Temporal Associations Among Body Mass Index, Fasting Insulin, and Systemic Inflammation

Key Points Question What are the temporal associations among higher body mass index (BMI) and chronic inflammation and/or hyperinsulinemia? Findings In this systematic review and meta-analysis of 5603 participants in 112 cohorts from 60 studies, the association between period 1 (preceding) levels of fasting insulin and period 2 (subsequent) BMI was positive and significant: for every unit of SD change in period 1 insulin level, there was an ensuing associated change in 0.26 units of SD in period 2 BMI. Meaning These findings suggest that adverse consequences currently attributed to obesity could be attributed to hyperinsulinemia (or another proximate factor).


Introduction
Obesity is associated with a number of noncommunicable chronic diseases (NCDs), such as type 2 diabetes, coronary disease, chronic kidney disease, and asthma. Although obesity is also purported to cause premature death, this association fails to meet several of the Bradford Hill criteria for causation. 1,2 First, the putative attributable risk of death is small (<5%). 3 Second, the dose-response gradient between body mass index (BMI) and mortality is U-shaped with overweight (and possibly obesity level I) as the minima. 3 Third, evidence from animal models comes largely from mice that have been fed high-fat diets; unlike humans, these animals did not normally have fat as part of their typical diet, and thus the experiments are potentially not analogous to those in humans. Fourth, evidence that people with obesity live longer than their lean counterparts in populations with acute or chronic conditions and older age is remarkably consistent. [4][5][6][7][8][9][10][11][12][13][14][15][16] Therefore, it is possible that rather than being a risk factor for NCDs, obesity is actually a protective response to the development of disease.
The putative links between obesity and adverse outcomes are often attributed to 2 potential mediators: chronic inflammation and hyperinsulinemia. These characteristics have been associated with several NCDs, including obesity as well as type 2 diabetes, cardiovascular disease, 17 and chronic kidney disease. 18 Existing data on the association of obesity with chronic inflammation and/or hyperinsulinemia are chiefly cross-sectional, making it difficult to confirm the direction of any causality. This systematic review and meta-analysis summarizes evidence on the temporality of the association between higher BMI and chronic inflammation and/or hyperinsulinemia. We hypothesized that changes in chronic inflammation and hyperinsulinemia would precede changes in higher BMI.

This systematic review and meta-analysis was conducted and reported according to Preferred
Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) 19 and Meta-analysis of Observational Studies in Epidemiology (MOOSE) 20 reporting guidelines. Research ethics board approval was not required because this is a systematic review of previously published research.

Data Sources and Searches
We performed a comprehensive search designed by a trained librarian (E.T.C.) to identify all longitudinal studies and randomized clinical trials (RCTs) that measured fasting insulin and/or an inflammation marker and weight with at least 3 commensurate time points. We included only primary studies published in the English language as full peer-reviewed articles. MEDLINE (1946 to August 20, 2019) and Embase (1974 to August 19, 2019) were searched; however, only studies published in 2018 were retained because of the high volume of results. No existing systematic reviews were found. The specific search strategies are provided in eTable 1 in the Supplement. The abstracts were independently screened by 2 reviewers (including N.W.). The full text of any study considered potentially relevant by 1 or both reviewers was retrieved for further consideration. The data analysis was conducted between January 2020 and October 2020.

Study Selection
Each potentially relevant study was independently assessed by 2 reviewers (N.W. and F.Y.) for inclusion in the review using the following predetermined eligibility criteria. Longitudinal studies and RCTs with men and nonpregnant and not recently pregnant women (Ն18 years of age) and at least 3 time points with 1 or more weeks of follow-up in which fasting insulin levels or a marker of inflammation and some measure of weight were included in this review. Disagreements were resolved by consultation.

Data Extraction and Risk of Bias Assessment
Data from eligible studies were extracted by a single reviewer (N.W.). A second reviewer checked the extracted data for accuracy. The following properties of each study were recorded in a database: study characteristics (country, era of accrual, design, duration of follow-up, populations of interest, intervention where applicable, and sample size), age and sex of participants, and the measures of interest (numbers, means, and SDs) for all time points: (1) fasting insulin, the homeostatic model assessment index, or the quantitative insulin sensitivity check index; (2) concentrations of C-reactive protein (CRP), interleukin cytokines, or tumor necrosis factor; and (3) weight, BMI (calculated as weight in kilograms divided by height in meters squared), fat mass, or fat mass percentage.
Risk of bias was assessed using items from Downs and Black 21 : clear objective, adequate description of measures, sample size or power calculation, intention to treat study design (in those studies that assigned the intervention), adequate description of withdrawals, adequate handling of missing data, and adequate description of results. Source of funding was also extracted, given its potential to introduce bias. 22

Statistical Analysis
Data were analyzed using Stata software, version 15.1 (StataCorp LLC). Missing SDs were imputed using interquartile ranges or using another SD from the same cohort. 23 Data were extracted from graphs if required.
To determine a likely temporal sequencing of fasting insulin level or chronic inflammation with obesity, we compared the associations of period 2 insulin level or inflammation regressed on period 1 BMI and period 2 BMI regressed on period 1 insulin or inflammation. A stronger association would support a particular direction of effect.
For each measure of interest, the change in means was calculated between adjacent time points and divided by the number of weeks between the measures. This slope or per week change in measure was then standardized by dividing it by the pooled SD, giving a standardized slope. Because of expected diversity among studies, we decided a priori to combine the standardized slopes using a random-effects models. Period 2 standardized slopes of weight measures were regressed onto period 1 standardized slopes of insulin or inflammation measures and vice versa. We regressed measures of insulin post hoc on measures of inflammation and vice versa.
The type I error rate for meta-regressions was set at a 2-sided P < .05. Statistical heterogeneity was quantified using the τ 2 statistic (between-study variance) 24 and the I 2 statistic. Differences in standardized slopes (βs) along with 95% CIs are reported.
We considered a number of sensitivity analyses. Because we included multiple standardized slopes at different intervals from the same studies (or same cohorts), we accounted for this nonindependence using a generalized linear model in which the family was gaussian and the link was identity, which allowed for nested random effects (results by intervals were nested within cohorts).
To estimate between-study heterogeneity, the coefficients for the within-cohort SEs were constrained to 1. We also performed 2 subgroup analyses: whether the study population had undergone bariatric surgery and the numbers of weeks between time points (>12 vs Յ12 weeks), reasoning that if the effects of one measure of interest acted quickly on the other, then shorter intervals might demonstrate stronger associations. We explored post hoc models with 2 measures of interest as period 1 independent variables.

Quantity of Research Available
The searches identified 1865 unique records identifying articles or abstracts published in 2018 ( Figure 1). After the initial screening, the full texts of 813 articles were retrieved for detailed evaluation. Of these, 753 articles were excluded, resulting in 60 that met the selection criteria and 5603 enrolled participants (of whom 5261 were analyzed). 25- 84 We decided to exclude 12 studies of children and adolescents post hoc because these studies used different BMI measures.
Disagreements about the inclusion of studies occurred in 2% of the articles (κ = 0.87). The 60 studies included 112 cohorts: 40 cohorts contained participants who had undergone bariatric surgery, 33 cohorts contained participants who were receiving diet therapies (all except 2 65,84 designed for weight loss or weight maintenance), 16 cohorts contained participants who received a medication or supplement, 7 cohorts contained participants who were following exercise regimens, 14 cohorts contained participants who were followed up for other reasons (ie, prostate

Risk of Bias Assessment
Studies were largely rated as low risk for description of the objectives (96.7%), the outcome measures (90.0%), and the results (98.3%) (Figure 2). Approximately half the studies were high risk because they lacked a sample size or power calculation (51.7%), they (in those studies that assigned the interventions) did not take an intention-to-treat approach (47.2%), they had a withdrawal rate greater than 20% or they did not adequately describe their withdrawals (50.0%), or they did not adequately explore the effect of missing data (50.0%). In addition, 38.3% of studies had an industry source of funding.

BMI and Fasting Insulin Level
There were 90 pairs of standardized slopes from 56 cohorts and 35 studies that measured BMI and fasting insulin ( Table 2). Most BMI and fasting insulin standardized slopes were negative (81% for BMI and 71% for fasting insulin), meaning that participants in most studies experienced decreases in BMI and insulin. The association between period 1 fasting insulin level and period 2 BMI was positive and significant (β = 0.26; 95% CI, 0.13-0.38; I 2 = 79%) (Figure 3), indicating that for every unit of SD change in period 1 insulin, there was an associated change in 0.26 units of SD in period 2 BMI. The association between period 1 BMI and period 2 fasting insulin level was not significant (β = 0.01; 95% CI, -0.08 to 0.10; I 2 = 69%) (Figure 3). The heterogeneities were large. The associations between insulin level and BMI increased in magnitude when studies that reported findings at 12 weeks or less were isolated from those that reported findings at greater than 12 weeks (eTable 3 in the Supplement). The magnitude of association between period 1 fasting insulin level and period 2 BMI was greater at 12 weeks or less than at greater than 12 weeks (β = 0.61; 95% CI, 0.38-0.84 vs β = 0.17; 95% CI, 0.05-0.30; I 2 = 76%, P = .001). The association between period 1 fasting insulin level and period 2 BMI was present in participants who had undergone bariatric surgery but not in participants who had not undergone bariatric surgery (β = 0.31; 95% CI, 0.19-0.44 vs β = -0.12; 95% CI, -0.41 to 0.18; I 2 = 76%, P = .007) (eTable 4 in the Supplement).

BMI and CRP
There were 57 pairs of standardized slopes from 39 cohorts and 22 studies that measured both BMI and CRP levels ( However, both β coefficients were positive and had similar magnitudes, and the β coefficient for BMI had larger heterogeneity. The associations between BMI and CRP level increased in magnitude when the studies that reported findings at 12 weeks or less were isolated from those that reported findings at greater than   12 weeks, when period 2 BMI was regressed on period 1 CRP level (eTable 3 in the Supplement).

Fasting Insulin and CRP
There were 42 pairs of standardized slopes from 27 cohorts and 16 studies that measured both fasting insulin and CRP levels ( Table 2)

Other Sensitivity Analyses
When we considered related measures of BMI (weight, fat mass, and fat percentage), homeostatic model assessment index, and the other inflammatory markers (ie, interleukin 6 and tumor necrosis factor α), the associations among these variables were similar to those for BMI or could not be

Discussion
This systematic review and meta-analysis suggests that decreases in fasting insulin are more likely to precede decreasing weight than are decreases in weight to precede decreasing levels in fasting insulin. After accounting for the association between preceding levels of fasting insulin and the subsequent likelihood of weight gain, there was no evidence that inflammation preceded subsequent weight gain (eTable 7 in the Supplement). This temporal sequencing (in which changes in fasting insulin precede changes in weight) is not consistent with the assertion that obesity causes NCDs and premature death by increasing levels of fasting insulin.

Support From Other Studies
In patients with type 2 diabetes, RCTs have found that introducing exogenous insulin and sulfonylureas (which increase endogenous insulin production) compared with lower doses or no drug therapy produce increases in weight. 85,86 Some patients with type 1 diabetes deliberately omit or reduce their insulin injections to lose weight. 87 Similarly, reports after bariatric surgery consistently indicate that insulin levels decrease before weight decreases in patients undergoing bariatric surgery. 88 Thus, the finding that changes in insulin levels tend to precede changes in weight rather than the other way around has been previously demonstrated in 3 different scenarios. To our knowledge, there is no clinical evidence demonstrating that weight gain or loss precedes increases or decreases in endogenous insulin.

Importance of the Findings
Obesity as a cause of premature death fails to meet several of the Bradford Hill criteria for causation: the strength of association is small 3 ; the consistency of effect across older and/or ill populations favors obesity [4][5][6][7][8][9][10][11][12][13][14][15][16] ; and the biological gradient is U-shaped, with overweight and obesity level 1 associated with the lowest risk 3 ; and if hyperinsulinemia is to be considered the mediator, then the temporal sequencing is incorrect.
Insulin resistance, a cause and consequence of hyperinsulinemia, 89 leads to type 2 diabetes and is associated with other adverse outcomes, such as myocardial infarction, chronic pulmonary disease, and some cancers, 90,91 and may also be indicated in diabetic nephropathy. 92 Despite the 3 scenarios described earlier, it is commonly believed that obesity leads to hyperinsulinemia. [93][94][95] If the converse is true and hyperinsulinemia actually leads to obesity and its putative adverse consequences, then weight loss without concomitant decreases in insulin (eg, liposuction) would not be expected to address these adverse consequences. In addition, weight loss would not address risk in people with so-called metabolically healthy obesity, that is, those without insulin resistance. 96 Of interest, insulin resistance is also present in lean individuals, in particular men and individuals of Asian descent. 97 These 2 groups are at heightened risk for type 2 diabetes 98 and cardiovascular disease, yet are more likely to be lean than women and individuals not of Asian descent. These observations are consistent with the hypothesis that hyperinsulinemia rather than obesity is driving adverse outcomes in this population. We speculate that the capacity to store the byproducts of excess glucose by increasing the size of fat cells (manifested as obesity) might delay the onset of type 2 diabetes and its consequences in some individuals, thus explaining the so-called obesity paradox of lower mortality among people with obesity. This idea, although not new, 99 fits better with the emerging evidence. If this speculation is correct, assessing the capacity to store such by-products at the individual level may be a useful step toward personalized medicine.
Although it is possible that hyperinsulinemia per se is not the causal agent that leads to adverse outcomes (but is rather a marker for another more proximate factor), this would not change the lack of support for recommending weight loss among people with obesity. Rather, other markers should be investigated that, although correlated with obesity, are more strongly associated with premature mortality because they also exist in lean individuals. Therapies that lower insulin levels (eg, moderate diets with fewer simple carbohydrates and metformin) may be sustainable if an intermediate marker

Conclusions
The pooled evidence from this meta-analysis suggests that decreases in fasting insulin levels precede weight loss; it does not suggest that weight loss precedes decreases in fasting insulin. This temporal sequencing is not consistent with the assertion that obesity causes NCDs and premature death by increasing levels of fasting insulin. This finding, together with the obesity paradox, suggests that hyperinsulinemia or another proximate factor may cause the adverse consequences currently attributed to obesity. Additional studies to confirm this hypothesis are urgently needed.