Visit-to-Visit Glucose Variability, Cognition, and Global Cognitive Decline: The Multi-Ethnic Study of Atherosclerosis

Abstract

Context

Higher visit-to-visit glucose variability (GV) is associated with dysglycemia and type 2 diabetes (T2D), key risk factors for cognitive decline.

Objective

Evaluate the association of GV with cognitive performance and decline in racially/ethnically diverse older populations with and without T2D.

Methods

We calculated the standard deviation of glucose (SDG), average real variability (ARV), and variability independent of the mean (VIM) among 4367 Multi-Ethnic Study of Atherosclerosis participants over 6 clinical examinations. Participants completed a cognitive assessment at the fifth examination, and a subset completed a second assessment 6 years later. We used multivariable linear regression to estimate the association of intraindividual GV with cognitive test scores after adjustments for cardiovascular risk factors and mean glucose level over the study period.

Results

Two-fold increments in the VIM and SDG were associated with worse Cognitive Abilities Screening Instrument (CASI) performance, while two-fold increments in VIM and ARV were associated with worse Digit Symbol Coding test score. GV measures were not associated with change in CASI performance among 1834 participants with repeat CASI data 6 years later. However, among 229 participants with incident T2D, the SDG and VIM were associated with decline in CASI (−1.7 [95% CI: −3.1, −0.3] and −2.1 [−3.7, −0.5] points, respectively). In contrast, single-timepoint glucose and HbA1c were not associated with CASI decline among participants with or without incident T2D.

Conclusion

Higher visit-to-visit GV over 16 to 18 years is associated with worse cognitive performance in the general population, and with modest global cognitive decline in participants with T2D.

Dysglycemia is a key, potentially modifiable risk factor for the development of age-related neurocognitive dysfunction in older adults. For example, type 2 diabetes (T2D) is associated with up to 50% greater risk for all-cause dementia and 150% greater risk for vascular dementia (1, 2). Prediabetes is also associated with increased risk for cognitive decline compared to normoglycemia (3, 4), suggesting that this risk occurs along a continuum of glucose control. However, which aspects of dysglycemia contribute to risk for neurocognitive dysfunction are unclear. Single-timepoint measures of fasting glucose and glycated hemoglobin (HbA1c) are inconsistently associated with cognitive impairment and incident dementia, even among adults with T2D (5-8). Furthermore, intensive glycemic control in adults with longstanding T2D has failed to prevent cognitive decline in randomized trials (9, 10), and even moderate hypoglycemia or frequency of hypoglycemic episodes in T2D are associated with incident dementia (11). These lines of evidence suggest that traditional markers of glycemic control are insufficient to describe the full impact of dysglycemia on age-related cognitive decline.

In addition to elevated glucose, T2D is characterized by impaired glycemic homeostatic mechanisms. These impairments can manifest as increased variability in fasting glucose levels that is measurable over months to years (“visit-to-visit”), including in the prediabetes latency period and after T2D diagnosis (12). Studies of visit-to-visit glucose variability (GV) have found associations with cardiovascular disease (CVD) mortality in adults with T2D and in the general population (13, 14), in addition to increased risk of T2D itself (15). Accordingly, higher GV in adults with and without T2D has been associated with incident all-cause dementia and Alzheimer's disease, and worse cognitive performance in cohort studies (16-18). However, fewer studies have examined the prospective association of midlife GV and later cognitive performance and decline in diverse groups. Our objective was to study the association of visit-to-visit GV with global cognitive performance, processing speed, working memory, and global cognitive decline in a general, multi-ethnic population of middle-aged and older US adults. We hypothesized that higher GV over 10 to 16 years in mid- to late-life would be associated with worse cognitive performance and greater global cognitive decline over 6 years.

Methods

Study Population

Participants were drawn from the Multi-Ethnic Study of Atherosclerosis (MESA), an ongoing, prospective cohort study (19). Briefly, 6814 men and women aged 45 to 84 years without known clinical CVD were recruited from 6 communities in the United States (Baltimore, MD; Chicago, IL; Forsyth County, NC; Los Angeles, CA; New York, NY; and St. Paul, MN). Participants self-identified with one of 3 races (Black, Chinese, White) or Hispanic ethnicity. Following a baseline examination (Exam 1; 2000-2002), 5 in-person clinical follow-up exams occurred between 2002 and 2018 that were supplemented by annual phone surveys to update health information. Participants provided written informed consent prior to enrollment, and clinical procedures were approved by the Institutional Review Board of each field center (Johns Hopkins University, Northwestern University, Wake Forest University, University of California at Los Angeles, Columbia University, and University of Minnesota).

Cognitive Test Outcomes

Cognitive performance was evaluated in MESA participants at the fifth examination (Exam 5) in 2010-2012 using 3 standardized tests: the Cognitive Abilities Screening Instrument (CASI, version 2), Digit Symbol Coding test, and Digit Span test (20, 21). Lower scores indicate worse performance on each test. The CASI (scored 0-100) is a test of global cognitive performance, comprising 25 items representing 9 cognitive domains (attention, abstraction/judgment, concentration, language, orientation, verbal fluency, visual construction, and short- and long-term memory). The Digit Symbol Coding test (scored 0-133) assessed processing speed, for which participants were presented a key of 9 digit-symbol pairs followed by a list of randomly ordered digits. Participants were tasked to write as many correct corresponding symbols beneath the digits as possible within 120 seconds. Working memory was assessed by the Digit Span test, for which participants were instructed to repeat increasing spans of random numbers delivered at 1 per second for a maximum of 16 trials in forward and 14 trials in backward order. One point was awarded for each correctly recalled span, and forward and backward scores were summed to provide a total score (range, 0-30).

The CASI was completed by 4591 MESA participants at Exam 5 (97%). A total of 199 (4%) CASI tests were excluded: 125 tests deemed invalid by the test administrator or with technically invalid scores <20, and 74 participants with documented history of dementia ascertained by review of International Classification of Diseases codes with external validation (22). Participants with dementia identified prior to or during the study period were excluded due to the absence of any cognitive evaluation in MESA prior to Exam 5. Participants with other potential neurological or psychiatric conditions were included if they provided valid cognitive test data. Complete and valid CASI data were obtained from the remaining 4392 participants. The Digit Symbol Coding was further completed by 4000 participants (91%), and the Digit Span was completed by 4377 (>99%) (Supplementary Fig. S1 (23)). Participants who completed the Exam 5 cognitive assessment were younger, more likely to be White, more likely to have completed high school, and had lower systolic blood pressure and fasting glucose at Exam 1 than those who died or were lost to follow-up prior to Exam 5, did not attend Exam 5, or did not complete the cognitive assessment (Supplementary Table S1 (23)).

The CASI was re-administered in a subset of participants (n = 1931) without clinically recognized dementia at Exam 6 in 2016-2018 (58% of the surviving cohort). Exclusion of incomplete or invalid tests yielded 1838 (95%) participants with valid Exam 6 CASI data. Compared to the Exam 5 participants who did not complete the CASI at Exam 6 (n = 2533), participants with repeat CASI data were younger, less likely to be Hispanic, more likely to have completed high school, less likely to be physically inactive, and had lower systolic blood pressure, plasma triglycerides, and fasting glucose compared with other Exam 5 participants (Supplementary Table S2 (23)).

Visit-to-Visit GV Measures

Venous blood was drawn in the morning following a ≥8 hour fast at baseline (Exam 1) and each subsequent MESA examination (Exam 2 to Exam 6; approximately 2, 4, 5, 10-12, and 16-18 years after Exam 1). Participants were instructed not to take oral diabetes medications or insulin in the morning prior to the blood draw. Serum glucose was measured after separation using the glucose oxidase method and Vitros 950 analyzer (Johnson & Johnson, Rochester, New York). In each participant with at least 3 glucose measurements through Exam 5, we calculated 3 indices of intraindividual GV: standard deviation of fasting glucose (SDG), average real variability (ARV), and variability independent of the mean (VIM). Intraindividual SDG (mg/dL) was calculated as the square root of the sum of the squared differences of individual glucose measurements and intraindividual mean glucose across exams, divided by the number of measurements minus 1. The ARV (mg/dL/year) was calculated as the mean of the absolute difference between successive glucose measurements divided by the duration in years between each (see Fig. 1) (18). The VIM is a transformation of the SDG that is uncorrelated with mean glucose (24). Intraindividual VIM (VIM_i) was calculated according to $VI{M}_i=k\frac{SD{G}_i}{mea{n}_i^{\beta 1}}$, where SDG_i refers to an individual's SDG, mean_i refers to an individual's mean fasting glucose over the measurement period, β₁ is the simple linear regression coefficient for the natural logarithm of the SDG fitted to the natural logarithm of the mean glucose (ie, Y_ln(mean) = β₀ + β₁·ln(SDG) + ε_i), and k is a constant chosen to give the overall VIM the same median value as the overall SDG so that values of VIM are on the same scale as SDG (25). Specific values for elements used in VIM calculations are listed in Supplementary Table S3 (23).

Figure 1.

Calculating the standard deviation (SDG) and average real variability (ARV) of fasting glucose. Intraindividual SDG_i (mg/dL) was calculated as the square root of the sum of the squared differences of individual glucose measurements (g_i) and the intraindividual mean glucose across exams (μ_g), divided by the number of measurements (n) minus 1: $SD{G}_i=\sqrt{\frac{\sum {(g_i-{\mu }_g)}^2}{n-1}}$. Annualized intraindividual ARV_i (mg/dL/year) was calculated as the mean of the absolute difference between successive glucose measurements (eg, δ₁ to δ₄) divided by the duration in years between each measurement (eg, τ₁ to τ₄): $AR{V}_i=\frac{1}{n}{\sum }_{i=1}^n\frac{{\delta }_i}{{\tau }_i}$.

Covariates and Other Measures

Clinical measures were obtained by trained study personnel using standardized procedures. Participants self-reported their sex, race/ethnicity, and highest educational attainment at baseline, and their smoking status and hypertension, lipid-lowering, and psychiatric medication use at each examination. Psychiatric medications included antidepressants (monoamine oxidase inhibitors, tricyclics, selective serotonin- and serotonin-norepinephrine reuptake inhibitors), typical and atypical antipsychotics/neuroleptics, benzodiazepines and other anxiolytics, and combinations thereof. Seated brachial blood pressure (mm Hg) was measured with an automated Dinamap Monitor Pro 100; 3 measures were taken and the mean of the second and third measurements was recorded. Body mass index (kg/m²) was calculated from measured height and weight. Fasting plasma triglycerides (mg/dL) were measured using the glycerol-blanked enzymatic method (Roche Diagnostics, Indianapolis, Indiana). Plasma high-density lipoprotein (HDL) cholesterol (mg/dL) was measured by the cholesterol oxidase method (Roche Diagnostics) after precipitation of non-HDL magnesium/dextran. Low-density lipoprotein (LDL) cholesterol (mg/dL) was estimated by the Friedewald equation. Depressive symptoms were assessed with the Center for Epidemiologic Studies Depression scale (CESD; range 0-60, with higher score indicating worse depressive symptoms). Physical activity was assessed using an adaptation of the Cross-Cultural Physical Activity Survey (26), with inactivity defined as a mean of <500 kcal/metabolic equivalents (METs)/week of moderate to vigorous activity over the Exam 1 to Exam 5 period. Isoforms of the Apolipoprotein E (APOE) ε4 allele, a genetic risk factor for dementia (27), were estimated from single nucleotide polymorphisms rs429358 and rs7412. The Cardiovascular Risk Factors, Aging, and Incidence of Dementia (CAIDE) risk score, a composite measure of all-cause dementia risk (28), was calculated according to published equations using Exam 5 clinical measures. Higher CAIDE score (range 0-18) indicates greater 20-year risk of dementia. HbA1c was measured from anticoagulated blood by high-performance liquid chromatography using a G7 Glycohemoglobin Analyzer (Tosoh Medics, San Francisco, CA). Diabetes was defined as fasting glucose ≥126 mg/dL or use of insulin or oral hypoglycemic medication. Incident T2D was identified based on new fasting glucose ≥126 mg/dL or the new use of insulin or oral hypoglycemic medications from baseline through Exam 5.

Statistical Analysis

First, we compared the distributions of our study population characteristics across quartiles of SDG over Exam 1 to Exam 5 using one-way analysis of variance or Kruskal-Wallis tests for continuous variables and Fisher exact test for proportions with a two-sided P < .05 threshold for significance testing. We then used multivariable linear regression with inverse probability weighting for attrition since baseline to evaluate the association of continuous SDG, ARV, and VIM over the Exam 1 to 5 period with performance on Exam 5 cognitive tests. Stabilized inverse probability weights conditioned on Exam 1 sample characteristics were generated for participants who completed Exam 5 cognitive testing relative to those who did not; a comparison of standardized mean differences in weighted and unweighted characteristics indicates that better balance was achieved by inverse probability weighting (Supplementary Table S4 (23)). Distributions of GV indices were normalized by log₂ transformation; therefore, we report main effects as the difference in cognitive test score per two-fold increment in GV with 95% CI. For comparison, we further report associations of glucose and HbA1c measured at Exam 5, as well as mean fasting glucose over the Exam 1 to Exam 5 period, with cognitive performance per SD increment in glucose or HbA1c. We fit 3 regression models: Model 1 was minimally adjusted for baseline age, sex, race, ethnicity, and educational attainment (completion of less than high school, high school, or education beyond high school). Model 2 was additionally adjusted for mean systolic blood pressure, body mass index, log-normalized CESD score, triglycerides, and HDL and LDL cholesterol over the period of follow-up from Exam 1 to Exam 5; physical inactivity; Exam 5 smoking status (current vs never/former); use of hypertension, lipid-lowering, and psychiatric medication at Exam 5; and APOE ε4 allele carrier status (0 vs ≥1 alleles vs missing). Model 3 included all covariates of Model 2 with the addition of mean fasting glucose over the period from Exam 1 to Exam 5. We then assessed moderation by sex, race/ethnicity, Exam 1 diabetes status, and incidence of T2D through Exam 5 among the GV measures by including appropriate interaction terms in Model 3, and stratified analyses by significant interactions.

Next, we used multivariable linear regression with inverse probability weighting for attrition between Exam 5 and Exam 6 to examine the association of GV indices recalculated to include Exam 6 glucose measurements (SDG_1–6, ARV_1–6, and VIM_1–6) with longitudinal performance on the CASI. Stabilized propensity weights conditioned on baseline and mean sample characteristics through Exam 5 were generated for participants who completed Exam 6 cognitive testing relative to those who did not; a comparison of standardized mean differences in weighted and unweighted characteristics indicates that better balance was achieved by inverse probability weighting (Supplementary Table S5 (23)). We report the 6-year point-change in CASI score associated with a 1-SD increment in glucose and HbA1c or two-fold increment in GV from multivariable models adjusted for Exam 5 CASI performance and Model 1, Model 2, and Model 3 covariates. We further tested for interactions by sex, race/ethnicity, Exam 1 diabetes status, and T2D incidence through Exam 6, and stratified analyses by significant interactions. Model 2 and Model 3 covariates were updated to include Exam 6 measurements.

Finally, we performed 3 sensitivity analyses. First, to account for the possibility that number of fasting glucose measurements or variable time between measurements impacted results, we imputed missing glucose measurements over the Exam 1 to Exam 6 period in the 4392 participants with Exam 5 cognitive data using multiple imputation by chained equations, and recalculated GV measures using imputed measurements. Missingness for each glucose measurement from Exam 1 to Exam 6 was ≤3.1%, while 7.3% of participants were missing at least 1 measurement. Results from 10 iterations of 20 imputed datasets generated by predictive mean matching were combined for validity. Next, to account for the potential influence of subsequent diabetes medication use on GV in 383 participants with new-onset T2D after Exam 1, we recalculated GV measures after censoring fasting glucose measurements at or after examinations at which incident T2D was identified so that these measurements did not contribute to GV. Finally, to determine if GV is differentially related to cognition in adults who have not yet developed T2D compared with adults who fulfill criteria for T2D, we stratified our analysis by person-time with or without a T2D diagnosis. We compared associations of the censored GV measures with cognitive performance after further excluding 440 participants with T2D at baseline with associations of GV measures calculated using only glucose measurements after T2D was identified among 823 participants with baseline or incident T2D after Exam 1.

Analyses were performed in R 4.1.2 (R Foundation, 2022). Multiple imputation was performed with the mice package (version 3.15.0) for R.

Results

Sample Characteristics

The final analytic sample included 4367 Exam 5 participants with cognitive testing and ≥3 fasting glucose measurements between Exam 1 and Exam 5 (Supplementary Fig. S1 (22)). Characteristics of these participants are presented by quartile of SDG in Table 1. Higher SDG was associated with higher mean systolic blood pressure, body mass index, plasma triglycerides, fasting glucose, HbA1c, worse depressive symptoms over the Exam 1 to Exam 5 period, and higher CAIDE dementia risk score and proportions of hypertension medication use, lipid-lowering medication use, and smoking at Exam 5. Diabetes prevalence at baseline and T2D incidence through Exam 5 were greater in the highest quartile of SDG. Black race, Hispanic ethnicity, and lower educational attainment were also associated with higher SDG. The SDG and ARV were significantly correlated with mean fasting glucose from Exam 1 to Exam 5 (Spearman ρ for fasting glucose vs SDG = 0.56; vs ARV ρ = 0.50; both P < .001), but the VIM was not (ρ = 0.02; P = .16) despite high correlation with the SDG (ρ = 0.78; P < .001) and ARV (ρ = 0.62; P < .001). Similar associations with quartiles of SDG_1–6 were observed among the 1834 participants with repeat CASI data at Exam 6, although women were less likely to be in the highest quartile of SDG_1–6 (Supplementary Table S6 (23)).

Table 1.

Characteristics by quartile of standard deviation of fasting glucose from Exam 1 (2000-2002) to Exam 5 (2010-2012)^a

	SD of fasting glucose				P valueb
	Quartile 1	Quartile 2	Quartile 3	Quartile 4
Range, mg/dL	0.6, 4.3	4.3, 6.0	6.0, 8.9	8.9, 168.3
n	1092	1092	1092	1091
Age, Exam 1, years	60.2 (9.4)	60.1 (9.5)	59.9 (9.5)	60.0 (9.2)	.58
Women, n (%)	580 (53.1%)	604 (55.3%)	587 (53.8%)	554 (50.8%)	.20
Race and ethnicity, n (%)					<.001
White	497 (45.5%)	493 (45.1%)	498 (45.6%)	321 (29.4%)
Chinese	141 (12.9%)	132 (12.1%)	95 (8.7)	111 (10.2%)
Black	258 (23.6%)	272 (24.9%)	269 (24.6%)	350 (32.1%)
Hispanic	196 (17.9%)	195 (17.9%)	230 (21.1%)	309 (28.3%)
Educational attainment, n (%)	964 (88.3%)	973 (89.1%)	954 (87.4%)	890 (81.6%)	<.001
<High school	131 (12.0%)	113 (10.3%)	136 (12.5%)	200 (18.3%)
Completed high school	176 (16.1%)	203 (18.5%)	176 (16.2%)	208 (19.1%)
Associate degree or higher	786 (71.9%)	780 (71.1%)	775 (71.3%)	683 (62.6%)
APOE genotype, carriage of 1 or 2 ε4 alleles, n (%)	251 (23.0%)	293 (26.8%)	270 (24.7%)	272 (25.0%)	.23
Systolic blood pressure, mean Exam 1 to Exam 5, mm Hg	121.7 (16.6)	122.0 (16.7)	122.0 (16.1)	126.6 (15.9)	<.001
Body mass index, mean Exam 1 to Exam 5, kg/m2	27.2 (4.9)	27.6 (4.9)	28.4 (5.2)	30.6 (5.9)	<.001
HDL cholesterol, mean Exam 1 to Exam 5, mg/dL	54.9 (14.9)	53.8 (14.0)	53.3 (15.1)	49.1 (12.9)	<.001
LDL cholesterol, mean Exam 1 to Exam 5, mg/dL	115.3 (24.9)	114.9 (25.7)	112.6 (25.6)	106.1 (26.4)	<.001
Triglycerides, mean Exam 1 to Exam 5, mg/dL	115.9 (54.6)	117.5 (57.3)	123.9 (63.2)	141.8 (82.5)	<.001
Physically inactive, < 500 kcal/METs/week average, Exam 1 to Exam 5, n (%)	711 (65.1%)	711 (65.1%)	717 (65.7%)	764 (70.0%)	.04
Smoker, Exam 5, n (%)	69 (6.3%)	100 (9.2%)	75 (6.9%)	107 (9.8%)	<.01
Center for Epidemiologic Studies Depression Scale, Exam 1 to Exam 5, median [25%, 75%]	5.3 [2.7, 9.3]	5.7 [3.0, 9.7]	6.0 [3.3, 10.7]	6.7 [3.3, 11.3]	<.001
Psychiatric medication use, Exam 5, n (%)	138 (12.6%)	142 (13.0%)	168 (15.5%)	163 (14.9%)	.144
Hypertension medication use, Exam 5, n (%)	513 (47.0%)	538 (49.3%)	589 (53.9%)	769 (70.5%)	<.001
Lipid medication use, Exam 5, n (%)	356 (32.6%)	381 (34.9%)	402 (36.8%)	566 (51.9%)	<.001
CAIDE dementia risk score, Exam 5, range 0-18	7.2 (1.8)	7.3 (1.8)	7.5 (1.9)	8.1 (2.0)	<.001
Fasting glucose, Exam 5, mg/dL	90.3 (7.3)	93.0 (8.5)	97.3 (10.2)	127.2 (46.1)	<.001
Fasting glucose, mean Exam 1 to Exam 5, mg/dL	89.1 (6.8)	90.6 (7.7)	93.4 (8.9)	121.7 (33.3)	<.001
Number of glucose measurements, Exam 1 to Exam 5	4.9 (0.4)	4.9 (0.3)	4.9 (0.3)	4.9 (0.3)	.12
HbA1c, Exam 5, %	5.6 (0.4)	5.7 (0.4)	5.7 (0.4)	6.8 (1.4)	<.001
Diabetes prevalence, Exam 1, n (%)	3 (0.3%)	7 (0.6%)	25 (2.3%)	405 (37.1%)	<.001
Diabetes incidence, Exam 1 to Exam 5, n (%)	25 (2.3%)	29 (2.6%)	48 (4.4%)	139 (12.7%)	<.001
Time to diabetes incidence, years, median [25%, 75%]	8.4 [4.0, 10.7]	7.6 [5.1, 9.9]	8.9 [5.9, 10.2]	6.2 [3.7, 9.0]	<.01
ARV of fasting glucose, Exam 1 to Exam 5, mg/dL/year, median [25%, 75%]	1.7 [1.2, 2.2]	2.6 [2.1, 3.3]	3.5 [2.7, 4.6]	7.4 [4.9, 13.7]	<.001
VIM of fasting glucose, Exam 1 to Exam 5, median [25%, 75%]	3.7 [2.9, 4.4]	5.5 [4.7, 6.4]	7.3 [6.1, 8.5]	9.6 [7.3, 12.5]	<.001
Cognitive performance, Exam 5
CASI (range 0-100)	88.7 (7.5)	88.4 (8.1)	88.4 (8.3)	86.3 (9.2)	<.001
Digit Symbol Coding (range 0-133)	52.7 (17.6)	52.3 (17.8)	51.7 (18.5)	46.3 (18.6)	<.001
Digit Span, forward and backward combined (range 0-28)	15.7 (4.6)	15.4 (4.4)	15.7 (4.6)	14.5 (4.4)	<.001

Abbreviations: ARV, average real variability of fasting glucose; APOE, apolipoprotein E; CAIDE, Cardiovascular Risk Factors, Aging, and Incidence of Dementia; CASI, Cognitive Abilities Screening Instrument; HDL, high-density lipoprotein; HbA1c, glycated hemoglobin; LDL, low-density lipoprotein; METs, metabolic equivalents; VIM, variability independent of the mean of fasting glucose.

^a Data are mean (SD) unless noted otherwise.

^b P value for global test: ANOVA for means, Kruskal-Wallis test for medians, and Fisher exact test for proportions.

Association of Visit-to-Visit GV With Cognitive Performance in Middle-aged and Older Adults

Table 2 presents associations of GV over 10 years with scores on the Exam 5 cognitive tests. Two-fold increments in each GV measure were associated with worse performance on the CASI, Digit Symbol Coding, and Digit Span tests after adjustment for age and sociodemographic factors. The associations of SDG, ARV, and VIM with the CASI and Digit Symbol Coding tests persisted after additional adjustment for cardiovascular risk factors, hypertension and lipid-lowering medication use, and APOE ε4 status (Model 2), although the association of each GV measure with the Digit Span test was attenuated after these adjustments. Associations among GV measures and Exam 5 fasting glucose and HbA1c with cognitive performance were of similar magnitude and direction. Finally, after adjusting for mean fasting glucose over the Exam 1 to Exam 5 period (Model 3), we observed a loss of precision rather than attenuation with respect to associations of the SDG and ARV with the CASI, and significant attenuation and loss of precision of associations of these GV measures with Digit Symbol Coding performance. In contrast, the point estimate and confidence interval range for the association of the VIM with the CASI were unaffected, and the association of the VIM with the Digit Symbol Coding test only mildly attenuated, by adjustment for mean fasting glucose.

Table 2.

Weighted multivariable associations of mean fasting glucose and visit-to-visit fasting glucose variability over 10 to 12 years with Exam 5 (2010-2012) cognitive performance in middle-aged and older adults

	Cognitive test score difference per SD increment in glucose or HbA1c and two-fold increment in GV (95% CI)
	Model 1a	Model 2b	Model 3c
CASI d
Mean glucose	−0.16 (−0.36, 0.05)	−0.14 (−0.36, 0.08)	-
Exam 5 glucose	−0.21 (−0.42, −0.003)	−0.17 (−0.39, 0.05)	-
Exam 5 HbA1c	−0.20 (−0.41, 0.01)	−0.18 (−0.40, 0.04)	-
SDG	−0.33 (−0.54, −0.11)	−0.31 (−0.53, −0.07)	−0.41 (−0.74, −0.08)
ARV	−0.23 (−0.46, 0.003)	−0.19 (−0.43, 0.05)	−0.16 (−0.50, 0.17)
VIM	−0.47 (−0.81, −0.13)	−0.40 (−0.74, −0.06)	−0.39 (−0.73, −0.05)
Digit Symbol Coding e
Mean glucose	−1.26 (−1.67, −0.85)	−0.97 (−1.39, −0.54)	-
Exam 5 glucose	−1.10 (−1.54, −0.66)	−0.74 (−1.20, −0.29)	-
Exam 5 HbA1c	−1.54 (−1.99, −1.09)	−1.21 (−1.67, −0.74)	-
SDG	−1.39 (−1.82, −0.95)	−1.05 (−1.50, −0.59)	−0.64 (−1.30, 0.02)
ARV	−1.44 (−1.89, −0.97)	−1.12 (−1.60, −0.65)	−0.80 (−1.46, −0.13)
VIM	−1.16 (−1.85, −0.47)	−0.82 (−1.51, −0.14)	−0.73 (−1.41, −0.04)
Digit Span f
Mean glucose	−0.18 (−0.29, −0.07)	−0.10 (−0.21, 0.02)	-
Exam 5 glucose	−0.18 (−0.30, −0.06)	−0.09 (−0.21, 0.04)	-
Exam 5 HbA1c	−0.13 (−0.25, −0.01)	−0.05 (−0.18, 0.07)	-
SDG	−0.13 (−0.25, −0.02)	−0.04 (−0.16, 0.08)	0.07 (−0.10, 0.25)
ARV	−0.09 (−0.21, 0.03)	0.003 (−0.12, 0.13)	0.15 (−0.03, 0.32)
VIM	0.01 (−0.17, 0.19)	0.08 (−0.11, 0.26)	0.09 (−0.10, 0.27)

Abbreviations: ARV, average real variability of fasting glucose; CASI, Cognitive Abilities Screening Instrument; GV, fasting glucose variability from Exam 1 to Exam 5; SDG, standard deviation of fasting glucose; VIM, variability independent of the mean of fasting glucose.

^a Adjusted for age, sex, race, ethnicity, and educational attainment.

^b Adjusted for Model 1 covariates, plus mean systolic blood pressure, body mass index, depressive symptoms, HDL cholesterol, LDL cholesterol, and plasma triglycerides from Exam 1 to Exam 5; smoking status at Exam 5; lipid-lowering, antihypertensive, and psychiatric medication use at Exam 5; physical inactivity over the Exam 1 to Exam 5 period; and APOE ε4 allele carrier status.

^c Adjusted for Model 1 and Model 2 covariates, plus mean fasting glucose from Exam 1 to Exam 5.

^d n = 4367.

^e n = 3979.

^f n = 4353.

Evidence for moderation by Exam 1 diabetes status was found with respect to the associations of SDG, ARV, and VIM with the Digit Symbol Coding (all interaction P < .05). Stratified analysis revealed that associations of higher GV with worse Exam 5 Digit Symbol Coding performance were driven more strongly by participants with diabetes at Exam 1 (Supplementary Table S7 (23)). Associations were consistent among sex, race and ethnic groups, and incident T2D status (all interaction P > .10).

Association of Visit-to-Visit GV With Global Cognitive Decline Over 6 Years

Among 1834 participants with repeat CASI data, the mean (SD) Exam 6 CASI score was 90.0 (7.7) and the overall 6-year mean (SD) change in CASI score from Exam 5 to Exam 6 was +0.1 (6.6). In weighted analyses adjusted for Exam 5 CASI performance and Model 2 or Model 3 covariates, two-fold increments in continuous SDG_1–6, ARV_1–6, or VIM_1–6 were not associated with CASI change, nor were fasting glucose or HbA1c after Model 2 adjustments (Table 3). Associations were modified by sex, prevalent diabetes at Exam 1, and T2D incidence over the Exam 1 to Exam 6 period (Supplementary Table S8 (23)). Specifically, ARV_1–6 was associated with decline in CASI in women (−0.8 [95% CI: −1.6, −0.1] points) but not men (0.5 [95% CI: −0.2, 1.2] points), and associations of all GV measures with CASI change were generally stronger in the negative direction in participants with baseline or incident T2D. For example, in 229 participants with incident T2D, two-fold increments in SDG_1–6 and VIM_1–6 were associated with −1.7 (95% CI: −3.1, −0.3) and −2.1 (95% CI: −3.7, −0.5) point decreases, respectively, compared with 0.5 (95% CI: −0.01, 1.1) and 0.6 (95% CI: 0.03, 1.1) point increases associated with SDG_1–6 and VIM_1–6, respectively, in participants without new-onset T2D during the Exam 1 to Exam 6 period (Supplementary Table S8 (23)). In contrast, fasting glucose and HbA1c over the Exam 1 to Exam 6 period were not associated with cognitive decline in participants with prevalent or incident diabetes. However, higher mean glucose from Exam 1 to Exam 5 was prospectively associated with decrements in CASI, while GV measures over Exam 1 to Exam 5 were not (Supplementary Table S9 (23)).

Table 3.

Weighted multivariable associations of mean fasting glucose and visit-to-visit glucose variability over 16 to 18 years with 6-year change in global cognitive performance in 1834 middle-aged and older adults with repeat CASI data

	CASI score change (2010-2012 to 2016-2018) per SD increment in glucose or HbA1c and two-fold increment in GV (95% CI)
	Model 1a	Model 2b	Model 3c
Mean glucose1–6	−0.32 (−0.61, −0.04)	−0.30 (−0.59, 0.001)	-
Exam 6 glucose	−0.14 (−0.40, 0.13)	−0.10 (0.38, 0.18)	-
Exam 6 HbA1c	−0.09 (−0.42, 0.25)	0.03 (−0.33, 0.39)	-
SDG1–6	−0.34 (−0.63, −0.04)	−0.25 (−0.57, 0.07)	−0.02 (−0.51, 0.47)
ARV1–6	−0.38 (−0.72, −0.05)	−0.32 (−0.68, 0.03)	−0.14 (−0.67, 0.39)
VIM1–6	−0.15 (−0.65, 0.36)	−0.04 (−0.55, 0.47)	0.01 (−0.51, 0.53)

Abbreviations: ARV, average real variability of fasting glucose; CASI, Cognitive Abilities Screening Instrument; GV, fasting glucose variability from Exam 1 to Exam 6; SDG, standard deviation of fasting glucose; VIM, variability independent of the mean of fasting glucose.

^a Adjusted for Exam 5 CASI score, age, sex, race, ethnicity, and educational attainment.

^b Adjusted for Model 1 covariates, plus mean systolic blood pressure, body mass index, depressive symptoms, HDL cholesterol, LDL cholesterol, and plasma triglycerides from Exam 1 to Exam 6; smoking status at Exam 6; lipid-lowering, antihypertensive, and psychiatric medication use at Exam 6; physical inactivity over the Exam 1 to Exam 6 period; and APOE ε4 allele carrier status.

^c Adjusted for Model 1 and Model 2 covariates, plus mean fasting glucose from Exam 1 to Exam 6.

Sensitivity Analyses

Analyses using GV calculated from glucose measurements after multiple imputation produced similar associations to the unimputed analyses for the Exam 5 cognitive tests (Supplementary Table S10 (23)) and change in CASI performance (Supplementary Table S11 (23)). Results with respect to the Exam 5 CASI and Digit Symbol Coding using GV calculated using glucose measurements censored after new-onset T2D was identified still indicated an association of higher GV with worse performance on these tests following adjustment for Model 2 covariates, but not with the Digit Symbol Coding after adjustment for mean fasting glucose (Supplementary Table S12 (23)). Associations with change in CASI performance from Exam 5 to Exam 6 did not change with censored GV measures (Supplementary Table S13 (23)), except in participants with incident T2D, in which associations were attenuated (Supplementary Table S14 (23)). Finally, associations of censored GV measures, calculated exclusively below a T2D threshold to account for the possible influence of new diabetes medication use, with the Exam 5 CASI were similar to results in Table 2, but largely attenuated with respect to Digit Symbol Coding test performance (Supplementary Table S15 (23)). Measures of GV calculated exclusively above a T2D threshold in 823 participants with baseline or incident T2D were not associated with Exam 5 cognitive performance or change in CASI performance (Supplementary Table S16 (23)).

Discussion

In a racially and ethnically diverse prospective cohort of middle-aged and older adults, we observed that greater intraindividual visit-to-visit GV over 10 years was associated with worse global cognitive performance and processing speed, independent of sociodemographic, cardiovascular, or dementia risk factors and mean fasting glucose levels over this period. Associations among each of the 3 GV measures calculated were broadly concordant in strength and direction. These associations were also generally equivalent in magnitude to associations with single-timepoint fasting glucose and HbA1c, as well as mean fasting glucose over the study period. Associations were consistent across race and ethnicity, although participants with prevalent diabetes at study baseline drove associations with worse processing speed. Further, higher visit-to-visit GV was associated with global cognitive performance decline only among participants who developed T2D after study baseline, but not when excluding glucose measurements after incident T2D was identified from GV calculations. Cross-sectional, single-timepoint measures of fasting glucose and HbA1c were not associated with cognitive performance decline, even among participants with incident T2D. Collectively, these results suggest that higher visit-to-visit GV, as an index of worse long-term glucose control, may be deleterious to cognitive function during aging beyond mean or single-timepoint fasting glucose levels, particularly among individuals with or at risk of developing T2D later in life.

While diagnosis of T2D is associated with increased risk for Alzheimer disease and related dementias, impairments in glucose control mechanisms that precede T2D may be a critical step in pathways to cognitive decline. This is illustrated by evidence suggesting risk for dementia is higher among individuals with prediabetes compared to the normoglycemic range (3, 4), and that intensive glucose-lowering interventions in adults with T2D has not led to reduced risk for CVD or cognitive impairment (9, 10). At present, single-timepoint “snapshot” measures of fasting glucose or HbA1c—often collected only at annual (or less frequent) doctor visits—are the most widely used measures for assessing glycemic control prior to diagnosis of T2D. Yet, evidence for associations between these measures and cognitive decline have been inconsistent. For example, an analysis of Atherosclerosis Risk in Communities Study participants found no prospective association of HbA1c measured at a single time point with cognitive decline over 6 years when stratifying by diabetes status, despite an overall strong association of T2D with decline (5). Similarly, neuroimaging studies have reported conflicting results regarding the association of snapshot measures of fasting glucose and HbA1c with brain structural abnormalities, as some show and others do not show dysglycemia-related differences in brain or lesion volumes (29-32). This heterogeneity in the literature suggests that standard measures of glucose control may not be sufficient to describe the full impact of dysglycemia on cognitive risk.

The long-term variation in a number of cardiovascular-related factors is increasingly recognized as physiologically relevant beyond their snapshot measurements. For example, increased visit-to-visit variability in fasting glucose, blood pressure, body mass, and plasma lipids were each associated with CVD and mortality, independent of their mean values and other relevant risk factors (33, 34). Visit-to-visit GV in particular has emerged as an independent risk factor for vascular diseases in individuals with and without diabetes (24). Furthermore, short- and long-term GV have been consistently associated with worse cognitive outcomes including incident dementia (16, 17), and lower brain gray matter integrity and volume and disrupted functional architecture (35, 36) in T2D and general population samples.

However, results from our current study and previous diabetes trials call into question the impact of diabetes treatment on GV and its association with clinical outcomes. For example, an association between GV and incident CVD was observed only in the intensive glucose-lowering arm of the Veterans Affairs Diabetes Trial (13). In the current study, we observed associations between GV and global cognitive decline only in participants who developed T2D after baseline, but not after removing glucose measurements that may have been affected by new diabetes treatment from GV calculations. Others have noted that diabetes treatment may lower average glucose without altering GV, leaving individuals with T2D vulnerable to both hyperglycemic peaks and hypoglycemic dips (37), both of which are associated with incident dementia in older adults with T2D (8, 38). Therefore, differential effects of GV at high or low ends of the mean glucose spectrum and the possible modifying effect of diabetes treatment on GV warrant further investigation.

Our primary results are broadly congruent with the findings of a study in the CARDIA cohort, the only previous study, to our knowledge, of the association of visit-to-visit GV and cognitive performance in a large, general population cohort (18). However, our results differ with respect to associations found in participants with vs without diabetes. Whereas the CARDIA study reported stronger associations of GV (calculated as the ARV and coefficient of variation about the mean) with worse cognitive test scores in T2D-free participants, we observed the opposite. We note several distinctions in study characteristics that may contribute to this incongruence: first, CARDIA participants were younger than MESA participants at study baseline (18-30 vs 45-84 years, respectively) and were younger at time of first cognitive testing (∼50 vs 69.5 years). MESA participants were free of known clinical CVD at enrollment during middle-to-later adulthood whereas CARDIA participants may have developed CVD by middle adulthood; therefore, these study populations likely have different cardiovascular-related contributions to cognitive morbidity due to study design and secular trends that may not be comparable with adjustment methods. Second, fasting glucose measurements in CARDIA were collected over a longer period than in MESA (25 and 30 vs 10-12 and 16-18 years, respectively), with longer periods in between measurements. Finally, GV in CARDIA was calculated exclusively below a diabetes threshold, similar to our sensitivity analysis in which we excluded glucose measurements taken after the identification of incident T2D from GV calculations. Nevertheless, it is possible that the association between GV and cognitive performance differs according to time-related factors (eg, cognition in earlier adulthood vs middle age or older, length of time or stage in the lifespan captured by GV measures, or duration of diabetes or prediabetes exposure), which should be explored further.

Several overlapping mechanisms may underlie the association between higher GV and worse cognitive performance. Among T2D patients, increased GV is known to contribute to vascular oxidative stress, endothelial dysfunction, and enhanced production of inflammatory cytokines in vascular cells beyond the effects of elevated blood glucose (39, 40). Resulting vascular injury and its cascading effects could negatively alter cerebral blood flow distribution and promote neural lesions or neurodegeneration. Indeed, acute fluctuations in glucose levels were associated with greater neuronal mitochondrial dysfunction and markers of endothelial dysfunction compared to constant high glucose concentrations in vitro (41, 42). Additionally, an increase in hyperglycemic peaks due to GV could disrupt homeostatic compensatory mechanisms for glucose control, leading to hypoglycemic dips due to excessive insulin secretion. This is proposed to exacerbate cerebral insulin resistance and impaired insulin signaling in the brain (43), which in turn could negatively moderate neurogenesis and neuronal survival, inflammation, changes in cerebral blood flow, and energy metabolism in the brain (44). Therefore, multiple factors secondary to increased GV could contribute to cognitive decline, independent of mean plasma glucose level. Interventions that reduce both GV and overall blood glucose could provide additional benefit for cardiovascular and cognitive outcomes (24).

Strengths, Limitations, and Other Considerations

Strengths of our study include a large, multi-ethnic sample, a prospective-longitudinal design, assessment of 3 distinct measures of intraindividual GV, and the use of statistical techniques to mitigate potential biases beyond confounding. We also note several limitations and other considerations: (i) Although T2D was represented in our sample, the MESA population was free from known clinical CVD at baseline and may not be fully generalizable; (ii) Cognitive performance was not evaluated until the fifth clinical examination, 10 to 12 years after baseline, precluding the assessment of cognitive trajectory concurrent to the period of GV assessment; (iii) We lack adequate data to clarify whether complications from new-onset T2D could be more impactful than GV itself; (iv) The cognitive tests we studied do not have established thresholds for defining cognitive impairment, however, the CASI was chosen because it is purportedly less influenced by cultural factors (20), an important consideration for a study as racially and ethnically diverse as MESA; (v) We restricted the traditional statistical hypothesis testing framework to the univariate analyses of sample characteristics in Table 1. Since our analyses were: (a) prespecified, (b) not independent (ie, the 3 measures of GV were expected to be correlated despite unique calculations for each), and (c) because several cardiovascular-related covariates in our models may be up- or downstream of glucose or GV in shared causal pathways to cognition, we purposefully avoided an analysis approach based on significance testing, opting to report effects on point estimates and 95% CI that do not require correction for multiple testing (45); and (vi) While visit-to-visit GV appears to be physiologically meaningful, it is a crude measure that may be influenced by changes in diet, lifestyle factors, exercise, socioeconomic or behavioral factors such as poor healthcare quality or medication compliance, which could also impact cognition (46). Although we adjusted for physical activity and hypertension, lipid, and psychiatric medication use over the study period, there may be residual confounding.

Conclusion

We found that greater visit-to-visit variability in fasting glucose during mid- to late-life was associated with worse performance on tests of global cognitive performance and processing speed in a diverse sample of adults, above and beyond mean glucose values over this period. Additionally, greater visit-to-visit GV was associated with modest cognitive decline over 6 to 8 years in older adults with or at risk for developing T2D, while cross-sectional, snapshot measures of fasting glucose and HbA1c were not. Our results add to a growing body of literature showing that the long-term variability in fasting glucose may provide information beyond single-timepoint or mean values in specific contexts. Future studies should determine whether interventions that improve overall glucose control in addition to lowering glucose levels in diabetes or prediabetes ranges provide added benefit for cardiovascular or brain health.

Abbreviations

APOE

apolipoprotein E

ARV

average real variability of fasting glucose

CAIDE

Cardiovascular Risk Factors, Aging, and Incidence of Dementia

CASI

Cognitive Abilities Screening Instrument

CESD

Center for Epidemiologic Studies Depression

CVD

cardiovascular disease

GV

visit-to-visit fasting glucose variability

HbA1c

glycated hemoglobin

HDL

high-density lipoprotein

LDL

low-density lipoprotein

MESA

Multi-Ethnic Study of Atherosclerosis

SDG

standard deviation of glucose

T2D

type 2 diabetes mellitus

VIM

variability independent of the mean of fasting glucose

Data Availability

Datasets analyzed in this study may be requested from https://www.mesa-nhlbi.org/ by individuals trained in human subject research after approval by the MESA Steering Committee.

Funding

This work was supported by the National Institute on Aging (grants K01AG073581 and R03AG064569 to C.L.S.; R01AG054069 to T.M.H.; R01AG058969 to T.M.H. and K.M.H.; RF1AG054474 to J.D.; and RF1AG070881 to M.M.C.) and the National Heart, Lung, and Blood Institute (grant R01HL127659 to S.R.H.). The Multi-Ethnic Study of Atherosclerosis is supported by the National Heart, Lung, and Blood Institute (contracts 75N92020D00001, HHSN268201500003I, N01-HC-95159, 75N92020D00005, N01-HC-95160, 75N92020D00002, N01-HC-95161, 75N92020D00003, N01-HC-95162, 75N92020D00006, N01-HC-95163, 75N92020D00004, N01-HC-95164, 75N92020D00007, N01-HC-95165, N01-HC-95166, N01-HC-95167, N01-HC-95168 and N01-HC-95169), and by the National Center for Advancing Translational Sciences (grants UL1-TR-000040, UL1-TR-001079, and UL1-TR-001420). The funders had no role in the study design, collection, analysis, or interpretation of data; drafting of the manuscript; or the decision to submit the manuscript for publication. No authors are or were affiliated with the funders.

Disclosures

The authors have nothing to disclose.