Independent Component and Graph Theory analyses reveal normalized brain networks on resting-state functional MRI after working memory training in people with HIV

Abstract

Background:

Cognitive training may partially reverse cognitive deficits in people with HIV (PWH). Previous functional MRI (fMRI) studies demonstrate that working memory training (WMT) alters brain activity during working memory tasks, but its effects on resting brain network organization remain unknown.

Purpose:

To test whether WMT affects PWH brain functional connectivity in resting-state fMRI (rsfMRI).

Study Type:

Prospective.

Population:

53 PWH (ages 50.7 ± 1.5 years, 2 women) and 53 HIV-seronegative controls (SN, ages 49.5 ± 1.6 years, 6 women).

Field Strength/Sequence:

Axial single-shot gradient-echo echo-planar imaging at 3.0–T Tesla was performed at baseline (TL1), at 1-month (TL2), and at 6-months (TL3), after WMT.

Assessment:

All participants had rsfMRI and clinical assessments (including neuropsychological tests) at TL1 before randomization to Cogmed WMT (adaptive training, n = 58: 28 PWH, 30 SN; non-adaptive training, n = 48: 25 PWH, 23 SN), 25 sessions over 5–8 weeks. All assessments were repeated at TL2 and at TL3. The functional connectivity estimated by independent component analysis (ICA) or graph theory (GT) metrics (eigenvector centrality, etc.) for different link densities (LDs) were compared between PWH and SN groups at TL1 and TL2.

Statistical Tests:

Two-way analyses of variance (ANOVA) on GT metrics and two-sample t-tests on FC or GT metrics were performed. Cognitive (e.g., memory) measures were correlated to eigenvector centrality (eCent) using Pearson’s correlations. The significance level was set at p < 0.05 after false discovery rate correction.

Results:

The ventral default mode network (vDMN) eCent differed between PWH and SN groups at TL1 but not at TL2 (p = 0.28). In PWH, vDMN eCent changes significantly correlated with changes in the memory ability in PWH (r = −0.62 at LD = 50%) and vDMN eCent before training significantly correlated with memory performance changes (r = 0.53 at LD = 50%).

Data Conclusion:

ICA and GT analyses showed that adaptive WMT normalized graph properties of the vDMN in PWH.

INTRODUCTION

Human immunodeficiency virus (HIV) infection not only interferes with the body’s immune system but also has a profound negative impact on the central nervous system ⁽¹⁾. When HIV spreads to the brain, 30% to 50% of people with HIV (PWH) experience a progressive decline in cognitive and motor functions such as deficits in working memory (WM), attention, and motor skills, which may lead to a diagnosis of HIV-associated neurocognitive disorders (HAND) ⁽²⁾. Effective combined antiretroviral therapy (cART) could markedly reduce the amount of HIV in the body, and the prevalence of HIV-associated dementia (HAD), the most severe form of HAND; however, the prevalence for milder forms of HAND remains the same, including asymptomatic neurocognitive disorder (ANI) and mild neurocognitive disorder (MND) ⁽³⁾.

WM training (WMT) has been applied as a form of behavioral therapy to improve cognitive deficits in PWH ⁽⁴⁾. For instance, a computer-assisted adaptive WMT program has shown to be effective for improving cognitive abilities in PWH with or without HAND ⁽⁴⁾. A previous study using functional MRI (fMRI) has discovered brain activation pattern changes after WMT in PWH during WM tasks, suggesting changes in the intrinsic brain functional networks ⁽⁴⁾. Changes in intrinsic functional networks, however, could be more readily measured during resting states when subjects are not performing a specific task or receiving any stimulation ⁽⁵⁾. To study intrinsic brain functional connectivity, spatially distinct brain regions can be identified, namely resting-state networks (RSNs) ^{(6, 7)}, using spontaneous blood-oxygen-level-dependent (BOLD) signals. Resting-state fMRI (rsfMRI) does not require an activation paradigm and may provide reduced variability in participant efforts during scanning; therefore, rsfMRI may be more easily applied in the clinical setting. Furthermore, RSNs might provide prognostic information that is complementary to that obtained in task fMRI, as shown for psychiatric disorders ^(7,8).

The independent component analysis (ICA) is a commonly used data-driven method that decomposes observations into ICs with a linear mixing model and an assumption of statistical independence among latent sources ⁽⁹⁾. Applying ICA on fMRI data analyses can result in spatial ICs (i.e., RSNs) and a corresponding mixing matrix (i.e., time courses of the ICs), the latter of which can be used to calculate functional connectivity between RSNs. To quantify the topological organization of RSNs, graph-theoretical (GT) analyses use a graph, composed of nodes and edges, to represent the brain networks and their pairwise relationships, and it can summarize the mathematical structural relationships between these nodes ⁽¹⁰⁾. In a brain graph, a node represents a functionally independent RSN, which can be identified by ICA, and an edge represents the functional connectivity between two RSNs, which can be calculated by the correlation between time courses of these RSNs ⁽¹¹⁾. Using brain graphs, the topological properties of each RSN can be characterized by using a set of graph measures.

This study aimed to examine group differences (HIV-seronegative (SN) vs. PWH) in the topological organization of brain networks from rsfMRI before and after WMT. We hypothesized that PWH would show abnormal topological organization at baseline. Furthermore, we hypothesized that WMT would cause more significant topological organization changes, especially with adaptive WMT in PWH compared to those in the SN group, and hence reduced group differences after training (i.e., normalization), accompanied by improved cognitive function in the PWH adaptive group.

MATERIALS AND METHODS

The clinical trial (GOV-NCT02602418) in this study was approved by the Committees on Human Studies at the University of Hawaii at Manoa. All participants provided written informed consent.

Participants

Participants were recruited from the local community (September 2010 to December 2015), screened through telephone interviews, and evaluated with detailed medical and neuropsychiatric examinations to ensure that they fulfilled the study criteria as reported ⁽¹²⁾. Specifically, the inclusion criteria for all participants were men or women of any ethnicity, ages ≥ 18 years and able to provide informed consent. PWH were required to have documented HIV seropositivity, and were stable on an antiretroviral regimen for six months or would remain off their antiretroviral medication during the duration of the WMT and follow-up periods. The SN controls were confirmed to be seronegative for HIV (by an HIV ELISA blood test if screened positive with ClearView^® COMPLETE HIV-1/2 test).

For all subjects, the exclusion criteria were: (1) a history of comorbid psychiatric illness that might confound the analysis of the study (e.g., schizophrenia, obsessive compulsive disorder, major depression, or bipolar disorder); (2) any confounding neurological disorder (including non-HIV brain infections, neoplasms, cerebral palsy, or significant head trauma with loss of consciousness >30 minutes); (3) significantly abnormal screening laboratory tests (>2 standard deviations) that indicated an uncontrolled chronic medical condition (e.g., diabetes, severe cardiac, renal or liver disorders) that might affect brain function; (4) any medications that could significantly alter functional brain imaging studies; (5) any current or history of severe substance use disorders within the previous 2 years, including amphetamines, cocaine, alcohol, and opiates, according to Diagnostic Statistical Manual of Mental Disorders-5 criteria (casual or recreational usage were allowed); (6) a positive test on urine toxicology screen (methamphetamine, amphetamine, cocaine, marijuana, benzodiazepine, barbiturates, and opiates, except for false positive tests due to medications) on the day of neuropsychological testing or fMRI studies; (7) pregnancy (screened with urine pregnancy test in women of childbearing age); (8) inability to read at an 8th-grade level (verified by the Wechsler Test of Adult Reading); (9) other contraindications for MRI studies, such as metallic or electronic implants in the body (e.g., pacemaker, surgical clips, or pumps), or severe claustrophobia ⁽¹²⁾.

SN controls were matched demographically to the PWH, i.e., they had similar age, sex, racial distribution, socioeconomic status (assessed using the Hollingshead Four Factor Index of Social Position), and depressive symptoms (assessed using the Center for Epidemiologic Studies-Depression Scale or CES-D).

All participants had baseline (TL1) rsfMRI scans along with the other MRI studies and clinical assessments for the current protocol ⁽¹²⁾ before randomization to 25 sessions of adaptive or non-adaptive Cogmed^® (Neural Assembly Int AB, Stockholm, Sweden) WMT over a period of five to eight weeks. The participants include 53 PWH (ages 50.7 ± 1.5 years, two women) and 53 SN (ages 49.5 ± 1.6 years, six women). A total of 58 participants (28 PWH, 30 SN) completed adaptive WMT and 48 participants completed non-adaptive WMT (25 PWH, 23 SN). One month after WMT (TL2), these 58 participants had the follow-up rsfMRI scan along with the other assessments. Additionally, six months after the WMT, a total of 46 participants (23 PWH, 23 SN) in the adaptive group and 22 participants (14 PWH, 8 SN) in the non-adaptive group returned for their additional 6-month follow-up scans (TL3). If the participants were randomized to the non-adaptive WMT initially, they were allowed to participate again in (crossover to) the adaptive WMT prior to the subsequent follow-up scans. A total of 13 (5 PWH, 8 SN) participants who completed the non-adaptive WMT also completed the adaptive WMT (i.e., crossover subgroup) and had their follow-up scans.

Across all participants and time, 323 scans, including the scans from TL1, TL2, TL3, and crossover, were carried out in total. We used all available 323 scans from all time points for extracting the RSNs. For the statistical comparison analysis of FNC and GT metrics, we only used fMRI data at TL1 and TL2, which were collected from participants scanned both at TL1 and TL2 (28 PWH, 30 SN for adaptive; 25 PWH, 23 SN for non-adaptive) because of the considerable attrition (drop-outs) at TL3 or crossover. Table 1 summarizes detailed information on the number of participants at TL1 and TL2.

Table 1.

Participant Characteristics.

	n	No. male sex (%)	Age (years) (mean ± standard deviation)
SN adaptive	28	24 (86)	50.7 ± 12.0
PWH adaptive	30	30 (100)	49.1 ± 11.2
SN non-adaptive	25	23 (92)	48.2 ± 12.1
PWH non-adaptive	23	21 (91)	52.8 ± 9.8

SN: seronegative controls; PWH: people with human immunodeficiency virus; n: number of participants.

MRI Acquisition and Preprocessing

Before and after WMT, each participant was scanned using a 3-Tesla MRI scanner (Tim Trio VB17, Siemens Healthineers, Erlangen, Germany). For structural MRI, a high-resolution three-dimensional T1-weighted anatomical sequence was acquired with magnetization prepared rapid gradient echo with a repetition time/echo time (TR/TE) = 2,200/4.47 ms, inversion time (TI) = 1,000 ms, echo spacing = 12 ms, flip angle (FA) = 12°, field of view (FOV) = 256 mm, matrix = 256 × 256, thickness = 1 mm, and 160 slices without interslice gaps. For rsfMRI, the entire brain was continuously scanned with an axial single-shot gradient-echo echo-planar imaging sequence using the following parameters: TR = 3,000 ms; TE = 30 ms; FOV = 192 mm; matrix 64 × 64; slice thickness = 3 mm; slice gap = 1 mm; 46 axial slices; no acceleration in slice direction; 120 time points. Participants were asked to keep their eyes open during the scan and stare passively at a central fixation cross.

To eliminate T1-related equilibrium effects ⁽¹³⁾, the first three time points from each fMRI scan were removed, resulting in 117 volumes per subject. The fMRI data were slice–timing-corrected to the center slice, realigned to the first image using the INRIalign algorithm embedded in DPARSF (DPARSF_V5.3_21010, http://rfmri.org/DPARSF), and spatially normalized to the standard Montreal Neurologic Institute (MNI) space ⁽¹⁴⁾ by using T1 unified segmentation, with a voxel size of 3 × 3 × 3 mm³, resulting in 61 × 73 × 61 voxels. Finally, the fMRI data were then spatially smoothed using a Gaussian kernel with a full-width at a half-maximum of 6 mm. The preprocessing was implemented using the Data Processing Assistant for rsfMRI (DPARSF, DPARSF_V5.3_21010, http://rfmri.org/DPARSF) ⁽¹⁴⁾.

WMT

Each participant was instructed to complete 20 to 25 sessions of Cogmed^® RM WMT over five to eight weeks. Each session took 30 to 40 minutes and comprised eight out of twelve possible modules (15 trials each), designed to train the verbal and visuospatial WM. The verbal WMT had four modules with different combinations of numbers or letters that were paired with lamps that lit up in different sequences. The visuospatial WMT had eight modules with different combinations of lamps that lit up and objects that rotated, moved, or required sorting. The participants were asked to recall the sequences with or without cues, and in forward or reverse orders. In the adaptive WMT sessions, these tasks became increasingly difficult, with more items to recall and adapting to the participants’ highest level of performance; however, in the non-adaptive WMT sessions, these tasks remained at the same low difficulty level with only three items to recall. More details about the WMT have been published earlier ^{(12, 13)}.

Cognitive Assessments

At TL1 and TL2, each participant also completed a battery of cognitive tests that allowed the assessments of seven cognitive domains required to assess HAND. Fluency was assessed using Delis-Kaplan Executive Function Syste (D-KEFS) or Ruff Figural Design Fluency and Verbal Fluency (with letters FAS). Learning ability was assessed using Rey Auditory Verbal Learning Test Trial 5 or Rey-Osterrieth Complex Figure Test-test Immediate Recall. Executive functions were assessed using D-KEFS-Color-Word Interference or Stroop Interference and Trail Making Test B. Speed of information processing was assessed using Symbol Digit, D-KEFS Trail-making Number Sequencing or Trail Making Test A, D-KEFS Color Naming or Stroop Color Naming, and California Computerized Assessment Package (CalCAP) Simple Reaction Time. As part of our cognitive battery to assess for HAND in PWH or HAND-equivalent deficits in the SN, the non-trained near transfer Attention/WM tests were further assessed using Arithmetic from Wechsler Adult Intelligence Scale-VI, Digit Span Backward, Letter-Number Sequencing, Arithmetic, and Paced Auditory Serial Addition Test 1. Memory was assessed using Rey Auditory Verbal Learning Test Delayed Recall (Trial 7), and Rey Complex Figure-Delayed Recall.

The raw scores for each test within each cognitive domain were combined to generate a standardized z-score, adjusted for age and education using a linear model, derived from a normative database of 481 seronegative healthy controls tested with the same protocol. The cognitive score summary of both PWH and SN is shown in Table 2 and Table 3.

Table 2.

Cognitive Performance at baseline and after one-month treatment (mean ± standard error) in adaptive group.

Cognitive z-scores	TL1 (n = 58)			TL2 (n = 58)
	SN (n = 28)	PWH (n = 30)	p	SN (n = 28)	PWH (n = 30)	p
Fluency	−0.10±0.14	−0.13±0.13	0.904	0.00±0.12	−0.12±0.13	0.500
Executive function	−0.15±0.15	−0.31±0.12	0.385	−0.06±0.15	−0.29±0.15	0.283
Speed	−0.10±0.13	−0.30±0.10	0.210	−0.13±0.14	−0.32±0.13	0.334
Attention	0.05±0.13	−0.42±0.14	0.016	0.20±0.16	−0.21±0.15	0.069
Learning	0.06±0.17	−0.21±0.12	0.192	0.21±0.17	−0.14±0.13	0.112
Memory	0.43±0.13	0.09±0.14	0.075	−0.03±0.15	−0.08±0.15	0.814

SN: seronegative controls; PWH: people with human immunodeficiency virus; TL1: baseline; TL2: one-month after treatment; n: number of participants; p: from two-sample t-test between PWH and SN.

Table 3.

Cognitive Performance at baseline and after one-month treatment (mean ± standard error) in non-adaptive group.

Cognitive z-scores	TL1 (n = 48)			TL2 (n = 48)
	SN (n = 25)	PWH (n = 23)	p	SN (n = 25)	PWH (n = 23)	p
Fluency	−0.13±0.14	−0.29±0.19	0.504	−0.08±0.13	−0.38±0.17	0.171
Executive function	−0.11±0.16	−0.25±0.13	0.521	−0.03±0.15	−0.24±0.17	0.237
Speed	−0.30±0.14	−0.27±0.15	0.906	−0.05±0.18	−0.27±0.16	0.364
Attention	−0.24±0.14	−0.41±0.14	0.392	−0.13±0.16	−0.24±0.13	0.602
Learning	−0.10±0.18	−0.35±0.20	0.370	−0.01±0.16	−0.17±0.21	0.526
Memory	0.04±0.20	−0.12±0.20	0.560	0.20±0.19	−0.14±0.20	0.212

SN: seronegative controls; PWH: people with human immunodeficiency virus; TL1: baseline; TL2: one-month after treatment; n: number of participants; p: from two-sample t-test between PWH and SN.

Group ICA

To estimate spatially independent RSNs across all subjects combined within a common subspace, we used group ICA to extract such components from multiple subjects (Figure 1a). First, to exclude non-brain voxels, we masked fMRI data for all subjects using Group ICA of fMRI toolbox (GIFT v3.0b) ⁽¹⁵⁾, resulting in 52,767 brain voxels for each subject. A matrix of each dataset with dimensions of 117 × 52,767 (number of time points × number of voxels) was constructed for the fMRI dataset. Second, a subject-level principal component analysis (PCA) was performed on the dataset collected from each subject to extract the signal subspace, which was followed by a group-level PCA on the PCs concatenated across all the subjects ⁽¹⁶⁾. Then, ICA was performed on the resulting group-level PCs. Finally, we performed back-reconstruction on the group-level ICs to obtain subject-level ICs and their associated time courses ⁽¹⁶⁾. The MATLAB 2019b (The MathWorks, Inc., Natick, MA) code provided in the Group ICA of fMRI Toolbox (GIFT v3.0b, https://trendscenter.org/software/gift/) was used for the implementation of the group ICA.

Figure 1:

Analytical methods used in this study. (a) Principal component analysis (PCA) was performed on the dataset of individual subjects (T × V; the number of time points × number of voxels) along time direction. The reduced matrices (T′ × V) were combined across subjects to produce a concatenated matrix (KT′ × V), which was subjected to a second PCA followed by ICA. Through back-reconstruction, the subject-level independent components and associated time courses were obtained for further analysis. (b) For each subject in one of two paired groups (e.g., SN vs. PWH and baseline (TL1) vs. one month after training (TL2)), the reduced subject-level time courses T × N′ (N′ is the number of resting-state networks; group 1: K₁ subjects, group 2: K₂ subjects) were used to compute the subject-level FNC (N′ × N′). Specifically, Pearson’s correlation coefficient between pair-wise network time courses was calculated and transformed to z-score using Fisher’s z-transformation. To compare the differences of the FNC between two paired groups, for each location in the FNC matrix, a t-statistic was calculated by using the z-scores obtained above between the two groups, which resulted in a t-statistic matrix (N′ × N′). (c) Through thresholding and binarization, the subject-level FNC was transformed into an adjacency matrix (N′ × N′), from which a graph was created with nodes and edges.

Order Selection

To select the order of signal subspace at both the subject and group levels of PCA, we used the entropy-rate based order selection by finite memory length model (ER-FM), an order estimation method that accounts for sample dependence and thus matches the fMRI data ⁽¹⁶⁾. When we applied this method to each subject’s fMRI data separately, the estimated order across subjects was 47.0 ± 5.6 (mean ± standard deviation). For the group-level PCA, we used an order equal to the mean plus one standard deviation, which was rounded up to 55 to account for variability across subjects. We then set the subject-level order to 1.5 times the group-level order (rounded it up to 85), which is the default setting in GIFT ⁽¹⁸⁾.

ICA Algorithm Selection

We used the Entropy Rate Bound Minimization (ERBM) algorithm regarding ICA ^{(19, 20)}. ERBM relaxes assumptions placed on the fMRI sources by assuming flexible source distributions and taking into account sample dependence (i.e., voxel-wise dependence) ⁽¹⁹⁾. The source density model in ERBM therefore matches the underlying properties of the fMRI sources.

Due to its iterative nature, ICA algorithms yield slightly different solutions with different initializations. Thus, we ran the ERBM algorithm 30 times with different random initializations and selected the most consistent run by using the cross-inter-symbol interference (ISI) metric ⁽²¹⁾.

RSN Identification

Each group’s ICs obtained from group ICA was transformed into z-scores to have zero mean and unit variance. We identified ICs of interest (i.e., RSNs) based on three criteria: (1) the coordinates of peak activation within clusters located in the grey matter; (2) minimal overlap with known vascular, susceptibility, ventricular, and edge regions associated with head motion; and (3) mean power spectra values of the corresponding time courses showing high low-frequency spectral power ⁽²²⁾. In this study, we calculated the ratio between the power spectra of the low-frequency band (below 0.10 Hz) and that of the high-frequency band (between 0.15 and 0.25 Hz) ⁽²³⁾. Components with a ratio lower than six was considered as an artifact component rather than the RSN. For a component to be considered an RSN, it has to pass all three criteria. As a result, 32 out of 55 ICs were identified as RSNs, which were classified into different categories of networks based on their anatomical locations and reported functional properties. These RSNs included default mode network (DMN), cognitive control network (CON), visual network (VIS), auditory network (AUD), sensorimotor network (SM), sub-cortical network (SC), and cerebellar network (CB). The spatial maps of the RSNs were converted to z-scores and thresholded at |z| ≥ 2 for visualization (Figure 2a).

Figure 2:

RSNs and associated time courses for t-statistic and graph-theoretical analyses. (a) Spatial maps of 32 RSNs. RSNs were classified into 7 categories of brain networks: default mode network (DMN), cognitive control network (CON), visual network (VIS), auditory network (AUD), sensorimotor network (SM), sub-cortical network (SC), and cerebellar network (CB). The number next to each network name is the number of ICs belonging to that network. Different components with each network were colored differently. (b) At the individual subject level, the time courses associated with the 32 RSNs were transformed to the functional network connectivity (FNC), which was further used for the t-statistical analysis, and graphical theory (GT) analysis. Ventral DMN and attention network are shown as two examples of RSNs. (c) For the GT analyses, through thresholding and binarization, the FNC was transformed to an adjacency matrix (left panel), which could be described in a graph consisting of nodes and edges (right panel). The same subject used in (b) and (c) was randomly selected for illustration.

FNC Extraction

We used the standard back-reconstruction method in GIFT to extract time courses of subject-level RSNs. We then constructed the subject-level FNC matrices by z-transformed pairwise correlations between time courses of RSNs. The pipeline consisted of (1) post-processing the time courses using linear, quadratic, and cubic detrending, and de-spiking using 3dDespike ⁽²⁴⁾; (2) temporal filtering with a fifth-order Butterworth low-pass filter with a high-frequency cutoff of 0.15 Hz ⁽²²⁾; (3) computing the pair-wise Pearson correlation between time courses for each pair of RSNs at the subject level; (4) Fisher’s z-transformation of the correlation coefficients, which resulted in an FNC matrix per subject. The FNC matrix represented functional connectivity between RSNs. Our subsequent analyses were all based on these matrices (Figure 1b).

To test whether the training has an overall effect on the FNC, we used FNC matrices from the 53 PWH and 53 SN controls who had usable rsfMRI data before (TL1; baseline) and after (TL2) the training (Table 1).

GT Analysis

The GT analysis is a mathematical approach that quantifies the property of associations between objects by modeling the associations as mathematical structures (i.e., graphs) ⁽²⁵⁾. The FNC matrix was used to construct a graph for each subject (Figure 1c), with each node being an identified RSN and edges being the functional connectivity between RSNs. To focus our analyses on only strong connectivity between nodes, before constructing graphs, a threshold was used to remove edges with weak connectivity (i.e., below the threshold) that represents lower correlation value in the FNC matrix. The percentage of survived edges after thresholding is defined as the link density (LD). The threshold was determined by a given LD. For instance, if the LD is set to be 70%, a certain threshold was used such that 70% of edges would survive while the remaining 30% would be removed from the FNC matrix. To avoid graphs being too sparse or too dense, we used a series of LDs ranging from 20% to 70% with a step size of 2% ⁽²⁰⁾. At each level of LD, the FNC matrix was binarized with a corresponding threshold, which resulted in an adjacency matrix (Figure 1c). The adjacency matrix was then used to form an undirected graph G, from which GT metrics (e.g., degree, eigenvector centrality, betweenness centrality, closeness centrality, local efficiency, clustering coefficient) were calculated. The definition and calculation of the GT metrics (degree, centrality, local efficiency, clustering coefficient) used can be found in supplementary material. The MATLAB 2019b code provided in the Brain Connectivity Toolbox (https://sites.google.com/site/bctnet/) ⁽²⁶⁾ was used for the implementation of GT analysis.

Statistical Analysis

All statistical analyses were performed using MATLAB 2019b. First, to summarize group effects for adaptive and non-adaptive training, we performed two-sample t-tests on the FNC matrices between the PWH and SN groups, separately at TL1 and TL2 (Figure 1b). A t-statistic was calculated between two groups of subjects (PWH and SN) in each entry of the subject-level FNC matrix, which resulted in a t-statistic matrix (32 × 32). To characterize the overall group differences of t-statistic matrices at baseline (TL1) and one month after training (TL2), separately in the adaptive or non-adaptive group, we first calculated the number of nodes above a changing threshold of t-statistic. We then performed an analysis similar to the receiver-operating characteristic (ROC) curve, in which we used a term ‘discrimination index (DI)’ to examine how the t-statistic matrices at TL1 and TL2 differ. Specifically, we plotted a curve with two axes—the Y-axis indicated true positive (corresponding to the sensitivity in ROC analysis), meaning the current t-statistic matrix was correctly identified as at TL1 based on the number of survived nodes being higher than a moving threshold; the X-axis indicated false positive (corresponding to the 1-specificity in ROC analysis), meaning the current t-statistic matrix was incorrectly identified as at TL1 based on the same criteria. The area under such a curve is the DI.

Next, we performed a two-way analysis of variance (ANOVA) test on the GT metrics for each RSN to screen for RSNs that were significantly affected by the disease condition, training, or their interactions. Specifically, to investigate the overall effect across GT metrics, we concatenated the six GT nodal metrics that were calculated at the 26 pre-determined LDs, which resulted in a matrix of 116 × 156 (subjects × metric values) for the adaptive group, and a matrix of 96 × 156 for the non-adaptive group. We performed a PCA on each matrix along the dimension of the metric values. In this study, we used the first two PCs, which together explained almost 70% of the variance for the two-way ANOVA tests. RSNs from the adaptive group or the non-adaptive group that were significantly affected by training (i.e., p < 0.05 for training or interaction) were then used for post-hoc analysis using t-tests to further examine the training effect on these RSNs.

Post-hoc two-sample t-tests were then used for the statistical comparisons between the PWH and SN groups, separately at TL1 and TL2. The p-values derived from multiple statistical tests at different LDs were corrected to reduce the false discovery rate (FDR) by using the Benjamini-Hochberg procedure. Corrected p-values less than 0.05 were used to reject the null hypothesis where there is no statistical difference between the two groups. Moreover, we performed an analysis similar to the ROC curve to examine the difference between PWH and SN groups in each GT metric at each LD. We used the metric DI to indicate the difference between PWH and SN groups in each GT metric at a specific LD.

To explore the relationships between brain function (characterized by the GT metrics) and cognitive abilities, we used Pearson correlation to evaluate the association between changes in GT metrics of an RSN and the changes in subjects’ cognitive domain z-scores (i.e., fluency, executive function, speed of information processing, attention, learning ability, and memory) from TL1 to TL2. Moreover, to examine whether the GT metrics before training could predict future changes in the cognitive abilities after training, we also calculated the correlation coefficients between the GT metrics at TL1 and the changes in cognitive scores from TL1 to TL2. For these exploratory analyses, p-values < 0.05 were considered statistically significant.

RESULTS

FNC Between PWH and SN Before and After WMT

The differences in FNC between PWH and SN groups are shown separately in adaptive (Figure 3a–c) and non-adaptive groups (Figure 3d–f) and at TL1 and TL2, respectively. The heat maps in Figure 3a show the resultant t-statistic matrices at TL1 and TL2 for the adaptive group, indicating that the overall difference between PWH and SN groups is greater at TL1 than at TL2. Correspondingly, more RSN pairs survived thresholding at p = 0.05 at TL1 compared to TL2. Moreover, with a sliding threshold, we found more surviving RSN pairs at TL1 than at TL2 in the adaptive group (Figure 3c). However, no such difference was observed in the non-adaptive group (Figure 3d–f).

Figure 3:

Comparisons of FNC between PWH and SN using t-statistic. A t-statistic was calculated between two groups of subjects (PWH and SN) in each entry in the subject-level FNC matrices, which resulted in a t-statistic matrix (32 × 32). (a) The two heatmaps show the t-statistic matrices at baseline (TL1) and one month after training (TL2) in the adaptive group. Warmer colors indicate higher t-statistic hence stronger connectivity between RSNs in the PWH group than that in the SN group. (b) Same as (a), but the t-statistic matrices are thresholded at p < 0.05, and uses the same color scale. (c) The number of survived nodes with a changing threshold of t-statistic. An analysis like the ROC was performed to compare the differences between TL1 and TL2, which resulted in the metric discrimination index (DI) shown in the panel. (d) The two heatmaps show the t-statistic matrices at TL1 and TL2 in the non-adaptive group. (e) Same as (d), but the t-statistic matrices are thresholded at p < 0.05, and uses the same color scale. (f) The number of survived nodes with a changing threshold of t-statistic. (a–c) Adaptive: SN (n = 28), PWH adaptive (n = 30). (d–f) Non-adaptive: SN (n = 25), PWH (n = 23).

GT Metrics Between PWH and SN Before and After WMT

A total of 13 RSNs (9 in the adaptive and 4 in the non-adaptive group) had at least one p < 0.05 (HIV-status, training, or interaction) on ANOVA (Figure 4). Overall, more RSNs in the adaptive group showed HIV-status or training effects than those in the non-adaptive group. In the adaptive group, RSNs within the DMN and CON were mainly affected by WMT. Among these 13 obtained RSNs, six from the adaptive group (parts of DMN or CON) and only two from the non-adaptive group (CON or VIS) showed training effects, which were then used for the post-hoc analyses.

Figure 4:

Identified RSNs that show statistical significance in two-way ANOVA tests. The name of the RSN and its IC number are shown above each image. The variance explained by each PC are described as mean ± standard deviation across all 32 RSNs. The three p-values obtained from ANOVA are for the two factors (HIV-serostatus and training) and their interaction, respectively. PC #1: the first principal component; PC #2: the second principal component; RSN: resting state network; vDMN: ventral default mode network; dDMN: dorsal default mode network; SMN: sensorimotor network; CON: cognitive control network; VIS: visual network.

For the adaptive group, two of six RSNs survived the FDR-corrected t-tests when PWH was compared to SN at TL1, but not at TL2 (p = 0.28, 0.70, at LD = 50% respectively), considering multiple LDs. Specifically, the eigenvector centrality of the ventral DMN (vDMN, IC #27; showing significant training effect on PC #1) is shown in Figure 5, and the local efficiency of dorsal DMN (dDMN, IC #32; showing significant training effect on PC #2) is shown in Figure 6.

Figure 5:

Eigenvector centrality of the ventral DMN (vDMN). (a) The eigenvector centrality of vDMN with link densities (LDs) from 0.2 to 0.7. The step size is 0.02. X-axes show the time at baseline (TL1) and one-month after training (TL2). Y-axes show the eigenvector centrality of vDMN. The p-value above each plot is the result of a two-sample t-test between PWH and SN groups at TL1. P-values are corrected with the Benjamini-Hochberg procedure. The green colors indicate statistical significance. (b) Measuring the difference between PWH and SN groups by using a discrimination index (DI, see Methods) at TL1 and TL2, with different LDs. Pink asterisks indicate statistical significance between PWH and SN at TL1 (p < 0.05, two-sample t-test). (c) The component associated with this GT analysis belongs to vDMN. (a, b) SN adaptive (n = 28), PWH adaptive (n = 30). Error bars denote the standard error of the mean.

Figure 6:

Local efficiency of the dorsal DMN (dDMN) in PWH and SN groups. (a) The local efficiency of the dDMN with link densities (LDs) from 0.2 to 0.7. The step size is 0.02. X-axes show the time at baseline (TL1) and one-month after training (TL2). Y-axes show the local efficiency of dDMN. The p-value above each plot is a result of a two-sample t-test between PWH and SN groups at TL1. P-values are corrected with the Benjamini-Hochberg procedure. The green colors indicate statistical significance. (b) Measuring the difference between PWH and SN groups by using a discrimination index (DI, see Methods) at TL1 and TL2, with different LDs. Pink asterisks indicate statistical significance between PWH and SN at TL1 (p < 0.05, two-sample t-test). (c) The component associated with this GT analysis belongs to dDMN. (a, b) SN adaptive (n = 28), PWH adaptive (n = 30). Error bars denote the standard error of the mean.

In contrast, the two RSNs in the non-adaptive group in the non-adaptive group (IC #41 with significant training effect on PC #1; IC #23 with significant training effect on PC #2) were no longer significant after the FDR correction for any GT metric (PWH versus SN) at TL1 or TL2. Specifically, for the IC #41: degree: p = 0.58 and 0.69 for TL1 or TL2; eigenvector centrality: p = 0.70, 0.32; betweenness centrality: p = 0.96, 0.57; closeness centrality: p = 0.90, 0.75; local efficiency: p = 0.55, 0.54; clustering coefficient: p = 0.74, 0.51.

For the IC #23: degree: p = 0.53 and 0.96 at TL1 or TL2; eigenvector centrality: p = 0.57, 1.00; betweenness centrality: p = 0.99, 0.91; closeness centrality: p = 0.77, 0.97; local efficiency: p = 0.62, 0.85; clustering coefficient: p = 0.72, 0.92. All p-values above shown are at LD = 50% for simplicity.

Eigenvector Centrality of vDMN

We identified one RSN that showed group differences in eigenvector centrality for the adaptive training group (Figure 5). In Figure 5a, representative plots of the eigenvector centrality for PWH and SN groups at TL1 and TL2 are plotted with different LDs. While 57.7% of LDs analyzed (i.e., LDs between 40% and 68%) yielded significant differences between PWH and SN groups at TL1, no significant differences were observed at TL2 (p = 0.28 at LD = 50%). The RSN identified was part of the vDMN (Figure 5c). Since the PWH adaptive group were all men while the SN adaptive group had four women, we performed two-sample t-tests on the eigenvector centrality of vDMN at LD = 50% with the four women removed. The resulting vDMN eigenvector centrality difference between the PWH and SN groups remained significant at TL1 but not at TL2 (p = 0.30).

Local Efficiency of dDMN

We found significant differences between PWH and SN in dDMN at TL1 but not TL2 (p = 0.70 at LD=50%) in the adaptive training group (Figure 6). With the four women removed, the local efficiency of dDMN still yielded significant group differences between PWH and SN at TL1 but not at TL2 (p = 0.28 at LD=50%).

Associations Between GT Metrics and Cognitive Abilities

In the PWH adaptive training group, the changes in eigenvector centrality of the vDMN before and one month after training were significantly and negatively correlated with the change in memory abilities from TL1 to TL2 (Figure 7). In the PWH adaptive training group, eigenvector centrality of the vDMN at baseline before training also showed significant positive correlations with the improvement in memory abilities after training (Figure 8).

Figure 7:

Associations between the change in ventral DMN (vDMN) eigenvector centrality and the change in memory ability in PWH adaptive group. (a) Illustration of the graphs built using different link densities (LDs) from 0.4 to 0.68 where the eigenvector centrality of the vDMN shows significance. The step size is 0.02. (b) The eigenvector centrality of the vDMN is measured with LDs from 0.4 to 0.68. X-axes show the change in eigenvector centrality (ΔeCENT) of the vDMN. Y-axes show the change in memory ability (Δmemory z-scores) from baseline to one month after treatment. (c) Pearson’s correlation between Δmemory and ΔeCENT in PWH adaptive group with different LDs. The r and p values were derived from the Pearson’s correlation analysis. The blue dot indicates statistical significance and blue asterisks indicate the statistical significance (p < 0.05). (b, c) SN adaptive (n = 19); PWH adaptive (n = 18).

Figure 8:

Association between the ventral DMN (vDMN) eigenvector centrality at baseline (TL1) and the change in memory ability in PWH adaptive group. (a) Illustration of the graphs built using different link densities (LDs) from 0.4 to 0.68 where the eigenvector centrality of the vDMN shows significance. The step size is 0.02. (b) The eigenvector centrality of the vDMN was measured with LDs from 0.4 to 0.68. X-axes show the eigenvector centrality (TL1 eCENT) of the vDMN at TL1. Y-axes show the change in memory ability (Δmemory) from TL1 to one month after treatment. (c) Pearson’s correlation between Δmemory and TL1 eCENT in PWH adaptive group with different LDs. The r and p values were derived from the Pearson’s correlation analysis. The blue dot indicates the statistical significance and blue asterisks indicate statistical significance (p < 0.05). (b, c) SN adaptive (n = 19); PWH adaptive (n = 18).

DISCUSSION

In this study, we aimed to understand the brain functional connectivity mechanisms that underlie improved cognitive abilities after WMT in PWH compared to those in SN. Using group ICA, we identified 32 spatial RSNs within a common subspace from 323 scans. Our analyses revealed several aspects of brain network topological changes after training. The major finding was that after adaptive WMT, the eigenvector centrality of the vDMN in PWH became more like that in SN; importantly, such changes also predicted the improvements in the memory ability in PWH. These results may suggest that the vDMN with high eigenvector centrality might be a key neural substrate leading to normalization of brain function during WMT.

The overall group differences (PWH versus SN) in terms of brain connectivity amongst the 32 RSNs became smaller after the adaptive training but not after the non-adaptive training. This finding suggests that adaptive training could lead to normalization of brain functional connectivity in PWH. Likewise, a task-activated fMRI study in the same participants found decreased brain activation on the WM task in PWH, suggesting improved neural efficiency after adaptive WMT ⁽⁴⁾. To further explore resting brain functional connectivity changes, we examined the topological properties of these RSNs before and one month after WMT in PWH and SN controls. Ventral and dDMNs showed group differences in corresponding topological properties, including eigenvector centrality and local efficiency. In particular, the eigenvector centrality of the vDMN showed statistically significant group differences between PWH and SN groups at TL1 but not at TL2. This finding may indicate that cognitive training could normalize the DMN at TL2 for PWH. Interestingly, WM training led to opposite changes in the vDMN eigenvector centrality, with primarily decreases in the initially higher vDMN eigenvector centrality, along with improved dDMN local efficiency in PWH groups, but increases or stable levels of the centrality and efficiency in healthy controls. These opposite changes suggest a possible compensatory process at baseline normalized in the PWH group but further improved or stable centrality and efficiency in the SN after WMT. Therefore, an appropriate or normal level of vDMN eigenvector centrality and dDMN local efficiency may be the most critical for normal memory ability. This interpretation is supported by the correlations between the changes in vDMN eigenvector centrality after adaptive training and the changes in the memory ability in PWH. Hence, this finding further indicate that brain network changes contribute to improved memory in PWH. Furthermore, the eigenvector centrality of the vDMN before training was significantly correlated with memory improvements in PWH. These correlations might further support the usefulness of the centrality measure of the vDMN as a predictor for the normalized brain function and suggest that topological properties of the vDMN might be used to predict future training effects.

Our results further suggest that the vDMN is a main driver for normalized brain network reorganization underlying improved cognitive abilities after WMT in PWH, in agreement with findings from prior studies ^{(27 – 31)}. For instance, brain regions within the DMN were hypothesized to play important roles in integrating and coordinating the communications across brain networks ⁽²⁷⁾. In addition, prior studies consistently showed abnormal DMN activity or connectivity in patients with HIV infections ^{(28, 29)}. Furthermore, using rsfMRI, one study found abnormal functional connectivity between the DMN and the cortices of the visual network ⁽³⁰⁾. Other studies reported altered functional connectivity within the DMN or between the DMN and other networks ^{(27 – 31)}, and that the closeness centrality of the DMN was altered in patients infected with HIV ⁽³²⁾. Lastly, increased centrality including eigenvector centrality of the DMN was also associated with cognitive impairments caused by other brain disorders such as multiple sclerosis ⁽³³⁾.

The normalization of the DMN was most prominent in its eigenvector centrality, one of many graphical properties that we examined in this study. Eigenvector centrality measures how important a node is within a network by taking into consideration not only the own but also the neighbors’ connections with other nodes. Thus, higher values indicate that a particular node plays a more central role in the network and required greater connections with other brain regions. This finding is consistent with prior task-activated fMRI studies in PWH that found greater than normal brain activation during WM or visual attention ^(34–38). The greater brain activation in PWH than SN represented increased usage of cognitive reserves due to the lower network capacity ^{(35, 36}) and with declined neural efficiency with aging ^{(37, 38)}. The normalization of the DMN eigenvector centrality could align with the idea that the DMN might play a central role in reversing the cognitive decline after WMT, especially in those affected by HIV. Our results suggest that the DMN might be an effective target brain region for monitoring future research using causal manipulation techniques to treat cognitive impairment or deficits in PWH, such as other cognitive behavioral interventions or repetitive transcranial magnetic stimulation.

LIMITATIONS

One limitation was the relatively small sample size of the subgroups of study participants in the adaptive versus the non-adaptive training groups. Although statistically significant effects were observed, such observation might be more reliable and generalizable with a larger sample size in each subgroup. Moreover, for participants that had both TL1 and TL2 scans, our PWH group consisted of only men, while the SN group had four women; although the findings were the same when the data from the women were removed, our findings could not be generalized to both men and women with HIV, given the well-documented gender differences in RSN ⁽³⁹⁾.

CONCLUSION

We applied group ICA and GT analyses to rsfMRI data to assess brain functional connectivity mechanisms underlying cognitive ability improvements in PWH and HIV-seronegative controls. The centrality of the vDMN normalized with improved cognitive performance after WMT, which might suggest that the vDMN plays an important role in the reversal of cognitive impairment or deficits in PWH.