Dysfunctional Architecture Underlies White Matter Hyperintensities with and without Cognitive Impairment

Abstract

Background:

White matter hyperintensities (WMH) are common in older adults and are associated with cognitive decline. However, little is known about the functional changes underlying cognitive decline in WMH subjects.

Objectives:

To investigate whole-brain functional connectivity (FC) underpinnings of cognitive decline in WMH subjects using univariate and multivariate analyses.

Methods:

Twenty-three WMH subjects with mild cognitive impairment (WMH-MCI), 43 WMH subjects with no cognitive impairment (WMH-nCI), and 55 healthy controls underwent resting-state functional MRI scans. Whole-brain FC was calculated using the fine-grained human Brainnetome Atlas, followed by performance of between-group comparisons and FC-cognition correlation analysis. A multivariate analysis using support vector machine (SVM) was performed to classify WMH-MCI and WMH-nCI subjects based on FC.

Results:

Both the WMH-MCI and WMH-nCI subjects exhibited characteristic impaired FC patterns. Markedly reduced FC involving subcortical nuclei and cortical hub regions of cognitive networks, especially the cingulate cortex, was identified in the WMH-MCI patients. In the WMH-MCI group, several connections involving the cingulate cortex were associated with cognitive decline. The exploratory mediation analyses indicated that FC alterations could partially explain the association between WMH and cognition. Furthermore, an SVM classifier based on FC distinguished WMH-MCI and WMH-nCI subjects with 78.8% accuracy. Connections that contributed most to the classification showed a similar distribution as the connections identified in the univariate analysis.

Conclusions:

This study provides a new window into the pathophysiology of cognitive impairment in WMH subjects and offer a novel and potential approach for early detection of the cognitive impairment in WMH subjects at the individual level.

Keywords

Functional connectivity mild cognitive impairment resting-state functional MRI support vector machine white matter hyperintensities

INTRODUCTION

White matter hyperintensities (WMH), characterized by high signals on fluid-attenuated inversion recovery (FLAIR) and T2-weighted MR images, are commonly observed in older adults and are considered one of the manifestations of cerebral small vessel disease [1, 2]. Convergent evidence suggests that WMH lead to cognitive decline and are associated with the risks of both vascular dementia and Alzheimer’s disease (AD)-related dementia [2 –4]. However, the mechanism underlying cognitive impairment in patients with WMH remains unclear. Previous studies have concentrated on structural abnormalities, especially white matter fiber disconnections and their relationships with cognitive decline in WMH subjects [5 –7], but little is known about the functional underpinnings. A few studies have suggested that WMH is correlated with remote cortical brain hypoperfusion [8], altered functional activity during specific tasks [9], and aberrant functional connectivity (FC) in the default mode network (DMN) [10]. Nevertheless, a comprehensive picture of the functional changes underlying cognitive decline in WMH has yet to be established.

Cognitive processes depend on information exchanges among brain regions [11, 12]. Reflecting the functional integration of the human brain, the whole-brain FC analysis has been proven to constitute a useful measurement for identifying the mechanisms of various cognitive disorders [13 –16] and may also provide new insights to elucidate the pathophysiology of cognitive impairment in subjects with WMH. Hence, to explore the functional underpinnings of cognitive decline in WMH, here we examined whole-brain FC patterns in 23 WMH patients with mild cognitive impairment (WMH-MCI), 43 WMH patients with no cognitive impairment (WMH-nCI), and 55 matched healthy controls (HCs) using the fine-grained Human Brainnetome Atlas [17]. In addition to comparisons of FC among the three groups, we explored associations between cognitive performance and FC in the WMH-MCI subjects and performed exploratory mediation analyses to evaluate whether FC changes can explain the relationship between WMH and cognition. Further, we hypothesized that FC in specific regions may constitute a biomarker of cognitive impairment in WMH patients. A support vector machine (SVM) classifier was trained to classify WMH-MCI and WMH-nCI subjects based on FC and important features contributing to the classification were identified (Fig. 1).

Fig.1

Schematic of the data analysis pipeline. Resting state fMRI data were preprocessed using the following steps: slice-timing, realignment, normalization, de-noising and filtering. The processed fMRI images were then segmented into 272 regions using the human Brainnetome Atlas, and Pearson’s correlation coefficients were calculated between each pair of the 272 regions and transformed into z-scores using Fisher’s r-to-z transformation, resulting in a 272×272 functional connectivity (FC) matrix for each subject. First, univariate analysis was applied to obtain the between-group differences in FC, and one-way ANOVA of the three groups and post hoc analysis of each pair among the three groups were employed. Next, correlation analysis was performed to evaluate the relationships between the altered connections and the z-scores of each cognitive domain and WMH load. Second, multivariate pattern analysis utilizing the SVM was conducted to classify the WMH-MCI and WMH-nCI groups based on FC using the leave-one-out-cross-validation (LOOCV) method. Classification performance was evaluated using both the area under the curve (AUC) and the classification accuracy metrics. The consensus features were used to identify the connections that contributed most to the classification.

MATERIALS AND METHODS

Participants and assessment of cognitive performance

The study was approved by the Ethics Committee of Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology. Written informed consent was obtained from each participant.

Between March 2014 and January 2016, we enrolled 80 consecutive right-handed patients with moderate to severe WMH at the outpatient service of the Department of Neurology, Tongji Hospital, Wuhan, China. Moderate to severe WMH was defined as a Fazekas rating score of three or higher (the sum of the deep WMH score and the periventricular WMH score) on FLAIR sequence images [18, 19]. In addition, 60 right-handed healthy controls without moderate to severe WMH and lacunes were recruited. The participants then underwent clinical interviews and a battery of neuropsychological tests, including the Mini-Mental State Examination (MMSE), the Clinical Dementia Rating (CDR), the Trail Making Test (TMT), the Symbol Digit Modalities Test (SDMT), the Digit Span Test (DST), the Verbal Fluency Test (VFT), the Auditory Verbal Learning Test (AVLT), and the Hamilton depression rating scale (24-items). For comparisons of performance across the tests, we calculated the z-score (e.g., (individual test scores - mean scores of HCs)/standard deviation of HCs) of each test [20]. In particular, as higher TMT scores indicate worse cognitive performance, we multiplied the individual z-scores for these tests by –1 [20]. To represent the cognitive abilities more robustly and to reduce random errors, the neuropsychological tests were composited into three cognitive domains [21]: processing speed (TMT-A and SDMT) [22], executive function (DST backward, VFT, and TMT-B) [23], and memory (AVLT immediate recall, short delay recall, long delay recall, and long delay recognition) [24]. The performance of the subjects on tests encompassing each cognitive domain was presented as the compound z-score calculated by averaging z-scores for the corresponding tests [7, 20].

The inclusion criteria for WMH patients were: 1) moderate to severe WMH with/without lacunes; 2) no dementia based on the criteria of the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV); a global clinical dementia rating (CDR) score ≤0.5; a MMSE score ≥20 (primary school) or ≥24 (junior school or above); and the ability to perform all activities of daily living (ADLs) independently [25]. The exclusion criteria were: 1) left-handedness; 2) <5 years of education; 3) the presence of any infarct with a diameter >20 mm on T1-weighted images, cortical infarcts, or cerebral hemorrhage; 4) WMH mimics (e.g., multiple sclerosis, hypoxic-ischemic encephalopathy, and leukodystrophy); 5) a history of Parkinson’s disease, epilepsy, or psychiatric diseases; 6) use of medications that may affect cognitive function; 7) insufficient cooperation with the study procedures for any reason; and 8) a Hamilton depression scale score (24 items) ≥20. The HCs should have no history of any neuropsychiatric disorders, no cognitive complaints and did not meet the exclusion criteria described above. Additionally, HCs were required to achieve an MMSE score ≥24 (primary school) or ≥27 (junior school or above) and score 0 on all CDR domains. Sixty-eight WMH subjects and 56 HCs with completed neuropsychological tests and MR scans fulfilled the criteria above. Additionally, two WMH participants with excessive head motion (greater than 3 mm translation or 3° angular rotation on any axis) during scanning and one HC without a complete set of MRI data due to technical problems were excluded.

Furthermore, we divided the WMH patients into two groups based on cognitive ability: 1) the WMH-MCI group, which included subjects with score ≥0.5 for at least one of the CDR domains, and objective evidence that one or more cognitive domains (processing speed, executive function, or memory) were impaired (age- and education-adjusted z-scores at least 1.5 SD below those of the HCs on one or more cognitive domains) [21, 25]; and 2) the WMH-nCI group, which included subjects with score = 0 on all six CDR domains, an MMSE score ≥24 (primary school) or ≥27 (junior school or above) and no objective evidence of impairment in any of the three cognitive domains. No subjects were excluded in this procedure. Finally, a total of 121 participants (23 WMH-MCI, 43 WMH-nCI, and 55 HC subjects) were included in this study.

MRI data acquisition

Data were collected on a 3.0T MR scanner (Discovery MR750, GE Healthcare, Milwaukee, WI, USA) using a 32-channel head array coil at Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology. The neuroimaging protocol included whole-brain T1-weighted, T2-weighted, fluid-attenuated inversion recovery (FLAIR), BRAin Volume (BRAVO), and resting-state fMRI sequences. FLAIR images were acquired using a repetition time (TR) = 8,000 ms, echo time (TE) = 160 ms, inversion time (TI) = 2,100 ms, flip angle (FA) = 111°, slice thickness = 5.0 mm, slice gap = 1.5 mm, data matrix = 512×512, and field of view (FOV) = 240×240 mm². High-resolution anatomical T1-weighted images were obtained using a sagittal BRAVO sequence with a TR = 8.16 ms, TE = 3.18 ms, TI = 50 ms, FA = 12°, number of slices = 188, slice thickness = 1.0 mm, data matrix = 256×256, and FOV = 256×256 mm². Resting-state fMRI data were acquired using an axial gradient echo planar imaging (EPI) sequence with a TR = 2,000 ms, TE = 35 ms, FA = 90°, number of slices = 36, slice thickness = 3.0 mm, slice gap = 1.0 mm, data matrix = 64×64, FOV = 220×220 mm², and scan time = 7 min. Subjects were asked to relax and keep their eyes closed during the scan. Foam pads and earplugs were used to minimize head movement and scanner noise.

Data processing

Structural MRI data

Structural MRI images were processed using the standard steps in the VBM8 toolbox (http://dbm.neuro.uni-jena.de/vbm/). The images were bias-corrected, segmented into gray matter, white matter, and cerebrospinal fluid, and registered in the Montreal Neurological Institute (MNI) space using sequential linear (affine) and non-linear transformations (warping). Finally, the modulated maps of gray matter volumes were smoothed with a 6-mm full-width at half-maximum Gaussian kernel for the final statistical analysis.

For the WMH volume, two independent raters (H.H. and Q.M.), blinded to the clinical information, manually segmented WMH from FLAIR images using ITK-SNAP 3.6.0 (http://www.itksnap.org/pmwiki/) and calculated the volume as the sum of the WMH of all the image layers. The correlation coefficient of the results from the two raters was 0.98 (p < 0.0001) and the inter-rater coefficient was 0.79 (p < 0.0001). The final WMH volume was established according to the overlapping results of the two raters. Furthermore, we segmented the total WMH into two subtypes: periventricular WMH (PWMH, defined as the WMH regions within 10 mm of the ventricular space) and deep WMH (DWMH, defined as the remaining WMH regions beyond the PWMH) [26], and then we calculated the volumes of PWMH and DWMH for each individual, respectively. Additionally, allowing for considerable differences in the number of lacunes among WMH patients, we defined a grading system for lacunes (grades 1 to 3) for further analysis, as reported in a previous study [27].

Functional MRI data

Resting-state fMRI data were processed using the Brainnetome fMRI toolkit (BRANT, http://brant.brainnetome.org) [28]. The first 10 volumes of each run were discarded due to the magnetization equilibration effect, and the remaining images were corrected for the acquisition time delay between different slices. Next, the images were realigned to the first volume for head motion correction. The images were then normalized to the MNI space and resampled to 2×2×2 mm³ voxels. Very small sub-millimeter head motion during data scanning has been shown to exert a substantial effect on some measurements of resting-state fMRI [29 –31]. In this context, it is notable that subjects with greater than 3 mm translation or 3° angular rotation on any axis when compared to the first recorded time point were excluded for further analysis. To further reduce the head motion effect, a linear regression analysis was then applied to remove nuisance variables, including head motion at the x, y, z axes, linear shift, and the mean signals from cerebrospinal fluid and white matter. Afterwards, a temporal band-pass filter (0.01–0.08 Hz) was used to eliminate the influence of low-frequency drifts and high-frequency noise.

Brain parcellation and functional connectivity

To characterize the whole-brain FC architecture of each subject, the human Brainnetome Atlas [17] was employed. The atlas was derived from a parcellation approach based on connectivity information and has been used in brain network studies [32, 33]. The atlas consists of 246 subregions in the cerebrum (210 cortical and 36 subcortical regions) and 28 subregions in the cerebellum (http://atlas.brainnetome.org). We excluded two regions in the cerebellum in which no fMRI signal was recorded in some participants in the study. The average time series of the 272 selected regions were extracted. We then calculated Pearson’s correlation coefficients between each pair of the 272 regions and applied Fisher’s r-to-z transformation to improve the normality of the coefficients, resulting in a 272×272 functional connectivity matrix for the subsequent analyses.

Statistical analyses

Demographic, neuropsychological, and structural MRI data

One-way analysis of variance (ANOVA), the Kruskal-Wallis test, and the χ² test were applied to evaluate differences in demographic, neuropsychological, and conventional structural MR metrics using SPSS 20.0 (IBM Corp., Armonk, NY, USA). Partial correlation analyses were implemented to test the associations between the composite z-scores of each cognitive domain and conventional structural MRI metrics within the WMH subjects, controlling for age, gender and years of education. The relationships among the conventional structure MRI metrics in the WMH subjects were determined using Spearman’s correlation.

Functional connectivity analyses and mediation analyses

The statistical analysis of FC was performed using an in-house MATLAB script. FC was recalculated in the generalized linear model with age, gender, years of education and total brain parenchymal volume (TBV) as covariates. We identified significant differences in FC according to the following two criteria: 1) the adjusted FC was significantly different from zero within each group at p < 0.0001 (Bonferroni-corrected); and 2) the sets of connections were significantly different from zero in at least one group. One-way ANOVA and post hoc analysis were employed to identify differences in FC among the three groups (p < 0.05, FDR corrected). Additionally, to reduce the potential effects of head motion and to validate the reproducibility of our main result, we also repeated the analyses by adding mean motion [31] as a covariate.

Pearson’s correlation analysis was performed to evaluate the relationships between all identified connections and the z-scores of each cognitive domain in the WMH-MCI group (p < 0.05). Furthermore, we performed exploratory analyses to illustrate whether the relationship between WMH load and early cognitive decline can be explained by FC changes in the WMH subjects. As reported in previous studies [34 –36], the mediation models were constructed with the mean z-score of the connections that were significantly altered between the WMH-MCI and WMH-nCI groups as a mediator, and cognitive z-scores (processing speed, executive function and memory, respectively) as outcomes. To evaluate the general and regional effect of WMH, here we entered total WMH volume, PWMH volume, and DWMH volume into the models respectively as predictors, since recently researches have argued that mediation hypothesis may be supported whatever a significant total effect exists [37, 38]. Mediation analyses were estimated using a bootstrap method (n = 5,000), and significant indirect effects were defined by a 95% confidence interval (CI) entirely above or below zero.

Group classification

We applied a non-linear support vector machine (SVM) algorithm with a radial basis function (RBF) kernel to classify WMH-MCI and WMH-nCI subjects using the Libsvm (http://www.csie.ntu.edu.tw/~cjlin/libsvm/) [39]. In the SVM, each subject was treated as a point with a label (e.g., with or without MCI) in a high-dimensional space (dimension = the number of selected features). The aim of the SVM is to identify a hyperplane or decision boundary with the maximal margin to separate subjects into different classes according to their labels, allowing for some misclassification. The SVM has two phases: training and testing. During the training phase, the training samples were used to identify a decision boundary that would be used to predict the class label of a new testing sample in the testing phase. In the present study, as the numbers of subjects in the two groups were imbalanced, we introduced an up-sampling method (for 1,000 permutations) to generate extra data based on the original data from the minority class [40]. Through spherical linear interpolation, we can randomly generate 19–21 new samples based on the 23 original WMH-MCI samples, resulting in 42–44 samples in the WMH-MCI group. In each time, the leave-one-out-cross-validation (LOOCV) method was applied to evaluate the classification accuracy of the SVM with the FC (z-score) as the input features. In consideration of relatively small samples with high dimension modeling in the present study, dimensionality reduction (feature selection) was conducted to accelerate computation. Briefly, in each fold of up-sampling iterations, a two-sample two-tailed t test was performed in each FC between the WMH-MCI and WMH-nCI subjects in the training dataset, and the FCs with top rank up to 150 t-scores were selected as the input features for training models [41 –43]. Next, we performed a grid search to choose the best C and best g for the classification model in the training sample, and the selected model was used to classify the testing sample [44]. The classification performance is evaluated by calculating the mean of accuracy, sensitivity and specificity for the 1,000 iterations (Supplementary Figure 3). These parameters are defined as follows: $accuracy = \frac{TP + TN}{TP + TN + FP + FN} \times 100 %$ $sensitivity = \frac{TP}{TP + FN} \times 100 %$ $specificity = \frac{TN}{TN + FP} \times 100 %$ where TP, TN, FP, and FN denote the true positive, true negative, false positive, and false negative, respectively. In the present article, we defined the WMH-MCI samples as the “positive” class and the WMH-nCI samples as the “negative” class. Furthermore, to visualize the performance, receiver operator characteristic (ROC) curves were plotted and the area under ROC curves (AUC) was also calculated.

As the training sample was different between iterations, the selected features differed in each iteration. However, some features were retained in most of the iterations, which were termed “consensus features” and were considered to contribute most to the classification [45]. Here we identified the consensus features as the features retained in all 1,000 up-sampling iterations. In addition, to further verify the stability of the classifier, we also used the 10-folds method instead of the LOOCV to evaluate the performance of classification.

Results

Demographic, structural MRI, and neuropsychological results

No significant differences in age, gender, and years of education were identified among the groups (p > 0.05). Both the WMH-MCI and WMH-nCI groups showed significantly higher Fazekas scores, total WMH volumes, PWMH volumes, DWMH volumes, and lacune grades compared with the HCs (p < 0.001). Although these metrics were all higher in the WMH-MCI group than in the WMH-nCI group, the differences were not significant (p > 0.05). TBV was lower in the WMH-MCI group than in the WMH-nCI group (p = 0.043). All three cognitive domains were affected in the WMH-MCI group compared to the other two groups (p < 0.001), whereas no differences were identified in cognitive performance between the WMH-nCI and HC groups (p > 0.05) (Table 1). In the WMH subjects, the Fazekas score, total WMH volume, PWMH volume, lacune grade, and TBV were all significantly correlated with processing speed and executive function (p < 0.05) but not with memory (p > 0.05) (Supplementary Table 2). WMH metrics (PWMH, DWMH, and total WMH volume) and lacune grade were all highly correlated with each other, while no correlations were detected between TBV and other structural metrics (Supplementary Table 3). No significant differences in mean head motion were observed among the three groups (p = 0.131) (Supplementary Table 1).

Table 1

Demographic, structural MRI, and cognitive features of all participants

	HC (n = 55)	WMH-nCI (n = 43)	WMH-MCI (n = 23)	Overall p
Demographic data
Age (y)	63.29±6.50	65.72±5.94	65.17±6.65	0.089^a
Gender (M/F)	32/23	27/16	15/8	0.813^b
Education (y)	10.75±3.29	10.28±3.60	9.34±3.89	0.251^a
Structural MRI feature
TBV (cm³)	1122.13±87.92	1143.30±88.35	1085.74±97.54^f	0.050^c
Fazekas WMH score	1.00±0.79	4.47±1.18^d	5.13±1.10^e	<0.001^a
Total WMH volume (cm³)	0.04±0.09	13.53±13.97^d	18.19±15.02^e	<0.001^a
PWMH volume (cm³)	0.02±0.07	6.76±7.62^d	10.96±8.23^e	<0.001^a
DWMH volume (cm³)	0.02±0.06	6.77±8.85^d	7.22±7.80^e	<0.001^a
Lacune grade	1.00±0.00	1.44±0.63^d	1.78±0.74^e	<0.001^a
Cognitive performance
MMSE	28.93±1.30	28.77±1.23	25.30±2.72^e,f	<0.001^a
Global CDR score	0	0	0.48±0.10^e,f	<0.001^a
Processing speed^g	0.00±0.91	0.04±0.82	–1.53±0.95^e,f	<0.001^c
Executive function^g	0.00±0.72	–0.26±0.57	–1.52±0.68^e,f	<0.001^c
Memory^g	0.00±0.85	–0.25±0.87	–1.71±0.63^e,f	<0.001^c

^aKruskal-Wallis test. ^bχ² test. ^cOne-way analysis of variance. ^dSignificant difference between the WMH-nCI and HC groups. ^eSignificant difference between the WMH-MCI and HC groups. ^fSignificant difference between the WMH-MCI and WMH-nCI groups. ^gZ-scores. The performance of processing speed, executive function and memory was presented as the compound z-scores for the related tests with scores of HCs as reference. TBV, total brain parenchymal volume; MMSE, Mini-Mental State Examination; CDR, Clinical Dementia Rating.

Univariate group analysis of whole-brain FC

A total of 27 altered connections were identified among the three groups (p < 0.05, FDR-corrected), and the terminal regions mainly included the cingulate cortex, subcortical nuclei (the thalamus and basal ganglia), and the fronto-insular cortex (Fig. 2A). The WMH-MCI patients clearly exhibited reductions in the strength of most connections compared to the other two groups, while the differences in FC values between the WMH-nCI and HC subjects varied across connections (Fig. 2B). Compared with the HCs, the WMH-nCI subjects exhibited decreased FC values mainly between several cortical regions and increased FC values involving regions of the parahippocampal gyrus and anterior cingulate cortex (ACC). Relative to the WMH-nCI subjects, the WMH-MCI patients showed more extensive and severe FC disturbances characterized by reduced connections involving the cingulate cortex and subcortical nuclei (especially connections within the cingulate cortex and cingulo-thalamic connections), several regions of the fronto-insular cortex, and the cerebellum (see Fig. 2C-E and Supplementary Table 4). Additionally, to further evaluate the head motion effect, we also performed the same analysis with the head motion metric as an additional control condition. Here we achieved a very similar results between the z-map obtained with and without head motion control (see Supplementary Figure 1). Hence, we considered that head motion is unlikely to provide a plausible explanation for the pattern of results we have reported.

Fig.2

Comparisons of functional connectivity among the three groups. A) Circle diagram illustrating the total differential connections among the three groups. The brain regions are shown in the circles and the connections are indicated by lines. B) Radar map showing that the absolute z-score values of the connections that differed across the three groups. Lines in different colors represent different groups, and most connections in the WMH-MCI group were lower than those in the other two groups. The altered connections between each pair of groups are separately shown in circle diagrams C-E). Solid lines indicate weakened connections in WMH-nCI subjects relative to those in HCs (C), in WMH-MCI subjects relative to those in WMH-nCI subjects (D), and in WMH-MCI subjects relative to those in HCs (E). Dotted lines indicate enhanced connections in WMH-nCI subjects relative to those in HCs (C) and in WMH-MCI subjects relative to those in HCs (E). BG, basal ganglia; CG, cingulate gyrus; Hipp, hippocampus; IFG, inferior frontal gyrus; INS, insular gyrus; ITG, inferior temporal gyrus; MFG, middle frontal gyrus; PhG, parahippocampal gyrus; PCL, paracentral lobule; PoG, postcentral gyrus; pSTS, posterior superior temporal sulcus; SFG, superior frontal gyrus; STG, superior temporal gyrus; Tha, thalamus; L, left; R, right. Detailed information about the brain regions can be accessed at http://atlas.brainnetome.org/.

Associations between FC, cognitive abilities, and WMH load

In the WMH-MCI group, several connections involving the cingulate cortex were found to be associated with cognitive decline, including connections between posterior cingulate cortex (PCC) and fronto-insular cortex with processing speed (r = 0.52, p = 0.011) and executive function (r = 0.48, p = 0.022), connections between the ACC and PCC with processing speed (r = 0.50, p = 0.015) and executive function (r = 0.50, p = 0.014), and connections between the ACC and dorso-medial thalamus with memory (r = 0.42, p = 0.049) (Fig. 3).

Fig.3

Significant correlation between altered functional connectivity and cognitive performance in WMH-MCI patients. Detailed information about the brain regions is shown as follows: CG_L_7_3, left pregenual area 32; CG_L_7_6, left caudal area 23; CG_R_7_5, right caudal area 24; IFG_R_6_5, right opercular area 44; Tha_R_8_1, right medial pre-frontal thalamus. CG, cingulate gyrus; IFG, inferior frontal gyrus; Tha, thalamus. L, left; R, right. Detailed information about the brain regions can be accessed at http://atlas.brainnetome.org/.

The exploratory mediation analyses revealed an overall correlation between the WMH load and the FC variable (the mean z-score of the altered FC clusters between the WMH-MCI and WMH-nCI groups) (β= 0.344, p = 0.024). However, the association was mainly observed for the relationship between PWMH and the FC variable, whereas DWMH increase was not significant correlated with FC changes. In all models, the FC variable showed significant correlations with cognition independent of the WMH variables (p < 0.05/3, since a compound FC variable was entered into the mediation models as the mediator and 3 separate analyses were performed for the 3 cognition variables). Significant mediation effects of PWMH on all the three cognition domains were observed via the FC variable, yet the FC variable was not a significant contributor to the DWMH-cognition relationship. In particular, the FC variable fully mediated the relationship between PWMH and memory. We still observed significant mediation effects in the model of the effect of the FC variable on the association between total WMH and cognition (Table 2).

Table 2

Mediation effect of functional connectivity on the relationship between WMH and cognition

	Total WMH volume		PWMH volume		DWMH volume
	β	p or 95% CI	β	p or 95% CI	β	p or 95% CI
Path a for all models	–0.286	0.024	–0.382	0.002	–0.123	0.338
Mediation effect of FC on the association between WMH and processing speed
Path b	0.328	0.001	0.282	0.006	0.382	<0.001
Total effect (path c)	–0.341	0.001	–0.412	<0.001	–0.187	0.087
Direct effect (path c^′)	–0.247	0.014	–0.305	0.003	–0.140	0.156
Mediation effect (path a*b)	–0.094	(–0.216, –0.010)	–0.108	(–0.237, –0.011)	–0.047	(–0.160, 0.033)
Mediation effect of FC on the association between WMH and executive function
Path b	0.403	<0.001	0.362	<0.001	0.447	<0.001
Total effect (path c)	–0.316	0.004	–0.394	<0.001	–0.164	0.146
Direct effect (path c^′)	–0.201	0.044	–0.256	0.013	–0.108	0.267
Mediation effect (path a*b)	–0.115	(–0.235, –0.018)	–0.138	(–0.259, –0.037)	–0.055	(–0.177, 0.043)
Mediation effect of FC on the association between WMH and memory
Path b	0.436	<0.001	0.405	<0.001	0.447	<0.001
Total effect (path c)	–0.140	0.239	–0.246	0.035	–0.003	0.981
Direct effect (path c^′)	–0.015	0.893	–0.092	0.421	0.052	0.623
Mediation effect (path a*b)	–0.125	(–0.256, –0.022)	–0.155	(–0.300, –0.049)	–0.055	(–0.180, 0.046)

Mediation analysis was performed to examine the potential indirect relationship between WMH variables (X) and cognition (Y) via the overall FC aberrations (presented as the mean z-score of the altered FC clusters identified between the WMH-MCI and WMH-nCI groups in the whole-brain FC analysis) (M). In each model, age, sex, years of education and TBV were entered as covariates. Path a and path b indicate the association between WMH variables and the FC variable, and the associations between the FC variable and cognition (when both WMH and FC variables were entered into the model as predicting variables), respectively. Path c represents the total effect of WMH variables on cognition, and path c^′ shows the direct effect of WMH variables on cognition after controlling for the FC variable as a mediating factor. The mediating role of the overall FC aberrations on the association between WMH and cognition is defined by the 95% bootstrap CI for 5,000 bootstrapping times. Bold values represent a significance path (p < 0.05), or a 95% CI that does not include 0 for the indirect effect. β, Standardized regression coefficient; CI, confidence interval; FC, functional connectivity; X, predictor variable; Y, outcome variable; M, mediator.

Multivariate analysis of whole-brain FC

We applied the SVM to classify the two WMH groups based on FC and evaluated the performance of the SVM with different numbers of features (50–150). We ultimately retained 78 features for further analysis with which the SVM achieved the highest mean accuracy of 78.8% (95% CI = 78.6%–79.1%) with the mean sensitivity of 77.6% (95% CI = 77.4%–77.9%), the mean specificity of 80.3% (95% CI = 80.0%–80.6%), and an AUC = 0.881, indicating high diagnostic power (Fig. 4A). The 26 identified consensus connections included 15 connections with subcortical nuclei (58%) and 8 connections involving the cingulate cortex (31%). Although these consensus connections were not precisely the same as the altered connections between the WMH-MCI and WMH-nCI groups, the distribution of the consensus connections also highlighted the importance of connections involving subcortical nuclei and the cingulate cortex (see Fig. 4B and Supplementary Table 5).

Fig.4

The ROC curve, the consensus connections and their terminal regions in the SVM classification. A) The ROC curve of functional connectivity for distinguishing individuals with WMH-MCI from WMH-nCI individuals, with an area under ROC curve (AUC) of 0.881. B) Graph displaying the consensus connections. The connections are indicated by lines and the brain regions are shown in the block diagrams. The red lines represent the full overlap of altered connections between the WMH-MCI and WMH-nCI groups and consensus connections identified in the SVM classification analysis. ROC, receiver operating characteristic; BG, basal ganglia; CG, cingulate gyrus; IPL, inferior parietal lobule; LOcC, lateral occipital cortex; MFG, middle frontal gyrus; MTG, middle temporal gyrus; MVOcC, medioventral occipital cortex; OrG, orbital gyrus; Pcun, precuneus; PoG, postcentral gyrus; SPL, superior parietal lobule; STG, superior temporal gyrus; Tha, thalamus; L, left; R, right.

Additionally, when we used the 10-folds method instead of the LOOCV, the SVM achieved the highest mean accuracy of 78.3% (95% CI = 78.1%–78.6%) with the mean sensitivity of 75.1% (95% CI = 74.9%–75.4%), the mean specificity of 82.7% (95% CI = 82.4%–83.0%), and an AUC = 0.868, indicating good stability of the classifier (see Supplementary Figure 2).

DISCUSSION

The present study for the first time systemically investigated whole-brain characteristic impaired FC patterns in both WMH-nCI and WMH-MCI subjects. Importantly, compared to the WMH-nCI and HC subjects, the WMH-MCI patients exhibited selective FC breakdown involving the cingulate cortex and subcortical nuclei. In the WMH-MCI group, several connections involving the cingulate cortex were also found to be correlated with cognitive impairment. Furthermore, the multivariate analysis revealed that the SVM was able to distinguish WMH-MCI patients from WMH-nCI subjects with a high level of accuracy based on FC. These findings provide compelling evidence that FC aberrations between specific regions of cognitive networks underlie cognitive decline in WMH and indicate that resting-state FC may serve as a novel neuroimaging marker for the early identification of cognitive impairment in WMH subjects.

Previous studies have suggested that the WMH load of some projection fibers connecting subcortical nuclei to cortical regions are related to cognitive performance in cerebral small vessel disease [5, 46]. In addition, WMH is a subcortical manifestation located not only in the white matter but also within subcortical grey matter nuclei [47]. Here we clearly revealed aberrant FC involving subcortical nuclei, especially subcortical-cortical disconnections in the both two WMH groups, providing strong evidence for structural abnormalities. Moreover, compared to WMH-nCI subjects, WMH-MCI patients showed more severe and extensive subcortical-cortical disconnections, including reduced connections between the thalamus or basal ganglia and several cortical regions involved in multiple cognitive processes [48, 49]. Both the thalamus and basal ganglia have been demonstrated to serve as network hubs and to constitute a core circuit supporting the integration of diverse networks linked to cognition [50, 51]. Emerging studies have also reported aberrant functional thalamic-cortical connections in MCI [52] and subcortical vascular cognitive impairment [53]. Consistent with these findings, our results suggested that a determination of the widespread functional involvement of subcortical nuclei, especially specific impaired subcortical-cortical connections, is essential for understanding the clinical-pathological correlations in cognitive impairment among WMH subjects.

Importantly, the present study identified an intriguing characteristic impaired FC pattern of WMH-MCI patients in a cluster located in the cingulate cortex, including connections between the PCC and ACC and between the cingulate cortex and several crucial regions (the fronto-insular cortex and the dorsomedial thalamus) of the salience network (SN). The human cingulate cortex, which is known to be a cortical hub [54, 55] and a core region of the DMN (both the PCC and ACC) [48 , 57], is recognized for its roles in a wide range of cognitive processes, such as processing speed, executive function and memory [58 –60]. Our findings are consistent with previous fMRI studies in subcortical mild vascular cognitive impairment [61, 62] and even extend several of those findings, emphasizing the key role of disconnections within/across the DMN and SN in cognitive decline among WMH subjects. Furthermore, in the WMH-MCI group, several cingulate-related connections among crucial regions of the DMN (the PCC and ACC) and the SN (the fronto-insular cortex, dorsal ACC, and dorsomedial thalamus) were associated with cognitive decline. Previous studies have demonstrated that several neuropsychiatric diseases displayed significantly but distinctively hub-concentrated lesion distributions [63, 64]. Our results added new evidence supporting this theory and showed a characteristic hub region damage pattern (selectively breakdown of hub regions in cognitive networks, especially the DMN and SN) of early cognitive decline in WMH subjects. Altogether, one could speculate that the dysfunction of hub regions in cognitive networks, especially the DMN and SN, may be a potential mechanism mediating the overall progression of normal cognition to MCI and subsequent dementia in subjects with high WMH load; however, this hypothesis requires confirmation in future longitudinal studies.

Compared with the HCs, both the WMH-MCI and WMH-nCI subjects exhibited altered FC in the insular, cerebellar, and sensorimotor regions, consistent with the clinical features of mood and motor disturbances in WMH subjects [65]. Interestingly, in the WMH-nCI group, we also observed some altered connections involved in cognitive networks. Most of these connections were found to be weakened in the WMH-nCI group, whereas several connections in the DMN (the ACC and parahippocampal gyrus) were increased. These findings are supported by a previous task-fMRI study that older individuals with normal cognition and high WMH load showed significantly reduced activity in the dorsolateral prefrontal cortex and greater activity in the ACC during cognitive control tasks, indicating that WMH may disturb the relationships between the DMN and other brain networks [66]. Since few studies have focused on functional underpinnings in WMH-nCI subjects [66, 67], further investigations should be implemented to evaluate the reorganization of different networks to maintain normal cognition in WMH subjects.

In line with prior investigations, our results showed that the total WMH load was negatively correlated with cognitive performance, especially processing speed and executive function [68]. When further taking WMH distribution into account, cognitive decline exhibited more closely association with PWMH than DWMH, as reported in previous studies, suggesting that these two categories of WMH may have different origins [69 –71]. Additionally, a significant correlation between PWMH and memory in the WMH subjects was also observed in the present study, supported by other previous researches [72 –74]. Inspired by previous studies [6, 75 –77], we further performed exploratory mediation analyses to evaluate whether the well-known relationships between conventional markers and clinical variables (WMH and cognition in the present study) was mediated by potentially novel factors (FC in the present study). An association between alterations in the whole-brain FC and increase in the WMH load (particularly PWMH) was observed, revealing a mediating role of the altered FC on the relationship between WMH load and cognitive decline in WMH subjects. Converging evidence has demonstrated that FC is associated with the integrity of white matter fibers [78, 79], and both the extent of the WMH load and disruption of specific white matter fibers, for instance, the anterior thalamic radiation [46] and cingulum bundle [80], are related to cognitive performance in WMH subjects. Consistent with and expanding on these previous findings, our results indicated that cognitive decline in WMH subjects could be attributed to not only structural but also functional disconnections independently and the characteristic changes in FC may provide additional evidence to improve our understanding of the pathogenesis of cognitive impairment in WMH.

The SVM algorithm is frequently employed for medical classification analyses since it is easy to use and understand. The current study is the first to conduct a multivariate analysis using SVM to evaluate whether FC is useful for discriminating WMH-MCI from WMH-nCI subjects. SVMs have been widely applied in the field of neuroimaging because of their ability to manage high-dimensional pattern classification problems and achieve high classification accuracy [81]. The high accuracy of the SVM classifier suggested that the clinical practicality of FC in identifying early cognitive dysfunction in WMH subjects. Furthermore, the consensus features showed a distribution similar to the connections identified in the univariate analysis, further supporting the importance of connections with the cingulate cortex and subcortical nuclei in maintaining cognitive function in WMH subjects.

Despite these promising findings, some limitations should be addressed. First, the sample size was relatively small, especially for the WMH-MCI group, which may explain why only a few altered connections were strongly correlated with cognitive performance in the WMH-MCI group. Second, since the WMH is usually accompanied with other abnormalities (such as lacunes), it is impossible to eliminate these concomitant structural changes of WMH. Hence, the FC alterations observed in our study cannot be solely attributed to WMH, although the lacunes were much less severe than WMH in the current study. Third, the SVM algorithm requires several user decisions, including the choice of the number of features and several key parameters (σ and γ in the present study) that must be set correctly to achieve the best classification results [39]. Fourth, we admitted that the up-sampling strategy may alter the distribution of data within the class and may create intrinsically similar samples that simplify the classification problem [82]; here we applied 1,000-times random permutation analyses to reduce this effect. In addition, to confirm our novel findings of characteristic functional disconnections underlying cognitive impairment in WMH subjects, future longitudinal studies incorporating samples across multiple centers are warranted.

In conclusion, the present study revealed whole-brain FC underpinnings of cognitive decline in WMH subjects and identified specific connections that offer the potential to serve as neuroimaging markers for cognitive impairment in these subjects. The findings provide insights for a better understanding of cognitive impairment in subjects with obvious WMH on MRI scans.

Footnotes

ACKNOWLEDGMENTS

We are grateful to Yafeng Zhan and Kaibin Xu for technical assistance, Prof. Jian Kong for helpful comments regarding the data analysis, and Prof. Min Zhang and Prof. Daishi Tian for recruiting the participants.

This work was supported by National Key Research and Development Program of China (Grant Nos. 2016YFC1300500 and 2016YFC1305904), the National Natural Science Foundation of China (Grant Nos. 91332108, 61327902-6, 81801146 and 81571062), NIH Grants (Nos.1R01NS091604 and P50MH106435), Beijing Municipal Science & Technology Commission (No. Z161100002616009), and the Natural Science Foundation of Hubei Province (No. 2017CFB392).

Authors’ disclosures available online ().

The supplementary material is available in the electronic version of this article: .

References

Wardlaw

, Smith

, Biessels

, Cordonnier

, Fazekas

, Frayne

, Lindley

, O’Brien

, Barkhof

, Benavente

, Black

, Brayne

, Breteler

, Chabriat

, Decarli

, de Leeuw

, Doubal

, Duering

, Fox

, Greenberg

, Hachinski

, Kilimann

, Mok

, Oostenbrugge

, Pantoni

, Speck

, Stephan

, Teipel

, Viswanathan

, Werring

, Chen

, Smith

, van Buchem

, Norrving

, Gorelick

, Dichgans

, STandards for ReportIng Vascular changes on nEuroimaging (STRIVE v1) (2013) Neuroimaging standards for research into small vessel disease and its contribution to ageing and neurodegeneration. Lancet Neurol 12, 822–838.

Prins

, Scheltens

(2015) White matter hyperintensities, cognitive impairment and dementia: An update. Nat Rev Neurol 11, 157–165.

Taylor

, Kambeitz-Ilankovic

, Gesierich

, Simon-Vermot

, Franzmeier

, Araque Caballero

, Muller

, Hesheng

, Ertl-Wagner

, Burger

, Weiner

, Dichgans

, Duering

, Ewers

, Alzheimer’s Disease Neuroimaging Initiative (2017) Tract-specific white matter hyperintensities disrupt neural network function in Alzheimer’s disease. Alzheimers Dement 13, 225–235.

Debette

, Markus

(2010) The clinical importance of white matter hyperintensities on brain magnetic resonance imaging: Systematic review and meta-analysis. BMJ 341, c3666.

Duering

, Gesierich

, Seiler

, Pirpamer

, Gonik

, Hofer

, Jouvent

, Duchesnay

, Chabriat

, Ropele

, Schmidt

, Dichgans

(2014) Strategic white matter tracts for processing speed deficits in age-related small vessel disease. Neurology 82, 1946–1950.

Lawrence

, Chung

, Morris

, Markus

, Barrick

(2014) Structural network efficiency is associated with cognitive impairment in small-vessel disease. Neurology 83, 304–311.

Tuladhar

, van Dijk

, Zwiers

, van Norden

, de Laat

, Shumskaya

, Norris

, de Leeuw

(2016) Structural network connectivity and cognition in cerebral small vessel disease. Hum Brain Mapp 37, 300–310.

Caroli

, Testa

, Geroldi

, Nobili

, Guerra

, Bonetti

, Frisoni

(2007) Brain perfusion correlates of medial temporal lobe atrophy and white matter hyperintensities in mild cognitive impairment. J Neurol 254, 1000–1008.

Patel

, Boada

, Price

, Sheu

, Tudorascu

, Reynolds Iii

, Aizenstein

(2012) Association of small vessel ischemic white matter changes with BOLD fMRI imaging in the elderly. Psychiatry Res 204, 117–122.

10.

Zhou

, Yu

, Duong

, Alzheimer’s Disease Neuroimaging Initiative (2015) White matter lesion load is associated with resting state functional MRI activity and amyloid PET but not FDG in mild cognitive impairment and early Alzheimer’s disease patients. J Magn Reson Imaging 41, 102–109.

11.

Bullmore

, Sporns

(2009) Complex brain networks: Graph theoretical analysis of structural and functional systems. Nat Rev Neurosci 10, 186–198.

12.

Fox

, Raichle

(2007) Spontaneous fluctuations in brain activity observed with functional magnetic resonance imaging. Nat Rev Neurosci 8, 700–711.

13.

Liu

, Yu

, Zhang

, Liu

, Duan

, Alexander-Bloch

, Liu

, Jiang

, Bullmore

(2014) Impaired long distance functional connectivity and weighted network architecture in Alzheimer’s disease. Cereb Cortex 24, 1422–1435.

14.

Bai

, Liao

, Watson

, Shi

, Wang

, Yue

, Teng

, Wu

, Yuan

, Jia

, Zhang

(2011) Abnormal whole-brain functional connection in amnestic mild cognitive impairment patients. Behav Brain Res 216, 666–672.

15.

Zhou

, Yao

, Wang

, Zhang

, Zhan

, Ma

, Xu

, Wang

, An

, Liu

, Zhang

(2015) Aberrant functional connectivity architecture in Alzheimer’s disease and mild cognitive impairment: A whole-brain, data-driven analysis. Biomed Res Int 2015, 495375.

16.

Sanz-Arigita

, Schoonheim

, Damoiseaux

, Rombouts

, Maris

, Barkhof

, Scheltens

, Stam

(2010) Loss of ‘small-world’ networks in Alzheimer’s disease: Graph analysis of FMRI resting-state functional connectivity. PLoS One 5, e13788.

17.

Fan

, Li

, Zhuo

, Zhang

, Wang

, Chen

, Yang

, Chu

, Xie

, Laird

, Fox

, Eickhoff

, Yu

, Jiang

(2016) The Human Brainnetome Atlas: A new brain atlas based on connectional architecture. Cereb Cortex 26, 3508–3526.

18.

Helenius

, Henninger

(2015) Leukoaraiosis burden significantly modulates the association between infarct volume and National Institutes of Health Stroke Scale in ischemic stroke. Stroke 46, 1857–1863.

19.

Fazekas

, Chawluk

, Alavi

, Hurtig

, Zimmerman

(1987) MR signal abnormalities at 1.5 T in Alzheimer’s dementia and normal aging. AJR Am J Roentgenol 149, 351–356.

20.

Papma

, de Groot

, de Koning

, Mattace-Raso

, van

der Lugt A

, Vernooij

, Niessen

, van Swieten

, Koudstaal

, Prins

, Smits

(2014) Cerebral small vessel disease affects white matter microstructure in mild cognitive impairment. Hum Brain Mapp 35, 2836–2851.

21.

Mirza

, Ikram

, Bos

, Mihaescu

, Hofman

, Tiemeier

(2017) Mild cognitive impairment and risk of depression and anxiety: A population-based study. Alzheimers Dement 13, 130–139.

22.

Brissos

, Dias

, Soeiro-de-Souza

, Balanza-Martinez

, Kapczinski

(2011) The impact of a history of psychotic symptoms on cognitive function in euthymic bipolar patients: A comparison with schizophrenic patients and healthy controls. Rev Bras Psiquiatr 33, 353–361.

23.

Peterson

, Savulich

, Jackson

, Killikelly

, Pickard

, Sahakian

(2016) The effect of shunt surgery on neuropsychological performance in normal pressure hydrocephalus: A systematic review and meta-analysis. J Neurol 263, 1669–1677.

24.

Jia

, Zhou

, Wei

, Jia

, Wang

, Li

, Wu

, Mok

, Gauthier

, Tang

, Chu

, Zhou

, Cui

, Wang

, Yin

, Hu

, Zuo

, Song

, Qin

, Wu

, Li

, Jia

, Song

, Han

, Xing

, Yang

, Li

, Qiao

, Tang

, Lv

, Dong

(2014) The prevalence of mild cognitive impairment and its etiological subtypes in elderly Chinese. Alzheimers Dement 10, 439–447.

25.

Jia

, Wang

, Wei

, Zhou

, Jia

, Li

, Tang

, Chu

, Zhou

, Cui

, Wang

, Yin

, Hu

, Zuo

, Song

, Qin

, Wu

, Li

, Jia

, Song

, Han

, Xing

, Yang

, Li

, Qiao

, Tang

, Lv

, Dong

(2014) The prevalence of dementia in urban and rural areas of China. Alzheimers Dement 10, 1–9.

26.

DeCarli

, Fletcher

, Ramey

, Harvey

, Jagust

(2005) Anatomical mapping of white matter hyperintensities (WMH): Exploring the relationships between periventricular WMH, deep WMH, and total WMH burden. Stroke 36, 50–55.

27.

Jokinen

, Schmidt

, Ropele

, Fazekas

, Gouw

, Barkhof

, Scheltens

, Madureira

, Verdelho

, Ferro

, Wallin

, Poggesi

, Inzitari

, Pantoni

, Erkinjuntti

, LADIS Study Group (2013) Diffusion changes predict cognitive and functional outcome: The LADIS study. Ann Neurol 73, 576–583.

28.

, Liu

, Zhan

, Ren

, Jiang

(2018) BRANT: A versatile and extendable resting-state fMRI toolkit. Front Neuroinform 12, 52.

29.

Power

, Barnes

, Snyder

, Schlaggar

, Petersen

(2012) Spurious but systematic correlations in functional connectivity MRI networks arise from subject motion. Neuroimage 59, 2142–2154.

30.

Satterthwaite

, Wolf

, Loughead

, Ruparel

, Elliott

, Hakonarson

, Gur

(2012) Impact of in-scanner head motion on multiple measures of functional connectivity: Relevance for studies of neurodevelopment in youth. Neuroimage 60, 623–632.

31.

Van Dijk

, Sabuncu

, Buckner

(2012) The influence of head motion on intrinsic functional connectivity MRI. Neuroimage 59, 431–438.

32.

Dresler

, Shirer

, Konrad

, Muller

, Wagner

, Fernandez

, Czisch

, Greicius

(2017) Mnemonic training reshapes brain networks to support superior memory. Neuron 93. 1227–1235 e1226.

33.

Kragel

, Kano

, Van Oudenhove

, Ly

, Dupont

, Rubio

, Delon-Martin

, Bonaz

, Manuck

, Gianaros

, Ceko

, Reynolds Losin

, Woo

, Nichols

, Wager

(2018) Generalizable representations of pain, cognitive control, and negative emotion in medial frontal cortex. Nat Neurosci 21, 283–289.

34.

Cheng

, Rolls

, Ruan

, Feng

(2018) Functional connectivities in the brain that mediate the association between depressive problems and sleep quality. JAMA Psychiatry 75, 1052–1061.

35.

Cocchi

, Harding

, Lord

, Pantelis

, Yucel

, Zalesky

(2014) Disruption of structure-function coupling in the schizophrenia connectome. Neuroimage Clin 4, 779–787.

36.

Zhan

YF YH

, Wang

, Zhou

, Zhang

, Guo

, An

, Ma

, Zhang

, Liu

(2016) Network-based statistic show aberrant functional connectivity in Alzheimer’s disease. IEEE J Sel Top Signal Process 10, 1182–1188.

37.

Mackinnon

, Fairchild

(2009) Current directions in mediation analysis. Curr Dir Psychol Sci 18, 16.

38.

Hayes

, Rockwood

(2017) Regression-based statistical mediation and moderation analysis in clinical research: Observations, recommendations, and implementation. Behav Res Ther 98, 39–57.

39.

Ben-Hur

, Ong

, Sonnenburg

, Scholkopf

, Ratsch

(2008) Support vector machines and kernels for computational biology. PLoS Comput Biol 4, e1000173.

40.

Golland

, Fischl

(2003) Permutation tests for classification: Towards statistical significance in image-based studies. Inf Process Med Imaging 18, 330–341.

41.

Dosenbach

, Nardos

, Cohen

, Fair

, Power

, Church

, Nelson

, Wig

, Vogel

, Lessov-Schlaggar

, Barnes

, Dubis

, Feczko

, Coalson

, Pruett

Jr. , Barch

, Petersen

, Schlaggar

(2010) Prediction of individual brain maturity using fMRI. Science 329, 1358–1361.

42.

Pereira

, Mitchell

, Botvinick

(2009) Machine learning classifiers and fMRI: A tutorial overview. Neuroimage 45 S199–209.

43.

Beheshti

, Demirel

(2016) Feature-ranking-based Alzheimer’s disease classification from structural MRI. Magn Reson Imaging 34, 252–263.

44.

Feng

, Wang

, Zhao

, Zhou

, Yao

, Meng

, Wang

, Zhang

, Ding

, Wang

, An

, Zhang

, Liu

(2018) Radiomic features of hippocampal subregions in Alzheimer’s disease and amnestic mild cognitive impairment. Front Aging Neurosci 10, 290.

45.

Greene

, Laumann

, Dubis

, Ihnen

, Neta

, Power

, Pruett

Jr , Black

, Schlaggar

(2014) Developmental changes in the organization of functional connections between the basal ganglia and cerebral cortex. J Neurosci 34, 5842–5854.

46.

Duering

, Zieren

, Herve

, Jouvent

, Reyes

, Peters

, Pachai

, Opherk

, Chabriat

, Dichgans

(2011) Strategic role of frontal white matter tracts in vascular cognitive impairment: A voxel-based lesion-symptom mapping study in CADASIL. Brain 134, 2366–2375.

47.

Jacqmin

, Herve

, Viswanathan

, Guichard

, During

, Dichgans

, Chabriat

(2010) Confluent thalamic hyperintensities in CADASIL. Cerebrovasc Dis 30, 308–313.

48.

, Zhou

, Liu

, Jiang

, Dong

, Zhang

, Walter

(2011) Functional segregation of the human cingulate cortex is confirmed by functional connectivity based neuroanatomical parcellation. Neuroimage 54, 2571–2581.

49.

Stoodley

(2012) The cerebellum and cognition: Evidence from functional imaging studies. Cerebellum 11, 352–365.

50.

Bell

, Shine

(2016) Subcortical contributions to large-scale network communication. Neurosci Biobehav Rev 71, 313–322.

51.

Arsalidou

, Duerden

, Taylor

(2013) The centre of the brain: Topographical model of motor, cognitive, affective, and somatosensory functions of the basal ganglia. Hum Brain Mapp 34, 3031–3054 .

52.

Wang

, Jia

, Liang

, Qi

, Yang

, Zhou

, Li

(2012) Changes in thalamus connectivity in mild cognitive impairment: Evidence from resting state fMRI. Eur J Radiol 81, 277–285.

53.

Zhou

, Hu

, Zhang

, Wang

, Zhu

, Xu

, Sun

, Yu

(2016) Aberrant functional connectivity and structural atrophy in subcortical vascular cognitive impairment: Relationship with cognitive impairments. Front Aging Neurosci 8, 14.

54.

Shackman

, Salomons

, Slagter

, Fox

, Winter

, Davidson

(2011) The integration of negative affect, pain and cognitive control in the cingulate cortex. Nat Rev Neurosci 12, 154–167.

55.

Leech

, Braga

, Sharp

(2012) Echoes of the brain within the posterior cingulate cortex. J Neurosci 32, 215–222.

56.

Fox

, Snyder

, Vincent

, Corbetta

, Van Essen

, Raichle

(2005) The human brain is intrinsically organized into dynamic, anticorrelated functional networks. Proc Natl Acad Sci U S A 102, 9673–9678.

57.

Leech

, Sharp

(2014) The role of the posterior cingulate cortex in cognition and disease. Brain 137, 12–32.

58.

Andrews-Hanna

, Snyder

, Vincent

, Lustig

, Head

, Raichle

, Buckner

(2007) Disruption of large-scale brain systems in advanced aging. Neuron 56, 924–935.

59.

Jeong

, Chung

, Kim

(2015) Episodic memory in aspects of large-scale brain networks. Front Hum Neurosci 9, 454.

60.

Amemori

, Graybiel

(2012) Localized microstimulation of primate pregenual cingulate cortex induces negative decision-making. Nat Neurosci 15, 776–785.

61.

Sun

, Qin

, Zhou

, Xu

, Qian

, Tao

, Xu

(2011) Abnormal functional connectivity in patients with vascular cognitive impairment, no dementia: A resting-state functional magnetic resonance imaging study. Behav Brain Res 223, 388–394.

62.

, Wang

, Jia

, Zhao

, Lu

, Li

, Jia

, He

, Jiang

, Han

(2012) Structural and functional changes in subcortical vascular mild cognitive impairment: A combined voxel-based morphometry and resting-state fMRI study. PLoS One 7, e44758.

63.

Crossley

, Mechelli

, Scott

, Carletti

, Fox

, McGuire

, Bullmore

(2014) The hubs of the human connectome are generally implicated in the anatomy of brain disorders. Brain 137, 2382–2395.

64.

Pievani

, de Haan

, Wu

, Seeley

, Frisoni

(2011) Functional network disruption in the degenerative dementias. Lancet Neurol 10, 829–843.

65.

Poggesi

, Pantoni

, Inzitari

, Fazekas

, Ferro

, O’Brien

, Hennerici

, Scheltens

, Erkinjuntti

, Visser

, Langhorne

, Chabriat

, Waldemar

, Wallin

, Wahlund

(2011) 2001-2011: A decade of the LADIS (Leukoaraiosis And DISability) Study: What have we learned about white matter changes and small-vessel disease? Cerebrovasc Dis 32, 577–588.

66.

Mayda

, Westphal

, Carter

, DeCarli

(2011) Late life cognitive control deficits are accentuated by white matter disease burden. Brain 134, 1673–1683.

67.

Reijmer

, Schultz

, Leemans

, O’Sullivan

, Gurol

, Sperling

, Greenberg

, Viswanathan

, Hedden

(2015) Decoupling of structural and functional brain connectivity in older adults with white matter hyperintensities. Neuroimage 117, 222–229.

68.

Bolandzadeh

, Davis

, Tam

, Handy

, Liu-Ambrose

(2012) The association between cognitive function and white matter lesion location in older adults: A systematic review. BMC Neurol 12, 126.

69.

Prins

, van Dijk

, den Heijer

, Vermeer

, Koudstaal

, Oudkerk

, Hofman

, Breteler

(2004) Cerebral white matter lesions and the risk of dementia. Arch Neurol 61, 1531–1534.

70.

Godin

, Tzourio

, Rouaud

, Zhu

, Maillard

, Pasquier

, Crivello

, Alperovitch

, Mazoyer

, Dufouil

(2010) Joint effect of white matter lesions and hippocampal volumes on severity of cognitive decline: The 3C-Dijon MRI study. J Alzheimers Dis 20, 453–463.

71.

De Groot

, De Leeuw

, Oudkerk

, Van Gijn

, Hofman

, Jolles

, Breteler

(2002) Periventricular cerebral white matter lesions predict rate of cognitive decline. Ann Neurol 52, 335–341.

72.

Burns

, Church

, Johnson

, Xiong

, Marcus

, Fotenos

, Snyder

, Morris

, Buckner

(2005) White matter lesions are prevalent but differentially related with cognition in aging and early Alzheimer disease. Arch Neurol 62, 1870–1876.

73.

de Groot

, de Leeuw

, Oudkerk

, van Gijn

, Hofman

, Jolles

, Breteler

(2000) Cerebral white matter lesions and cognitive function: The Rotterdam Scan Study. Ann Neurol 47, 145–151.

74.

Kim

, Hwang

, Kim

, Lee

, Rhee

, Park

(2011) Regional white matter hyperintensities in normal aging, single domain amnestic mild cognitive impairment, and mild Alzheimer’s disease. J Clin Neurosci 18, 1101–1106.

75.

Heinen

, Vlegels

, de Bresser

, Leemans

, Biessels

, Reijmer

(2018) The cumulative effect of small vessel disease lesions is reflected in structural brain networks of memory clinic patients. Neuroimage Clin 19, 963–969.

76.

Drakesmith

, Dutt

, Fonville

, Zammit

, Reichenberg

, Evans

, Lewis

, Jones

, David

(2016) Mediation of developmental risk factors for psychosis by white matter microstructure in young adults with psychotic experiences. JAMA Psychiatry 73, 396–406.

77.

, Vorburger

, Gazes

, Habeck

, Stern

, Luchsinger

, Manly

, Schupf

, Mayeux

, Brickman

(2016) White matter integrity as a mediator in the relationship between dietary nutrients and cognition in the elderly. Ann Neurol 79, 1014–1025.

78.

van den Heuvel

, Mandl

, Kahn

, Hulshoff Pol

(2009) Functionally linked resting-state networks reflect the underlying structural connectivity architecture of the human brain. Hum Brain Mapp 30, 3127–3141.

79.

Damoiseaux

, Greicius

(2009) Greater than the sum of its parts: A review of studies combining structural connectivity and resting-state functional connectivity. Brain Struct Funct 213, 525–533.

80.

van der Holst

, Tuladhar

, van Norden

, de Laat

, van Uden

, van Oudheusden

, Zwiers

, Norris

, Kessels

, de Leeuw

(2013) Microstructural integrity of the cingulum is related to verbal memory performance in elderly with cerebral small vessel disease: The RUN DMC study. Neuroimage 65, 416–423.

81.

Orru

, Pettersson-Yeo

, Marquand

, Sartori

, Mechelli

(2012) Using support vector machine to identify imaging biomarkers of neurological and psychiatric disease: A critical review. Neurosci Biobehav Rev 36, 1140–1152.

82.

Dubey

, Zhou

, Wang

, Thompson

, Ye

(2014) Analysis of sampling techniques for imbalanced data: An=648 ADNI study. Neuroimage 87, 220–241.