PubMed, PsycInfo, and Cochrane Library (1950 to January 2020) was searched using the following terms: “(Diabetes or T2D) AND (brain or cerebrum or cerebral or cerebellum or cerebellar or hippocampus or hippocampal or subcortical or (cerebral ventricle) or ventricular or thalamus or thalamic or (basal ganglia) or striatum or (grey matter) or (white matter)) AND ((magnetic resonance imaging) or MRI or (computed tomography) or neuroimaging or morphometry or (diffusion tensor imaging) or volume or volumetric or thickness or atrophy or shrinkage).” Search included all text, all dates, full text papers in English (all article types for PubMed and PsycInfo; trials for Cochrane Library). Both literal and Medical Subject Heading searches were performed when possible. Titles and abstracts of the paper were screened by two reviewers (T.Z. and N.C./M.S.) for full text review. Full text and supplemental material of qualified studies were retrieved and examined by two reviewers against the inclusion and exclusion criteria. Disagreements between two reviewers were resolved by consensus or by a third reviewer. Citation maps of retrieved papers, previous reviews and previously identified journals were examined to identify additional journal articles that meet the criteria.

Studies were included if they had: (1) human adult participants; (2) at least a control group consisting of healthy participants without T2DM; (3) at least one comparison between participants diagnosed with T2DM and controls; (4) brain grey matter or white matter volume data from structural MRI or computed tomography (CT) scan data of T2DM and control participants; (5) data acquired using a validated automatic or manual segmentation method; (6) for longitudinal studies, longitudinal measurements at a minimum of two time points. Studies were excluded if they had: (1) participants with only type 1 diabetes mellitus (T1DM) or that did not differentiate T1DM and T2DM; (2) participants with only subclinical diabetes, including impaired fasting glucose, impaired glucose tolerance, insulin resistance, in the disease group or control group; (3) T2DM participants that all have major conditions other than diabetes, e.g., mental illness, behavioural problems, substance abuse, systemic illness or major brain structural abnormalities; (4) T2DM participants that are all under diabetes treatment (including placebo treatments) and being compared with those without such treatment; (5) only case studies and small samples with less than 10 participants in one group; (6) duplicate samples (for identified duplicates, the sample that best fits the study criteria will be included and the other samples will be excluded); (7) review articles, theses, unfinished studies, or entries with abstract only; (8) for longitudinal studies, samples where the total MRI follow-up period is less than 12 months.

Studies meeting the inclusion and exclusion criteria were assessed for quality using the Newcastle-Ottawa scale. Each study was evaluated on eight items classified into three categories including the selection of the study groups, the comparability of the groups, and the ascertainment of outcome of interest. Each quality item was awarded by a star (except two for comparability) and for each study up to nine stars in total.

Two of the authors extracted data (T.Z. and N.C.) and discrepancies were resolved by consensus. Data extracted consisted of (1) study design and number of participants in each group; (2) participants’ demographics including age, sex ratio, diagnostic criteria for T2DM, duration of T2DM, medication status, glycosylated hemoglobin (HbA1c) levels, fasting plasma glucose levels and body mass index; (3) measurement details including MRI parameters, structural measurements and segmentation method; and (4) study results including areas of interest (left and right) and effect sizes (left, right, and total). Data from studies reporting ratios relative to intracranial volume were included if a volumetric difference could be computed based on group statistics.

Multiple reports on the same cohort but on different brain structures were considered independent studies and included. Where a particular sample was reported in multiple studies on the same brain structure, the study that fit the selection criteria and provided data compatible for meta-analysis was included and other studies were excluded. Studies that reported effect sizes (or provided them after contact) were considered and from those the most recent study with the largest sample size was selected. If there was more than one study similar in sample size and time, the one with the highest quality rating was selected.

R version 3.1.1 (R Foundation for Statistical Computing, Vienna, Austria) [18] was used for statistical analysis. Meta-analyses were performed using the Metafor version 1.9–4 R package (https://www.metafor-project.org) [19]. Volumes of brain structures and annual percentage mean atrophy rate was considered as the effect size, and calculation of required standard error (SE) for meta-analysis was based on the standard deviation and number of participants in each group for each individual study. Availability of volume of a brain structure, either reported or computed based on other reported results, was the essential requirement for the meta-analysis. Atrophy rate was calculated using the formula: atrophy=[(volume_time1–volume_time2)/volume_time1]/(time1–time2) if not provided. Where insufficient data were available for inclusion in the meta-analysis, authors were contacted directly to seek additional information.

A random-effects model using a restricted maximum likelihood estimator was used for all meta-analyses. A random effects model was chosen based on the assumption that included studies are heterogeneous because they sample populations with different characteristics using a range of methodologies and therefore one cannot assume that there is a single effect size [20]. A random effects meta-analysis estimates the mean of a distribution of effects rather than estimating a unique effect [20]. We assessed heterogeneity across studies with the Q statistic (with P<0.01 being suggestive of significant heterogeneity) and the I² statistic (values of 25%, 50%, and 75% were indicative of low, medium, and high heterogeneity). Separate meta-analyses were performed for different brain structures.

Meta-regression was used to investigate the impact of demographic and diabetes characteristics on differences in brain volumes between metabolically healthy individuals and individuals with diabetes, if there were at least 10 studies providing information of a covariate, including age (centred at 60 years), sex, diabetes duration, ratio of diabetes patients taking medication, plasma insulin levels, HbA1c levels and fasting plasma glucose levels, using linear mixed-effects models.

Sensitivity analyses were conducted using the leave-one-out method to identify studies contributing excessively to heterogeneity. Visual evaluation of asymmetry of the funnel plots was used to assess the bias in the meta-analyses results toward publication of studies with significant outcomes. The trim-and-fill method was used to estimate the number of missing studies (representative of unreported effect sizes) in the meta-analysis to estimate adjusted effect sizes.

The systematic search identified 10,360 entries from PubMed, 664 from PsycInfo, and 3,641 from Cochrane Library; 60 duplicate entries were identified and excluded. Of these 14,605 studies, 355 studies passed title screening, and 109 studies remained after abstract screening; 64 studies remained after applying inclusion criteria in full-text assessment (Fig. 1).

Of the included studies, 40 were cross-sectional case-control, 17 cross-sectional population-based, one longitudinal case-control, and six longitudinal population-based studies. Forty-two studies (n=31,630; mean age 71.0 years; 44.4% male; 26,942 controls; 4,688 diabetes) with enough volumetric data could be considered for meta-analysis. Of these, 25 used automated segmentation, six used manual tracing, and 15 studies applied voxel-based morphometry (Tables 1 and 2) [11,21–84].

Twenty-eight studies reported disease duration, 17 medication status (number of patients taking medication, or specifically taking oral medication or insulin), five plasma insulin levels, 29 HbA1c levels, and 22 fasting plasma glucose levels (for specific demographic and T2DM-related information) (Table 2).

Of all the brain regions investigated, only total brain volume, total grey matter, total white matter, hippocampus, thalamus, caudate, putamen, globus pallidus, amygdala, nucleus accumbens, frontal lobe, superior temporal gyrus, total cerebrospinal fluid, and white matter hyperintensity were reported in a sufficient number of studies to be included in meta-analyses. Total brain volume annual atrophy rate was also reported in five longitudinal studies (see Table 1 for details).

Significant inhomogeneity was observed in Q tests of the meta-analysis, except in thalamus, caudate, putamen, globus pallidus, amygdala, nucleus accumbens, and in T2DM-normal difference in superior temporal gyrus and frontal lobe (Table 3).

Evidence of some publication bias was also detected for most brain regions investigated. Visual inspection of funnel plots revealed that studies were likely missing for total brain volume (13.3% of total), globus pallidus (40%), superior temporal gyrus (16.7%), frontal lobe (28.6%), cerebrospinal fluid (20%), and white matter hyperintensity (20%). Although asymmetry and presence of missing studies suggested some publication bias toward studies reporting higher atrophy rates, trim and fill test indicated that the differences between corrected and reported volumetric differences were generally small and therefore publication bias is unlikely to have significantly influenced the present results (Fig. 3).

The effect of age, sex ratio, diabetes duration, medication, fasting glucose and HbA1c on the association between T2DM and volume of total brain, grey matter, white matter and hippocampus was investigated by meta-regression analysis. A significant negative association between decreasing total brain volume difference and increasing age was detected such that every additional year in age above 60 was associated with a 4.4% smaller volumetric dif ference between individuals with and without diabetes (individuals with diabetes at age 60 years are 28.45cm³ smaller; this difference decreases by 1.24 cm³ per year) (Supplementary Table 1). A significant positive association between decreasing grey matter volume difference and increasing diabetes duration was also detected such that every additional year above mean diabetes duration (10.5 years) in age above 60 was associated with an 8.8% larger volumetric difference between individuals with and without diabetes. No significant effects were observed in other analyses (Supplementary Table 1, Supplementary Fig. 3).

This study’s results are consistent with the findings of previous reviews reporting that brain volumes are smaller in those who live with T2DM, but it also substantially extends our understanding of the scope of this effect. The present study was able to convincingly demonstrate volumetric differences in four brain structures while also precisely summarising their magnitude. It demonstrated that diabetes-related volumetric brain differences were substantial (total brain: 1.88%; grey matter: 2.81%; white matter: 2.15%; hippocampus: 4.4%). Indeed, in normal ageing, total brain volume shrinks by about 0.5% every year from the 40s onwards with further acceleration after age 70 [21]. Similarly, the hippocampus shrinks by about 0.3%/year before 55, 0.85%/year between 55 and 70, and 1.1%/year thereafter in those cognitively intact [16]. Thus, the differences observed in T2DM correspond to about 4 to 5 years of normal ageing, and possibly more. It is also worth noting that while these effects were relatively large across these brain regions, they were particularly strong in the hippocampus. This is noteworthy because subcortical atrophy in the hippocampus is a hallmark of Alzheimer’s disease and is one of the strongest predictors of conversion from mild cognitive impairment to Alzheimer’s disease [85,86]. Importantly, the rate of hippocampal atrophy in mild cognitive impairment has been estimated in a recent meta-analysis to be approximately 2.5%/year [87]. This may suggest that the hippocampal volumetric difference observed in T2DM might lead to an earlier conversion to Alzheimer’s disease by almost 2 years.

Furthermore, the rate of total brain atrophy in T2DM was significantly higher than in metabolically healthy individuals by 13.4%. This is consistent with cross-sectional results, and that cross-sectional volumetric differences may increase with diabetes duration according to meta-regression analysis. However, meta-regression results also revealed that the difference in total brain volume between those with T2DM and metabolically healthy individuals decreased with age (4.4% for every year above 60). This implies that a disease onset before age 60 years may have a greater impact on brain volumes. Moreover, similar consistent trends were observed for grey matter, white matter, and hippocampal volumes. Together, these may suggest that diabetes-related neurodegeneration occurs before onset of diabetes and slows down with increasing age. Hyperglycaemia is the main characteristic of T2DM pathology, and studies on metabolically healthy individuals and individuals with prediabetes have found association between higher blood glucose levels, smaller brain volumes and poorer cognitive functions [21,88,89]. Known mechanisms that may contribute to T2DM-related brain changes, such as hyperglycaemia, vascular disorders and insulin resistance, were likely present before clinical diagnosis [5]. An implication of these findings is that it will be important for future research to investigate brain changes leading to T2DM diagnosis, and to conduct more longitudinal studies following individuals with T2DM over longer periods of time to clarify these issues. From a health policy perspective, it may also suggest that more resources should be directed towards risk reduction interventions in those at risk before the disease develops, rather than mitigate the effects of T2DM when much of the damage has already taken place.

T2DM is a known risk factor of dementia, with increased risk of dementia by two-to-three fold [90]. The current study not only shows substantial brain volume differences that might result in earlier conversion to mild cognitive impairment and Alzheimer’s disease, but also indicates potential predictors and structural basis of cognitive deficit among individuals with T2DM. Previous studies showed evidence of cognitive decline in important functions that may affect self-caring ability of diabetes patients, such as executive function [91] and processing speed [89]. Our study has also found significant associations between brain volume of some subcortical structures and diabetes, even though studies on local volumes were fewer. Indeed, studies on specific cognitive deficits in T2DM are relatively few with inconsistent results, and even fewer studies on both brain volumes and cognition. Further studies with larger sample size are needed to understand how changes at local brain structures are related to cognitive functions of diabetes patients.

We conducted meta-regression analyses to investigate moderators of volumetric differences between metabolically healthy individuals and individuals with T2DM. In this study, meta-regression analyses to examine the moderating effects of age, sex ratio, diabetes duration, ratio of medication, fasting glucose and HbA1c were only possible for some brain structures. These showed that volumetric differences decreased with age, which suggests early pre-clinical occurrence of diabetes-related brain changes. However, an increase in volumetric differences was also observed with diabetes duration in grey matter, and a similar trend in white matter. A likely explanation may be that effects related to T2DM are attributable to pathogenic mechanisms but that they become obscured by the increasingly prevalent, and often related, effects of other risk factors for neurodegeneration including cardiovascular diseases, hyperlipidemia, obesity, and others. The cause of the brain differences reported in the present study may be a combination and interaction of pathogenic mechanisms of T2DM and genetic factors, environmental exposures, age, sex, comorbidities, and medication. Some anti-diabetes oral medications, such as dipeptidyl peptidase 4 inhibitors, metformin, thiazolidinediones and sulfonylurea have potential neuroprotective effects for individuals with diabetes, whereas insulin may be associated with increased risk of dementia [92]. Although lifestyle factors such as high cholesterol diet, smoking, etc. are known risk factors for cognitive decline common among people with T2DM, no conclusive evidence is available on whether cardiovascular risk factor management via lifestyle change, controlling blood pressure and cholesterol levels may reduce the risk of cognitive dysfunction in people with diabetes compared with those without diabetes [93]. While these factors are sometimes reported by studies included in our meta-analysis, usually only some of these factors were reported in one study; reports on medication were often unspecific. This limited the covariates we could control for in our meta-analysis, the types of meta-regression we could conduct and the number of studies that could be included.

The main strengths of this review were an extensive search of the literature using a wide range of search terms across multiple databases, and inclusion of both cross-sectional and longitudinal studies across many brain structures. The main limitation was the relatively small number of studies which could be included in meta-analyses of regional volumes and particularly longitudinal atrophy. This also limited the number of moderators that could be tested in meta-regressions.