Identification of Two Distinct Working Memory-Related Brain Networks in Healthy Young Adults

Synthesis

Both reviewers agreed that the paper represents an impressive undertaking - a large dataset, thoughtful analyses as well as a commitment to data sharing. However, both reviewers had concerns about the conclusions and the scholarship of the paper.

First, please address more carefully the claim that the two networks are actually distinct. One reviewer was concerned that Figure 3 seems to show a lot of apparent overlap between the two networks but that this overlap is largely ignored. If you mean functionally distinct, then some qualifications may be necessary as ICA does not guarantee the sources are completely independent; it just identifies a space in which the sources are maximally independent. Furthermore it was noted that both ICs have a relationship with 0-back and 2-back performance, which should be discussed.

Related to this point, there appears to be a lack of appropriate discussion of prior related work that discusses distinct parietal and prefrontal contributions to WM and attention (for recent examples, see Kundu et al., 2015, NeuroImage; Sarma et al., 2016, Nature Neuroscience; Li et al., 2017, NeuroImage). Given this large body of existing work, it is important that the authors address what advance the current approach provides.

Second, please be more clear about the logic of relating ICs to networks. An independent component is a basis function (a set of voxel weights) that when scaled appropriately explains variance in the voxel weights across subjects. A network is a set of locations that interact with each other. It is not self-evident that these are the same things, and it would strengthen the paper if the authors could be more explicit about why ICs with large eigenvalues are interpreted as networks.

Also, it was not clear why you did subtraction of the fMRI signal (2B - 0B) but not the task performance? Since ICA was done on the subtracted regression coefficients, the ICA would be expected to reveal components that related to the difference in performance between the two tasks. There might have been a good reason to do this, but an explicit statement in the text was not found. This makes the correlation hard to interpret. Please clarify.

In several places, you point out that ICA is able to discover networks that would not be discoverable with a voxel-wise analysis however what is the basis for this statement? At least by visual inspection, the two analyses look highly similar. Figure 5a looks like 3a and 5b looks like 3b. It is fine, of course, if two methods arrive at similar conclusions. But it was surprising to read in multiple places how essential the ICA was when it appears that both types of analyses lead to the same conclusion. Perhaps you could state the limits of the conclusion they would have been able to draw had they only done the voxel-wise analysis.

Finally, the diffusion analysis produced an interesting result which, at least superficially, appears to conflict with the main results (or at least the way the main results are summarized in the text). Namely,IC3, the network the authors interpret as being most closely related to working memory activity, is mostly parietal. This is a central claim in the abstract and throughout. Yet the brain areas whose FA is most correlated with the IC3 weights is not parietal but rather precentral cortex. Perhaps it would make more sense for you to refer to IC3 as a parietal-precentral network, rather than a parietally-centered network? Please address this.

Minor:

(1) It is not clear what the FA results add to the story. It seems like an afterthought that is not appropriately incorporated into the rest of the claims about WM networks. How should one interpret the association between the parietal IC and global FA (which, if I understand correctly, has FA measures that mostly come from brain regions outside of the parietal IC network)? The abstract claims an association between FA and WM performance, but that's not explicitly addressed in the results.

(2) The manuscript needs copy editing. The references are not formatted correctly and there are duplications.

(3) The authors claim that the large sample size allowed them to describe networks in a way that would not be possible with fewer than 100 subjects (Line 585). What is the evidence for this claim? Didn't the authors demonstrate that these same ICs can be identified with 100 subjects?

(4) Line 234-235 “Since several changes on the scanner hardware and/or software over time were known...” What does this mean?

(5) Figure 7: I could not understand how values on x-axis of 7B and 7C were computed.

(6) The authors refer to ‘false-alarm corrected d-prime measures’. Since d-prime already accounts for false alarms, what does this mean? Do the authors just mean they used d-prime as a measure?

(7) The anatomical brain images were aligned volumetrically, and were smoothed quite a bit (8-mm Gaussian). Since all of the analyses were cortical structures, I wonder if they would have achieved better alignment using a surface-based registration, which tends to show better success for functional analyses on the cortex (though not sub cortex) (e.g., Wang et al, 2014, Cer Cortex, Probabilistic Maps of Visual Topography in Human Cortex)

(8) At one point, the authors talk about ‘IC activity’ (l 327). What does this mean? The ICs are static spatial patterns.

Author Response

Dear reviewers,

Thank you very much for your valuable comments on our paper “Identification of two independent working memory related brain networks in healthy young adults”, which have helped to clarify and improve the paper.

Major points:

“First, please address more carefully the claim that the two networks are actually distinct. One reviewer was concerned that Figure 3 seems to show a lot of apparent overlap between the two networks but that this overlap is largely ignored. If you mean functionally distinct, then some qualifications may be necessary as ICA does not guarantee the sources are completely independent; it just identifies a space in which the sources are maximally independent. ”

Thank you for pointing out that our description of distinct / independent networks requires clarification throughout the paper.

With respect to the statistical definition of independency in ICA, we added now more methodological details on ICA decomposition in general and how we interpret the results:

Introduction section (line 78): “ICA decomposition is a data-driven unbiased approach to retrieve a low-dimensional representation of a dataset, resulting in statistically independent signals (Kong, Vanderburg, Gunshin, Rogers, & Huang, 2008).”

Methods-Section (line 263): “Applied to a matrix X of m observations (subjects) and n variables (voxels), ICA estimates a matrix of k × n latent sources S that underlie the variables, holding the source estimates (referred to as voxel loadings throughout the paper) as statistically independent from each other as possible (Engreitz, Daigle, Marshall, & Altman, 2010). In addition to the source estimates, ICA also yields a matrix of m × k mixing coefficients A (referred to as subjects scores throughout the paper) for each independent component (IC). The mixing coefficients of a particular component depict the projection of the original data onto this component's estimated source, such that X = AS (Hyvärinen & Oja, 2000). By applying ICA decomposition to a matrix of 2-back–0-back contrast estimates, containing rows of voxels and columns of subjects, our source estimates (voxel loadings) described independent latent sources that underlie the contrast estimates. Accordingly, each component's mixing coefficients described the activity strength of each component for each subject (Chiapetta, Roubaud, & Torrésani, 2004). Subjects with high scores on an IC showed higher contrast estimates in voxels with high voxel loadings in this IC. Hence, we interpreted the subject scores as a measure of co-activation in the voxels that loaded onto the IC.”

Results section (line 495): “ICA decomposition is a data-driven unbiased approach that models observations as a linear combination of latent components (Engreitz et al., 2010), which are as statistically independent and uncorrelated as possible (Hyvärinen & Oja, 2000). ICA applied onto the 2-back–0-back contrast estimates resulted in one score per subject and one loading per voxel for each IC. If a subject scored highly onto one IC, it also showed increased activation in the voxels that loaded highly onto that particular IC. Accordingly, the subject scores of an IC represented a measure of co-activation across the voxels that loaded onto this IC. We therefore interpreted the estimated ICs as networks of co-activated brain regions.”

With respect to the overlap between the two networks IC3 and IC4 we added now a more detailed analysis that shows the variation in overlap, depending on the threshold used to identify the most-extreme voxel loadings. Within the scope of this more detailed analysis, we also recognized that our initial decision rule that we used when applying a threshold to the ICA source estimates was suboptimal, as it impeded a uniform interpretation of illustrations and annotations. This is now updated for both, the WMN ICA and the NeuroSynth ICA, throughout the paper: we illustrate now the 10% most extreme voxel loadings across all ICs (see for example Fig. 3 and Fig. 5).

Methods section (line 288, WMN-ICs): “For visualization purpose and for anatomical annotation, we determined the voxel loadings with the 10% most extreme absolute values (|z| > 1.47), when considering all six ICs.”

Methods section (line 433, NeuroSynth-ICs): “We tested the overlap between the 16 NeuroSynth networks and the WMN derived from our functional MRI data by calculating the percentage of voxels that show high loadings on the NeuroSynth networks (|z| > 0.70; i.e. the 10% most extreme absolute values across all NeuroSynth ICs; the same threshold was used to visualize the NeuroSynth ICs) and were additionally located in the WMN.”

“Furthermore it was noted that both ICs have a relationship with 0-back and 2-back performance, which should be discussed.”

We thank the reviewers and the editor for pointing out the necessity for commenting on the reported associations of WMN-ICs with 0-back and 2-back performances. We had previously reported nominal p-values as well as p-values that were corrected for multiple testing. The association of IC4 with 2-back performance was only nominally significant, but did not survive multiple-testing correction. In order to be consistent throughout the paper, we now report for all association analyses p-values that are corrected for multiple testing (FDR, α = 0.05). When applying FDR-correction, only WMN-IC3 was significantly associated with both 2-back and 0-back performances; of note, the associations of WMN-IC3 with 2-back and 0-back performances showed opposite directions of effect, with a larger effect size for 2-back performance (5.8% in comparison to 2.2% for 0-back performance; see Table 1).

We added this information in the result section (line 518): “In these analyses, WMN-IC3 was significantly associated with both the d-prime 2-back performance (pFDR = 2.8 × 10-18, see Fig. 3b) and the d-prime 0-back performance with opposite direction of effects (pFDR = 7.6 × 10-7).”

We also added this information in the discussion section (line 641): “A network with prominent parietal and smaller frontal features was mainly associated with WM-related performance (5.8% variance explained), the associations with attention-related performance were smaller (2.2% variance explained) and in the opposite direction of effect.”

We now also applied FDR-correction to the results from our white matter analyses (result section, lines 400 and 403): “We tested separately for each WMN-IC whether the p-values of the associations between the 70 FA-values and WMN-IC scores deviate from the uniform distribution that is expected for continuous data under a simple null hypothesis (Murdoch, Tsai, & Adcock, 2008). The resulting p-values were FDR-corrected for six tests (α < 0.05). We additionally calculated empirical p-values based on the number of nominally significant associations for each WMN-IC after permuting the WMN-IC scores 10000 times, applying FDR-correction (α < 0.05) for six independent tests.”

“Related to this point, there appears to be a lack of appropriate discussion of prior related work that discusses distinct parietal and prefrontal contributions to WM and attention (for recent examples, see Kundu et al., 2015, NeuroImage; Sarma et al., 2016, Nature Neuroscience; Li et al., 2017, NeuroImage). Given this large body of existing work, it is important that the authors address what advance the current approach provides.”

We thank the reviewers and the editor for drawing our attention to this literature. We updated the Introduction and Discussion accordingly:

Introduction section (line 56): “Recent studies suggest that frontal and parietal regions differ regarding their contributions to WM. Neuronal recordings in the prefrontal cortex (PFC) and the lateral intraparietal (LIP) region of monkeys showed that encoded stimuli were retained in both regions, with more task-specific mnemonic encoding in the LIP as compared to the PFC (Sarma, Masse, Wang, & Freedman, 2016). Another study provided causal evidence for differing roles of parietal and frontal regions in attentional aspects of WM processing, by applying transcranial direct current stimulation. Stimulating the right parietal cortex increased the amount of information maintained in the visual WM, whereas stimulating the right PFC improved focusing on relevant information and directing attention away from irrelevant stimuli (Li et al., 2017). In addition, measuring the directed connectivity between the DLPFC and superior parietal lobule (SPL) during a visual WM task hinted towards a top-down drive from DLPFC to SPL that increased with WM load (Kundu, Chang, Postle, & Van Veen, 2015). These insights were based on a priori defined regions of interest (ROIs) and therefore described functional properties of separate brain regions.”

Discussion section (line 647): “Our findings of a frontally-centered and a parietally-centered network involved in different aspects of WM-task performances are in line with recent ROI-based studies that have reported distinct functional roles of frontal and parietal regions on WM (Kundu et al., 2015; Li et al., 2017; Sarma et al., 2016).”

We furthermore emphasize now on the differences between ROI-based approaches (as applied in these papers) and the ICA decompositions used in our analysis (discussion section, line 652): “Notably, using ICA-decomposition to estimate brain networks resulted in several advantages as compared to voxel-wise or ROI-based analyses. Both, voxel-wise and ROI-based approaches require prior knowledge, either about the subject's task performance or the anatomical regions of interest. In contrast, the ICA decompositions applied here were based on the WMN contrast estimates without including performance measures into the network estimation and without anatomical pre-selection of voxels. Thus, the WMN-ICs constitute data-driven and unbiased measures of brain networks that underlie the task performances.”

“Second, please be more clear about the logic of relating ICs to networks. An independent component is a basis function (a set of voxel weights) that when scaled appropriately explains variance in the voxel weights across subjects. A network is a set of locations that interact with each other. It is not self-evident that these are the same things, and it would strengthen the paper if the authors could be more explicit about why ICs with large eigenvalues are interpreted as networks.”

Many thanks for this important comment. We added now more information with respect to the ICA results and how we interpreted them in the context of fMRI data and networks:

Methods section (line 271): “By applying ICA decomposition to a matrix of 2-back–0-back contrast estimates, containing rows of voxels and columns of subjects, our source estimates (voxel loadings) described independent latent sources that underlie the contrast estimates. Accordingly, each component's mixing coefficients described the activity strength of each component for each subject (Chiapetta, Roubaud, & Torrésani, 2004). Subjects with high scores on an IC showed higher contrast estimates in voxels with high voxel loadings in this IC. Hence, we interpreted the subject scores as a measure of co-activation in the voxels that loaded onto the IC.”

Discussion section (line 663): “Subjects with high scores on a WMN-IC showed increased co-activation of the voxels that loaded highly onto this WMN-IC, we thus interpreted WMN-ICs as co-activating brain networks.”

“Also, it was not clear why you did subtraction of the fMRI signal (2B - 0B) but not the task performance? Since ICA was done on the subtracted regression coefficients, the ICA would be expected to reveal components that related to the difference in performance between the two tasks. There might have been a good reason to do this, but an explicit statement in the text was not found. This makes the correlation hard to interpret. Please clarify.”

We fully agree that associating the fMRI signal's contrast estimates with the difference in performance between the two tasks is more obvious than associating them with the individual performances. Based on this comment, we added now additionally the subtracted performances (2-back – 0-back) as a measure of WM performance (see Table 1).

Results section line 512: “Two of the six components were associated with the difference of WM-related and attention-related performances (d-prime 2-back – d-prime 0-back; see Table 1). WMN-IC3 was positively associated with the performance difference (pFDR = 2.3 × 10-19, R2 = 0.06) and WMN-IC4 was negatively associated with the performance difference (pFDR = 2.8 × 10-6, R2 = 0.02).”

“In several places, you point out that ICA is able to discover networks that would not be discoverable with a voxel-wise analysis however what is the basis for this statement? At least by visual inspection, the two analyses look highly similar. Figure 5a looks like 3a and 5b looks like 3b. It is fine, of course, if two methods arrive at similar conclusions. But it was surprising to read in multiple places how essential the ICA was when it appears that both types of analyses lead to the same conclusion. Perhaps you could state the limits of the conclusion they would have been able to draw had they only done the voxel-wise analysis.”

This comment helped us realize that our initial statements appear confusing, hence they were removed. We now compared univariate analyses with the ICA results in a more objective manner, and highlighted the advantages of the WMN-IC estimates:

Introduction section (line 78): “ICA decomposition is a data-driven unbiased approach to retrieve a low-dimensional representation of a dataset, resulting in statistically independent signals (Kong, Vanderburg, Gunshin, Rogers, & Huang, 2008).”

Discussion section (line 650): “The functional results from our estimated networks were also consistent with the voxel-wise results from our data. Notably, using ICA-decomposition to estimate brain networks resulted in several advantages as compared to voxel-wise or ROI-based analyses. Both, voxel-wise and ROI-based approaches require prior knowledge, either about the subject's task performance or the anatomical regions of interest. In contrast, the ICA decompositions applied here were based on the WMN contrast estimates without including performance measures into the network estimation and without anatomical pre-selection of voxels. Thus, the WMN-ICs constitute data-driven and unbiased measures of brain networks that underlie the task performances. Importantly, ICA-decomposition optimized the detection rate of true effects by considerably decreasing the number of tests performed, from N = 26542 voxel-wise tests to 6 association analyses with the WMN-ICs. Furthermore, ICA-decomposition enabled us to represent intrinsic brain networks with distinct functional characteristics. Subjects with high scores on a WMN-IC showed increased co-activation of the voxels that loaded highly onto this WMN-IC, we thus interpreted WMN-ICs as co-activating brain networks. The identification of distinct functional networks within the WM brain activation is in line with numerous recent studies demonstrating that the brain activation at rest as well as during different tasks is most likely based on independent but possibly spatially overlapping networks (Cole et al., 2014, 2016; Power et al., 2011; Xu et al., 2016). Importantly, due to the large sample size used here we can provide robust network estimates that can also be applied to samples with smaller sample-sizes.”

“Finally, the diffusion analysis produced an interesting result which, at least superficially, appears to conflict with the main results (or at least the way the main results are summarized in the text). Namely, IC3, the network the authors interpret as being most closely related to working memory activity, is mostly parietal. This is a central claim in the abstract and throughout. Yet the brain areas whose FA is most correlated with the IC3 weights is not parietal but rather precentral cortex. Perhaps it would make more sense for you to refer to IC3 as a parietal-precentral network, rather than a parietally-centered network? Please address this.”

Many thanks for this important comment. We added now some explanations on how white matter properties in the posterior ACC (i.e. not situated within WMN-IC3) could contribute to the co-activations in the parietally-centered network (discussion section line 701): “Subjects in our study with higher FA-values in the posterior cingulum, superior parietal, and precentral regions showed more co-activation of voxels in the parietally-centered network during the WM task. Interestingly, white matter bundles in the cingulum have been shown to connect the superior parietal lobe with the medial parts of the superior frontal gyrus (Wu, Sun, Wang, Wang, & Ou, 2016). Positive associations of FA with fMRI measurements or with connectivity measures could represent better transmission and stronger functional connections (Warbrick, Rosenberg, & Shah, 2017). Thus, white matter integrity might contribute to the observed WM performance-relevant co-activation.”

Furthermore we updated the anatomical annotation of the WMN-ICs, based on a probabilistic anatomical atlas. Because WMN-IC3 showed high loadings (z > 1.47) in 63% of the left and 52% of the right superior parietal cortex but only 12% of the left precentral cortex (see Fig. 3-1), we continued referring to WMN-IC3 as a parietally-centered network.

Methods section new paragraph (Line 292) – Cortical and subcortical labeling of the WMN-ICs: “Labeling of grey matter brain regions was based on a population-averaged probabilistic atlas (see below). The atlas comprises a total of N = 87 distinct cortical and subcortical brain regions from both hemispheres. Each of the N = 55614 voxels was assigned to one of these anatomical brain regions. Voxels for which the probability to belong to a given brain region was below 25% (N = 2926) or that were not located within cortical or subcortical regions (N = 21451) were excluded, resulting in N = 31237 voxels used for anatomical labeling. For each WMN-IC, we grouped voxels that showed the 10% most extreme values (see above) into clusters of adjacent voxels (WMN-IC clusters). Within each WMN-IC cluster and for each anatomical brain region, we determined the absolute number of voxels that belonged to this cluster and were annotated with this region. We report only brain regions comprising more than 10 voxels of a WMN-IC cluster. We also calculated the percentage of voxels per WMN-IC cluster and anatomical brain region by dividing the absolute number of voxels by the total number of voxels labeled with the anatomical brain region across the N = 31237 voxels.

The used population-average probabilistic anatomical atlas was built by automatic segmentation of the subjects' T1-weighted images. Each participant's T1-weighted image was first automatically segmented into cortical and subcortical structures using FreeSurfer (v4.5, http://surfer.nmr.mgh.harvard.edu/; Fischl et al., 2002). Labeling of the cortical gyri was based on the Desikan–Killiany Atlas (Desikan et al., 2006), yielding 35 regions per hemisphere. We also labeled 17 subcortical regions, following Fischl et al. (2002). The segmented T1 image was then normalized to the study-specific anatomical template space using the subject's previously computed warp field, and affine-registered to the MNI (Montreal Neurological Institute) space (see above, fMRI preprocessing). Nearest-neighbor interpolation was applied, to preserve labeling of the different structures. The normalized segmentations were finally averaged across subjects, to create a population-average probabilistic atlas. Each voxel of the template could consequently be assigned a probability of belonging to a given anatomical structure, based on the information of N = 1000 subjects that are part of the samples included in this study.”

Minor points:

Comment 1, part 1:

“It is not clear what the FA results add to the story. It seems like an afterthought that is not appropriately incorporated into the rest of the claims about WM networks. How should one interpret the association between the parietal IC and global FA (which, if I understand correctly, has FA measures that mostly come from brain regions outside of the parietal IC network)? ”

Many thanks for this comment. We now discussed how the FA in a brain region not situated within WMN-IC3 (i.e. the posterior ACC) could contribute to the co-activations in the parietally-centered network (discussion section line 704): “Interestingly, white matter bundles in the cingulum have been shown to connect the superior parietal lobe with the medial parts of the superior frontal gyrus (Wu, Sun, Wang, Wang, & Ou, 2016). Positive associations of FA with fMRI measurements or with connectivity measures could represent better transmission and stronger functional connections (Warbrick, Rosenberg, & Shah, 2017). Thus, white matter integrity might contribute to the observed WM performance-relevant co-activation.”

Comment 1, part 2:

“The abstract claims an association between FA and WM performance, but that's not explicitly addressed in the results.”

Apparently, the wordings ‘WM-dependent performance’ and ‘WM-related performance’ in the abstract were ambiguously formulated as they had been interpreted as ‘white matter dependent’ by one of the reviewers or editors. We therefore explicitly introduced the term ‘WM-dependent’ as ‘task measures related to WM performance’ in the abstract (line 10).

Comment 2:

“The manuscript needs copy editing. The references are not formatted correctly and there are duplications.”

We carefully went through the manuscript and the references to address these issues.

Comment 3:

“The authors claim that the large sample size allowed them to describe networks in a way that would not be possible with fewer than 100 subjects (Line 585). What is the evidence for this claim? Didn't the authors demonstrate that these same ICs can be identified with 100 subjects?”

Many thanks for this comment. We removed this claim and instead highlighted the robustness of the WMN-ICs in the discussion section (line 669): “Importantly, due to the large sample size used here we can provide robust network estimates that can also be applied to samples with smaller sample-sizes.”

We furthermore added the extended data figure Table 1-3 to illustrate the gains in statistical power when conducting analyses on WMN-ICs as compared to running voxel-wise analyses.

Comment 4:

“Line 234-235 ‘Since several changes on the scanner hardware and/or software over time were known...’ What does this mean?”

We replaced this sentence with the following sentence (Methods, line 248): “Across the timespan of the data acquisition, the gradient coils were changed twice (hardware batches), and parts of the scanner's software configuration were changed once (software batches). Additionally, the scanner console displayed irregularities during the data acquisition in a small group of subjects (processing batches). We regressed out these potential group-effects from the voxel-signal.”

Comment 5:

“Figure 7: I could not understand how values on x-axis of 7B and 7C were computed.”

We specifically explain now the Quantile-Quantile-plots in Fig. 8B and Fig. 8C in line 1090: “The Quantile-Quantile-plot compares the distribution of -log10(p)-values expected at random (x-axis) with the distribution of the observed -log10(p)-values (y-axis).”

Comment 6:

“The authors refer to ‘false-alarm corrected d-prime measures’. Since d-prime already accounts for false alarms, what does this mean? Do the authors just mean they used d-prime as a measure?”

We agree with the reviewer or editor that the wording “false-alarm corrected d-prime measures” was misleading. We changed the respective sentences in the methods section (line 151): “Task performances were defined as d-prime measures (Macmillan & Creelman, 1990). These measures account for false alarms and were calculated separately for the 0-back and 2-back conditions.”

Comment 7:

“The anatomical brain images were aligned volumetrically, and were smoothed quite a bit (8-mm Gaussian). Since all of the analyses were cortical structures, I wonder if they would have achieved better alignment using a surface-based registration, which tends to show better success for functional analyses on the cortex (though not sub cortex) (e.g., Wang et al, 2014, Cer Cortex, Probabilistic Maps of Visual Topography in Human Cortex)”

We agree that using a surfaced based registration, as suggested by a reviewer or editor, might improve the alignment of anatomical brain images in cortical regions. However, we were aiming at estimating components of WM task activation across the whole brain, including cortical and subcortical regions, in order to retrieve unbiased brain networks. We further emphasized this aim in the introduction (line 81): “We included both cortical and subcortical regions into the ICA decomposition to retrieve maximally unbiased estimates of brain networks.”

We furthermore highlighted in line 219 (methods section) that DARTEL had been shown to perform well in volume-based alignment when compared to other tools: “Subject-to-template normalization was done using DARTEL (Ashburner, 2007), which allows registration to both cortical and subcortical regions and has been shown to perform well in volume-based alignment (Klein et al., 2009).”

Comment 8:

“At one point, the authors talk about ‘IC activity’ (l 327). What does this mean? The ICs are static spatial patterns.”

‘IC activity’ had erroneously been used to describe the subject-wise WMN-IC scores. We changed the used term to ‘WMN-IC scores’ (methods section line 401).