Variabilities in Childhood Cardiovascular Risk Factors and Incident Diabetes in Adulthood: The Bogalusa Heart Study

Measures of cardiovascular risk factors (CVRFs), including BMI, blood pressure (BP), and atherogenic lipids, are typically used to assess an individual’s risk for cardiovascular events and diagnosis of obesity, hypertension, and dyslipidemia and subsequently guide the need for antihypertensive or lipid-lowering drugs (1–3). However, these CVRFs can fluctuate due to genetic, clinical, physiological, behavioral, and environmental factors (4–8). Loss of physiological homeostasis, for example, through disease or the aging process, would lead to disturbances in levels of these CVRFs and variability over time compared with persistent healthy levels. These intraindividual variabilities in CVRFs have recently sparked interest as potential novel markers that can be used in addition to mean levels of CVRFs to help in risk prediction. Long-term follow-up data show that higher variability in BMI, BP, or lipids has been consistently associated with future adverse outcomes such as cardiovascular events, cognitive dysfunction, and mortality, independent of their respective absolute levels and other traditional risk factors (9–17).

Variability in CVRFs as a possible marker of residual risk for target-organ damage has been extensively studied in middle-aged/older persons or high-risk populations (9–17). However, to date, there has been little examination of the effects of variability in CVRFs during childhood, beyond that due to growth and development, on the risk of future diabetes. In middle-aged and older adults, CVRF variability itself could be influenced substantially by common confounding risk factors such as older age, autonomic dysfunction, and antihypertensive/antidiabetes/lipid-lowering agents (dosing, adherence, etc.), as well as by preexisting comorbidities such as atherosclerosis and chronic liver/kidney disease. Whether the significant association between CVRF variability and adverse outcomes is the result of confounding risk factors and preexisting comorbidities or due to true the adverse effect of variability itself is not clear. The Bogalusa Heart Study (BHS), which is unique in having repeatedly measured and prospectively collected CVRFs from childhood through young adulthood to midlife, provides an opportunity to determine whether variability in CVRFs during childhood, a time when few confounding risk factors or comorbidities exist, is associated with diabetes in adulthood.

Exploring the contribution of variability in CVRFs during childhood could increase our understanding of how dysfunction in CVRF homeostasis impacts the development of diabetes later in life and may have implications for early prevention and intervention. Thus, using BHS data, we aimed to examine whether variability in CVRFs during childhood is an independent risk factor for diabetes in adulthood.

The BHS is a series of long-term studies initiated in 1973 and designed to increase our understanding of the early natural history of cardiovascular disease since childhood (18). Between 1973 and 2016, 9 cross-sectional surveys of children aged 4–19 years and 11 surveys of adults aged 20–58 years, who had been previously examined as children, were conducted in the semirural biracial (65% white and 35% black) community of Bogalusa, Louisiana. This panel design of repeated cross-sectional examinations, conducted approximately every 3–4 years, resulted in serial observations from childhood to adulthood. There were 3,173 children who were initially examined in the BHS and were followed from childhood to adulthood. Approximately two-thirds of the adults did not have four measurements of fasting glucose in childhood. To maximize the sample size and to increase the statistical power, we included 1,718 adults (1,043 whites and 675 blacks, 46.0% men, mean age 37.9 years at follow-up, with an age range of 20.1–56.7 years) who had exam data, including BMI, systolic/diastolic BP, total cholesterol (TC), triglycerides (TG), LDL cholesterol (LDL-C) and HDL cholesterol (HDL-C), at least four times during childhood (on average five times) and for fasting plasma glucose (FPG) at least once in adulthood in the present analysis. We conducted sensitivity analyses among 2,401 individuals who had at least three measures of these risk factors in childhood. The number of individuals who had been examined four, five, six, seven, and eight times during childhood was 793, 574, 221, 108, and 22, respectively. Participants were excluded if they had a diagnosis of diabetes or FPG measurement ≥7.0 mmol/L (126 mg/dL) at baseline. The initial date for follow-up was set at the last measurement of CVRFs in childhood prior to 20 years of age. The mean follow-up period was 20.5 years.

Written informed consent was obtained from parents or guardians in childhood and from the participants themselves in adulthood. Study protocols were approved by the Institutional Review Board of the Tulane University Health Sciences Center.

All participants in each survey were asked to complete a structured questionnaire that collected information on demographics, household income, parents’ educational attainment, parental history of diabetes, medical history and use of medications, and smoking and drinking habits.

Standardized protocols were followed by trained personnel across all surveys. Participants were instructed to fast for 12 h before the screening. For each participant, replicate measurements of height and weight were obtained, and the mean values were used for analysis. BMI was calculated as weight in kilograms divided by the square of height in meters. BP was measured on the right arm in a relaxed sitting position by two trained technicians (triplicate each) using mercury manometers. The fourth Korotkoff phase was used for diastolic BP in children and adults to avoid bias resulting from using different phases for diastolic BP. The six readings were averaged.

Between 1973 and 1986, TC and TG were determined with Technicon AutoAnalyzer II (Technicon Instrument Corp., Tarrytown, NY) according to the laboratory manual of the Lipid Research Clinics Program. From 1987, these variables were measured using an Abbott VP instrument (Abbott Laboratories, Abbott Park, IL) by enzymatic procedures. Both chemical and enzymatic procedures met the performance requirements of the Lipid Standardization Program of the Centers for Disease Control and Prevention. Measurements on Centers for Disease Control and Prevention–assigned quality control samples showed no consistent bias over time within or between surveys. Serum lipoprotein cholesterols were analyzed by using a combination of heparin-calcium precipitation and agar-agarose gel electrophoresis procedures. Between 1978 and 1991, FPG was determined with a glucose oxidase method using a Beckman glucose analyzer (Beckman Instruments, Fullerton, CA). Since 1992, FPG has been measured enzymatically as part of a multichemistry profile. HbA_1c was measured with high-performance liquid chromatography using a National Glycohemoglobin Standardization Program–certified automated analyzer.

All statistical analyses were performed with SAS, version 9.2 (SAS Institute, Cary, NC). The primary outcome, diabetes, was defined based on FPG ≥7.0 mmol/L (126 mg/dL) or use of insulin or oral antidiabetes medications over the follow-up (1). Diabetes was also defined based on FPG ≥7 mmol/L or HbA_1c ≥6.5% (48 mmol/mol) or use of insulin or oral antidiabetes medications.

The representativeness of the study population participating in the study was examined by comparing the baseline differences between participants and nonparticipants (Supplementary Table 1).

Due to their skewed distribution, TGs were natural log transformed before conducting of growth curve analysis and regression analysis. The primary exposures of interest were intraindividual variabilities in CVRFs. For each CVRF, we calculated four indices of intraindividual variability for each participant: SD, coefficient of variation (CV), deviation from age-predicted values (DEV), and residual SD (RSD). For each CVRF, CV for each participant was calculated as SD divided by mean. DEV, which was similar to the random variability of BP in previous studies (19–21), was calculated as ⁠, where xi is the observed value at time i, x_i is the predicted value at time i according to the growth curve derived from the mixed model for each individual, and n is the number of data points of the relevant CVRF during childhood. RSD (22,23) was calculated as ⁠, where xi is the observed value at time i, x_i is the growth curve–predicted value at time i, and n is the number of data points for respective CVRF during childhood. The growth curve represents a nonlinear trend of increase in each CVRF with age for each individual. Growth curves of CVRFs measured repeatedly during childhood were constructed for each race and sex group using a random-effects model in SAS Proc MIXED. Three different models were fitted that included 1) age only, 2) age and its quadratic term, and 3) age and its quadratic and cubic terms. Age and its higher-order terms were included one by one for model building. Model selection was based on the Akaike information criterion. The most parsimonious growth curve model was selected. This model allows for repeated measurements and different numbers of unequally spaced observations across individuals. A quadratic curve was fitted for all CVRFs.

Logistic regression analyses, in which variability in each CVRF was examined individually as the main exposure, were performed to assess the association with adult diabetes. To examine whether sex or race modifies these associations, we tested the possible effect of modification caused by sex or race. The associations between measures of each CVRF variability in childhood and adult diabetes were of similar magnitude in both sexes and in whites and blacks. Therefore, models were sex and race pooled to increase precision. Four models were applied (separately for each CVRF): model 1 was adjusted for childhood mean age, sex, race, and adulthood age, and model 2 was further adjusted for parental history of diabetes, drinking and smoking status, and use of antihypertensive and lipid-lowering agents. In order to examine the independent effect of CVRF variability in childhood on adult diabetes, the relevant mean levels of multiple measurements of CVRF in childhood should be considered as a covariate. Hence, model 3 was further adjusted for the respective mean CVRF values during childhood. Model 4 was adjusted further for adulthood BMI and BMI variability in childhood. Model 5 was adjusted additionally for parents’ educational attainment and household income. We chose these covariates because of their significant correlations with CVRF variability or as relevant factors associated with diabetes. To facilitate comparisons of effects sizes of different CVRF variability, we standardized all CVRF variability to z scores (mean = 0; SD = 1) by race-sex groups prior to the regression analysis. BMI variability in childhood and the above covariates except for categorical variables were also standardized into Z scores. BMI variability was adjusted for mean childhood BMI by regression residual analyses and then standardized with Z transformation. Significance was accepted at a two-tailed P < 0.05. Correction for multiple comparisons was not performed in this observational study, which is inherently exploratory as opposed to confirmatory or an intervention as in a randomized controlled trial.

The baseline characteristics of the study population are shown in Table 1 by sex and race. Participants’ levels of studied CVRFs were similar to those of nonparticipants, indicating that participants included in the present analysis are likely to be representative of the initial cohort (Supplementary Table 1).

BMI variability was positively associated with other CVRF variabilities (Table 2). For instance, Pearson correlations with SD-BMI were 0.16 for SD–systolic BP, 0.06 for SD–diastolic BP, 0.11 for SD-TC, 0.11 for SD-TG, 0.21 for SD–LDL-C, and 0.16 for SD–HDL-C.

During the 20.5-year follow-up period, 133 participants developed diabetes. Associations between CVRF variability in childhood and diabetes developed in adulthood are shown in Table 3. Increased BMI variability and HDL-C variability in childhood associated with diabetes risk in all models (Table 3). The adjusted odds ratio (OR) for the development of diabetes in adulthood with every 1 standardized unit increase in BMI variability during childhood using SD, CV, DEV, and RSD was 1.55 (95% CI 1.32–1.81), 1.50 (1.26–1.80), 1.42 (1.24–1.63), and 1.43 (1.24–1.63), respectively, after adjustment for childhood mean age, sex, race, and adulthood age (model 1). Further adjustment for parental history of diabetes, drinking and smoking status, and use of antihypertensive and lipid-lowering agents attenuated the associations, but BMI variability remained strongly associated with diabetes (model 2). These associations still remained even after further adjustment for the mean levels of BMI during childhood (model 3). In model 3, a 1–standardized unit increment in HDL-C variability in childhood using SD, DEV, and RSD was associated with 1.39-, 1.17-, and 1.23-fold risk for diabetes, respectively. Additional adjustment for adulthood BMI, BMI variability in childhood (model 4), parents’ educational attainment, and household income (model 5) attenuated the associations, but HDL-C variability remained significantly associated with diabetes risk. No clear associations were observed for variability in BP, TC, TG, or LDL after adjustments for other covariates.

We then examined the independent effects of these CVRF variabilities on diabetes in adulthood. As shown in Supplementary Table 2, both BMI variability and HDL-C variability, irrespective of the indices used, were independently associated with an increased diabetes risk.

Since different BMI-related pathological processes might be involved in the development of diabetes, we examined the combined effects of BMI levels and BMI variability on diabetes risk (Fig. 1A). Participants were categorized into four mutually exclusive groups based on mean levels of BMI during childhood (low versus high, which was defined as below and above the fourth quartile, respectively) and levels of BMI variability (low versus high, which was defined as below and above the fourth quartile, respectively). Compared with the reference group with both low mean levels of BMI and low BMI variability using SD as the measure, after adjustment for childhood mean age, sex, race, adulthood age, parental history of diabetes, drinking and smoking status, parents’ educational attainment, household income, and use of antihypertensive and lipid-lowering agents, the group with both high BMI value and high BMI variability are OR 2.91 (95% CI 1.78–4.75) times more likely to suffer from diabetes. No significant increased diabetes risk was noted in the group with high mean BMI but low BMI variability OR 1.54 (95% CI 0.86–2.77). Similar results were noted when BMI variability was assessed by the other three indices.

Participants were also categorized into four groups based on mean levels of HDL-C (low versus high, which was defined as below and above the first quartile, respectively) and levels of HDL-C variability during childhood (low versus high, which was defined as below and above the fourth quartile, respectively). We used the group with both high mean levels of HDL-C and low HDL-C variability using SD as the measure as the reference group. No significant increased diabetes risk was noted in the group with high mean HDL-C but high HDL-C variability OR 1.72 (95% CI 0.75–3.96) or in the group with low levels of HDL-C but low HDL-C variability 1.46 (0.80–2.69) after adjustment for childhood mean age, sex, race, adulthood age, parental history of diabetes, drinking and smoking status, parents’ educational attainment, household income, use of antihypertensive and lipid-lowering agents, BMI variability in childhood, and adulthood BMI. The group with both low HDL-C and high HDL-C variability in childhood was OR 2.11 (95% CI 1.06–4.20) times more likely to suffer from diabetes in adulthood. When other indices of HDL-C variability were used for analysis, results were remarkably similar (Fig. 1B).

Analyses were additionally performed using cut points of 70th, 80th, and 85th percentiles for determining the high or low levels. Qualitatively similar results were noted (data not shown).

Last, we explored the combined effects of BMI variability and HDL-C variability on diabetes risk because BMI and HDL-C variabilities were independent of each other in predicting diabetes risk. We divided participants into four groups based on BMI variability and HDL-C variability during childhood (low and high, which was defined as below and above the fourth quartile, respectively). Diabetes risk was evaluated for the four groups using the group with low variabilities in both BMI and HDL-C as a reference. Compared with the reference group, after adjustment for childhood mean age, sex, race, adulthood age, parental history of diabetes, drinking and smoking status, parents’ educational attainment, household income, and use of antihypertensive and lipid-lowering agents, OR of diabetes was 1.96 (95% CI 1.14–3.38) for the group with only high BMI variability using SD, 1.85 (1.11–3.15) for the group with only high HDL-C variability using SD, and 2.20 (1.18–4.09) for the group with high variability in both BMI and HDL-C (Fig. 1C). Pairwise comparisons between the three nonreference groups showed that only high BMI variability had an effect on diabetes risk similar to that of only high HDL-C variability; the simultaneous presence of high BMI variability and high HDL-C variability bore the greatest risk of diabetes. Interaction between BMI variability and HDL-C variability on diabetes risk was tested, and no significant interactions were noted. For other measures of variability in BMI or HDL-C, similar patterns were noted (Fig. 1C).

Similar results were observed in the sensitivity analysis (Supplementary Table 3).

Results did not differ when diabetes was defined as FPG ≥7.0 mmol/L or HbA_1c ≥6.5% (48 mmol/mol) or use of insulin or oral antidiabetes medications (data not shown).

In this community-based biracial cohort of participants with repeated measurements of CVRFs from childhood to midlife, higher BMI variability and HDL-C variability during childhood were significantly associated with increased diabetes risk in adulthood independent of their respective mean levels in childhood and other possible confounding factors, indicating that BMI variability and HDL-C variability during childhood may contain uniquely valuable predictive information for diabetes. This is, to the best of our knowledge, the first report to describe the association of variability in BMI and HDL-C during childhood with the risk of incident diabetes in adulthood. Our study also demonstrated that the diabetes risk increased significantly in individuals with both high HDL-C variability and low mean HDL-C values in childhood and in individuals with high BMI variability and high mean BMI values in childhood. Furthermore, we observed that diabetes risk was comparatively higher among individuals with both high BMI variability and HDL-C variability than that among individuals with either only high BMI variability or only high HDL-C variability. Taken together, these findings offer new insights into the childhood origin of adult-onset diabetes and highlight the importance of variability in BMI and HDL-C for assessing later-life diabetes risk.

Overwhelming evidence has shown the prognostic value of childhood CVRFs, including BMI and HDL-C, for later-life diabetes (24–26). A snapshot or a few measurements of CVRFs may not fully characterize an individual’s phenotype of CVRFs throughout childhood that are linked to diabetes later in life. Variability in CVRFs has been increasingly recognized as a novel metric that provides information unique from that provided by the mean CVRF values, improves risk stratification, and has clinical relevance apart from random variation or measurement error (9–17). Associations between CVRF variability and adverse outcomes have been well described but only in adults or high-risk populations (9–17). However, CVRF variability in these populations can be confounded by comorbidities and medication use. The design of our study minimized these potential confounding issues. Given the accumulated information on the childhood origin of insulin resistance that culminates in adult diabetes, it is important to know whether CVRF variability during childhood also affects future risk of adult diabetes. The current study provides the first prospective evidence that increased BMI or HDL-C variability was consistently linked to diabetes later in life. Hence, the childhood period may be a crucial age window for the development of diabetes in later life.

The pathophysiological mechanisms that underlie the observed intraindividual childhood BMI variability–diabetes association or HDL-C variability–diabetes association remain unclear. Increased variabilities in BMI and HDL-C may be markers for unhealthy lifestyle, such as dietary excess nutrition and physical inactivity, or they may represent an essential trait that reflects responses or adaptations to alterations in neuroendocrine signals. Childhood growth and development undergo consecutive, programmed periods with neurohormonal regulation playing a major role. Experimental studies indicated that fluctuating levels of BMI were associated with a greater increase in adipose tissue inflammation and insulin resistance than a persistent obese state (27). A human study showed that fluctuations in BMI induced by a complex interplay among various hormones, such as hunger hormones (ghrelin) and satiety hormones (leptin, peptide YY, and amylin), caused an unfavorable body composition (more fat mass and less lean mass) (28–31). Hence, pathways associating BMI variability and later-life diabetes may be mediated by increased inflammation and insulin resistance, which are core pathological feature of diabetes, and alteration of neuroendocrine signals. High childhood HDL-C variability may be a reflection of fluctuations in the composition of lipid deposits, which can possibly induce instability of the vascular wall, inflammation, and oxidative stress and thereby increase the diabetes risk. Further reports on how childhood growth and development programming influences variabilities in BMI and HDL-C are warranted.

We also found that children with only high BMI variability or only high HDL-C variability had significant increase in adulthood diabetes risk compared with children with low variability for both measures. BMI variability was as strong as HDL-C variability in predicting diabetes. High BMI variability combined with high HDL-C variability in childhood presented the greatest diabetes risk. No significant interaction between BMI variability and HDL-C variability indicates that the combined effect of the simultaneous presence of high BMI variability and HDL-C variability on the increased risk of diabetes equals the sum of their individual effects. A positive modest correlation existed between BMI variability and HDL-C variability. Hence, high BMI variability tends to coexist with high HDL-C variability, and combining information from both parameters may help to improve prediction of which children are at risk for developing diabetes in adulthood.

Our study has several important strengths. First, this study using data from BHS is one of the first to investigate the effect of CVRF variability in childhood on later-life diabetes, which requires a life span approach. Conducting life span studies is challenging because of the requirement of sufficient follow-up period. The BHS is unique in having repeatedly measured, prospectively collected CVRFs starting in childhood among participants extending over 40 years through midlife. Another strength is the inclusion of information on diverse potential confounding factors, as well as a vigorous quality assurance program over the entire study period. Furthermore, we assessed the role of CVRF variability in the development of adulthood diabetes, using four different definitions for each CVRF variability measure, and identified consistent associations.

Limitations of the current study require careful consideration. First, a standard 75-g oral glucose tolerance test was not performed, which might result in an underestimation of the prevalence of diabetes. Second, the most important CVRF variability with respect to diabetes, FPG variability, was not analyzed. Third, 46% of potentially eligible individuals were excluded due to missing information on childhood CVRFs or covariates, which may have slightly reduced the power to detect significant relationships. However, participants included in the present analysis were comparable with those who were excluded, and our sensitivity analysis among a larger group did not substantially change the results. Fourth, although we tried to adjust for diverse confounding variables, the possibility of measurement error and residual confounding resulting from unmeasured factors is always possible in observational studies. Finally, the sample is community based; therefore the generalizability of the results to other groups not studied may be limited.

In summary, the current study identified associations of high BMI variability and HDL-C variability during childhood, independent of their mean levels, with increased risk of later-life diabetes. Moreover, the magnitude of the association with diabetes risk was similar for high childhood BMI variability and high childhood HDL-C variability. These findings raise the possibility that variabilities in BMI and HDL-C would be useful for identifying future risk of diabetes. Further research is required to confirm these findings, identify potential physiological mechanisms by which childhood variabilities in BMI and HDL-C influence the development of diabetes, and examine the possibility of variabilities in BMI and HDL-C as targets for intervention to reduce adult diabetes risk burden.

Acknowledgments. The BHS is a joint effort of many investigators and staff members, whose contribution is gratefully acknowledged. The authors especially thank the Bogalusa, LA, school system, and most importantly the children and adults who have participated in this study over many years.

Funding. This work was supported by National Institutes of Health grants R01AG041200, R01HD069587, R01ES021724, and R01AG016592. L.A.B. was supported in part by National Institutes of Health grants R01AG062309, P20GM109036, and K12HD043451. T.D. was supported by Natural Science Foundation of China grant 81700762.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. T.D. and L.A.B. conceived the study design, wrote the first draft of the manuscript, researched data, contributed to interpretation of results, commented on drafts, and approved the final version. C.F. and R.B. researched data, contributed to interpretation of results, and approved the final version. V.F. and W.C. contributed to interpretation of results, commented on drafts, and approved the final version. L.A.B. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.