Pre-treatment graph measures of a functional semantic network are associated with naming therapy outcomes in chronic aphasia

Abstract

Naming treatment outcomes in post-stroke aphasia are variable and the factors underlying this variability are incompletely understood. In this study, 26 patients with chronic aphasia completed a semantic judgment fMRI task before receiving up to 12 weeks of naming treatment. Global (i.e., network-wide) and local (i.e., regional) graph theoretic measures of pre-treatment functional connectivity were analyzed to identify differences between patients who responded most and least favorably to treatment (i.e., responders and nonresponders) and determine if network measures predicted naming improvements. Responders had higher levels of global integration (i.e., average network strength and global efficiency) than nonresponders, and these measures predicted treatment effects after controlling for lesion volume and age. Group differences in local measures were identified in several regions associated with a variety of cognitive functions. These results suggest there is a meaningful and possibly prognostically-informative relationship between patients’ functional network properties and their response to naming therapy.

1. Introduction

Most patients with post-stroke aphasia experience some recovery of speech-language functions, but the degree of recovery is highly variable (Lazar & Antoniello, 2008). Evidence indicates that lesion size, demographics, and other factors contribute to this variability (Dignam et al., 2017; Lambon Ralph et al., 2010a; Lazar & Antoniello, 2008; Plowman et al., 2012; Votruba et al., 2013; Watila & Balarabe, 2015), but predicting behavioral outcomes, especially treatment-related outcomes, remains challenging. However, previous investigations of the brain bases of post-stroke cognitive-linguistic deficits suggest that network analyses have the potential to enhance our understanding of language recovery and, potentially, improve prognostic accuracy.

Neurological damage changes functional connectivity and the functional organization of brain networks, and these changes relate to various cognitive-linguistic deficits (Sandberg, 2017; Siegel et al., 2016, 2018; D. Zhu et al., 2014; Y. Zhu et al., 2017). In several studies, such changes were identified via resting-state fMRI. For example, Siegel and colleagues found that subacute stroke patients had reduced functional connectivity between homologous regions in the ipsilesional and contralesional hemispheres relative to healthy controls (Siegel et al., 2016). Additionally, functional connectivity was as effective as structural damage at predicting severity of attention and language impairment, and an even better predictor of visual and verbal memory. In another study, subacute stroke patients had lower functional connectivity involving the left fronto-parietal network (LFPN) than healthy controls, and patients with lower connectivity had more severe comprehension deficits (D. Zhu et al., 2014). A follow-up two months post-onset showed that improvements in comprehension were associated with increased LFPN functional connectivity, reflecting correspondence between a shift toward more normal functional connectivity and the restoration of more typical language functions. In chronic aphasia, patients had abnormally reduced functional connectivity in several resting state networks and a semantic network, and less severe aphasia was associated with higher functional connectivity in the default mode, dorsal attention, executive control, salience, and sensorimotor networks (Sandberg, 2017). Collectively, these studies indicate that functional synchronization between various brain regions is altered in the immediate wake of neural insult and may remain so even in the later stages of recovery. Moreover, abnormal functional connectivity is associated with the severity of behavioral deficits.

Graph theory has also been applied to investigations of the effects of focal damage on functional brain networks. Graph measures characterize regional and global network properties, such as the extent to which a network is comprised of smaller subsystems (e.g., clustering coefficient, local efficiency, modularity) and the ease with which information can be transmitted throughout the entire network (e.g., shortest path length, global efficiency) (Bullmore & Sporns, 2009; Rubinov & Sporns, 2010).

A resting-state study by Y. Zhu et al. (2017) found that 29 acute stroke patients had greater global efficiency and lower shortest path length than healthy controls (notably, these metrics are inversely related to each other and summarize the extent of connectivity among all of the nodes in a network based on direct and indirect connections between them), perhaps reflecting a trend toward more random network organization and a disruption in the balance between efficiency of local processing and global information transfer. Patients also had abnormally high node degree, which quantifies the number of connections between a given node and all other nodes in the network, and efficiency in several default mode network (DMN) regions. Of note, the degree and efficiency of left middle frontal gyrus (LMFG) were positively associated with performance on the Mini-Mental State Examination. Siegel et al. (2018) recently found that stroke patients had significantly lower modularity than healthy controls two weeks post-stroke, indicating they had poorer integration within subsystems and less distinct partitioning between them. Patients’ modularity increased over the first year post-stroke onset, and larger increases were associated with better recovery of language, attention, and spatial memory skills. Thus, a shift toward a more control-like balance between efficient local processing and efficient global communication appears to underlie the restoration of high-level cognitive functions.

Several treatment studies have also examined recovery from a network perspective using graph metrics. Naming therapy increased the node degree of the bilateral inferior frontal gyri (IFG) and left superior medial gyrus (Sandberg et al., 2015), and more improvement in narrative production after imitation-based treatment was associated with a larger increase in modularity from pre- to post-treatment (E. S. Duncan & Small, 2016). These studies demonstrate an association between behavioral outcomes from therapy and certain patterns of functional neural changes. However, such changes can only be identified after treatment has been administered. On the other hand, characteristics that distinguish between patients who are most and least likely to benefit from treatment before it begins would be of great value to clinical rehabilitation for the purpose of predicting potential outcomes. To this end, recent intervention studies suggest that outcomes after aphasia therapy may be associated with several pre-intervention brain network properties. Tao and Rapp showed that patients’ average within-module degree z-score, which characterizes a network’s propensity for segregation or local processing, was positively correlated with improvement after spelling treatment in 15 patients with aphasia (Tao & Rapp, 2019). Network integration measures have also been linked to treatment outcomes: in a recent study of eight patients who completed an intensive comprehensive treatment program, higher pre-treatment resting-state global efficiency was associated with larger reductions in aphasia severity (Baliki et al., 2018). Similarly, Marcotte et al. found a trending (albeit non-significant) relationship between higher pre-treatment DMN integration and more improvement in naming after semantic feature analysis therapy in nine individuals with aphasia (Marcotte et al., 2013).

Guided by the literature above, we aimed to explore relationships between functional brain network properties and recovery from aphasic language deficits by conducting a graph analysis of a semantic processing network during a feature judgment task in 26 patients who received naming therapy. We examined pre-treatment global network properties and hypothesized that higher indices of network integration (i.e., average network strength and global efficiency) and segregation (i.e., local efficiency and average clustering coefficient) would be associated with larger gains in naming after treatment. We also sought to determine if these measures explained additional variance in treatment outcomes above and beyond that explained by other traits that were previously linked to aphasia outcomes (namely, age and lesion volume). Finally, to determine how the functional contributions of specific regions in the network might be related to treatment response, we compared the local (i.e., region-specific) graph properties of patients who responded best to treatment (i.e., responders) and those who responded less favorably (i.e., nonresponders).

2. Material and Methods

This study utilized functional imaging and behavioral data collected as part of a larger investigation by the Center for the Neurobiology of Language Recovery (https://cnlr.northwestern.edu/). Data from some or all participants in the present study have been utilized to address other research questions (Gilmore et al., 2018; Johnson et al., 2019; Kiran et al., 2015; Meier et al., 2018, 2019b), but the aims, analyses, and results of the present study are unique and, with the exception of some demographic and behavioral data, have not been reported elsewhere.

2.1. Participants

Twenty-six adults with chronic (i.e., onset >6 months prior to enrollment) post-stroke aphasia were included in this study. These individuals formed the treatment group in a prior study that also included 10 untreated patients with aphasia and 17 healthy control subjects (Johnson et al., 2019). Because the present study aimed to identify predictors of treatment outcomes, only the subset of treated patients from Johnson et al. (2019) were included here; however, readers interested in the efficacy of our treatment and differences in behavioral outcomes in treated and untreated patients are referred to our previous work (Gilmore et al., 2018; Johnson et al., 2019).

In addition to chronic aphasia subsequent to stroke, enrollment criteria for this investigation included pre-stroke English language proficiency and the presence of anomia confirmed via performance on a 180-item study-specific picture naming battery. Exclusion criteria included multiple stroke events, other neurological (i.e., besides stroke) or psychiatric disorders, and contraindications for MRI. Participant demographics and other information are reported in Table 1 (note that some of this information and additional participant data have also been reported in Johnson et al., 2019). All participants provided written confirmation of informed consent, in accordance with IRB protocols at Boston University and Massachusetts General Hospital.

Table 1.

Patient demographics and behavioral scores.

	Sex	Age (years)	MPO	Lesion volume (cm3)	WAB AQ	# Tx sessionsǂ	Tx cat.	Avg. baseline score* (SD)	Effect Size	Avg. Effect Size	fMRI task accuracy† (%)
Responders
BU03	F	63	62	175.38	52.00	24	B	0.33 (0.58)	9.81	9.53	72.22
							F	4.67 (0.58)	9.24
BU04	M	79	13	84.78	74.10	14	B	7.67 (2.52)	4.11	8.70	91.67
							F	10.33 (0.58)	13.28
BU06	M	49	113	298.97	66.60	14	C	12.50 (1.29)	8.96	5.84	91.67
							F	7.75 (0.96)	2.71
BU07	M	55	137	181.97	48.00	24	V	2.67 (1.53)	2.18	4.27	80.56
							F	6.33 (0.58)	6.35
BU08	M	49	57	87.59	82.80	7	B	14.67 (1.15)	2.89	3.63	94.44
							F	11.33 (1.53)	4.36
BU09	F	71	37	11.66	95.20	24	B	3.33 (1.53)	6.76	8.87	88.89
							F	11.33 (0.58)	10.97
BU10	F	53	12	76.55	80.40	13	B	6.33 (0.58)	16.74	9.82	88.89
							F	14.00 (1.15)	2.89
BU13	M	42	18	12.13	92.70	24	B	8.00 (1.41)	7.07	11.04	83.33
							V	9.00 (0.00)	15.01
BU14	F	64	24	96.93	64.40	17	V	6.00 (1.00)	8.00	8.33	88.89
							F	6.67 (1.15)	8.66
BU15	F	71	74	189.31	87.20	24	B	5.67 (0.58)	6.35	3.66	75.00
							V	10.33 (2.08)	0.96
BU17	M	61	152	163.49	74.30	11	B	9.00 (1.00)	8.67	11.84	94.44
							F	9.33 (0.58)	15.01
BU18	F	70	152	69.64	78.00	25	V	12.33 (0.58)	5.77	5.47	97.22
							F	7.00 (2.00)	5.17
BU19	M	80	22	89.03	28.90	25	V	0.33 (0.58)	3.46	5.20	41.67
							C	4.33 (0.58)	6.93
BU20	F	48	14	164.33	13.00	25	V	0.00 (0.00)	3.52	5.08	38.89
							C	0.00 (0.00)	6.64
BU22	M	62	12	100.02	65.40	24	V	0.33 (0.58)	4.62	3.18	66.67
							F	1.33 (0.58)	1.73
BU27	M	65	17	34.15	84.30	24	V	6.00 (1.00)	7.00	4.59	94.44
							F	12.33 (1.53)	2.18
BU28	M	63	15	76.65	56.00	26	B	4.00 (1.00)	8.33	5.43	75.00
							F	2.67 (2.52)	2.52
Mean (SD)		61.47 (10.94)	54.76 (51.99)	112.50 (74.11)	67.25 (22.22)	20.29 (6.13)		6.70 (4.29)	6.73 (4.03)	6.73 (2.79)	80.23 (17.48)
Nonresponders
BU11	M	78	22	32.11	92.1	24	V	4.33 (0.58)	1.15	2.10	16.67
							F	9.33 (2.08)	3.04
BU12	M	68	104	186.85	40	17	V	0.33 (0.58)	1.15	1.15	58.33
							C	1.00 (1.73)	1.15
BU16	M	50	71	317.07	33.6	25	V	0.33 (0.58)	2.31	2.31	58.33
							C	0.00 (0.58)	2.31
BU21	M	65	16	247.59	11.7	24	B	0.00 (0.00)	0.58	1.74	44.44
							C	0.00 (0.00)	2.89
BU23	M	60	24	172.81	45.2	23	V	1.00 (1.00)	0.33	0.67	58.33
							C	1.00 (1.00)	1.00
BU24	M	69	170	183.45	40.4	24	V	1.67 (0.58)	1.73	2.07	66.67
							F	1.00 (0.00)	2.40
BU25	F	76	33	184.39	37.5	24	B	0.00 (0.00)	2.02	0.72	63.89
							C	0.67 (0.58)	−0.58
BU26	F	64	115	127.7	58	24	V	2.67 (0.58)	3.46	3.23	41.67
							C	7.00 (1.00)	3.00
BU30	M	59	29	186.52	60	24	V	1.00 (1.00)	3.00	2.29	72.22
							C	8.00 (3.61)	1.57
Mean (SD)		65.44 (8.66)	64.89 (54.01)	182.05 (77.54)	46.50 (22.17)	23.22 (2.39)		2.19 (2.96)	1.81 (1.10)	1.81 (0.84)	53.39 (16.88)

Abbreviations: MPO, months post-onset of aphasia; Tx cat., treatment categories (B, birds; C, clothing; F, furniture; V, vegetables); WAB AQ, Western Aphasia Battery-Revised Aphasia Quotient (a measure of overall aphasia severity).

^ǂFor participants BU04, BU06, BU08, BU10, BU14, and BU17, treatment was terminated early per the criteria described in section 2.3 of the text. Treatment was also terminated early for BU12 at the participant’s request for personal reasons.

^*Average baseline score indicates the average number of correctly named items (out of 18 total items) in a given category across all pre-treatment administrations of the naming battery.

^†Task accuracy reflects performance on items from participants’ assigned training sets.

2.2. Study design

Participants in this study completed pre-treatment behavioral evaluations and fMRI scans and then received naming therapy for up to 12 weeks. After treatment cessation, behavioral evaluations were re-administered to allow for comparisons between pre- and post-treatment naming performance. Details of each of these aspects of the study are provided below in sections 2.3-2.5. Post-treatment fMRI scans were also obtained, but those data were not analyzed for the present study.

2.3. Behavioral assessment, stimuli assignment, and naming therapy

Pre-treatment naming performance was assessed via a 180-item picture naming battery (see Gilmore et al., 2018, for additional details) in order to determine participants’ study eligibility, establish their baseline naming accuracy, and identify treatment stimuli. The battery was administered to 23 participants on three occasions and to three participants on four occasions prior to the initiation of treatment. The average timespan from first to last administration of the pre-treatment naming battery was 46.32 days (SD = 55.67)¹. The average timespan from the last administration of the pre-treatment battery to fMRI scanning was 18.81 days (SD = 13.46), and from the last administration of the pre-treatment battery to the start of treatment was 27.12 days (SD = 13.38).

Based on their performance on the pre-treatment naming battery, patients were assigned 36 items from the battery to be targeted in treatment. Assigned items consisted of 18 nouns per category from two of the following experimental categories: vegetables, birds, clothing, and furniture. Categories were balanced for frequency, concreteness, and familiarity (Coltheart, 1981; van der Wouden, 1990). Participants’ category assignments were determined based on average accuracy in the categories listed above, while also counterbalancing for category and typicality of trained and untrained items across participants (the latter of which was pursuant to the goal of examining the effects of item typicality on generalization to untrained items, as addressed in Gilmore et al., 2018). Participants qualified for treatment if their average accuracy in at least one category was below 65%. Most participants (i.e., 21 of 26) averaged below 65% in multiple categories and were assigned two such categories. However, five participants (BU06, BU08, BU10, BU18, and BU27) performed relatively well in several categories at pre-treatment and, after accounting for the additional constraints on which we counterbalanced, were each assigned one category on which they averaged below 65% and another on which they averaged above 65%. Participants’ average raw pre-treatment scores in their assigned categories are shown in Table 1.

The Western Aphasia Battery – Revised (WAB-R; Kertesz, 2007) was administered to assess overall aphasia severity prior to treatment. The WAB-R was re-administered after treatment, but only pre-treatment scores are reported in the present study.

Naming therapy was initiated after completion of all pre-treatment testing and fMRI scanning. Therapy was generally provided in two two-hour sessions per week for up to 12 weeks, though five patients received one or two additional sessions due to scheduling delays between their 24^th treatment session and their post-treatment fMRI scans. During each session, therapy tasks included sorting pictured items into their corresponding categories, naming trained items, determining if semantic features accurately described items from the training set, and a generative naming activity. At every other session, patients completed a naming probe in which all items from their training set were presented. Treatment continued for 12 weeks or until patients reached 90% accuracy on trained items on two successive probes.

Following each patient’s final treatment session, the complete 180-item naming battery was readministered three times to assess for changes in naming accuracy on trained and untrained items (only trained items are addressed in this study; readers interested in behavioral change on untrained items are referred to Gilmore et al., 2018). The average time between the final treatment session and the first post-treatment administration of the naming battery was 7.46 days (SD = 7.27) and the average time between the first and last administrations of the post-treatment battery was 9.5 days (SD = 6.69).

2.4. Behavioral analysis

Consistent with our prior studies (Gilmore et al., 2018; Johnson et al., 2019), participants’ treatment outcomes were quantified by computing effect sizes for each trained category by subtracting the average pre-treatment score (across three or four administrations) from the average post-treatment score (across three administrations) and dividing the difference by the pre-treatment standard deviation (Beeson & Robey, 2006)². An important benefit of this particular effect size measure is that it has been used widely enough in the aphasia treatment literature that there are established benchmarks for the magnitude of patients’ responsiveness to naming interventions (Beeson & Robey, 2006). Thus, participants with an effect size of at least 4.0 (a small effect per Beeson and Robey, 2006) in at least one category were classified as treatment responders; all others were classified as nonresponders (Gilmore et al., 2018; Johnson et al., 2019).

2.5. fMRI methods

2.5.1. Data acquisition

Scanning was conducted at the Athinoula A. Martinos Center for Biomedical Imaging in Charlestown, MA. A Siemens 3T Trio Tim scanner with a 20-channel head/neck coil was used for acquisition of T1-structural images (176 sagittal slices, 1mm³ voxels, TR = 2300ms, TE = 2.91ms, flip angle = 9°, matrix = 256x256, FOV = 256x256mm) and T2*-weighted functional images (40 axial slices, 3mm slices, 2x2x3mm voxels, TR = 2570ms, TE = 30ms, flip angle = 90°, matrix = 80x78mm, FOV = 220x220mm, parallel acquisition).

2.5.2. fMRI task

The fMRI task was an event-related semantic feature judgment task (Meier et al., 2018). For each patient, stimuli presented in the scanner included pictures of their assigned 36-item training set, 36 untrained items from the same semantic categories as the trained items, 36 items from an unrelated control category, and 36 scrambled pictures, for a total of 144 items. Stimuli were split in half and presented in random order in two separate runs with 72 items in each. The total duration of each run was 9 minutes and 36 seconds. For each trial, a picture was presented on the screen (Figure 1). One second later, a written semantic feature appeared beneath it and remained for four seconds. Patients were instructed to indicate whether or not the feature related to the picture by pressing corresponding buttons on a response box using their left index and middle fingers. Responses were recorded from the onset of the written feature until the onset of the subsequent fixation. Trials were separated by a randomly alternating two- or four-second ISI, during which a black fixation cross was presented. fMRI task accuracy for the 36 items in participants’ assigned training sets is reported in Table 1.

Figure 1. fMRI semantic feature judgment task.

Pictures were presented for five seconds. One second after the onset of a picture (e.g., cauliflower), a written feature appeared below it (e.g., “is juicy”), and the participant had four seconds to indicate via button press whether the feature accurately described the item. For the scrambled condition, distorted pictures were presented in color or black and white and accompanied by the feature “color” or “black and white.” A black fixation cross appeared on the screen for two or four seconds between each trial.

2.5.3. Preprocessing and network nodes

Initial preprocessing was performed in SPM12 (Wellcome Trust Centre for Neuroimaging). Slice timing correction with reference to the middle slice was conducted to account for differences in the timing of slice acquisition. Realignment to the mean functional image was used to correct for head motion in the scanner, and functional images were coregistered with the T1 structural scan. Based on SPM12’s tissue probability maps, structural scans were segmented into gray matter, white matter, and cerebrospinal fluid. Images were normalized to MNI space with a 4^th degree b-spline interpolation. Using MRIcron (Rorden & Brett, 2000), lesion masks, wherein lesioned voxels were deleted, and lesion maps, wherein lesioned voxels were preserved, were drawn in subjects’ native space by author YP or a trained research assistant. Lesion masks and maps were incorporated during the aforementioned coregistration, segmentation, and normalization (Brett et al., 2001). A group-level map depicting the overlap of patients’ lesions is provided in Figure 2 and individual lesions are shown in Supplementary Figure 1.

Figure 2. Overlap of patient lesions.

The legend in the lower right-hand corner shows correspondence between a given color and the approximate number of participants whose lesions intersect in a given voxel. Black and dark red represent areas with minimal lesion overlap (i.e., ≤ 3 participants); bright white represents areas with maximal lesion overlap (i.e., ~12 participants).

Additional preprocessing was conducted with the CONN toolbox (Whitfield-Gabrieli & Nieto-Castanon, 2012). Preprocessed T1-structural images, segmented tissue maps, and functional images were imported into CONN. fMRI task onsets and durations were entered in CONN, separated by condition (i.e., trained pictures, scrambled pictures, untrained pictures, control pictures, fixation), and convolved with the canonical hemodynamic response function. Motion outliers in functional volumes were identified via the Artifact Detection Tool (https://www.nitrc.org/projects/artifact_detect) using the intermediate settings in CONN (i.e., outliers were flagged based on a composite motion threshold of .9 mm and a global signal threshold of 5 SDs from the mean). Using CONN’s denoising process, linear regression and high-pass filtering were performed to remove motion, physiological noise, and other artifacts from the BOLD time series. Specifically, a high-pass filter of .01 Hz was applied and aCompCor (Behzadi et al., 2007) was used to decompose the BOLD time series in white matter and CSF into 16 principal components each. Realignment parameters, motion outliers (i.e., scrubbing parameters), and the first five white matter and first five CSF components were regressed from the BOLD signal. Additionally, per recommendations from Whitfield-Gabrieli and Nieto-Castanon (2012), the main effects of task conditions were regressed out to reduce the impact of region-to-region coactivation on measures of functional connectivity.

The network of interest in the present study comprised 44 anatomical regions of interest (ROIs) previously linked to semantic processing in healthy controls and/or patients with aphasia (Supplementary Table 1). ROIs were specified using the AAL atlas (Tzourio-Mazoyer et al., 2002) and subsequently modified on a subject-by-subject basis to account for patients’ lesions, in accordance with procedures previously used by our group (Johnson et al., 2019; Meier et al., 2018, 2019a; Sims et al., 2016). This process involves overlaying a patient’s lesion map with all the selected atlas-based ROIs. Any portion of an ROI that intersects with the lesion map is deleted, resulting in a set of ROIs that are unique to that patient and represent only the spared tissue in each region.

Pairwise Fisher-transformed Pearson correlations in the average residual BOLD time series were computed between all ROIs for the condition of interest (i.e., items from patients’ training sets, including both correct and incorrect trials). This resulted in a 44x44-region correlation matrix for each patient that represented functional connectivity in the semantic network during judgments related to the specific items on which patients would be trained during therapy. Correlation matrices were also generated for each of the other task conditions, though they were not further analyzed in this study.

2.5.4. Network measures

Patients’ functional connectivity matrices for the trained condition were entered in BRAPH (Mijalkov et al., 2017) and converted to weighted, undirected graphs (Barrat et al., 2004; Newman, 2004; Rubinov & Sporns, 2010) comprised of the 44 regions described above (i.e., nodes) and all of the positive correlations (i.e., edges) between them. Following these steps, there were no disconnected nodes in any of the participants’ graphs. Edge counts for each participant and summary statistics by subgroup are reported in Supplementary Table 2; there was not a significant difference in number of edges between responders and nonresponders; t₂₁ = 1.06, p = .30. Next, the weighted variants of five local (i.e., node-level) and four global (i.e., network-level) graph measures were calculated in BRAPH:

2.5.4.1. Local graph measures

Node strength is an integration measure that summarizes a node’s functional connectivity within a weighted network. It is computed as the sum of the weights of all a node’s edges (Barrat et al., 2004). Thus, a node with high strength has more and/or stronger connections to other nodes than a node with low strength.

Nodal global efficiency is an integration measure that refers to a node’s capacity to exchange information with all other nodes in the network. In a weighted graph, this is calculated as the inverse of the shortest path from one node to another, measured via the weights of all the edges between them (Latora & Marchiori, 2001). Therefore, a given node’s global efficiency is obtained by averaging the inverse shortest path from that node to every other node in the network. A node with high global efficiency should be capable of sharing and receiving information from across the network more easily than a node with low global efficiency.

Clustering coefficient is a segregation measure that indicates whether a node is part of a tightly interconnected cluster that may be capable of specialized processing (Rubinov & Sporns, 2010). If two of a node’s neighbors are connected to each other, they form a closed triangle with that node. The clustering coefficient is computed as the proportion of triangles around a node relative to the number of triangles that could possibly be formed around that node (Watts & Strogatz, 1998). In BRAPH’s weighted variant, triangles are represented by the geometric mean of their edge weights. A larger clustering coefficient indicates that a node is part of a highly interconnected cluster.

Local efficiency of a given node is a segregation measure that is calculated similarly to global efficiency, but only for the subnetwork comprised of that node’s neighbors. As such, it describes the flow of information between pairs of these neighbors, either through direct or indirect connections with one another (Fornito et al., 2016). High local efficiency indicates that a node’s neighbors are in close communication with one another.

Betweenness centrality is the proportion of shortest paths between all nodes in the network that pass through a particular node (Rubinov & Sporns, 2010). The more shortest paths that pass through a node, the higher its betweenness centrality and, presumably, the more integral it is to the efficient transfer of information across the network (i.e., the more likely it can be considered a central hub within the network).

2.5.4.2. Global graph measures

Average strength (i.e., network strength) refers to the average of the strengths of every node in the network. Thus, it describes the overall functional connectivity of a weighted network; the higher the average strength of the network, the greater the connectivity among its nodes.

Network global efficiency, computed as the average of each node’s global efficiency, summarizes the capacity for information to be transferred and combined from regions dispersed throughout the network (Rubinov & Sporns, 2010). Greater average global efficiency is indicative of a more highly integrated network.

Average clustering coefficient is the average of the clustering coefficients of all the network’s nodes. A network with high average clustering coefficient would be comprised of more densely connected clusters than a network with low average clustering coefficient, and therefore, would likely have an advantage with respect to efficient local processing.

Network local efficiency is the average local efficiency of all the nodes in the network. Similar to the average clustering coefficient, the higher the local efficiency, the more the network consists of highly interconnected clusters of regions.

2.5.5. Statistical analyses

Relationships between global network measures and treatment outcome were first investigated by separating patients according to the treatment response classifications utilized in our previous investigations (Gilmore et al., 2018; Johnson et al., 2019). Group differences in each of the global measures were analyzed in BRAPH via permutation tests with 1,000 replications each (Mijalkov et al., 2017). Results were FDR-corrected for four comparisons (i.e., one per measure); FDR-corrected p-values are reported as q-values.

Next, to determine if any global measures predicted treatment outcomes when controlling for lesion volume and age, which have previously been associated with recovery in aphasia, four hierarchical regression analyses were conducted. In all models, the dependent variable was improvement in naming from pre- to post-treatment as indicated by average effect size (i.e., averaged across the two trained categories). All variables (dependent and independent) were standardized before the regression models were constructed. First, a reduced/base model was constructed with lesion volume and age as the independent variables. Next, four full models were constructed, each including one global graph measure (i.e., average network strength, network global efficiency, network local efficiency, or average clustering coefficient) as an independent variable, plus age and lesion volume as covariates. F-tests were used to examine each model’s performance individually and determine if the addition of graph measures improved the fit of the models relative to the base model. Finally, because we found that responders and nonresponders differed in fMRI task accuracy (as reported in section 3.2 of the results), a follow-up analysis was conducted in which task accuracy (i.e., an index of semantic feature judgment ability) was added as an independent variable to the full model that best explained the variance in average effect size. An FDR correction was applied to control for multiple comparisons (i.e., six individual model analyses plus five between-model comparisons). Regression analyses were conducted using R (R Core Team, 2019) and RStudio (RStudio Team, 2019). Global assumptions and potential multicollinearity in the models were assessed with the ‘gvlma’ (Pena & Slate, 2019) and ‘car’ (Fox & Weisberg, 2019) packages, respectively, and the ‘sjPlot’ package (Lüdecke, 2019) was used for presentation of regression results.

Local network measures were analyzed in a manner comparable to the initial analysis of global measures. That is, permutation tests with 1,000 replications per comparison were performed for the five local measures in all 44 nodes of the network, allowing for the identification of nodes in which there were statistically significant differences in local measures between responders and nonresponders. Results were FDR-corrected for multiple comparisons across nodes.

3. Results

3.1. Behavioral results

Six participants (BU04, BU06, BU08, BU10, BU14, and BU17) completed treatment early (i.e., in fewer than 24 sessions) because they reached 90% accuracy in naming trained items on two consecutive probes; additionally, BU12 withdrew from treatment early for personal reasons, though he completed all post-treatment evaluations as planned. The number of treatment sessions completed by each participant is reported in Table 1.

Analyses of treated and untreated patients described in Johnson et al. (2019) and Gilmore et al. (2018) confirmed the efficacy of the semantic treatment utilized in this study at a group-level. As shown in Table 1, effect sizes varied among treated patients, 17 of whom were classified as responders and nine of whom were classified as nonresponders.

3.2. fMRI task accuracy

A Welch’s two-sample t-test indicated that responders were significantly more accurate on the fMRI semantic feature judgment task than nonresponders (t₁₇ = 3.81, p < .01). See Table 1 for individual and mean scores for each group.

3.3. Global graph results

As described in section 2, Materials and Methods, the relationship between global graph measures and treatment response was initially investigated by comparing responders and nonresponders on each measure. Permutation tests conducted in BRAPH indicated that responders had significantly higher average strength and network global efficiency than nonresponders, while differences in mean clustering coefficient and network local efficiency were not significant (Table 2.).

Table 2.

Mean values for global and local network measures and between-group differences. For local measures, only nodes with significant group differences after FDR correction are reported. Bold text indicates q < .05. n.s. = not significant.

Measure	Resp.	Nonresp.	Difference (Resp. - Nonresp.)	p	q
Global Measures
Average strength	7.37	6.08	1.29	0.013	0.038
Network global efficiency	0.26	0.23	0.03	0.019	0.038
Mean clustering coefficient	0.18	0.15	0.03	0.120	0.120
Network local efficiency	0.56	0.40	0.16	0.058	0.077
Local Measures & Nodes
Node strength
LACC	7.64	4.87	2.77	0.002	0.044
RIFGtri	7.39	5.15	2.24	0.003	0.044
RIFGop	6.71	4.50	2.21	0.003	0.044
Node global efficiency
LIFGop	0.22	0.15	0.07	0.007	0.044
LMFG	0.29	0.23	0.06	0.002	0.035
LPCUN	0.25	0.21	0.04	0.005	0.037
RIFGop	0.26	0.21	0.05	0.002	0.035
RMFG	0.29	0.25	0.04	0.004	0.035
RSFG	0.29	0.25	0.05	0.004	0.035
RAG	0.24	0.20	0.04	0.004	0.035
Node local efficiency
RIFGop	0.48	0.31	0.17	0.001	0.044
Clustering coefficient	-	-	-	n.s.	n.s.
Betweenness centrality	-	-	-	n.s.	n.s.

Next, to determine if any of the global measures predicted treatment outcomes when controlling for other potential explanatory variables, several regression models were constructed, as follows (complete regression results are provided in Table 3). First, a reduced model that included only the a priori-selected covariates lesion volume and age as independent variables was constructed (i.e., covariates-only model) to serve as a basis for comparison for the models that also included graph measures. The covariates-only model was significant and explained 20.0% of the variance in effect size (F_2,23 = 4.13, q < .05, adjusted R² = .200). There was a significant negative association between lesion volume and effect size, such that a 1 standard deviation increase in lesion volume was associated with a .49 standard deviation decrease in effect size (β = −.49, p = .01) (i.e., those with larger lesions had smaller effect sizes).

Table 3.

Results of reduced (covariates-only), full (covariates + graph measures), and follow-up (covariates + strength + task accuracy) regression models predicting average treatment effect size in 26 patients with aphasia.

	covariates-only				covariates + strength				covariates + global efficiency				covariates + clustering coefficient				covariates + local efficiency				covariates + strength + accuracy
Pred.	β	SE	t	p	β	SE	t	p	β	SE	t	p	β	SE	t	p	β	SE	t	p	β	SE	t	p
Int.	0.00	0.18	0.03	0.98	0.00	0.16	0.02	0.98	0.00	0.16	0.02	0.98	0.00	0.17	0.02	0.98	0.00	0.17	0.02	0.98	0.00	0.15	0.02	0.99
Lesion Vol.	−0.49	0.19	−2.64	0.01	−0.32	0.19	−1.73	0.10	−0.33	0.19	−1.73	0.10	−0.40	0.19	−2.16	0.04	−0.39	0.18	−2.09	<0.05	−0.29	0.17	−1.65	0.11
Age	−0.32	0.18	−1.73	0.10	−0.39	0.17	−2.25	0.03	−0.39	0.17	−2.24	0.04	−0.39	0.18	−2.15	0.04	−0.40	0.18	−2.22	0.04	−0.26	0.17	−1.51	0.15
Avg. Str.	-	-	-	-	0.43	0.18	2.33	0.03	-	-	-	-	-	-	-	-	-	-	-	-	0.30	0.18	1.62	0.12
NGE	-	-	-	-	-	-	-	-	0.41	0.19	2.15	0.04	-	-	-	-	-	-	-	-	-	-	-	-
MCC	-	-	-	-	-	-	-	-	-	-	-	-	0.32	0.19	1.68	0.11	-	-	-	-	-	-	-	-
NLE	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	0.36	0.18	1.94	0.06	-	-	-	-
Acc.	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	0.35	0.17	2.03	0.06
Model R/adj. R2	0.264/0.200				0.409/0.329				0.392/0.309				0.348/0.259				0.372/0.286				0.506/0.412
Model RSS	18.400				14.770				15.204				16.307				15.705				12.346
Model F (p/q)	4.125 (0.029/0.047)				5.079 (0.008/0.040)				4.725 (0.011/0.040)				3.909 (0.022/0.047)				4.34 (0.015/0.042)				5.381 (0.004/0.040)
Model Comp. F (p/q)*	-				5.408 (0.030/0.047)				4.626 (0.043/0.059)				2.823 (0.107/0.107)				3.776 (0.065/0.071)				4.122 (0.055/0.067)

Abbreviations: Acc., accuracy on fMRI task; Avg. Str., average strength; Int., intercept; Lesion Vol., lesion volume; MCC, mean clustering coefficient; Model Comp., model comparisons; NGE, network global efficiency; NLE, network local efficiency; Pred., predictors (i.e., independent variables); q, q-value (i.e., FDR-corrected p-value); RSS, residual sum of squares.

^*Values reflect results of F-tests for between-model comparisons. Note that the first four entries in this row are based on comparisons between the covariates-only and covariates + graph measure models, while the last entry in the row reflects the comparison between the covariates + strength and covariates + strength + accuracy models.

Next, we fit four full models, each of which included one of the global graph measures plus lesion volume and age as predictors (i.e., covariates + strength model, covariates + global efficiency model, covariates + clustering coefficient model, and covariates + local efficiency model). All four models were significant (all q < .05). The covariates + strength model explained 32.9% of the variance in effect size (F_3,22 = 5.08, adjusted R² = .329), covariates + global efficiency explained 30.9% of the variance (F_3,22 = 4.73, adjusted R² = .309), covariates + clustering coefficient explained 25.9% of the variance (F_3,22 = 3.91, adjusted R² = .259), and covariates + local efficiency explained 28.6% of the variance (F_3,22 = 4.34, adjusted R² = .286). An F-test comparing the covariates + strength and covariates-only models indicated that adding average strength significantly improved the fit of the model (q < .05). Comparable tests showed trending improvements in model fit when global efficiency (q = .06) and local efficiency (q = .07) were added, whereas the addition of clustering coefficient did not improve the model (q = .11).

Within their respective models, average strength (β = .43, t = 2.33, p = .03) and global efficiency (β = .41, t = 2.15, p = .04) were significantly associated with effect size, and there was a trending association between local efficiency and effect size (β = .36, t = 1.94, p = .06). Thus, effect size would be expected to increase as each of these graph measures increased. Clustering coefficient was not significantly associated with effect size (p = .11). Additionally, age was significantly negatively associated with effect size in all four models (β range = −.39 to −.40, t range = −2.25 to −2.15, all p < .05), and lesion volume was significantly negatively associated with effect size in the covariates + clustering coefficient model (β = −.40, t = −2.16, p = .04) and the covariates + local efficiency model (β = −.39, t = −2.09, p < .05), indicating that patients who were older and/or had larger lesions had poorer treatment outcomes.

As stated above in section 3.2 of the Results, responders were significantly more accurate than nonresponders on the fMRI task. Thus, we conducted a follow-up analysis to determine if task performance (i.e., semantic processing/feature judgment skills) predicted patients’ effect sizes, given that prior literature suggests a link between lexical-semantic processing and naming treatment outcomes in aphasia (Dignam et al., 2017). We added task accuracy as an independent variable to the network measure model that best fit the data (i.e., the covariates + strength model). As shown in Table 3, the new model (covariates + strength + accuracy) was significant (F_4,21 = 5.38, q = .04) and explained more variance in effect size than any of the models described previously (adjusted R² = .412), and there was a trending increase in model fit relative to the covariates + strength model (q = .07). Finally, although effect size was not significantly associated with any of the individual predictors in the covariates + strength + accuracy model, there was a trending association with task accuracy (β = .35, t = 2.03, p = .06). Tests of model assumptions performed in R with the ‘gvlma’ (Pena & Slate, 2019) and ‘car’ (Fox & Weisberg, 2019) packages indicated none of the models presented in Table 3 violated linear model assumptions.

We observed that across the regression models, the magnitude of the relationship between lesion volume and average effect size was diminished whenever a graph measure was added as a predictor, suggesting there might be a direct relationship between the graph measures and lesion volume. Correlation tests indicated there were weak-to-moderate trending but non-significant relationships among these variables, such that as lesion volume increased, the graph measures decreased (r ranged from −.33 to −.44, all q > .05; complete results are included in Supplementary Table 3).

In summary, including global graph measures in the regression analyses improved the overall fit of the models relative to the covariates-only model. Average network strength had the strongest association with treatment outcome of all global measures, significantly improved the fit of the model, and explained an additional 12.9% of variance in effect size that was not already accounted for by age and lesion volume. Similar, albeit less substantial and non-significant, contributions were found for global and local efficiency, while clustering coefficient was not associated with treatment outcomes and did not significantly improve the explanatory power of the model. Finally, the follow-up analysis indicated that the fit of the model was improved even further when fMRI task accuracy was included as a predictor, and there was a trending association between accuracy and average effect size, controlling for age, lesion volume, and average strength.

3.3. Local graph analyses

Significant group differences based on nonparametric permutation tests of local measures are shown in Figure 3 and corresponding statistics are provided in Table 2. Relative to nonresponders, responders had significantly greater node strength in left anterior cingulate cortex (LACC) and right IFG pars triangularis (RIFGtri) and pars opercularis (RIFGop); significantly greater global efficiency in left precuneus (LPCUN), right superior frontal gyrus (RSFG) and angular gyrus (RAG), and bilateral IFGop and middle frontal gyrus (MFG); and significantly greater local efficiency in RIFGop. There were no differences in clustering coefficient or betweenness centrality after FDR-correction.

Figure 3. Local (i.e., regional) differences in strength, global efficiency, and local efficiency between responders and nonresponders.

For all significant differences, responders had larger values than nonresponders (after FDR correction). Sphere size and corresponding values indicate the magnitude of the difference between groups. Small blue spheres represent nodes in which there was not a significant group difference. Spheres generally indicate the location of nodes, but do not reflect their precise shape or size. L/R: left hemisphere/right hemisphere; n.s.: not significant.

4. Discussion

We used graph analysis to examine a network of regions previously linked to semantic processing in healthy adults and patients with aphasia in 26 patients who went on to receive naming treatment. Patients with higher pre-treatment graph metrics, particularly those reflecting global network functional connectivity and integration (i.e., average strength and network global efficiency), responded better to treatment than those with lower network measures. Comparisons between responders and nonresponders revealed nodal differences predominantly involving frontal and parietal regions in both hemispheres. The implications of these findings are discussed below.

4.1. The relationship between global network measures and treatment outcome

In line with our hypothesis, patients with higher network strength and global efficiency improved more in naming therapy than those with lower strength and global efficiency. Additionally, average clustering coefficient and local efficiency contributed to models that predicted treatment outcomes and were positively (though not significantly) related to behavioral improvement. The summary measures of network integration (i.e., strength and global efficiency), in particular, differed significantly between responders and nonresponders and were associated with improvement (i.e., average effect size) when controlling for other variables that have also been shown to relate to language recovery (i.e., age and lesion volume). This suggests that poorer treatment outcomes and lower indices of network functional connectivity and efficiency are not merely coincidental effects of more damage or older age; rather, network integration in the chronic stage of aphasia may have a direct, explanatory relationship with treatment response.

The finding that indices of integration were more informative than indices of segregation (as indicated by model fit and unique variance in effect size explained) is consistent with Baliki et al. (2018), who found a significant correlation between pre-treatment resting-state global efficiency and post-treatment language improvement in eight PWA, but no such association with modularity, which reflects the extent to which a network is segregated into distinct subsystems. On the other hand, Tao and Rapp (2019) recently found a significant relationship between treatment outcome and a different measure related to network segregation, average within-module degree z-score, in 15 patients with aphasia (though in this study as well, modularity did not correlate with recovery). There are numerous methodological differences between Tao and Rapp’s study and our own that may explain why our findings are somewhat at odds. Even still, in the present study, two measures of network segregation (i.e., clustering coefficient and local efficiency) contributed to models that significantly predicted effect size, and their independent relationships with effect size went in the direction that would be expected based on Tao and Rapp’s findings (i.e., higher measures of segregation corresponded to higher effect sizes), even though those associations were not significant.

Indeed, a balance between integration and segregation reflects small-world topology, a potentially optimal organizational structure for brain networks, as it supports specialized processing within closely connected clusters but also allows for the efficient distribution of information between disparate regions or clusters (Bullmore & Sporns, 2009; Bullmore & Bassett, 2011; van den Heuvel & Hulshoff Pol, 2010). Thus, while our results most strongly highlight the relationship between integration measures and naming improvement, they are not incongruous with prior findings linking higher segregation to behavioral gains after aphasia treatment. Furthermore, they suggest that the relative retention of small-world properties in the semantic network is an indicator for a successful response to lexical-semantic naming treatment. Given the demands of the treatment we employed (i.e., processing multi-modal input, accessing stored semantic representations, comparing novel features to those representations to judge their veracity, etc.), it may well be advantageous to have network topology that supports efficient localized processing in closely integrated clusters and the capacity for efficient global integration of information between these clusters. In the present study, the latter capability—ease of network-wide communication—may have been especially critical for the act of semantic feature processing, or for sufficiently strengthening semantic representations to the point that picture naming improved.

While our planned analyses demonstrated a relationship between global graph properties and treatment outcome, the follow-up regression model that included fMRI task accuracy as an additional predictor was the best-fit model identified in this study. Though average effect size was not significantly associated with any of the individual predictors in the follow-up model, it did have a trending relationship with task accuracy. This is notable because the fMRI task required participants to judge the veracity of semantic features pertaining to the items on which they would eventually be trained; as such, it was essentially an index of their pre-treatment semantic processing skills for those items. Moreover, the task closely approximated the predominant activity in which participants were engaged throughout treatment: analyzing and judging semantic features. Thus, our results suggest that participants’ ability to access lexical-semantic items and make accurate judgments about their features is related to their capacity for improvement in naming the same items subsequent to a semantically oriented feature verification treatment. Furthermore, prior studies have demonstrated an association between naming therapy outcomes and lexical-semantic processing abilities (Dignam et al., 2017), as well as cognitive and linguistic skills more generally (Gilmore et al., 2019; Lambon Ralph et al., 2010b). Thus, future research focused on predicting treatment outcomes and the development of prognostic clinical tools should account for baseline cognitive-linguistic abilities, and perhaps even baseline lexical-semantic processing with respect to patients’ specific treatment stimuli.

4.2. A network of frontal and parietal regions underlies treatment success

Given that treatment outcomes were associated with some global network measures, and that global measures represent averages of nodal properties, it was not unreasonable to expect that responders and nonresponders would differ on at least some nodal measures. Indeed, node-level differences between the groups were identified in nine frontal and parietal regions distributed across the left and right hemispheres. That the other 35 regions in the network did not differ between groups suggests that these nine regions may be particularly relevant to patients’ response to treatment.

Most differences between responders and nonresponders were based on node strength and global efficiency, which reflect the extent to which a region is integrated within the larger network. Responders had higher node strength in LACC, RIFGtri and RIFGop; and higher global efficiency in LPCUN, RSFG, RAG, and bilateral IFGop and MFG. Notably, there were no group differences in temporal node properties, which may suggest that core semantic regions, such as the anterior temporal lobes, were similarly situated within the networks of responders and nonresponders; however, given our sample size and potential differences in variability across measures and regions, further research is warranted to verify this interpretation.

In contrast to the lack of differences among core semantic regions, differences on local measures largely involved regions that have been associated with a range of processes including: lexical-semantic control/selection and phonological processing (i.e., IFG [Thompson-Schill, D’Esposito, Aguirre, & Farah, 1997; Vigneau et al., 2011]); cognitive control, monitoring, and effort (i.e., ACC [Shenhav, Cohen, & Botvinick, 2016]); domain-general cognitive functions (i.e., MFG [Fedorenko, Duncan, & Kanwisher, 2013], ACC and AG [J. Duncan, 2010]); working memory, cognitive and motor control, and attention (i.e., SFG [Li et al., 2013]); visuo-spatial and self-referential processing, and episodic memory (i.e., PCUN [Cavanna & Trimble, 2006]); and default mode processes (i.e., AG [Humphreys & Lambon Ralph, 2015; Seghier, 2010] and PCUN [Cavanna & Trimble, 2006; Utevsky, Smith, & Huettel, 2014]). Thus, responders and nonresponders did not exclusively differ in the nodal properties of “classic” left perisylvian language regions; rather, differences emerged in the integration of areas that support a wide range of extra-linguistic cognitive functions, suggesting these regions may play a critical compensatory role in language recovery.

Additionally, three of these regions—LPCUN and bilateral MFG—overlap with nodes in the left frontoparietal cognitive-linguistic network examined by Zhu et al. in subacute stroke patients (D. Zhu et al., 2014). Those authors found an association between reduced LFPN functional connectivity and poorer reading and auditory comprehension skills, and a marginal association between an increase in LFPN functional connectivity and improvements in comprehension from one to two months post-stroke. The recruitment³ of a similar, though not identical, left fronto-temporo-parietal language network was also previously observed to correspond to speech production abilities in post-stroke aphasia (Geranmayeh et al., 2016). Therefore, our results lend further support to the view that the LFPN or its constituent parts are associated with the restoration of speech-language functions in post-stroke aphasia.

The bilateral distribution of regions whose network characteristics differentiate responders and nonresponders is also of interest. A longstanding question in the aphasia literature is whether language recovery is best when it is underpinned by the left hemisphere or the right hemisphere. In this study, semantic naming treatment was most effective for patients in whom regions in both hemispheres were efficiently linked to the broader semantic network. Thus, these results align well with studies that have highlighted bilateral contributions to behavioral improvement (Fridriksson et al., 2006, 2007; Kiran et al., 2015; Menke et al., 2009; Nardo et al., 2017).

Finally, responders had greater local efficiency than nonresponders in RIFGop, indicating that this region was part of a more efficient local processing subsystem in patients with better treatment outcomes than in those with poorer outcomes. Functional activation studies have demonstrated an association between recruitment of RIFGop and treatment-related naming improvement in patients with aphasia (Crosson et al., 2009; Nardo et al., 2017). Perhaps, the capacity to recruit RIFGop during naming is dependent on its local efficiency; that is, when it is part of a highly interconnected cluster, as in responders, it may be better positioned to contribute to successful word retrieval. However, this is a speculative interpretation of the results; further research is required to determine the extent to which regional activation and graph properties are associated in post-stroke aphasia.

The results of this study are encouraging, as they point to a real and meaningful relationship between functional connectivity-based graph theoretic measures and naming therapy outcomes in a patient population whose response to intervention is notoriously heterogeneous. However, a limitation is that the strongest regression model we identified (i.e., the covariates + strength + accuracy model) explained only about 41% of the variance in effect size, serving as a reminder that other critical factors must be identified and investigated as well. Bonilha et al. (2016) found that structural network properties (i.e., normalized small-worldness and left temporal betweenness centrality) also predict naming improvement; thus, future investigations should consider the relative contributions and combined explanatory power of structural and functional network properties.

Another limitation is that this study did not include an analysis of post-treatment fMRI data, so we cannot comment on whether participants’ behavioral changes after treatment corresponded with changes in functional connectivity-based network measures in the context of semantic processing. However, we previously identified changes in pre- vs. post-treatment functional activation associated with a naming task in responders, as well as a relative lack of change in nonresponders, in a study involving the same participants (Johnson et al., 2019), and we intend to address treatment-related changes in functional connectivity in the future.

We would also note that the functional connectivity analyses in this study were based on timeseries data for both correct and incorrect responses. Given that the responders and nonresponders differed in their accuracy on the fMRI task, it is possible that our results reflect differences in cognitive processes such as effort or error awareness in addition to differences in the functional organization of the semantic network. Lastly, our analyses were based on fMRI data associated with participants’ trained items, and it is possible that including other conditions (i.e., untrained items or scrambled pictures) might have produced different results.

5. Conclusions

Pre-treatment global network properties differed between responders and nonresponders and predicted patients’ response to semantic naming therapy. Measures of network integration were positively associated with behavioral improvement when controlling for lesion volume and age. Network strength, especially, explained a significant proportion of additional variance in treatment outcome that was not captured by the covariates. Thus, higher strength and global efficiency in the semantic network appear to be favorable prognostic indicators for improvement from semantic naming treatment. We also identified region-specific differences between patients who improved most and least in treatment that suggest the functional connectivity of frontal and parietal regions that support a wide range of cognitive functions may be critical for a favorable outcome. Future research should investigate the extent to which prognostic accuracy can be improved by combining functional graph measures with other metrics, including structural connectivity.

Statement of significance to the neurobiology of language

This is one of the first and largest studies to examine associations between pre-treatment functional connectivity-based network characteristics and post-treatment language outcomes in individuals with aphasia, a highly heterogenous clinical population. Results indicate that specific functional properties of the semantic network and individual regions therein are associated with the extent to which patients respond to semantically focused therapy for naming impairments. Critically, global network measures, particularly integration measures, were associated with treatment outcomes even after controlling for demographic and stroke factors previously linked to aphasia recovery, and model fit was further improved by the addition of fMRI task accuracy, an index of semantic processing skills.

Abbreviations

ACC

anterior cingulate cortex

AG

angular gyrus

DMN

default mode network

FDR

false discovery rate

IFG

inferior frontal gyrus

IFGop

inferior frontal gyrus pars opercularis

IFGtri

inferior frontal gyrus pars triangularis

ISI

interstimulus interval

L/R (as prefix to region acronyms)

left/right

LFPN

left fronto-parietal network

MFG

middle frontal gyrus

PCUN

precuneus

ROI

region of interest

SFG

superior frontal gyrus