How Can Graph Theory Inform the Dual-stream Model of Speech Processing? A Resting-state Functional Magnetic Resonance Imaging Study of Stroke and Aphasia Symptomology

In the past two decades, the dual-stream model of speech processing ¹ (Hickok & Poeppel, 2000, 2004, 2007) has, arguably, emerged as the prominent functional anatomical model of speech perception and production (Keator, Yourganov, Faria, Hillis, & Tippett, 2022; Rogalsky & Hickok, 2009; Spitsyna, Warren, Scott, Turkheimer, & Wise, 2006; Crinion & Price, 2005; Vandenberghe, Nobre, & Price, 2002; Humphries, Willard, Buchsbaum, & Hickok, 2001). During this same time, the connectivity of brain regions with similar response properties has emerged as a valuable predictor of behavioral performance and neurological disease (Keator et al., 2021; Zhang, Kucyi, et al., 2021; Battistella et al., 2020; Stone et al., 2020; Ferguson et al., 2019; Fitzhugh, Hemesath, Schaefer, Baxter, & Rogalsky, 2019; Walsh, Baxter, Smith, & Braden, 2019; Avants, 2018; Zhu et al., 2016). However, it remains unclear if the dual-stream model of speech processing represents an intrinsic, coherent brain network, and, if so, the nature of its organizational structure and how damage to one part of the network affects overall network function and speech performance remains unclear. Resting-state fMRI is a particularly promising, emerging tool to explore the dual-stream network as it relates to poststroke aphasia severity and specific impairments.

It is well established that low-frequency BOLD signal fluctuations at rest indicate coherent language-related neural activity (Battistella et al., 2020; Xu, Huang, Cui, & Yu, 2020; Zhao, Ralph, & Halai, 2018; Tie et al., 2014; Binder et al., 1999). For instance, a recent study by Battistella and colleagues (2020) identified several language-related networks with resting-state fMRI in young neurotypical adults, including dorsal articulatory-phonological and ventral semantic networks. Functional connectivity studies have also found significant functional connectivity across language-related areas in control participants, such as between sensorimotor areas and expressive and receptive language regions (Hampson, Peterson, Skudlarski, Gatenby, & Gore, 2002; Cordes et al., 2000). Thus, there is strong evidence that nodes of language networks are connected at rest; however, the architecture of these networks at rest have not been well mapped, particularly in older adult control participants and stroke survivors.

Previous work has also begun to explore resting-state functional connectivity in individuals with stroke and aphasia, generally finding that functional connectivity differences are associated with treatment outcomes and behavioral measures of speech-processing abilities (e.g., Stockbridge et al., 2023; Xie et al., 2023; Chen et al., 2021; Klingbeil, Wawrzyniak, Stockert, & Saur, 2019; Siegel et al., 2016; Carter, Shulman, & Corbetta, 2012). A recent study conducted during the subacute phase of stroke revealed that participants who responded more to treatment exhibited significantly increased connectivity between the left fusiform gyrus and both ipsilateral and contralateral regions such as Broca's area and the superior temporal gyrus, compared with their lower responding counterparts. This finding suggests a potential restoration of proximal connectivity and compensatory reorganization in contralateral regions (Stockbridge et al., 2023). Another study observed that individuals with aphasia displayed lower local efficiency in nodes across the whole brain compared with the control group, and local efficiency was positively correlated with language performance (Chen et al., 2021). The authors' interpretations of their findings are informed by the dual-stream model of speech processing, but it remains unclear if the dual-stream model is in fact is a network at rest and, if so, how its connectivity is affected by stroke.

The functional connectivity studies described above typically employed anatomically defined ROIs as seeds to identify correlations with other ROIs or voxels that predicted speech processing abilities. In addition to these traditional functional connectivity methods to characterize networks in resting-state fMRI, graph-theoretical approaches may be a powerful method to better understand the network properties of the dual-stream model's regions because they can identify hub regions that play pivotal roles in the network organization, and they can characterize additional network-level properties, which are not reflected in simple functional connectivity correlation measures (He & Evans, 2010; He, Chen, & Evans, 2008). Typically, graph-theoretical analyses consist of a series of nodes and edges corresponding to the constituent units and interactions of a network (Liao, Vasilakos, & He, 2017; Barabási, 2012; Boccaletti, Latora, Moreno, Chavez, & Hwang, 2006). Unlike traditional functional connectivity that focuses on the relationship between a few regions, graph theory tends to describe the organization of the whole network or the local node cluster. After building a graph, hub regions can be defined as the average shortest path length (Achard, Salvador, Whitcher, Suckling, & Bullmore, 2006). The shortest path length is the number of edges comprising the shortest path between any pair of nodes; the average shortest path length is the mean of the n – 1 minimum pathway between the index node and all other nodes in the networks (Achard et al., 2006).

A few studies related to language in adult control participants and children have applied the graph theory to depict the interactions between brain regions known to be involved in speech processing, with greater performance on various language measures being related not only just to increased functional connectivity but also to increased connectivity of hub regions, as defined by graph theory approaches (Xiao, Friederici, Margulies, & Brauer, 2016; Muller & Meyer, 2014). In the area of degenerative brain disease, hub regions have been shown to be more vulnerable (Achard et al., 2006; Bassett & Bullmore, 2006; Albert, Jeong, & Barabási, 2000), in that damage to or disruptions of hub regions are typically more correlated with more severe neurocognitive impairments than damage or disruption to nonhub regions of a network (Roger et al., 2019; Stam, De Haan, Daffertshofer, Jones, & Manshanden, 2009; He et al., 2008). However, to our knowledge, no previous studies have examined the functional network connectivity of the dual-stream model using graph theory, and only a few studies have investigated stroke-induced language impairments using a graph theory more broadly (Mazrooyisebdani, Nair, Garcia-Ramos, & Prabhakaran, 2018; Duncan & Small, 2016; Bohland, Kapse, & Kiran, 2014). These previous studies found that some network-level properties were different in individuals with aphasia compared with control groups and that these properties were correlated with aphasia severity or specific types of language impairments. However, the discussion of functional connectivity between the hubs was absent. By moving beyond the more typical functional connectivity analyses and using graph theory, we will be able to examine network organization and reorganization in stroke survivors in new ways that can potentially help us better understand poststroke compensation and language impairments.

In the present study, we not only extend the study of the dual-stream model to resting-state fMRI but also implement graph theory approaches in both stroke survivors and neurotypical control participants to characterize how the dual-stream network differs because of stroke and is related to language abilities poststroke. We have three hypotheses: (1) The neurotypical control participants will exhibit strong resting-state functional connectivity patterns across dual-stream nodes that have been defined by previous task-based findings; (2) the stroke survivors, in comparison to the control group, will exhibit functional connectivity and graph theory parameter differences; and (3) in stroke survivors, differences to the hub regions defined by the graph theory will predict impairments in our behavioral language measures.

Twenty-nine chronic left-hemisphere stroke survivors were recruited and tested at the University of South Carolina as part of an ongoing study of the neurobiology of speech and language. Sample size was based on prior studies using similar outcome measures that yielded robust findings with about 30 or less participants in each group (Chen et al., 2021; Fitzhugh, Schaefer, Baxter, & Rogalsky, 2021; Fitzhugh et al., 2019; Khazaee, Ebrahimzadeh, & Babajani-Feremi, 2015; Braun et al., 2012; Liu et al., 2008). One of the stroke participants' data was removed from the present study's analyses because of substantial head motion during scanning. Remaining stroke survivors range in age from 35 to 78 years (M = 60.00, SD = 11.47). Inclusion criteria included chronic stroke (>6 months before testing), right-handed prestroke, native speakers of American English, 18+ years of age, and with no history of neurological disease, head trauma, or psychiatric disturbances before their stroke (self-report and supported by review of MRI scans by a clinical neurological radiologist). Aphasia classification was determined by the Western Aphasia Battery–Revised (WAB; Kertesz, 2007); each stroke participant's aphasia diagnosis is reported in Table 1. The lesion overlap map (Figure 1) and proportion of damaged ROIs (Table 4) characterize the distribution of lesion locations for the stroke group.

Twenty-eight neurotypical adults ranging in age from 20 to 79 years (M = 58.79 years, SD = 18.72 years) who were also right-handed, native speakers of American English, 18+ years of age, have a minimum score of 27 on the Mini-Mental State Examination (Tombaugh & McIntyre, 1992), with no history of neurological disease, head trauma, or psychiatric disturbances were recruited from the greater Phoenix, Arizona, area. There were no significant differences between groups on age or sex, but the control group did have significantly more years of education than the stroke group, 15 versus 17 years, t(54) = −2.42, p = .02 (see Table 2).

This study was approved by Arizona State University institutional review board. All participants provided written informed consent.

MRI data of the stroke group were acquired on a 3 T Siemens scanner equipped with a 12-channel radiofrequency head coil at Prisma Health Richland Hospita l (Columbia, South Carolina). A T2 structural image was collected with the following parameters: echo time = 57 msec, image size: (176 256 256), voxel size = 1 × 1 × 1 mm. Resting-state fMRI data were collected using EPI with the following parameters: one 11-min run, 427 total volumes, repetition time (TR) = 1650 msec, percent phase field of view (FOV) = 100, 2-mm slice thickness, 90 × 90 matrix size of 2.4-mm voxels.

MRI data of the control group were acquired on a 3 T Phillips Ingenia MRI scanner equipped with a 32-channel radiofrequency head coil located at the Keller Center for Imaging Innovation at the Barrow Neurological Institute in Phoenix, Arizona. A T1 image was collected with the following parameters: FOV = 270 × 252, TR = 6.74 s, echo time = 3.10 msec, flip angle = 9, voxel size = 1 × 1 × 1 mm. Resting-state fMRI data were collected using EPI with the following parameters: one 10-min run, 197 total volumes, TR = 3000 msec, FOV = 217 × 217, matrix = 64 × 62, 3.39-mm slice thickness, in-plane resolution = 3.39 × 3.39 mm. Although the stroke and control groups' data were acquired from different scanners, there is evidence that interscanner variability effects on functional connectivity, such as we use here, is limited (Belleau et al., 2020; Marek et al., 2019; Noble et al., 2017).

The language abilities of the stroke group were evaluated by the administration of the WAB (Kertesz, 2007). The WAB's aphasia quotient indicating overall aphasia severity, as well as the scores of the following subtests, were used in subsequent analyses: Spontaneous Speech, Auditory Verbal Comprehension, Repetition, Naming, and Word Finding.

All resting-state fMRI data were preprocessed using Statistical Parametric Mapping 12 (https://www.fil.ion.ucl.ac.uk/spm). The first 10 time points of each run were discarded to ensure that the magnetization reached a steady state and the participants adapted to the environment. Slice timing was adjusted to compensate for the interleaved acquisition in the remaining 187 volumes for control and 427 volumes for stroke participants. Next, realignment was conducted to correct the head motion using the six standard head motion parameters. Then, the structural image (i.e., the T1-weighted image for the control group and the T2-weighted image for the stroke group) was reoriented to the mean functional image. Diffeomorphic anatomical registration through exponentiated Lie algebra normalization (Ashburner, 2007) was used to segment the structural image to white matter (WM), gray matter, and cerebral spinal fluid and normalize it to the Montreal Neurological Institute (MNI) template. Using the normal parameters of the structural image generated by diffeomorphic anatomical registration through exponentiated Lie algebra normalization, the functional images were spatially normalized to MNI space. Nuisance covariates including the WM signal, the cerebral spinal fluid signal, and head motion parameters were regressed out from the functional signal. To further ensure that our results are not driven by motion effects, we examined the relationship between motion and functional connectivity by computing correlations between the network-level functional connectivity and the mean motion measurement (from the six head motion parameters extracted during preprocessing; Van Dijk, Sabuncu, & Buckner, 2012), for each part of the dual-stream speech-processing network that we investigated. No significant correlations were found, and thus, no further action was taken regarding motion effects. We then applied band-pass filtering between 0.01 and 0.1 Hz and spatial smoothing with an 8-mm FWHM Gaussian kernel to facilitate group analyses. For the stroke group's fMRI data processing, we also added a cost function masking step (Andersen, Rapcsak, & Beeson, 2010; Brett, Leff, Rorden, & Ashburner, 2001) to avoid the risk that a stretched or altered perilesion tissue would impair normalization. The mask used in this step was generated from a manual demarcation of the chronic stroke lesion as seen on the T2-weighted images by researchers trained in neuroanatomy and lesion mapping procedures.

There are several meta-analysis studies and reviews that identify left dominant regions that are involved in speech processing, including regions in the left frontal, temporal, and parietal lobes (Klingbeil et al., 2019; Mesulam et al., 2014; Margulies & Petrides, 2013; Price, 2012; Dronkers, 2011; Vigneau et al., 2006), which is in line with the dual stream model. To depict these comprehensive speech processing areas, we used 32 nodes from a previous task-based fMRI study in the left hemisphere, which used three contrasts including sentence production, listening, and reading, with corresponding word-list reference tasks (Labache et al., 2019). These regions are highly consistent with typical language areas found in other studies and reviews. Out of the Labache and colleagues coordinates, the coordinates within the precentral sulcus, superior frontal gyrus, inferior frontal sulcus, Broca's area, anterior insula, supramarginal gyrus, and posterior superior temporal sulcus were classified as dorsal; other coordinates in the superior temporal gyrus, middle temporal gyrus, superior temporal sulcus, and angular gyrus were classified as ventral, based on the location. All the nodes in the left hemisphere are listed in Table 3. Table 4 indicates the proportion of the participants in the stroke group who have lesions within the boundary of each node (i.e., a 6-mm radius sphere around the peak coordinates), as defined by the lesion demarcation procedure based on the T2 structural MRIs. To examine the influence of structural lesions within the functional connectivity nodes on subsequent analyses, we also identified stroke participants with lesions in more than 50% of the left hemisphere dual-stream nodes' centers of mass (i.e., lesions on the Labache et al., 2019, coordinates used to define the nodes). Four participants were identified using this threshold; the between-group comparisons and regression analyses within the stroke group described below were then conducted with and without these four participants.

Functional connectivity was computed between nodes. The nodes were defined as 6-mm radius spheres around the peak coordinates reported in Labache and colleagues (2019). To better fit the dual-stream model's bilateral ventral streams, we also included the homologue of temporal and parietal nodes in the right hemisphere to study the activity in the right ventral (RV) stream. We then calculated the Pearson correlation coefficients between each pair of nodes; correlation coefficients were then Fisher transformed. To explore possible functional reorganization in the right hemisphere in response to the left dorsal (LD) stream damage, we also included potential right dorsal (RD) stream nodes using the homologue coordinates of the left hemisphere's dorsal stream's nodes. All the functional connectivity analyses were performed using in-house scripts executed in MATLAB. The average functional connectivity matrices of the control and stroke groups are shown in Figure 2.

To confirm that the neural correlates of the dual-stream model do in fact form a significant network in our neurotypical control participants, we computed the following paired-samples t tests within the control group: (1) comparison of the average functional connectivity within the entire, bilateral dual-stream network, as well as the average connectivity in just the left-hemisphere nodes of the dual-stream network (i.e., between and within the LD and left ventral [LV] streams), with that of within the well-defined default mode and visual networks, and (2) comparison of the average functional connectivity between each stream of the dual-stream network (i.e., bilateral ventral and LD) to the average functional connectivity between the dual-stream nodes and those of the default mode network and visual network. The coordinates of the default mode network nodes (Table 5) used have been well defined by numerous previous resting-state and task-related fMRI studies (Zhu et al., 2016; Gao & Lin, 2012; Vincent, Kahn, Snyder, Raichle, & Buckner, 2008). The visual network coordinates (Table 5) were taken from a previous task-based fMRI study (Gao & Lin, 2012), which aligns well with numerous other works identifying this network (Tedeschi et al., 2016; Lee, Smyser, & Shimony, 2013; Schöpf et al., 2010; Van Den Heuvel & Pol, 2010). False discovery rate (FDR) correction was used to control the false positives at p < .05. All the statistical analyses including paired t tests, two-sample t tests, regression models, and multiple comparison corrections were coded in R (https://www.r-project.org/).

For the stroke group (in the same fashion as for the control group, described above), we computed the overall connectivity of the dual stream model network, as well as the left-hemisphere streams only and each stream separately, and compared them to that of the visual and default mode networks using paired-samples t tests with FDR correction to control false positives at p < .05.

To compare the functional connectivity of the dual-stream nodes of the stroke group to those of the control group, we computed independent-samples t tests of the average functional connectivities between the stroke and control groups for the following groups of nodes: within left ventral, within LD, between LV and LD, between RD and RV, within RD, within RV, between left and RV, between LD and RD, and between LV, LD and RV. An FDR correction was used to control the false positives at p < .05.

Using the functional connectivity correlations described above as the edges between the nodes, we used graph theory methods to extract the average shortest path lengths for each node to identify hub regions. The parameters for the graph construction are summarized in Table 6 (Achard et al., 2006; Humphries, Gurney, & Prescott, 2006). To obtain the average shortest path length, first, a binary undirected graph at varying thresholds of the functional connectivity was constructed, to identify an appropriate threshold for the subsequent analyses using small-world properties. In order for a graph to be classified as a small-world network, it must satisfy three requirements: (1) the ratio of cluster coefficient γ = C_net/C_ran > 1, (2) the ratio of path length λ = L_net/L_ran ∼ 1, and (3) the “small worldness” ratio σ > 1 (Achard et al., 2006; Humphries et al., 2006; Montoya & Solé, 2002; Watts & Strogatz, 1998). To compute C_ran and L_ran, we applied randomized manipulation by a Markov-chain algorithm (Liu et al., 2008; Maslov & Sneppen, 2002) 100 times for each potential threshold and obtained C_ran and L_ran from the mean of 100 random networks. Then, we found an optimized threshold to attain a best balance between γ, λ, and σ (γ > 1, λ ∼ 1, and σ > 1), using a search of the proportion of possible connections that actually exist between all the nodes. To ensure that the network is not too sparse and thus causing the small-world properties to be inestimable (Achard et al., 2006), the search began at K*N/T = 13.2% (see Table 6 for variables). The current study has 64 nodes for the graph theory analysis. The minimal edge number should be K * N, thus, in this case, equal to 266. Then, the maximum possible of edges (T) of a 64 × 64 matrix is 2016. Finally, K * N/T = 266/2016 is the minimum percentage (i.e., the beginning of the search). The end of the search is 20%, which was the maximum threshold to account for the known sparsity of a small-world organization (Roger et al., 2019; Achard & Bullmore, 2007).

We then identified “hub” nodes based on the mean shortest path lengths computed. Again, the mean shortest path length formula of a node is: $Li=1N−1∑i≠jminLi,j$⁠. To be identified as a hub of the dual-stream network, a node must satisfy two criteria: (1) be one of the two nodes with the smallest values in the LD, right dorsal, left ventral, or RV streams, and (2) be a node with functional connectivity to at least half of the other nodes either directly or mediated by one or more nodes that passes the threshold defined above. Then, to continue our exploration of a possible right dorsal stream involvement in aphasia, we labeled the right homologues of the hubs in the LD stream as right dorsal hubs. The hub nodes defined above in the control group were used to calculate the functional connectivity between hub nodes in the stroke group. Independent-samples t tests then were used to detect differences between the control and stroke groups for functional connectivities within and between the streams' hub nodes. Once again, the significant results were FDR corrected to p < .05.

The functional connectivities within each stream were used as predictors in multiple regression models (lm command in R) to predict each stroke survivor's performance on each WAB behavioral measure. The selection of the predictors included to predict each WAB measure was based on previous lesion-symptom mapping and task-based fMRI work (Kertesz, 2022; Fridriksson et al., 2018; Thye & Mirman, 2018; Baldo, Arévalo, Patterson, & Dronkers, 2013; Kümmerer et al., 2013; Shulman et al., 1997) and is summarized in Table 7. Lesion size, age, sex, and years of education were also included in each model as covariates. We applied a reduction of predictor variables using Akaike-information-criterion-based elimination in R (step order) to remove the predictors or covariates that contribute little to each model—a standard procedure to remove predictors that are not related to the response variable that could lead to an inflated prediction error (Chambers & Hastie, 1992). To identify potential collinearity between the hub functional connectivity predictors, we (1) implemented a multicollinearity test such that any pairs of predictors with a |r| > 0.7 (Dormann, Elith, Bacher, et al., 2013) would be considered as too highly correlated and one of the predictors should be removed to not inflate results, and (2) computed the variance inflation factor (VIF) for each predictor in each model; VIF > 5 would lead to the exclusion of that predictor in that model (Shrestha, 2020).

The paired-samples t tests computed within the control group to compare the mean functional connectivity within the dual-stream network with that of the other reference networks yielded the following results: The functional connectivity of the visual network was significantly higher than that of the dual-stream network, t(27) = 7.84, FDR p < .01; the default mode network, t(27) = 5.97, FDR p < .01; and the left hemisphere-only dual-stream network, t(27) = 7.65, FDR p < .01 (Figure 3). There were no significant differences between the functional connectivity of the dual-stream and default mode networks, or between the dual stream and just the left hemisphere dual stream. The functional connectivity between LD and bilateral ventral streams was significantly higher than the connectivity between dual-stream regions and default mode network, t(27) = 6.09, FDR p < .01; the functional connectivity between LD and bilateral ventral streams also was significantly higher than the connectivity between dual-stream regions and visual network, t(27) = 3.98, FDR p < .01.

Similar to the control group, the stroke group exhibited greater functional connectivity within the visual network than in the dual-stream network, t(27) = 6.00, FDR p < .01; the default mode network, t(27) = 4.61, FDR p < .01; or the left hemisphere-only dual-stream network, t(27) = 6.16, FDR p < .01 (Figure 4). No significant results were found between the dual-stream network, default mode network, and the left hemisphere-only dual-stream network.

There were no significant differences between the stroke and control groups regarding the average functional connectivity in the dual-stream speech network as a whole or in any of the streams (Figure 5). The comparisons excluding the four stroke group participants with more than 50% lesioned left-hemisphere nodes (Figure A1) also yielded no significant differences between the stroke and control groups.

To construct a graph that satisfies the small world network assumption, we need γ > 1, λ ∼ 1, and σ > 1 (Achard et al., 2006). Figure 6 indicates the trend of λ, γ, and σ as the proportion of connections is increased. The λ is closest to 1 at proportion 0.132, and the highest value of γ and σ also occur at threshold 0.132. Therefore, the best proportion of connections in our control group data is 0.132. At this threshold, λ is equal to 1.02, γ is equal to 1.06, and σ is equal to 1.04.

Eight nodes of the dual-stream network were identified as hubs (defined by the top two ranked shortest path length nodes in the LD, left ventral, and RV streams; right dorsal hubs are the homologue of the LD hubs). The hubs identified were the following nodes: left superior temporal sulcus and left pars triangularis in the LD stream, their homologues in the right dorsal stream, the middle portion of the left superior temporal gyrus and posterior portion of the left middle temporal gyrus in the left ventral stream, and the middle portion of the right middle temporal gyrus and the middle portion of the right superior temporal sulcus in the RV stream (Table 8 and Figure 7).

Independent t tests examining group functional connectivity differences of hub regions resulted in the following significant results (Figure 8): The functional connectivity within the LD stream of the control group was significantly higher than that of the stroke group, t(54) = 2.84, FDR p = .03, and the functional connectivity between the left ventral and RV streams of the control group was significantly higher than that of the stroke group, t(54) = 2.30, FDR p = .03. The between-group hub comparisons excluding the four stroke participants with more than 50% lesioned left-hemisphere nodes yielded no significant differences, although the values follow similar trends as in the whole data set, and three t tests approach significance after FDR correction: whole model, t(50) = 2.34, FDR p = .087; LV, t(50) = 2.25 FDR p = .087; and LV to RV, t(50) = 2.45, FDR p = .087 (Figure A2).

Our tests of multicollinearity indicate that multicollinearity does not need to be addressed further in the regression models: (1) No pair of predictors have a correlation of |r| > 0.7, and (2) no VIF > 5 was found (Table 9). Table 10 lists all of the results from the regression models predicting each WAB measure from the functional connectivities of the dual-stream network's hubs in the stroke group. Significant results were as follows: Spontaneous Speech's significant predictors were LD (β = −15.44, t = −3.51, p < .01) and right dorsal–ventral (RDV, β = −7.21, t = −2.21, p = .05) functional connectivity, as well as age (β = −0.19, t = −2.64, p = .02), education (β = 0.78, t = 2.36, p = .03), and lesion size (β = −1.0e−04, t = −2.28, p = .03). RDV functional connectivity (β = −3.17, t = −2.13, p = .04) and lesion (β = −4.0e−05, t = −2.54, p = .02) were the only significant predictors of the Auditory Verbal Comprehension score. Repetition scores were significantly predicted by LD (β = −4.92, t = −2.92, p < .01), bilateral dorsal (between the left and right dorsal [L&RD], β = 6.55, t = 2.58, p = .02), RV (β = 2.11, t = 2.44, p = .03), and RDV (β = −5.37, t = −2.84, p = .01) functional connectivity, as well as education (β = 0.34, t = 2.20, p = .04). Naming and Word Finding scores were significantly predicted by LD (β = −5.38, t = −3.08, p < .01), L&RD (β = 5.38, t = 2.11, p = .05), RV (β = 2.80, t = 3.45, p < .01), and RDV (β = −6.88, t = −3.75, p < .01) functional connectivity, as well as education (β = 0.37, t = 2.34, p = .03). Aphasia Quotient yielded the following significant predictors: LD (β = −56.83, t = −3.13, p < .01), L&RD (β = 58.76, t = 2.93, p < .01) and RDV (β = −41.74, t = −2.67, p = .02) functional connectivity, as well as education (β = 3.76, t = 2.88, p < .01). The regression models excluding the four stroke participants with more than 50% lesioned left-hemisphere nodes also yielded significant F tests for the prediction of each WAB measure; there were some differences in significant predictors compared with the models with all of the participants, but the overall patterns are very similar (e.g., LD and RDV as frequent negative significant predictors, L&RD as a frequent positive significant predictor; Table A1).

The purpose of this study was twofold: (1) characterize the dual-stream model's network characteristics in a neurotypical control group and (2) determine the feasibility of using these dual-stream network characteristics to predict speech and language impairments in stroke survivors. Our findings indicated that: (1) the regions within the dual-stream model of speech processing do perform as an intrinsic network in resting-state fMRI; (2) stroke survivors exhibited significantly less functional connectivity of the hub regions identified by the graph theory measures compared with the neurotypical control group; and (3) functional connectivities of hub regions are significant predictors of poststroke language performance.

In previous studies, the intrinsic connectivity of a network is defined as an ongoing neural and metabolic activity that occurs in the resting basal state (Raichle, 2015; Kilpatrick et al., 2011; Napadow et al., 2010). After the default mode network was discovered in the resting-state (Raichle et al., 2001), many other functional networks were studied with resting-state fMRI, including but not limited to executive, salience, dorsal attention, and visual networks (Tessitore et al., 2017; Zhu et al., 2016; Napadow et al., 2010). To test if the regions identified by the dual-stream model of speech processing also can be treated as an intrinsic functional network at rest, we computed two sets of comparisons: (1) the functional connectivity within the dual-stream network compared with functional connectivity within two well-known networks—the default mode network and the visual network, and (2) the functional connectivity between each of the three main components of the dual-stream model (i.e., LD, LV, and RV) compared with between the dual-stream, default-mode, and visual networks. Our results indicated no significant difference between the strength of the functional connectivity within the default mode and within the dual-stream network, but the visual network exhibited significantly higher functional connectivity than the default mode and dual-stream network (Figure 3). It is not surprising that the visual network had the strongest functional connectivity because the locations of the visual nodes are spatially very close to one another compared with the other two networks. Although null results should be interpreted cautiously, the nonsignificant results between the default mode and dual-stream networks indicate that the strength of connections within the dual-stream network are similar to those of the default mode network. This finding suggests the intensity of the intrinsic activity within the dual-stream regions is similar to the most well-known resting-state network, which provides robust evidence for the existence of an intrinsic language network as outlined by the dual-stream model of speech processing. In addition, the connections between each of the streams were significantly stronger than the connections between the streams and the visual and default-mode networks. If the dorsal and ventral streams do not belong to the same intrinsic network, these between-stream connections would be the same or even weaker than the functional connectivity between the dual-stream regions and regions in these other networks. Based on these findings, we highly recommend that the dual-stream model can be interpreted as an intrinsic network in neurotypical adults, providing further evidence that there is potential to use resting-state functional connectivity measures of the dual-stream network to study the integrity of language circuitry in stroke survivors (Battistella et al., 2020). Our resting-state functional connectivity findings are highly consistent with previous task-based fMRI studies that have shown that regions in the dual-stream model are co-activated during a variety of language tasks (Saur et al., 2008, 2010; Hickok & Poeppel, 2004).

To further characterize the dual-stream network, we identified hub regions of the resting-state dual-stream network using graph theoretical parameters. The hub regions identified in the current study were in superior temporal and inferior frontal (pars triangularis) cortices for the dorsal stream, in the middle temporal gyrus and superior temporal gyrus for the left ventral stream, and in the superior temporal gyrus and middle temporal gyrus for the RV stream. These regions are all implicated in numerous lesion-symptom mapping and neuroimaging studies of speech processing (Mesulam et al., 2014; Margulies & Petrides, 2013; Price, 2012; Dronkers, 2011; Vigneau et al., 2006). The left pars triangularis, that is, the anterior portion of Broca's area, is reliably implicated in speech production, including naming and phonological awareness (Kibby, Kroese, Krebbs, Hill, & Hynd, 2009; Foundas, Eure, Luevano, & Weinberger, 1998; Foundas, Leonard, Gilmore, Fennell, & Heilman, 1996). The other hub node in the dorsal stream in posterior superior temporal cortex is near areas known to be involved in sensorimotor and audiovisual integration (Rogalsky et al., 2015; Heim, Eickhoff, & Amunts, 2008). The ventral stream hubs include bilateral nodes in the mid and posterior middle temporal gyrus, overlapping with bilateral middle temporal regions that are reliably implicated in speech comprehension, particularly lexical-semantic processing (Acheson & Hagoort, 2013; Wallentin et al., 2011). Lesion-symptom mapping studies of stroke consistently identify damage in our left hemisphere middle temporal hubs as associated with auditory comprehension deficits (Rogalsky et al., 2022; Pillay, Binder, Humphries, Gross, & Book, 2017; Hickok & Poeppel, 2007; Dronkers, Wilkins, Van Valin, Redfern, & Jaeger, 2004; Bates et al., 2003). Thus, the hub nodes identified by our graph theory approach are not surprising, but it is noteworthy that the RV hubs are in the middle temporal gyrus and not the superior temporal gyrus. Most task-based neuroimaging and neurostimulation studies of speech identify more superior temporal than middle temporal regions in the right hemisphere, which is often attributed to spectrotemporal acoustic processing that is not specific to speech (Wolmetz, Poeppel, & Rapp, 2011; Hickok & Poeppel, 2007). Our findings suggest that multiple regions of the right middle temporal gyrus are highly connected with other nodes of the dual-stream network and require further study as to their role in speech processing and aphasia recovery.

Numerous previous studies indicate that structural damage to dual-stream cortical regions or the WM connecting them is related to various speech and language impairments in poststroke aphasia (Matchin et al., 2022; Baboyan et al., 2021; Yang et al., 2017; Agosta et al., 2013; Kümmerer et al., 2013). We see our study using functional connectivity as complementary to these structural network findings, as our graph theory analysis identified critical regions in the network, without the restraints of direct physical WM tract connectivity. Previous voxel-based lesion symptom mapping studies have identified several left ventral stream regions to be critical for auditory speech comprehension, including posterior superior temporal and middle temporal gyrus regions (Na, Jung, Tench, Auer, & Pyun, 2022; Rogalsky et al., 2022; Sul et al., 2019; Gajardo-Vidal et al., 2018; Dronkers et al., 2004; Bates et al., 2003). Speech production deficits have been associated with damage to the dorsal stream, including inferior frontal, insular, premotor, and inferior parietal regions (Na et al., 2022; Sul et al., 2019; Bates et al., 2003). Performance on tasks that involve both dorsal and ventral stream functions, such as naming or repetition, have implicated numerous dorsal and ventral regions (Na et al., 2022; Døli, Andersen Helland, Helland, & Specht, 2021; Sul et al., 2019; Baldo et al., 2013; Rogalsky, Rong, Saberi, & Hickok, 2011). Our current study corroborates these findings from voxel-based lesion symptom mapping research and further pinpoints potentially crucial nodes within these extensive brain areas. For example, pars triangularis and posterior superior temporal regions are the hubs for the dorsal stream, which implies that they are particularly critical for speech production. In addition, our functional analyses are not only limited to areas of lesion; we also can examine the potential compensation within the right hemisphere and between hemispheres. For instance, functional connectivity of the right (but not the left) ventral's hubs is a significant predictor for repetition and naming. In addition, the RDV hub connectivity is a significant predictor of overall aphasia severity, whereas LD-ventral hub connectivity is not a significant predictor of any of our aphasia measures.

We found no significant differences between the groups when examining the functional connectivity of all nodes in each stream. However, the hub nodes' functional connectivity comparisons did find significantly weaker connectivity among the hub regions of the LD stream (pars triangularis and posterior superior temporal), as well as among the hub regions of the bilateral ventral stream (left middle superior temporal gurus and posterior middle temporal gyrus, as well as right superior temporal sulcus and middle temporal gyrus), in the stroke group compared with the control group. Based on these group-level comparisons, we conclude that graph theory approaches provide meaningful information regarding network-level dysfunction poststroke and that these dual-stream hubs defined by the graph theory should be investigated further for possible sites of neurostimulation to promote language rehabilitation in stroke survivors. The majority of our stroke group has a diagnosis of aphasia (n = 25 of 28). Thus, it is likely that weakened hub node functional connectivity, rather than overall node connectivity in the dual-stream network, is strongly related to an aphasia diagnosis post left-hemisphere stroke, although future studies are needed to explore this further.

Because of the significantly lower average functional connectivity between the hub nodes in the stroke group compared with the control group, we tested the ability of these hub functional connectivities to predict performance on subtests of the WAB, after accounting for variance because of lesion size, age, sex, and education years (Table 8).

In the ventral stream, hubs were identified bilaterally in posterior middle temporal and mid superior temporal regions. Notably, neither the functional connectivity of the hub nodes within the left ventral stream nor the functional connectivity of the left ventral hub nodes with other streams' hubs were significant predictors of performance on any WAB subtest. Perhaps partial damage to the left ventral stream does not yield a deficit on clinical measures because of the RV stream's contributions, providing additional evidence of the bilateral nature of the ventral stream. We did expect that a major damage to the left ventral nodes would yield comprehension impairments (Bonilha et al., 2017; Robson et al., 2014), but our stroke group contained only 3.5% of individuals with Wernicke's aphasia, which is often associated with large left temporal lesions. Thus, left ventral effects perhaps did not reach significance at the group level because of this small sampling of Wernicke's aphasia.

It also is noteworthy that our regression models did not find functional connectivity between the LD and ventral streams to be a significant predictor of performance on any WAB subtest, including repetition and naming, which both have been previously shown to reliably engage both left frontal and left posterior temporal cortices (Trimmel et al., 2018; Cappa, Sandrini, Rossini, Sosta, & Miniussi, 2002). However, it should be noted that for Spontaneous Speech scores, LD-ventral connectivity approaches significance (p = .052). To our knowledge, there is no clear theoretical basis or finding in previous work that could explain these null results, but we will speculate here, in hopes that it may be potentially helpful to direct future work. According to the dual-stream model, the dorsal stream and ventral stream communicate mainly via the auditory-motor interface in the posterior superior temporal gyrus (i.e., area Spt) and the arcuate fasciculus, as well as via WM connections between the anterior temporal lobe and inferior frontal gyrus. These connections in the dual-stream model are supported by numerous structural connectivity studies (Glasser & Rilling, 2008; Rilling et al., 2008; Saur et al., 2008). However notably, our hub nodes identified by our graph theory approach did not include any regions in the anterior temporal lobes. Thus, some of the typical communications between the two streams may not be reflected in our functional connectivity measures between the dorsal and ventral hubs. Another possibility is that the functional connectivity between the LD and ventral hub nodes may not be necessary for successful language abilities, and perhaps even the synchronization of these hub nodes may suggest network dysfunction that leads to lower performance on some language measures.

Our findings related to the right-hemisphere homologue of the LD stream (i.e., the “right dorsal” stream) may provide some clarity as to why previous work investigating potential benefits of the right hemisphere in poststroke recovery have been equivocal (François et al., 2019; Xing et al., 2016). Notably, across all WAB subtests, greater RDV functional connectivity predicted poorer performance. This finding coincides with previous evidence implying that the activity in the right hemisphere following a left-hemisphere stroke is not helpful to language rehabilitation (Richter, Miltner, & Straube, 2008; Karbe et al., 1998), that is, seemingly supporting a negative view of the compensation activity in the right hemisphere. However, our results provide a more nuanced view of the right hemisphere's contributions to poststroke speech and language abilities: While greater RDV functional connectivity was related to poorer performance and more severe aphasia, greater functional connectivity L&RD streams were significant predictors of more mild aphasia symptoms overall (i.e., a higher WAB aphasia quotient). These findings suggest that right dorsal stream contributions to poststroke speech could possibly be beneficial, but not by replacing the functionality of the LD stream, but rather supporting the remaining LD stream via more domain-general cognitive resources that have been previously located to right lateral frontal cortex (Huang et al., 2022; Reuter et al., 2011). Right frontal regions have been shown to be activated by a variety of speech and language tasks, particularly as task demands increase (van Ettinger-Veenstra et al., 2010; Just, Carpenter, Keller, Eddy, & Thulborn, 1996). It may be that the right hemisphere's dorsal stream is able to contribute to some aspects of speech-processing for which these cognitive processes may be particularly helpful, such as word selection or sentence building, but unable to support others, such as speech motor planning and control. These motor speech deficits (i.e., elements of dysarthria and apraxia of speech) are common in individuals with aphasia and contribute to performance on the WAB's expressive subtests. Future studies with more sensitive measures of motor speech deficits are needed to test this possibility.

One possible explanation to consider for the right dorsal findings is lesion size—perhaps larger left-hemisphere lesions lead to both greater speech impairments and also greater right hemisphere involvement in speech tasks, which could be reflected in greater RDV connectivity. However, this is not the case in our study: We found lesion size to be a significant predictor of performance on only one WAB subtest, Spontaneous Speech. Thus, future studies are needed to investigate the possibility of how to promote functional connectivity between left and right dorsal hubs, but not between right dorsal and ventral hubs, in individuals with aphasia to maximize this potential compensatory strategy.

Previous task-based fMRI studies have found that greater right-hemisphere activations predict greater aphasia severity (LaCroix, James, & Rogalsky, 2021; Postman-Caucheteux et al., 2010; Richter et al., 2008). Conversely, in the present study, we found that stronger RV functional connectivity was a significant predictor of better naming and repetition performance—two functions known to engage the ventral stream in neurotypical control participants (Ross, McCoy, Wolk, Coslett, & Olson, 2010; Saur et al., 2008; Hickok & Poeppel, 2004; Wise et al., 2001). These findings provide further evidence for the bilaterality of the ventral stream of the language network. However, it is also notable that performance on the Auditory Verbal Comprehension subtest and overall aphasia severity (as measured by WAB-R Aphasia Quotient) were not predicted by RV connectivity. These differences in the predictive ability of RV connectivity across the WAB subtests may be because of the nature of the WAB's Comprehension subtest, which includes many stimuli of increasing complexity at the phrasal and sentence levels. The RV stream has been found to support single-word comprehension and phonological encoding (Hickok & Poeppel, 2007; Gorno-Tempini et al., 2004), but are not modulated by sentence structure or morphosyntactic manipulations (Novén, Schremm, Horne, & Roll, 2021; Humphries et al., 2006). Although the RV connectivity did not predict overall aphasia severity, our findings related to naming and repetition suggest that RV connectivity may help support some lexical-semantic and phonological functions in poststroke aphasia. Future studies of the RV stream's connectivity in relation to other more specific receptive speech measures are needed to characterize how the RV stream is contributing to receptive and productive speech abilities in individuals with aphasia.

Lastly, we would like to discuss our consistent finding of LD stream functional connectivity being a significant negative predictor of WAB subtest performance. On the surface, this finding was surprising: How could less coherence within the LD stream predict better performance on speech tasks known to be supported by these LD regions (e.g., spontaneous speech and repetition)? Upon further inspection of the data, it is apparent that individuals with aphasia with large left frontal lesions encompassing the LD stream hub nodes are driving this negative correlation between LD hub connectivity and WAB performance. The functional connectivity of these individuals' injured LD nodes (i.e., nodes that overlap with the lesion area) are very high compared with the functional connectivity from the intact nodes, because of the lack of fluctuation in the BOLD response of these “lesioned” nodes. In other words, the functional connectivity of their LD nodes is very strongly correlated because all of these nodes exhibit very little to no BOLD response. To our knowledge, this paradox has not been addressed in the previous literature of functional connectivity after a stroke, and thus previous characterizations of functional connectivity changes may need to be interpreted with this issue in mind (Liu et al., 2017; Zhu et al., 2014). Perhaps novel approaches need to be developed to determine how to best consider both correlation and amplitude of BOLD fluctuations to maximize the utility of using functional connectivity predictors to better understand the neurobiology of poststroke speech and language impairments.

Our study has a few important limitations to consider. First, the effects of interscanner and intersite differences are of ongoing debate in the field (Badhwar et al., 2020; Noble et al., 2017), and our stroke and control groups' data were collected on two different scanners. The difference in number of channels in the head coil or head coil type may also bring some variability. However, in the current study, only the between-group comparisons may be affected by possible interscanner effects; the hub nodes were defined only within the control group, and the functional connectivity predictors of speech-processing abilities were only within the stroke group. Future studies can use a single-scanner data set and also consider other alternative normalization procedures for the stroke MRI data, such as using a cohort-specific template, to determine the optimal procedures for these between-group comparisons (Pappas et al., 2021).

Another methodological item that requires further consideration is how to best treat functional ROIs that include areas of structural damage (i.e., within the binary structural lesion map). Klingbeil and colleagues (2019) suggest that areas of lesion do not provide a “meaningful BOLD signal” and thus should be excluded from analyses (Klingbeil et al., 2019). Yet, previous studies (e.g., Shahid et al., 2017; Siegel et al., 2016) and visual inspection of our own data indicate that blood flow within areas delineated as having structural damage is often nonzero, and the neural activity within lesion penumbra has been associated with greater functional recovery in both chronic and acute stroke (Kroll, Zaharchuk, Christen, Heit, & Iv, 2017; Tsai et al., 2014; Zhang, Ke, Li, Yip, & Tong, 2013). Some studies have explored the impact of lesioned areas on functional connectivity analyses (Griffis, Metcalf, Corbetta, & Shulman, 2019; Saenger et al., 2018). However, there is no gold standard for how to treat lesions in functional ROIs, and numerous exclusion criteria can be found in the resting-state fMRI literature, ranging from none (e.g., Wang et al., 2021; Zhang, Xia, et al., 2021; Riccardi, Yourganov, Rorden, Fridriksson, & Desai, 2020; Balaev, Orlov, Petrushevsky, & Martynova, 2018), to preset thresholds of amount of damage (e.g., Griffis et al., 2019; Siegel et al., 2016), to exclusion of any lesioned ROIs (e.g., Klingbeil et al., 2019). Given that a focus of the present study's analyses within the stroke group was to determine if resting-state functional connectivity of the dual-stream speech network can be used to predict poststroke speech processing impairments, our primary analyses within the stroke group did not take into account areas of structural lesion; we wanted to avoid having to make assumptions about the integrity of blood flow or the quality of the BOLD signal within the lesion, or attempt to delineate areas of infarct versus penumbra. Nonetheless, we wanted to ensure that performance of stroke participants with large areas of damage encompassing most of our dual-stream nodes were not driving results. To this end, overall lesion size was included as a predictor in our regression models. We also computed the stroke group analyses after excluding stroke participants with severe damage to the LD and ventral streams (defined as >50% of nodes damaged). The results (Figures A1, A2; Table A1) after excluding the four participants meeting this criteria remain very similar to our primary results. Our findings suggest that large lesions encompassing a majority of our functional nodes are not driving our results, although certainly future work is needed to identify optimal approaches to integrating lesion maps and functional connectivity for predicting aphasia measures.

We computed t tests between the stroke and control groups to identify group differences in the functional connectivities of the dual-stream speech network (all nodes, and then for hub nodes only). An alternative method to examine group differences is network-based statistics (NBS: Zalesky, Fornito, & Bullmore, 2010; https://www.nitrc.org/projects/nbs/). NBS provides more power than mass univariate testing by controlling FWE rate across all tests that are performed for each connection in the connectivity graph. NBS was not the focus of our analyses because we aimed to compare across groups the overall and hub node connectivities within the different subportions of the dual-stream network. However for comparison purposes, an exploratory post hoc NBS was computed for the entire dual-stream network. The NBS results (Figure A3) generally coincide with our original between-group comparisons of our hub nodes, with LD (precentral gyrus and pars triangularis) and left ventral (middle temporal gyrus) nodes exhibiting weaker connections in the stroke group than in the control group.

Lastly, alternative approaches to our graph theory methods to define hub nodes should also be considered. For example, normalized participation coefficients (NPCs; Pedersen, Omidvarnia, Shine, Jackson, & Zalesky, 2020) consider relative intranetwork versus internetwork connectivity, normalized by network size, to identify hub nodes. The present study focused on hubs within different portions of the dual-stream network, but considering interstream connectivity when determining hub nodes is an interesting future direction. Post hoc, for exploratory comparison purposes, we computed the NPC for each node and identified hubs as the nodes with the top two NPCs in each stream (Table A2). There is some overlap between the hubs defined by NPC versus our original average shortest path length method, but the hubs are not identical. For example, the NPC nodes in the left ventral stream are more anterior in the superior temporal lobe than our original nodes, and portions of the inferior frontal gyrus (pars opercularis) are defined as LD hubs by both methods (NPC hub in pars opercularis, original method hub in pars triangularis). Some differences are expected given that NPC considers internetwork connectivity and shortest path length does not. We also computed between-group t tests and regression models for these NPC-defined hubs as we did for our original hub nodes identified by the shortest path length (Table A3). Unlike for our original nodes, there were no significant differences between the stroke and control groups' functional connectivity for these NPC-defined hubs. However, like our original regression models, the NPC-defined hub models did yield significant results for predicting each aphasia measure, albeit with some different significant predictors for some of the aphasia measures.

In conclusion, our study has five main findings: (1) The dual-stream model is representative of an intrinsic network as measured by resting-state fMRI; (2) the functional connectivity of the hub nodes of the dual-stream network, but not overall network connectivity, is significantly lower in our stroke group than in the control group; (3) the intrinsic connectivity of the dorsal and ventral stream hubs identified in neurotypical control participants significantly predict performance on clinical aphasia measures (e.g., WAB subtests) in individuals with poststroke left hemisphere damage; (4) greater connectivity of the right dorsal with the LD stream, and less connectivity with the RV stream, significantly predicts better performance in expressive and receptive speech tasks and lower overall aphasia severity; (5) greater functional connectivity within the RV stream significantly predicts better naming and repetition performance in individuals with poststroke left hemisphere damage.

Corresponding author: Corianne Rogalsky, Speech and Hearing Science, Arizona State University, P.O. Box 870102, Tempe, AZ 85287-0102, e-mail: [email protected].

Materials, data, and analysis scripts are available via an email request to the corresponding author.

Haoze Zhu: Conceptualization; Formal analysis; Methodology; Software; Visualization; Writing—Original draft; Writing—Review & editing. Megan C. Fitzhugh: Data curation; Investigation; Writing—Review & editing. Lynsey M. Keator: Data curation; Investigation; Writing—Review & editing. Lisa Johnson: Data curation; Investigation; Writing—Review & editing. Chris Rorden: Data curation; Funding acquisition; Writing—Review & editing. Leonardo Bonilha: Data curation; Investigation; Writing—Review & editing. Julius Fridriksson: Data curation; Funding acquisition; Supervision; Writing—Review & editing. Corianne Rogalsky: Conceptualization; Funding acquisition; Supervision; Writing—Original draft; Writing—Review & editing.

This work is supported by Arizona State University and National Institute on Deafness and Other Communication Disorders (https://dx.doi.org/10.13039/100000055), grant number: P50 DC014664 to J. F.

Retrospective analysis of the citations in every article published in this journal from 2010 to 2021 reveals a persistent pattern of gender imbalance: Although the proportions of authorship teams (categorized by estimated gender identification of first author/last author) publishing in the Journal of Cognitive Neuroscience (JoCN) during this period were M(an)/M = .407, W(oman)/M = .32, M/W = .115, and W/W = .159, the comparable proportions for the articles that these authorship teams cited were M/M = .549, W/M = .257, M/W = .109, and W/W = .085 (Postle and Fulvio, JoCN, 34:1, pp. 1–3). Consequently, JoCN encourages all authors to consider gender balance explicitly when selecting which articles to cite and gives them the opportunity to report their article's gender citation balance.