HDL and the M and ARV values of FPG. These findings reconfirmed the important role of postload PGs, past alteration of measurements, and mutual interactions in T2DM risk calculation.

Abbreviations

Abbreviations
 
• ARV

  average real variability

 
• AUC

  area under the curve

 
• B

  baseline measurement at visit 3

 
• BMI

  body mass index

 
• BW

  body weight

 
• CV

  coefficient of variation

 
• FPG

  fasting plasma glucose

 
• FPI

  fasting plasma insulin

 
• HbA_1c

  glycated hemoglobin A_1c

 
• HDL

  high-density lipoprotein cholesterol

 
• HOMA-IR

  homeostasis model assessment of insulin resistance

 
• HR_range

  adjusted hazard ratio range

 
• HTN

  hypertension

 
• logH

  log-hazard

 
• M

  mean value of measurements during Time_var

 
• MP

  metabolic parameter

 
• NGT

  normal glucose tolerance

 
• OGTT

  oral glucose tolerance test

 
• PG

  plasma glucose

 
• PG_ave

  mean of FPG, 1-hour PG, and 2-hour PG at each visit

 
• PI

  plasma insulin

 
• PI_ave

  mean of FPI, 1-hour PI, and 2-hour PI at each visit

 
• SA

  sensitivity analysis

 
• T2DM

  type 2 diabetes mellitus

 
• TC

  total cholesterol

 
• TG

  triglycerides

 
• Time_fol

  10-year follow-up time

 
• Time_var

  the first 6-year follow-up time for variability calculation

 
• VIM

  variability independent of the mean

Acknowledgments

The authors thank the Korean Genome and Epidemiology Study National Research Institute of Health, Centers for Disease Control and Prevention, Ministry for Health and Welfare, and the Republic of Korea for sharing data.

Financial Support

This study was supported by the Korea Institute of Industrial Technology as “Development of core technology for smart wellness care based on cleaner production process technology” (No. PEH21050), and the program of the National Research Foundation (NRF) was funded by the Ministry of Science, ICT, and Future Planning (No. 2016M3C1A6936606).

Author Contributions

D.D.P. and C.H.L. conceptualized the hypothesis and designed the work. D.D.P. analyzed the data and drafted the manuscript. J.S., Y.J., and I.H. contributed partly to data management. All authors revised and reviewed the manuscript critically. All authors read and approved the final version of the manuscript before submission.

Disclosures

The authors have nothing to disclose.

Data Availability

Data in this study are from the Korean Genome and Epidemiology Study (KoGES; 4851-302), National Research Institute of Health, Centers for Disease Control and Prevention, and the Ministry for Health and Welfare, Republic of Korea, and are available at https://nih.go.kr/contents.es?mid=a40504010000.

References