Relationship Between Glycemia and Cognitive Function, Structural Brain Outcomes, and Dementia: A Mendelian Randomization Study in the UK Biobank

Observational evidence largely suggests that hyperglycemia, diabetes, and insulin resistance are associated with poorer brain health, including worse cognitive function, risk of cognitive decline, and dementia (1–4). The exact mechanisms remain elusive (5), as is how to best treat those with a diagnosis of both diabetes and cognitive dysfunction (6). These factors limit intervention attempts, as it is unclear whether hyperglycemia per se is the culprit or whether instead vascular risk factors (e.g., hypertension, dyslipidemia, inflammation) mediate the association between diabetes and poorer brain health outcomes. It is also unclear whether the associations between hyperglycemic conditions and brain outcomes are causal in nature. Some evidence also supports a bidirectional relationship (5), implicating a vicious cycle whereby diabetes may result in dementia and dementia could then trigger further diabetes complications (7).

Mendelian randomization (MR) overcomes some of the limitations of causal interpretation in observational studies. So far, three MR studies have focused solely on Alzheimer dementia (AD), with all reporting no impact of diabetes (8–10). Pathways to cognitive decline and dementia involve a combination of vascular and neurocognitive mechanisms that may act either independently or in concert (11). Diabetes is more related to the vascular pathways, but there is evidence that it also has neurotoxic consequences (12). There have been no previous MR studies that have investigated glycated hemoglobin (HbA_1c) and a range of brain health measures, such as cognitive function or structural brain abnormalities. No previous MR studies have investigated whether the bidirectional association may be causal in nature. Thus, the current study used 1) genetic instruments for type 2 diabetes and HbA_1c to examine the relationship with cognitive function, structural brain measures, and AD and, where possible, 2) genetic instruments for cognitive function to investigate whether the relationship with diabetes or HbA_1c might be bidirectional.

Two-sample MR (a design that exploits genome-wide association [GWA] summary statistics derived in nonoverlapping samples) was used to mitigate biased results that are due to the “winner’s curse” (the overestimation of genetic associations that are common in the one-sample MR setting) because it is neither necessary nor desirable that the genetic variants to be instrumented be derived from the same sample as the one under study (13). An important advantage of using two-sample MR is that it allows sensitivity analyses to identify unbalanced (directional) horizontal pleiotropy (described under statistical analyses), which is crucial to satisfy MR assumptions. In our MR analyses, there was some sample overlap for diabetes and cognitive function (reaction time [RT]) but not for the HbA_1c genetic variants.

Full details of the UK Biobank (UKB) cohort have been described elsewhere (14). Briefly, UKB consists of 500,000 males and females from the general U.K. population, aged 40–69 years at baseline (2006–2010). There was a maximum of 349,326 participants of European ancestry with both genotype and all the phenotypes of interest in the current study (Fig. 1).

A total of 487,409 UKB participants were genotyped using one of two customized genome-wide arrays that were imputed to a combination of the UK10K, 1000 Genomes phase 3, and the Haplotype Reference Consortium reference panels, which resulted in 93,095,623 autosomal variants (15). We then applied additional variant-level quality control (QC) and excluded genetic variants with: Fishers exact test <0.3, minor allele frequency (MAF) <1%, and a missing call rate of ≥5%. Individual-level QC meant that we excluded participants with: excessive or minimal heterozygosity, >10 putative third-degree relatives as per the kinship matrix, no consent to extract DNA, sex mismatches between self-reported and genetic sex, missing QC information and non-European ancestry (on the basis of how individuals had self-reported their ancestry and the similarity with their genetic ancestry, as per a principal component analysis of their genotype).

UKB administered five baseline cognitive assessments to all participants via a computerized touch screen interface, all of which are described in detail elsewhere (16). In the visual memory assessment, respondents were asked to correctly identify matches from six pairs of cards after they had memorized their positions. Then, the number of incorrect matches (number of attempts made to correctly identify the pairs) was recorded, with a greater number reflective of a poorer visual memory. RT (in milliseconds) was recorded as the mean time taken by participants to correctly identify matches in a 12-round game of the card game “Snap.” A higher score on this test indicated a slower (poorer) RT. Both of these variables were positively skewed, and therefore RT scores were transformed using the natural logarithmic function ln(x), while visual memory was transformed using ln(x + 1).

Structural brain MRI scans were performed by UKB in a subsample of participants using standard protocols, as published previously (17). The postprocessed measures derived by UKB and used in this study included mean hippocampal volume (HV) (mm³) and white matter hyperintensity volumes (WMHV) (mm³). WMHV was log-transformed, as it was positively skewed. The number of participants with WMH volume was 32,506, and 32,407 had HV data available (Fig. 1). This was after excluding 114 individuals who were outliers (+3 SD from the mean) and who were not included in the genetic sample after QC. We checked whether there was any overlap between participants with AD and those with white matter hyperintensities, but as there was only one with both AD and neuroimaging data, we did not consider this an issue for our analyses. We report results in mm³ for HV and exponentiated β (expβ)/percentages for WMHV. UKB provided algorithmically defined AD. AD (2006–2017) was captured using ICD-10 codes (alphanumeric codes to classify symptoms, diseases, injuries, infections, and disorders) in linked hospital episode statistics data, as well as from death certification, primary care, self-report, and nurse interview. These algorithmically defined outcomes were provided by UKB. Coded diagnoses were compared with clinical expert adjudication of full-text medical records. Details of ICD-10 and primary care Read codes are presented in Supplementary Tables 6 and 7, and more in-depth information on the algorithm by Wilkinson et al. (18) can be found elsewhere.

Analyses were performed using a combination of the mrrobust package in STATA, version 15, the MendelianRandomization R package, using RStudio version 1.1.456, and PLINK version 2.0.

For diabetes, 157 independent (via linkage disequilibrium [LD] clumping performed in PLINK, using r² = 0.2 and a 250kb window) genetic variants were chosen from the 2018 GWA study (GWAS) by Mahajan et al. (19), in which they combined data across 32 studies, including 74,124 people with diabetes and 824,006 control subjects of European ancestry. In our sample these variants had an F-statistic of 27.43 and explained ∼1.5% (pseudo-R² = 0.015) of the variance in 14,010 diabetes cases (defined using a validated [against primary care data] algorithm of self-reported physician diagnosis and/or medication [20]). The 51 HbA_1c single nucleotide polymorphisms (SNPs) we used were from the latest transancestry GWAS by Wheeler et al. (21). These variants explained 2.8% (R² = 0.028) of the variance in HbA_1c in our sample and had an F-statistic of 164.6. In UKB, HbA_1c assays were performed using five Bio-Rad Variant II Turbo analyzers (22)). LD clumping in PLINK confirmed that the 51 SNPs were independent (r² = 0.2, 250-kb window). For bidirectional MR analyses, we used 43 SNPs associated with RT from a recent GWAS (23) of 330,069 White European UKB participants with both phenotype and genotype data available. The RT variants explained 0.3% of the variance in RT in our study, and the instrument had an F-statistic of 24.0. As with the diabetes and HbA_1c SNPs, the RT SNPs were also confirmed to be independent via clumping in PLINK. We harmonized genetic variants from the published GWAS with UKB by aligning the effect alleles. Full details of all the SNPs are in Supplementary Table 1. Our selection process for genetic instruments is detailed in Fig. 2. Briefly, in relation to MAF for T2DM and HbA_1c, the authors excluded any (rare) variants with an MAF <1%. For the reaction time variants, as we were uncertain of MAF filtering in the discovery GWAS, we inspected the MAF for each SNP and found that one variant had an MAF of 0.3%. However, when we performed a leave-one-out analysis excluding this variant (rs141885450), our results remained identical (data not shown). More details on the discovery GWAS for our exposures can be found in the original articles (19,21,23).

We firstperformed linear/logistic regression to examine the associations between SNPs for HbA_1c/diabetes and all of our outcomes in PLINK. Second, we fitted logistic/linear models to examine the associations between RT SNPs and diabetes and HbA_1c. Then, inverse-variance–weighted (IVW) MR was implemented as our main model. This approach calculates the effect of a given exposure (e.g., diabetes) on an outcome of interest (e.g., visual memory) by taking an average of the genetic variants’ ratio of variant-outcome (SNP→Y) to variant-exposure (SNP→X) relationship estimated using the same principles as a fixed-effects meta-analysis (24). We also performed standard MR sensitivity analyses, including MR-Egger regression (which yields an intercept term that indicates the presence or absence of unbalanced horizontal pleiotropy) (25) and the weighted median estimator (WME) (which can yield more robust estimates when up to 50% of the genetic variants are invalid) (26). Identical MR analyses were performed for diabetes (157 SNPs), HbA_1c (51 SNPs), and RT, visual memory, WMHV, HV, and AD. Additionally, for RT, visual memory, WMHV, and HV, we repeated the MR analyses using only the 16 glycemic HbA_1c SNPs and then the 19 erythrocytic SNPs (16 SNPs were unclassified, as per the discovery GWAS). We did not perform these analyses for the AD outcome because of the likelihood of imprecision as a result of a substantially reduced sample size. For bidirectional analyses, we used the RT SNPs to investigate associations with HbA_1c and diabetes. Results are presented as expβ-coefficients (multiplicative effect size) for RT/visual memory/WMHV, AD risk, and unit differences in HV (mm³); per unit increase in HbA_1c (mmol/mol); and 1-log-odds of diabetes. For bidirectional MR analyses, results are expressed as diabetes risk and unit differences in HbA_1c (mmol/mol) per unit increase in RT (milliseconds). To ensure that our results were not affected by residual population stratification, we performed all of our MR analyses with adjustment for 10 genetic principal components. These results were qualitatively identical to the main results, and thus, we present them in Supplementary Table 9.

MR has three strict assumptions that must be met for study results to be valid

1. The association between the genetic variants for the exposure and the exposure itself must be strong and robust (meaning that these associations have usually been replicated and validated via GWAS). This assumption was met because our genetic variants for diabetes, HbA_1c, and RT were all from large-scale recently published GWAS. However, for the RT SNPs only, as there was some concern about weak instrument bias, we additionally included an MR-Egger Simulation Extrapolation (SIMEX) (27) sensitivity analysis, which we report in the results section.

2. The association between the genetic variants (for the exposure) and the outcome must only be via the exposure under study; otherwise, this is known as unbalanced horizontal pleiotropy and may bias MR results. This assumption was assessed using the methods detailed below, including MR-Egger.

3. There should not be an association between the genetic variants (for the exposure) and common confounders of the relationship under study (e.g., the diabetes SNPs should not be associated with factors such as smoking). We checked this assumption by regressing multiple confounders (BMI, deprivation, systolic blood pressure, total cholesterol, triglycerides, C-reactive protein [for which outliers >3 SD were removed]), smoking and stroke) on the diabetes, HbA_1c, and RT SNPs. We applied a Benjamini-Hochberg false discovery rate (BH-FDR) of 0.25 to account for multiple testing.

As mentioned earlier, UKB contributed to the Mahajan et al. (19) diabetes GWAS, and thus, we performed some analyses in addition to the main MR analyses to understand whether our diabetes and brain health results may be subject to “winner’s curse” bias. Thus, we turned to the earlier 2014 GWAS by Mahajan et al. (28), as this study did not include UKB. We looked up the 157 diabetes SNPs used in our instrument and found 77 of them in the Mahajan et al. 2014 GWAS summary statistics; this reduced number of SNPs may be due to differences in coverage of imputation panels (i.e., the 2014 GWAS imputed to phase II/III of HapMap, and the 2018 GWAS used the Haplotype Reference Consortium). We took the corresponding log(β) and SE for this 77-SNP diabetes instrument (F-statistic = 30.88) from the Mahajan et al. 2014 GWAS so, that the estimates would not include UKB. Third, we performed all of our MR analyses (IVW, MR-Egger and WME) with this instrument, and as results were qualitatively the same as when we used the 157-SNP instrument, we present these in Supplementary Table 8.

The UKB data are publicly available to all bona fide researchers at https://www.ukbiobank.ac.uk.

Sample characteristics are presented in Supplementary Table 2. In our sample 54% of participants were male, and the mean age was 56.7 years; 27% participants reported ever smoking, and 20% were in the most deprived group. Mean HbA_1c was 35.9 mmol/mol, mean RT was 554.6 ms, and the mean number of visual memory errors (i.e., number of incorrect matches) was 4.1. Mean HV was 3,830 mm³, while the median WHMV was 2,824 mm³. Mean systolic blood pressure and BMI were 138.2 mmHg and 27.3 kg/m², respectively. Median values for triglycerides and C-reactive protein were 1.5 and 1.3 mmol/mol, respectively, while mean total cholesterol levels were 5.7 mmol/mol. There were 14,010 participants with diabetes, 746 with AD, and 6,301 with stroke at baseline. On average, participants engaged in 3.6 days of moderate physical activity for >10 min.

Diabetes was not associated with RT or visual memory using IVW, and these results were consistent with MR-Egger and WME approaches (Table 1). HbA_1c was not associated with RT using IVW, MR-Egger, or WME. However, the MR-Egger intercept P value was <0.05; thus, we performed leave-one-out analyses and found that rs10774625 was pleiotropic. When we removed this SNP from the model, the intercept P value changed to >0.05 and results remained consistent. When restricted to the 16 glycemic and subsequently 19 erythrocytic SNPs, there was no evidence of an association with RT (Table 1). Using all 51 SNPs, none of the three MR approaches used showed evidence of an association between HbA_1c and visual memory (Table 1). When restricted to the 16 glycemic SNPs, there was also no association with visual memory (Table 1). Finally, when restricting to the 19 erythrocytic SNPs, we also observed no associations with visual memory across all MR approaches (Table 1).

Diabetes was not associated with HV, WMHV, or AD using IVW, MR-Egger, or WME (Table 2). For HbA_1c, the 51-SNP genetic instrument showed no evidence of associations with WMHV (Table 2). When we restricted analyses to only the 16 glycemic SNPs and 19 erythrocytic SNPs, we also saw no evidence of associations between HbA_1c and WMHV (Table 2). For HV, the 51-SNP HbA_1c instrument showed no associations across the IVW, MR-Egger, and WME (Table 2). When analyses were restricted to the 16 glycemic SNPs and, subsequently, the 19 erythrocytic SNPs, there was also no evidence of associations with HV (Table 2). Both HbA_1c (using all 51 SNPs) and diabetes were not associated with AD using conventional IVW MR (odds ratio [OR] 1.09 [95% CI 0.42; 2.83]) and 1.15 [0.87; 1.52], respectively).

In the bidirectional analyses, we observed no associations between RT and diabetes with all MR approaches producing consistent results (Table 3). However, our results suggested an association between RT and HbA_1c, but the MR-Egger intercept P value was <0.05, indicating unbalanced horizontal pleiotropy. Thus, we performed leave-one-out analyses and found that the pleiotropic SNP was rs10775404. The results for the WME after exclusion of this variant suggested an association such that slower RT was associated with lower HbA_1c (β-coefficient = −1.11 mmol/mol [95% CI −1.95; −0.28]). Also, the MR-Egger intercept P value was >0.05 after exclusion of this variant.

We performed MR-Egger SIMEX alongside the conventional IVW MR to address issues with weak instruments in relation to the RT SNPs. Our SIMEX analyses were consistent with all the other MR approaches for RT and diabetes (Table 3). However, for RT and HbA_1c, MR-Egger SIMEX suggested the presence of unbalanced horizontal pleiotropy, and we decided to perform leave-one-out analyses. We identified rs10775404 as the pleiotropic SNP, and after excluding this variant, the MR-Egger SIMEX intercept P value increased to >0.05. Additionally, we checked to see whether our genetic instruments for diabetes, HbA_1c, and RT were associated with common confounders. When we regressed BMI, socioeconomic deprivation, systolic blood pressure, total cholesterol, smoking, stroke at baseline, triglycerides, and C-reactive protein on the diabetes, HbA_1c, and RT SNPs, we observed some associations between our genetic variants and confounders, using a BH-FDR of 0.25 to account for multiple testing (Supplementary Tables 3–5).

In the first comprehensive MR study of HbA_1c/diabetes and brain health, we show that overall there is unlikely to be a causal relationship. In bidirectional MR analyses, we found no relationship between RT and diabetes or HbA_1c.

No previous studies have attempted to investigate, using MR, the association between HbA_1c and any of the outcomes reported here. In addition, we did not find any association when the instrument was restricted to the glycemic variants, providing little support for a true association. Bidirectional findings of RT and diabetes showed no evidence of causal relationships across IVW, MR-Egger, and WME MR approaches. The IVW and MR-Egger also showed no associations between RT and HbA_1c. However, our WME result showed a marginal association, in an unexpected direction between RT and HbA_1c, such that slower RT was associated with lower HbA_1c. However, as this was inconsistent with the IVW (conventional MR) and MR-Egger, we believe that this finding should be interpreted with utmost caution.

We are the first to investigate diabetes/HbA_1c and HV, and WMHV using an MR approach, but we observed no evidence of associations among these phenotypes. Although UKB has the largest brain imaging study in the world, perhaps a larger sample size would allow for more precise estimation of the relationships with these structural brain outcomes. However, the weak association between diabetes and AD only is at least supported by previous MR studies, which reported no impact of diabetes on AD (8–10), and thus, taking all of this evidence together, it is likely that diabetes does not exert a causal influence on risk of AD. Additional support for these findings comes from a recent study that suggested that, using a polygenic risk score for diabetes, the association between diabetes and cognitive state shown by observational studies (2) may be explained by early life socioeconomic factors and childhood cognition, as well as by educational attainment (29).

Our MR findings were validated by checking that we met all three core assumptions. Assumption 1 was met by ensuring that we selected the best available genetic variants for our exposures (diabetes, HbA_1c, and RT) from the latest and most robust GWAS. Assumption 2, which relates to horizontal pleiotropy between exposure SNPs and the outcomes, was checked by performing standard sensitivity analyses under the MR-Egger and weighted median estimator models. Where necessary (Egger intercept P < 0.05), we performed additional leave-one-out analyses to exclude a SNP that was identified as pleiotropic and reran our MR analyses. Finally, we checked assumption 3 by performing linear/logistic regressions between our genetic instruments for diabetes, HbA_1c, and RT and unobserved confounders. As we found evidence of associations with some confounders after applying a BH-FDR, we believe that these warrant further investigation but are beyond the scope of our study. The reasons for these associations could be first related to the fact that these traits are all polygenic in nature, and/or second, it could also be that some of these associations indicate vertical rather than horizontal pleiotropy. Future research could investigate whether any of these SNPs are vertically pleiotropic by performing MR mediation analyses, either using multivariable MR or two-step MR (30).

Our study design had some limitations in terms of the RT and diabetes genetic variants, as the GWAS from which we selected these SNPs both contained UKB in their samples. However, we also performed sensitivity analyses using a diabetes instrument and estimates from a previous GWAS that did not include UKB (28), and results remained qualitatively identical. For HbA_1c, a two-sample MR design with no overlap was used. We had lower precision for MR analyses with AD, HV, and WMHV, and larger samples are required for more robust conclusions. It is also possible that the lack of evidence for causal relationships in the current study may indicate that other cognitive function and neuroimaging outcomes should be studied in the future. Cognitive decline is also an important outcome that we did not investigate, but there would have been very few individuals for this analysis, as only a subsample underwent repeat cognitive testing. The time between tests is also not likely to be sufficient for cognitive decline to manifest (mean = 6 years for visual memory and 4 years for RT), as participants were, on average, aged 57 years at baseline. In relation to other exposures of interest, duration of diabetes as well as other glycemic exposures could be considered in future. However, there are currently no genetic variants for duration of diabetes, and instruments for traits such as insulin resistance are not particularly strong (e.g., HOMA for insulin resistance has only two validated SNPs with small effect sizes). It would also be of value to test other mechanisms, as it is possible that the observational association between hyperglycemia and brain health is not due to elevated peripheral glucose levels. Moreover, the UKB cognitive tests are novel and specific to this cohort and have thus, not been extensively validated (16). The AD diagnoses may also be problematic, as accurate dementia diagnoses are extremely challenging to clinical experts, particularly among patients in the age range of UKB. However, previous UKB studies have used similar dementia diagnoses (31,32), although the algorithm we relied on here additionally incorporates primary care data alongside hospital episode statistics, mortality, self-report, and nurse interview data. Our findings are unlikely to suffer from issues related to population stratification, as all of the individuals in our sample were of White European descent and in sensitivity analyses we adjusted for 10 principal components, which yielded the same results. However, MR studies should also be performed to investigate the associations we report here in other ethnic groups, particularly given that the SNPs we used were derived using transethnic GWA approaches.

In conclusion, our MR study of glycemia and cognitive function, structural brain MRI measures, and AD suggests that these associations are not likely to be causal. However, all of our findings should be triangulated using other methods, in particular those relevant for causal inference.

Acknowledgments. This work was conducted under the approved UKB project number 7661. The authors thank the volunteer participants of the UKB and the UKB researchers.

Funding and Duality of Interest. This work was jointly funded by Diabetes UK and British Heart Foundation (BHF) grant 15/0005250. K.B. holds a Sir Henry Dale Fellowship funded by Wellcome and the Royal Society (grant number 107731/Z/15/Z). R.M. is funded by a Sir Henry Wellcome Postdoctoral Fellowship (201375/Z/16/Z). L.S. reports grants from BHF and Diabetes UK during the conduct of the study and grants from Wellcome, MRC, National Institute for Health Research, GlaxoSmithKline, and BHF outside the submitted work. L.S. is a trustee of the BHF. N.C. reports grants from Diabetes UK and BHF during the conduct of the study and personal fees from AstraZeneca and grants from the Medical Research Council outside the submitted work. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. V.G. performed all statistical analyses and wrote the manuscript. V.G. and N.C. conceived the idea and design of the study. All authors contributed to the interpretation of the results, provided important intellectual input, and approved the manuscript. V.G. 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.

Prior Presentation. A non–peer-reviewed version of this article was submitted to the medRxiv preprint server (https://doi.org/10.1101/2020.05.07.20094110) on 12 May 2020.