Network-based statistics distinguish anomic and Broca aphasia

Participants were recruited from the local community in Columbia, South Carolina, as part of the Predicting Outcome of Language Recovery in Aphasia (POLAR) study of post-stroke aphasia. Only participants with a single ischemic or a hemorrhagic stroke in the left hemisphere were included. The participants with lacunar infarcts or with isolated damage in brainstem or cerebellum were excluded. The research was approved by the Institutional Review Board (IRB) at the University of South Carolina.

Aphasia types were classified based on the Western Aphasia Battery-Revised (WAB-R) (kertesz2007). Among the 49 participants included in the study sample, 15 were diagnosed with anomic aphasia, and 34 were diagnosed with Broca’s aphasia. Figure 1 shows a lesion overlap map of the participants in the two groups. Demographic statistics of the two groups are summarized in Table 1. The mean age in the anomic group was 62.73 y.o. (s.d. $=11.97$; range $=41$) and the mean age in the Broca’s group is 59.82 y.o. (s.d. $=10.35$; range $=39$). Respectively $60\%$ and $68\%$ of the participants in the anomic and Broca’s group were male. There was no significant difference in age and gender between the anomic and Broca’s group (age: $p$-value $=0.42$ by two-sample $t$-test; gender: $p=0.74$ by $\chi^{2}$-test). The mean WAB-R score for the anomic group was $85.74$ (s.d. $=6.38$; range $=22.1$) and $46.44$ (s.d. $=16.93$; range $=59.1$) for the Broca’s group. The WAB-R score for the anomic group is significantly higher than the Broca’s group ($p$-value $<0.01$ by two-sample $t$-test), since the participants with Broca’s aphasia tend to have a lower score in the section of fluency and repetition than the participants with anomic aphasia.

The rs-fMRI data were acquired on a Siemens Prisma 3T scanner with a 20-channel head coil located at the Center for the Study of Aphasia Recovery at the University of South Carolina. The following imaging parameters of images were used: a multiband sequence (x2) with a $216\times 216$ mm field of view, a $90\times 90$ matrix size, and a 72-degree flip angle, 50 axial slices (2 mm thick with $20\%$ gap yielding 2.4 mm between slice centers), repetition time TR =1650 ms, TE=35 ms, GRAPPA=2, 44 reference lines, interleaved ascending slice order (Grigori2018). During the scanning process, the participants were instructed to stay still with eyes closed. A total of 370 volumes were acquired. The preprocessing procedures of the fMRI data include motion correction, brain extraction and time correction. By applying the atlas of intrinsic connectivity of homotopic areas (AICHA) (Joliot2015), 384 regions of interest (ROIs) were created.

We used the NBS method to identify a subnetwork that differentiates connectivity patterns of the anomic and Broca’s aphasia groups. An illustration of the analytical pipeline is summarized in Figure 2. It builds on edge-level mass univariate testing with multiple comparison adjustment but refines the approach through a breadth-first search algorithm.

We first constructed functional connectivity matrices based on Pearson’s correlation between BOLD signals in the ROIs. The Fisher’s $z$-transformation was applied to the coefficients to enforce normality:

where $r$ was a correlation coefficient and $z$ was the corresponding normalized correlation coefficient. After normalizing the coefficients, a $p\times p$ connectivity matrix, with $p=384$ being the number of ROIs in the AICHA atlas, was constructed for each participant in the two aphasia groups. The $(i,j)$-th entry of such a matrix denotes the weight of an edge or connection between the $i$th and $j$th ROIs of the corresponding network. As the connectivity matrices are symmetric, we only included the $p(p-1)/2$ upper or lower triangular entries in the analysis.

To identify differences in the resting-state connectivity of the anomic and Broca’s groups, we tested the null hypothesis

where $\mu_{1j}$ and $\mu_{2j}$ are the respective $j$-th population means of the vectorized $p(p-1)/2$ edge weights. That is, we assumed no difference in the resting-state connectivity between the two groups under the null hypothesis. Under the alternative hypothesis, we assumed that at least one pair of mean edge weights are different between the two groups, i.e.

The test statistic for comparing the $j$-th edge weights is

where

with $\bar{x}_{j}$ and $\bar{y}_{j}$ being the respective $j$th sample means of the $p(p-1)/2$ edge weights, $s_{1j}$ and $s_{2j}$ being the respective standard deviations, and $n_{1}$ and $n_{2}$ being the respective sample sizes of the two groups. In total, $p(p-1)/2$ test statistics were computed.

We then obtained a binary adjacency matrix by applying a chosen threshold to the edge-level test statistics. If a test statistic exceeded the threshold, the corresponding edge weight of the adjacency matrix was set to $1$ and otherwise 0. The edges that remained after the thresholding are called the suprathreshold edges. Connected component(s) or subnetwork(s) (there may be more than one), where all nodes in each subnetwork were linked by suprathreshold edges, were identified from the adjacency matrix through the breadth-first search algorithm (Ahuja1993; Hopcroft1973). An illustration of the algorithm is included in Figure 2.

Similar to cluster-based thresholding of statistical parametric maps (Nichols2002; Hayasaka2004), we conducted a random permutation test to determine the statistical significance of the subnetwork(s) and control the family-wise error rate (FWER). In each permutation, participants in the two groups were shuffled and randomly divided into two new groups. Test statistics were computed by (2) based on the new grouping. The threshold imposed on the test statistics was the same for all permutations. Denoting the maximum of all possible subnetwork sizes in a permutation as $S^{*}$, the $p$-value for any observed subnetwork of size $S$ was calculated by $Pr\left(S^{*}\geq S\right)$, depicting how likely the size of the largest subnetwork in a permutation exceeds the size of the observed subnetwork. When the $p$-value was less than a certain significance level, which is set to $5\%$ in this study, we concluded that this subnetwork is significantly different between the two aphasia groups.

For baseline comparison with the NBS method, mass univariate permutation testing with multiple comparison was performed. The $p$-value for comparing the $j$th edge weights with (2) was calculated by $Pr\left(|t^{*}_{j}|\geq|t_{j}|\right)$, where $t^{*}_{j}$ and $t_{j}$ were the test statistics computed from permuted data and original data respectively. This $p$-value describes how likely the absolute value of test statistic from permuted data exceeds the absolute value of test statistic from original data. Then, the following multiple comparison procedures were applied to correct multiple $p$-values and control the FWER: Bonferroni correction (Bonferroni1936), Holm’s Bonferroni correction (Holm1979), and false discovery rate (FDR) control (Benjamini1995).

We used three simulation studies to evaluate the empirical performance of the NBS method against mass univariate testing with multiple comparison. We assessed two aspects of the performance: (1) sensitivity or true positive rate (TPR): the proportion of connections or edges containing group differences that are correctly identified; (2) $1-$ specificity or false positive rate (FPR): the proportion of edges without differences that are misclassified. Ideally, TPR $=1$ (all edges that differ between the two groups are identified), and FPR $=0$ (all edges that do not differ between the two groups are not identified). Suppose $H$ is the set of edges that differ between the groups, $R$ is the set of edges that do not differ between the groups, and $\hat{h}$ is the set of edges comprising the subnetwork identified by a specific method (NBS, baseline mass univariate testing with multiple comparison). The TPR was then calculated by $|H\cap\hat{h}|/|H|$ and the FPR by $|R\cap\hat{h}|/|R|$.

In each of the three studies, we generated two groups of $p$-node networks. The group sizes are $n_{1}=n_{2}=10$ for all three studies. In each network, the weight of the edge between node $i$ and $j$ was generated by $w_{ij}\sim N(arctanh(r_{ij}),\sigma_{w}^{2})$ with $r_{ij}\sim U(-1,1)$ and $\sigma_{w}=1/\sqrt{p(p-1)/2-3}$. We randomly chose $C_{r}\%$ of $p(p-1)/2$ edges to differ in weights between Group 1 and 2, and refer to these edges as contrast edges. The weights of contrast edges were generated independently with

where $w^{*}\sim N(0.03,0.01)$. We compared the performance of the methods via different $p$ and $C_{r}$ values in the three studies.

Study 1.

$p=20$ and $C_{r}\%=10\%$.

Study 2.

$p=40$ and $C_{r}\%=10\%$.

Study 3.

$p=40$ and $C_{r}\%=5\%$.

After the networks were generated, NBS with threshold 2.5 and mass univariate testing with multiple comparison were performed (Figure 3). We repeated the simulation process 5,000 times for each study. Average TPR and FPP were computed respectively.

We used complex network measures to quantify the topological properties of the subnetwork (Rubinov2010) (Figure 4). We focused on the functional segregation and centrality measures. Functional segregation is the ability of processing a certain task to occur within densely interconnected ROIs (Rubinov2010). The clustering coefficient is a basic measure of functional segregation based on the number of triangles around a given ROI, where each triangle is formed by the given ROI and two other distinct ROIs, and the three edges connecting them. The clustering coefficient of the ROI quantifies how well connected this ROI’s neighbors are, which is equivalent to the number of triangles around the ROI divided by the number of edges that could possibly exist between the ROI and its neighbors. The ROI with a clustering coefficient that is close or equal to $1$ implies that the other ROIs in this subnetwork cluster around it. On the other hand, central nodes or hubs are the ROIs with central placement in the overall subnetwork structure. They play an important role in communication and integration in the subnetwork (Vandenheuvel2013). Degree, a simple measure of centrality, is defined as the number of edges directly linked to a given ROI. The larger the degree is, the more central the ROI is. Betweenness is another measure of centrality defined at a given ROI as the fraction of all shortest paths in the subnetwork, which are paths with the smallest numbers of edges between two ROIs in the subnetwork, passing through the ROI. An ROI with high betweenness is regarded as a bridge connecting the other ROIs in the subnetwork.

We used three traditional mass univariate methods: voxel-, region-, and connectivity-based lesion symptom mapping (VLSM, RLSM, CSLM). Whole brain V- and RLSM was used to identify brain damage associated with aphasia type (anomic or Broca’s). VLSM shows the statistical likelihood that damage to a given voxel is associated with aphasia type group membership, where each voxel in each patient is binarily demarcated as either damaged or undamaged (Bates2003). RLSM differs from VLSM in that instead of using binary voxel-wise values, it uses the percent of voxels damaged within each ROI as the predictor of aphasia type. This sacrifices spatial specificity while providing the advantage of analyzing the effects of damage over an entire region without requiring overlapping damage at level of an individual voxel. We conducted RLSM using the AICHA ROIs. We then conducted CLSM (Gleichgerrcht2017) using resting-state functional connectivity based on the AICHA atlas, including all left-to-left, left-to-right, and right-to-right connections in the analysis. Only voxels (or regions for RLSM) where at least 5 patients had damage were considered, based on the minimum overlap recommendation of  10% of the patient sample (Baldo2022). All tests were two-tailed, with $\alpha=0.05$, and significance was determined via permutation testing, where stability of $p$-value were tested in increments of 1000 permutations, ranging from 1,000 permutation to 10,000 permutations.

On top of whole-brain analysis, we also restricted the analysis to the ’dorsal stream’ areas, i.e. frontoparietal and superior temporal areas that are involved in form-to-articulation during speech (Fridriksson2016). These areas would be hypothesized to be especially disrupted in individuals with Broca’s aphasia who struggle with many aspects of speech production compared to the relatively mild anomic cases where the individuals just have occasional word-finding difficulties. We included the AICHA ROIs corresponding to supramarginal gyrus, primary sensory and motor cortices, inferior frontal gyrus (Broca’s area), superior temporal gyrus, and rolandic operculum. This allowed us to restrict the # of connections while also allowing us to use a one-tailed analysis since we specifically hypothesized these connections would be associated with Broca’s aphasia. It is worth noting that we also tried the alternate analysis, using a different set of language regions that might be implicated in anomic aphasia more than Broca’s, but this did not reveal any significant results. This is likely because anomic aphasia as a behavioral syndrome may be caused by deficits at various functional levels within the language production system (conceptual, lexical, semantic, phonological, for example), so that similar surface behavior may result from different patterns of neural damage. In addition, in our own sample anomic aphasia was ’less severe’ than Broca’s aphasia, on average, which would also make the detection of areas specifically related to the anomic group more difficult.

Table 2 summarizes results of the three simulation studies. In Study 1, the NBS method has the largest TPR and FPR, while the mass univariate testing with FDR detects a desirable proportion of contrast edges and contains a small number of false discoveries. In Study 2, as the network size expands, the TPRs by the NBS and mass univariate testing methods increase, whereas the FPRs increase by the NBS and FDR methods and decrease by the Bonferroni and Holm’s Bonferroni methods. In Study 3, when fewer contrast edges are placed in either of the two groups, the TPRs by the NBS and FDR methods decrease and the FPRs by the NBS and the two Bonferroni corrections stay similar to Study 2.

In summary, compared with mass univariate testing, the NBS method detects small group differences well under various network sizes and proportions of contrast edges. For the mass univariate testing, the FDR has the highest power and a favorable FPR; Bonferroni and Holm’s Bonferroni corrections are highly conservative in detecting contrast edges. Additionally, we find that the computation speed is mainly affected by network size.

Applying the NBS method to the POLAR rs-fMRI, we identified a subnetwork distinguishing the anomic and Broca’s groups with respect to each of the thresholds 3, 3.5, 4 and 5. As the threshold increased, the subnetowrk size and the number of ROIs contained in the subnetwork decreased (Table 3). Only one subnetwork was identified when the threshold was set to 4. Subsequent results all pertain to this subnetwork. Figure 5 shows connections comprising this subnetwork. Except for one connection within the left precentral sulcus (between middle and superior precentral sulcus), the majority of connections in the networks from the group with anomic aphasia were stronger than the networks from the people with Broca’s aphasia. The subnetwork was nearly bilaterally symmetric and primarily involved the ROIs in the premotor, primary motor, primary auditory, and primary sensory cortices. A few ROIs in the left Broca’s area and cingulate cortex were also included in this subnetwork. In comparison, mass univariate testing with multiple comparison also identified several connections that were significantly different between the two aphasia groups. By comparing the results in the two methods, we found that the majority of significant connections declared by mass univariate testing existed in the subnetwork identified by the NBS with threshold 4 and above. When the threshold in the NBS was less than 4, a large amount of connections in the subnetwork were not detected by mass univariate testing.

For topological properties of the subnetwork identified by NBS with threshold 4, the clustering coefficient, degree, and betweenness as described in the previous section were computed (Figure 6). The clustering coefficient of the ROIs in the superior temporal gyrus (STG) and auditory cortex (Aud) in the right hemisphere reached the maximum value of 1. The clustering coefficient of the ROIs in the left STG and Aud was also higher than other ROIs in this subnetwork. Regarding the centrality, we found there were more than 10 connections linked to the ROIs in the left premotor, primary sensory, and primary motor cortices and thus these ROIs had a larger degree than others. Furthermore, the ROIs in the left premotor, primary motor cortices and the right supramarginal gyrus exhibited a high betweenness.

Whole-brain VLSM revealed a small cluster of voxels in the left superior temporal gyrus where damage was significantly associated with Broca’s aphasia group membership (peak $z=4.55$, $p<.001$; Figure 7). RLSM showed that percent of voxels damaged within 9 AICHA ROIs on the left hemisphere was associated with Broca’s aphasia group membership, including portions of STG, MTG, and insula (Figure 7). On the other hand, no connections were significant in the CLSM analysis. Using the set of AICHA ROIs hypothesized to be more involved in Broca’s aphasia than anomic, only the connection from G_Rolandic_oper_1-R to G_Temporal_Sup_1-L survived and is significantly weaker in the Broca group ($z=-4.12$). Restricting the analysis to just interhemispheric (left to right) connections revealed two additional connections (left posterior insula to right superior temporal gyrus, and left superior temporal gyrus to right superior temporal gyrus). Overall, significant results remained unchanged as we increased the # of permutations from 1,000 up to 10,000. Critical values only changed slightly depending on # of permutations.

The research was approved by the Institutional Review Board (IRB) at the University of South Carolina.

None.