Introduction

Type 2 diabetes mellitus (T2DM), a chronic metabolic disorder characterized by hyperglycemia due to insulin resistance, is growing in prevalence worldwide1. T2DM not only impairs metabolic-related bodily functions, but it is also associated with multidimensional cognitive deficits, such as deficits in execution, attention and memory2,3. Although cognitive decline may not be severe enough to affect the normal life quality of T2DM patients at an early stage2,4, compelling evidence suggests that T2DM can accelerate the aging process and increase the risk for dementia5,6. Recently, early detection and prevention have been considered the most promising strategies in co** with progressive dementia7; therefore, it is of great significance to identify the signs of early cognitive decline in T2DM patients before they develop dementia.

Neuroimaging methods can provide important insights into the neural mechanisms of cognitive impairment in patients with T2DM. For example, T2DM patients have been shown to exhibit distributed gray matter atrophy in the hippocampus, amygdala and prefrontal and parietal cortices and this atrophy might contribute to cognitive impairment8,9,10,11. T2DM is also frequently accompanied by brain vascular lesions, which have been associated with cognitive deficits4,12. Moreover, T2DM patients have been shown to exhibit alterations in regional spontaneous brain activity that are correlated with poor cognitive performance38,39. Furthermore, significant correlations were found between the activation amplitudes of fronto-parietal networks in T2DM patients and WM performance, suggesting that patients with stronger fronto-parietal activation possess better WM abilities. As a result, these correlation results further supported our hypothesis of the presence of compensatory mechanisms of brain activation during WM processing in T2DM patients.

Reduced functional connectivity in working memory networks

Our T2DM patients showed decreased intra-network FC in the bilateral lingual gyri of the visual network (IC 10), the vlPFC of the left fronto-parietal component (IC19) and the left IPL of the posterior parietal component (IC21). In addition, inter-network FC between the right fronto-parietal network (IC22) and visual network (IC10) were also decreased in T2DM patients.

Early studies reported that both the inferior frontal and the inferior parietal regions were involved in WM40. The vlPFC is closely related to both temporary information maintenance41,42,43 and comparison44 and is particularly critical for WM31,45. Recently, increased neural activity in the vlPFC was found in older adults compared with young adults during the performance of more complex tasks relative to a baseline; additionally, a higher BOLD signal in the vlPFC was shown to predict the efficiency of WM performance46. In diabetes, the vlPFC was considered to be vulnerably damaged19. IPL was frequently reported as playing a role in short-term memory storage which is an important content of 1-back working memory task47,48. In diabetes, decreased resting state FC in bilateral IPL was demonstratedhttp://www.fil.ion.ucl.ac.uk/spm/software/spm8). First, slice timing correction was performed to correct for the inter-slice time delay within each volume. Second, motion correction was estimated and corrected using rigid co-registration. Subjects who had head shift greater than 2.0 mm or rotation greater than 2.0° were excluded from the analyses. Third, images were spatially normalized into Montreal Neurological Institute (MNI) space using a standard EPI template provided by SPM8 and were resliced into a voxel-size of 3 mm × 3 mm × 3 mm. Finally, the data were spatially smoothed using a 6-mm full width at half-maximum Gaussian kernel.

Independent component analyses

Group spatial ICA was applied using GIFT software (http://icatb.sourceforge.net/, version 2.0d) to decompose the preprocessed fMRI images into spatial ICs as follows. First, 23 ICs were automatically estimated using the minimum description length criteria64. After that, a two-step principal component analysis was used to decompose the fMRI timecourses of the whole brain voxels into 23 principal components. This analysis was followed by group-level IC estimation using an Informax algorithm65. The most stable estimation of ICs was achieved by re-running the ICA analysis 100 times using the ICASSO method. Then, a spatial-temporal algorithm was used to back reconstruct the subject-level ICs from the group-level ICs. The spatial-temporal regression algorithm was introduced as following equations:

In which variable Y represents the group-level spatial components and variable X represents the preprocessed fMRI timecourses of a certain subject, TC represents the temporal components of this subjects and SC represents the spatial components of this subjects.

This step produced subject-level spatially IC maps (spatial components) as well as the featured time courses of these components (temporal components). Finally, the subject-specific spatial and temporal components were transformed into z-scores to create a normal distribution.

Component selection

We first used spatial correlation analysis to exclude ICs that were not contributed by gray matter. To accomplish this, we employed the following steps: first, maximum probability maps (MPMs) of gray matter, white matter and cerebral blood flow were generated using the tissue prior templates provided by SPM. A certain tissue class (gray matter, white matter, or cerebral blood flow) was defined for each voxel in the MPM based on the maximum probability among the three tissue prior templates. Then, the values for the spatial IC map and the MPM were rearranged to include one vector for each map and Pearson correlation analysis was performed to test the association between the vector of each spatial IC map and the vector of each MPM map. Finally, the highest spatial correlation coefficient of each IC was identified and the corresponding tissue class was retrieved. Following these steps, 15 ICs were identified as gray matter components. We further excluded 2 ICs that belong to the cerebellum. Then, the remaining 13 candidate components were further tested by a first-level canonical general linear model (GLM) to clarify whether the featured time course of a specific component was statistically related to the WM task. Specifically, the time course for each component of each subject was regressed against the design matrix of the WM task using GLM and canonical hemodynamic response function (HRF). The resulting β-estimate of each component represents the activation of this component in response to WM loading (1-back versus 0-back) in the subject. The β-estimates were further assessed using a second-level random effect one-sample t-test to clarify whether the mean β-estimate of each component was statistically significant (P < 0.05, Bonferroni correction). In total, eight of the 13 candidate components were found to be statistically associated with the WM task and were chosen for further analysis (Fig. 2).

WM-related intra-network and inter-network FC calculation

It should be noted that aforementioned subject-level spatial and temporal components were back-reconstructed using the full fMRI timecourses. Thus the components contained both 0-back and 1-back information. Because 0-back mainly reflects attention information but not the WM capacity, these subject-level components are not suitable for evaluate the WM-related FC changes in the T2DM. Thus, in order to obtain the spatial and temporal components that are specific for 1-back WM condition, we removed the timecourses that were abstained during the 0-back blocks from the raw fMRI data. We also corrected the HRF effect by delaying the task timing for 3 volumes (6 s). Then the remaining fMRI blocks were concatenate to form the timecourses that are specific for 1-back WM condition and they were considered as the variable X in the dual-regression function. Finally the spatial components and temporal components were scaled using Z-score.

The value of each voxel within a spatial component reflects the temporal coherence between the BOLD timecourses of each voxel and its temporal component. Thus, we termed the value of the spatial component as the intra-network functional connectivity. We also calculated the inter-network functional connectivity as the Pearson correlation between the temporal components of each IC pair, which were further Fisher r-to-z transformed to satisfy parametric statistics.

Statistical analyses

Activation analysis

Differences in global activation were tested by comparing the β-estimates of eight WM networks between the two groups using two-sample t-tests after controlling for the effects of age, gender and education (P < 0.05, Bonferroni correction). Finally, partial correlation coefficients were used to test the possible association between activation strength and clinical/cognitive variables after controlling for the effects of age, gender and education (P < 0.05).

Functional connectivity comparisons

A voxel-wise one-sample t-test was performed on the spatial components of the WM networks to recognize the spatial distribution pattern of the FC of each network. Multiple comparison corrections were made using family-wise error (FWE) correction (P < 0.05). The Cohen’s d effect size of each comparison was also calculated. Brain regions in each network showing statistically positive FC were binarized and used as mask for further intergroup comparisons. Intergroup differences in FC within eight networks were examined using two-sample t-tests after controlling for the effects of age, gender and education. Correction for multiple comparisons was performed using an AlphaSim algorithm, resulting in a corrected threshold of P < 0.05 at the cluster level (parameters: single voxel uncorrected P = 0.01, 1000 simulations, full width at half maximum = 6 mm, cluster connection radius r = 5 mm) and within the mask of the spatial distribution of each component. We further compared intergroup differences in FC between each pair of the eight WM networks using a two-sample t-test after controlling for the effects of age, gender and education (P < 0.05, uncorrected). The Cohen’s d effect size of each comparison was also calculated. Finally, partial correlation was used to assess possible associations between FC and clinical/cognitive variables after controlling for the effects of age, gender and education (P < 0.05).

Statistical analysis for demographic data

SPSS 21.0 (SPSS, Inc, Chiago.IL) was used to analyze demographic data. Shapiro-Wilk tests were performed to assess the distributions of demographic variables. Intergroup differences in demographic variables were tested either with Student’s t-test (normal distribution) or the Mann-Whitney U-test (non-normal distribution). Achi-squared (χ2) test was used to assess intergroup differences in gender. The significance level was set as P < 0.05.

Additional Information

How to cite this article: Zhang, Y. et al. Altered brain activation and functional connectivity in working memory related networks in patients with type 2 diabetes: An ICA-based analysis. Sci. Rep. 6, 23767; doi: 10.1038/srep23767 (2016).