Abnormally weak functional connections get stronger in chronic stroke patients who benefit from naming therapy

Abstract

Language recovery in aphasia is likely supported by a network of brain regions, but few studies have investigated treatment-related changes in functional connectivity while controlling for the absence of treatment. We examined functional connectivity in a 38-region picture-naming network in 30 patients with chronic aphasia who did or did not receive naming therapy. Compared to healthy controls, patients had abnormally low connectivity in a subset of connections from the naming network. Linear mixed models showed that the connectivity of abnormal connections increased significantly in patients who benefited from therapy, but not in those who did not benefit from or receive therapy. Changes in responders were specific to abnormal connections and did not extend to the larger network. Thus, successful naming therapy was associated with increased connectivity in connections that were abnormal prior to treatment. The potential to strengthen such connections may be a prerequisite for a successful treatment response.

1. Introduction

Stroke-related damage to language regions in the brain causes aphasia, a communication disorder that often persists into the chronic stage of stroke recovery. Speech-language rehabilitation can improve language and communication skills (Brady et al., 2016), yet treatment outcomes and the extent of recovery vary from patient to patient, and the neural mechanisms that explain these differences and that underlie recovery in general are incompletely understood. Some evidence suggests that better outcomes are associated with recruitment of spared tissue in the ipsilesional (usually left) hemisphere of the brain (Fridriksson, 2010; Saur et al., 2006), but improvement has also been linked to right hemisphere and bilateral activation (Crosson et al., 2009; Johnson et al., 2019; Menke et al., 2009; Nardo et al., 2017). Collectively, these results suggest language recovery may be subserved by a distributed network of spared resources, the composition of which likely varies across patients depending on the nature of their lesions. As such, examining the brain as a network of functionally connected nodes may provide new insights into the neural mechanisms that underlie or influence language recovery.

To this end, the field of aphasiology has seen a growing interest in techniques such as functional connectivity (i.e., correlations in brain activity between brain regions) derived from resting-state or task-based functional neuroimaging, effective connectivity (i.e., analyses of directional relationships between brain regions), and graph theoretical analyses, which characterize connectivity-based brain network properties at local (i.e., regional) and global (i.e., network-wide) levels. As described in further detail below, these approaches have begun to provide insight into the course and nature of network-level changes following stroke and subsequent rehabilitation.

Studies of acute/subacute stroke have shown that stroke-related brain damage results in abnormalities in functional connectivity, and that these abnormalities are associated with corresponding behavioral changes. This was clearly demonstrated by Siegel et al. (2016), who found that, relative to healthy control subjects, a group of 100 early stroke survivors had abnormally low inter-hemispheric connectivity within most of the resting-state networks they examined, but abnormally high intra-hemispheric connectivity between the ipsilesional dorsal attention and default mode networks. Importantly, in addition to identifying this “physiological network phenotype of stroke injury,” as the authors referred to it, Siegel et al. (2016) found a significant relationship between patients’ language performance and functional connectivity, such that those who performed best on language measures had higher inter-hemispheric functional connectivity and higher intra-hemispheric functional connectivity in the left hemisphere, a result consistent with previous studies suggesting language functions rely on canonical left-hemisphere language areas, as well as the engagement of some right hemisphere regions.

Subsequent studies focusing specifically on patients with post-stroke aphasia have identified similar patterns as those reported by Siegel et al. (2016). In a recent review, Klingbeil and colleagues (2017) found that cross-sectional studies of resting-state functional connectivity in people with aphasia—including those in the chronic stage of recovery (i.e., months or years post-onset)—have predominantly reported that patients have abnormally low connectivity among language and other networks and within and between the ipsilesional and contralesional hemispheres (but see Yang et al., 2017, for an alternative finding). More recently, in a study of background connectivity (i.e., functional connectivity derived from the residual timecourse of task-based fMRI after task-related responses have been removed) in patients with chronic post-stroke dysgraphia, Tao and Rapp (2020) found that patients had lower interhemispheric functional connectivity than healthy controls and a mix of higher- and lower-than-normal intra-hemispheric connections. Our own investigation of task-based effective connectivity also highlighted significant differences between patients with chronic aphasia and healthy controls, including lower connectivity in patients than controls among left-hemisphere (ipsilesional) connections in a circumscribed bilateral semantic processing network (Meier et al., 2019a). Also consistent with Siegel et al. (2016), studies of those with post-stroke language deficits have reported associations between various measures of functional connectivity and cognitive-linguistic skills (Geranmayeh et al., 2016; Sandberg, 2017; Tao & Rapp, 2019, 2020). Collectively, the literature on functional connectivity and network-level characteristics in individuals with post-stroke aphasia indicates a tendency for patients to have abnormally low connectivity, particularly—though not exclusively—between the left and right hemispheres. Furthermore, there appears to be a meaningful relationship between behavioral performance and network-level brain function.

While the findings above offer insight into the early and long-lasting effects of stroke on neural function at the network level, several longitudinal intervention studies have explored connectivity-based changes in brain function following aphasia rehabilitation. Briefly, the most consistent findings from this literature to date are that A) treatment engenders changes—especially increases—in functional connectivity or shifts in network properties that bring patients into closer alignment with healthy controls; and B) that such changes correlate with the extent of treatment-related behavioral improvement that patients experience (Duncan & Small, 2016; Gili et al., 2016; Marangolo et al., 2016; Sandberg et al., 2015; Santhanam et al., 2018; Tao & Rapp, 2019, 2020; van Hees et al., 2014). Still, despite these trends, the location and nature of treatment-induced changes in functional connectivity and the relationship between connectivity changes and behavior are nuanced and vary across studies. This variability likely stems from several sources, including differences in neuroimaging methods, network delineation techniques, the particular connectivity or network metrics utilized, statistical approaches, and rehabilitation targets and strategies. Furthermore, despite the important contributions of these studies, they are small in number, and while several compared patients to healthy control subjects (Santhanam et al., 2018; Tao & Rapp, 2019, 2020; van Hees et al., 2014), only Sandberg et al. (2015) reported comparisons to untreated patients as a control for changes in connectivity that might be driven by factors other than treatment (and even then, there were only three such patient controls).

In the present study, we examined functional connectivity in 30 patients with chronic aphasia, 20 of whom received language therapy for naming impairments and 10 of whom were untreated but followed longitudinally for a period of time comparable to the treated patients. Unlike previous studies, we aimed to explicitly investigate how treatment affected connections that were functionally abnormal in patients prior to intervention (note that we refer to this set of connections as an “abnormal connectome” throughout this paper). We also sought to determine if changes in the abnormal connectome varied as a function of treatment outcome, as numerous aphasia treatment studies have reported differential outcomes across individual patients and such variability may obscure the presence or magnitude of group-level neural changes. In previous studies involving patient samples that overlap with the present study, we identified subsets of patients who did and did not respond favorably to naming treatment and found corresponding differences in functional activation in brain regions involved in picture naming and graph-theoretic properties of a semantic network (Johnson et al., 2019, 2020). While those studies highlighted meaningful relationships between the rehabilitation of naming deficits and neural function in chronic aphasia, they did not examine network-level effects of treatment as indicated by changes in functional connectivity. Thus, the present study aimed at addressing the following three aims: 1) to identify the presence and extent of functional connections within a picture-naming network that differed between patients with aphasia and healthy controls (i.e., the abnormal connectome) at the start of the study; 2) to determine if naming therapy affected functional connectivity in the abnormal connectome; and 3) to determine if changes in the abnormal connectome differed depending on treatment response (i.e., connectivity differences between responders and nonresponders). Based on the literature on functional connectivity in aphasia described above, we hypothesized that (a) patients would have abnormally low connectivity relative to healthy controls throughout the picture-naming language network, (b) that connectivity would increase in patients who received treatment but remain unchanged in those who did not, and (c) that greater changes in the abnormal connectome would be observed in treatment responders than nonresponders.

2. Methods

2.1. Participants

This study was conducted under the Center for the Neurobiology of Language Recovery (CNLR; https://cnlr.northwestern.edu/). Most research activities occurred in and around Boston, Massachusetts, though two individuals from metropolitan Chicago, Illinois, also participated. Most or all participants described in this study have also been included in related investigations of treatment efficacy and the neural bases of language rehabilitation (Gilmore et al., 2018; Johnson et al., 2019, 2020; Kiran et al., 2015; Meier et al., 2016, 2018, 2019a, 2019b), but the objectives, analyses, and results of the current study are original. Informed consent was obtained in writing from all participants and all research activities were conducted in accordance with Boston University, Massachusetts General Hospital, and Northwestern University’s IRB protocols. Complete details pertaining to recruitment, screening, and eligibility were described in Johnson et al. (2019).

Thirty-five patients with chronic aphasia due to left hemisphere stroke at least six months prior to screening were enrolled in this study. Inclusion was based on premorbid English proficiency and naming impairments identified via an extensive picture-naming assessment. One in four patients was asked to join an untreated patient control group, though they had the option to enroll directly in the treatment group instead. Twenty-three patients were enrolled in the treatment group and 12 were enrolled in the untreated group. However, three treatment participants were excluded from the present study due to acquisition errors during fMRI data collection and two untreated participants voluntarily withdrew in the course of the study. Thus, the present study includes data from 20 treated patients and 10 untreated patients¹; these groups did not differ significantly on any demographic or personal variables (i.e., age, years of education, lesion volume, or months post-onset of aphasia) (Table 1).

Table 1.

Demographic and assessment data in treated and untreated patients. Bold text indicates p < .05.

	Treated Mean (SD)	Untreated Mean (SD)	t	df	p
Age	63.10 (10.00)	59.00 (11.79)	0.94	15.67	.360
Education (years)	15.23 (2.04)	14.65 (2.24)	0.68	16.69	.504
Lesion Volume	132.67 (76.68)	112.66 (94.60)	0.58	15.11	.570
MPO	57.45 (51.63)	85.20 (141.91)	−0.60	10.21	.562
AQ	61.59 (24.23)	65.80 (24.57)	−0.44	17.88	.662
BNT	24.30 (20.03)	25.67 (20.62)	−0.17	15.09	.870
PPT	45.45 (5.21)	48.80 (2.20)	−2.47	27.57	.020
fMRI Task Accuracy – pre	25.65 (20.72)	32.93 (24.38)	−0.81	15.71	.430
fMRI Task Accuracy – post	45.05 (36.12)	30.64 (24.09)	1.30	25.40	.206
Change in fMRI Task Accuracy	19.40 (19.80)	−2.29 (9.80)	4.02	27.99	< .001
Average Treatment Effect Size*	5.53 (3.57)	0.65 (1.58)	5.19	28	< .001

^*Single-subject effect sizes were calculated according to Beeson and Robey (2006), as described in section 2.4 and footnote 3.

Abbreviations: AQ, Western Aphasia Battery Aphasia Quotient; BNT, Boston Naming Test; MPO, months post onset of aphasia; PPT, Pyramids and Palm Trees Test

Seventeen right-handed adults (also described in Johnson et al. 2019) with no history of neurological illness (mean age = 60.41 years, SD = 10.81; mean years of education = 16.29, SD = 2.16; 11 males) completed the fMRI task at a single time point only. There were no significant differences in age or education between the healthy controls and the full group of patients (age: t_32.8 = −0.41, p = 0.69; education: t_24.7 = −1.83, p = 0.08).

2.2. Study design

For all patient participants, the study involved three stages (Figure 1), which we refer to as pre-treatment/hold (i.e., pre-treatment or pre-hold stage for treated and untreated participants, respectively); treatment/hold; and post-treatment/hold (i.e., post-treatment or post-hold stage for treated and untreated participants, respectively). During pre-treatment/hold, participants were seen at a research speech-language pathology clinic for a series of appointments during which several cognitive-linguistic assessments were administered. A 180-item naming probe was administered at least three times to obtain a measure of baseline naming performance. Aphasia severity was quantified via the Aphasia Quotient (AQ) from the Western Aphasia Battery-Revised (WAB-R) (Kertesz, 2007), naming abilities were measured via the Boston Naming Test (BNT) (Kaplan et al., 2001), and nonverbal semantic processing was assessed with the Pyramids and Palm Trees test (PPT) (Howard & Patterson, 1992). As shown in Table 1, the untreated patient group performed significantly better on the PPT than the treated group, though both groups performed relatively well, on average. No significant differences were found on the BNT or WAB-AQ. Once all assessments were administered, patients completed two fMRI scans (see section 2.5 for details).

Figure 1.

Participant enrollment and study stages (A); overview of naming treatment activities (B) and the fMRI naming task (C).

During the treatment/hold stage, participants in the treatment group underwent a course of behavioral naming therapy administered over approximately 12 weeks (see sections 2.3 and 2.4, below), while no therapy was provided to untreated patients. After completion of all treatment sessions (for the treatment group) or 12 weeks without treatment (for the untreated group), the patients entered the post-treatment/hold stage, during which they underwent two more fMRI scans and three more administrations of the naming probe.

2.3. Naming probe and treatment stimuli

As described in our prior studies (Gilmore et al., 2018; Johnson et al., 2019, 2020; Meier et al., 2016), the picture-naming probe consisted of 36 nouns from each of four experimental categories (vegetables, birds, clothing, furniture) and a control category (fruit), which were balanced on lexical frequency, concreteness and familiarity (Coltheart, 1981; van der Wouden, 1990). Pictures were presented in random order at least three times during pre- and post-treatment/hold. Based on their performance on the naming probe, each patient was assigned 36 items from two experimental categories (18 items per category), which were presented during the fMRI task and, for patients in the treatment group, trained during treatment sessions. The remaining items from the assigned categories (n = 36, 18 per category) were also presented during naming probes and in the scanner to address questions related to treatment-driven generalization (i.e., improvement on untreated but related items). These questions are beyond the scope of the present study; interested readers are referred to Gilmore et al. (2018). The control category (fruit) was also presented to all participants to control for repeated exposure to items during pre- and post-treatment testing, a potential confound for identifying treatment effects in aphasia studies. Stimuli assignments were counterbalanced to ensure relatively equal distribution of the categories across participants, while also accounting for their accuracy within their assigned categories. A more in-depth description of stimuli assignment procedures is provided in Johnson et al. (2020). Healthy control participants were assigned 36 items from two experimental categories in a similar counterbalanced fashion.

2.4. Treatment

The intervention in this study has been described in detail elsewhere (Gilmore et al., 2018). Briefly, patients sorted pictured items into superordinate categories, attempted to name trained items and judge whether semantic features applied to them, and performed a generative naming task (Figure 1). Naming performance was monitored via weekly probes and treatment was typically discontinued after a maximum of 24 two-hour sessions² or when a patient correctly named 90% of trained items in both experimental categories on two consecutive probes.

As in our previous work (Gilmore et al., 2018; Johnson et al., 2019, 2020), longitudinal changes in naming accuracy for the treatment group were quantified by calculating single-subject effect sizes³, which have been widely used in the aphasia literature and for which there are established benchmarks for improvement from naming treatment (Beeson & Robey, 2006). Thus, patients who had a small or larger effect size (≥4.0, per Beeson and Robey, 2006) in at least one category were classified as responders, and those who did not have a small effect size were classified as nonresponders. See Table 2 for individual patient data, including effect sizes and associated treatment response classifications.

Table 2.

Patient demographics, assessment scores, effect sizes reflecting changes in naming in two categories (ES1/ES2), and fMRI task accuracy. Note that patients are grouped by treatment assignment (treated or untreated), and treated patients are further subdivided into responders and nonresponders based on the criteria described in the text.

ID	Sex	Age	Ed	MPO	Lesion volume (cm3)	AQ	BNT	PPT	ES1	ES2	fMRI Task Accuracy (%)
											Pre	Post	Chg.
Treated Patients
Responders
BU03	F	63.00	16.00	62.00	175.40	52.00	10.00	46.00	9.81	9.24	16.70	19.40	2.70
BU04	M	79.00	16.00	13.00	84.80	74.10	52.00	49.00	4.11	13.28	55.60	100.00	44.40
BU06	M	49.00	16.00	113.00	299.00	66.60	44.00	48.00	8.96	2.71	52.80	88.90	36.10
BU07	M	55.00	16.00	137.00	182.00	48.00	6.00	46.00	2.18	6.35	22.20	41.70	19.50
BU09	F	71.00	16.00	37.00	11.70	95.20	45.00	50.00	6.76	10.97	44.40	63.90	19.50
BU10	F	53.00	16.00	12.00	76.60	80.40	37.00	49.00	16.74	2.89	61.10	91.70	30.60
BU13	M	42.00	13.50	18.00	12.10	92.70	43.00	49.00	7.07	15.01	44.40	77.80	33.40
BU14	F	64.00	13.00	24.00	96.90	64.40	41.00	49.00	8.00	8.66	36.10	83.30	47.20
BU15	F	71.00	12.00	74.00	189.30	87.20	43.00	44.00	6.35	0.96	44.40	33.30	−11.10
BU17	M	61.00	16.00	152.00	163.50	74.30	54.00	51.00	8.67	15.01	30.60	88.90	58.30
BU18	F	70.00	16.00	152.00	69.60	78.00	24.00	50.00	5.77	5.17	41.70	75.00	33.30
BU20	F	48.00	16.00	14.00	164.30	13.00	0.00	40.00	3.52	6.64	2.80	11.10	8.30
BU22	M	62.00	16.00	12.00	100.00	65.40	1.00	37.00	4.62	1.73	11.10	8.30	−2.80
BU28	M	63.00	12.00	15.00	76.70	56.00	21.00	51.00	8.33	2.52	18.50	59.30	40.80
Responders Mean		60.79	15.04	59.64	121.56	67.66	30.07	47.07	7.22*		34.46	60.19	25.73
Responders SD		10.36	1.62	55.75	78.16	21.29	19.26	4.18	4.20*		17.82	31.79	20.43
Nonresponders
BU11	M	78.00	18.00	22.00	32.10	92.10	41.00	49.00	1.15	3.04	22.20	33.30	11.10
BU12	M	68.00	12.00	104.00	186.80	40.00	1.00	46.00	1.15	1.15	0.00	0.00	0.00
BU21	M	^65.00	18.00	16.00	247.60	11.70	0.00	43.00	0.58	2.89	0.00	5.60	5.60
BU23	M	60.00	16.00	24.00	172.80	45.20	6.00	42.00	0.33	1	0.00	0.00	0.00
BU25	F	76.00	18.00	33.00	184.40	37.50	2.00	34.00	2.02	−0.58	0.00	0.00	0.00
BU26	F	64.00	12.00	115.00	127.70	58.00	15.00	36.00	3.46	3	8.30	19.40	11.10
Nonresponders Mean		68.50	15.67	52.33	158.57	47.42	10.83	41.67	1.60*		5.08	9.72	4.63
Nonresponders SD		7.09	2.94	44.75	72.86	26.62	15.77	5.75	1.27*		9.02	13.78	5.46
All Treated Mean		63.10	15.23	57.45	132.67	61.59	24.30	45.45	5.53*		25.65	45.05	19.40
All Treated SD		9.99	2.04	51.63	76.68	24.23	20.03	5.21	4.41*		20.72	36.12	19.80
Untreated Patients
BUc01	M	49.00	12.00	49.00	87.60	85.50	53.00	49.00	−0.78	−0.61	66.70	61.10	−5.60
BUc02	M	79.00	18.00	10.00	89.00	26.90	3.00	47.00	−0.58	0.87	11.30	0.00	−11.30
BUc05	M	49.00	12.00	67.00	317.10	32.30	3.00	44.00	−0.29	−2.02	8.30	0.90	−7.40
BUc06	M	69.00	16.00	164.00	183.40	39.30	5.00	48.00	1.73	−1.73	5.60	0.00	−5.60
BUc07	M	39.00	16.00	17.00	26.20	71.30	36.00	52.00	0.87	1.31	47.20	44.40	−2.80
BUc08	M	64.00	12.00	13.00	34.10	79.60	−	50.00	0	1.09	41.70	50.00	8.30
BUc09	M	62.00	16.00	21.00	1.60	91.50	39.00	49.00	5.77	−1.46	47.20	52.80	5.60
BUc10	M	68.00	13.50	21.00	80.30	78.60	31.00	49.00	−2.31	−0.38	22.20	22.20	0.00
BUc11	M	58.00	14.00	23.00	186.50	61.80	10.00	51.00	0.13	6.42	10.40	25.00	14.60
BUc12	M	53.00	17.00	467.00	120.80	91.20	51.00	49.00	2.31	2.6	68.70	50.00	−18.70
Untreated Mean		59.00	14.65	85.20	112.66	65.80	25.67	48.80	0.65*		32.93	30.64	−2.29
Untreated SD		11.79	2.24	141.91	94.60	24.57	20.62	2.20	2.31*		24.38	24.09	9.80
All Patients Mean		61.73	15.03	66.70	126.00	62.99	24.72	46.57	3.90*		28.07	40.24	12.17
All Patients SD		10.61	2.09	90.41	81.98	24.00	19.85	4.67	4.47*		21.86	32.90	19.87

^*Because the order of treatment categories is arbitrary, summary statistics for effect sizes represent means/SDs across both categories for the specified patient group.

Abbreviations: AQ, Western Aphasia Battery Aphasia Quotient (scores out of 100); BNT, Boston Naming Test (scores out of 60); Chg., change in accuracy (Post - Pre); Ed, education (years); ES1/ES2, effect sizes in trained categories; MPO, months post onset of aphasia; PPT, Pyramids and Palm Trees test (scores out of 52).

2.5. fMRI methods

2.5.1. Data acquisition

All but two participants were scanned on a Siemens 3T Skyra scanner with a 20-channel head/neck coil at the Athinoula A. Martinos Center for Biomedical Imaging in Charlestown, MA. Verbal responses were recorded with a Fibersound Fiber Optic microphone (Micro Optics Technologies, Cross Plains, WI). Patients BUc07 and BUc10 were scanned on a Siemens 3T Prisma Fit scanner with a 20-channel head/neck coil at the Center for Translational Imaging in Chicago, IL⁴; their responses were recorded with an Avotec audio/mic system (Avotec Incorporated, Stuart, FL).

Acquisition parameters for T1-weighted structural images were as follows: TR = 2300ms, TE = 2.91ms, 176 sagittal slices, 1mm³ voxels, matrix = 256×256, FOV = 256×256mm, flip angle = 9°; and parameters for T2*-weighted functional images were as follows: interleaved parallel acquisition, TR = 2570ms, TE = 30ms, 40 axial slices, 2×2×3mm voxels, FOV = 220×220mm, flip angle = 90°.

2.5.2. Task and stimuli

Patients completed two runs of an overt picture-naming task at the end of pre-treatment/hold, with half of the stimuli (see below) presented in each run; these runs were concatenated for the neuroimaging analysis. Two runs were also completed and concatenated at post-treatment/hold. Healthy controls completed two runs of the task at a single time point.

An in-depth overview of the fMRI task can be found in Meier et al. (2016) and Johnson et al. (2019). In brief, an event-related naming task was utilized in this study. For a given participant, stimuli included color photographs of their assigned/trained items (n = 36); the unassigned/untrained items from the same categories as their assigned items (n = 36); a set of scrambled pictures (n = 36); and items from the control category (i.e., fruit, n = 36). Stimuli were presented in random order and were separated by a randomly varying interstimulus interval (ISI) of two or four seconds, an approach that has been shown to reduce the effects of motion associated with overt verbal responses in event-related designs (Birn et al., 2004). A black fixation cross on a white background was presented during the ISI. Participants were instructed to say the name of intact pictures and “skip” for scrambled pictures (Figure 1). As shown in Table 1, there was no significant difference in naming accuracy in the scanner between treated and untreated patients at pre or post (individual patient scores at both time points are provided in Table 2); however, treated patients improved significantly more from pre- to post-treatment than did untreated patients. Average accuracy for healthy controls was 73.9% (SD = 22.5%).

2.5.3. Preprocessing

Initial preprocessing was completed in SPM12 (http://www.fil.ion.ucl.ac.uk/spm/software/spm12) and included slice timing, motion correction, and coregistration of the T1 structural image to the mean functional image. Segmentation into gray and white matter and cerebrospinal fluid was performed using SPM12’s tissue probability maps. Structural and functional images were normalized to MNI space. Subject-specific lesion masks and maps were hand-drawn in native space using MRIcron (Rorden & Brett, 2000) and incorporated into coregistration, segmentation, and normalization. See Supplementary Figure 1 for group-level lesion overlap across all patients and in relevant subgroups.

Further preprocessing was performed with the CONN Toolbox (Whitfield-Gabrieli & Nieto-Castanon, 2012). The Artifact Detection Tool (https://www.nitrc.org/projects/artifact_detect) was used to identify motion and global-signal outliers in the functional scans according to CONN’s intermediate detection setting. Using CONN’s denoising routine, potential confounding effects were removed from the fMRI data via linear regression at the subject level. These effects included: six motion parameters and their first-order temporal derivatives; outlier volumes flagged by the Artifact Detection Tool (i.e., scrubbing); the first five principal components of BOLD signal in each of the white matter and cerebrospinal fluid per the aCompCor method (Behzadi et al., 2007); and the main effects of task conditions (i.e., trained/assigned items, untrained items, scrambled items, control items (fruit), and fixation) convolved with the canonical hemodynamic response function (HRF) and their first-order temporal derivatives, a procedure utilized in prior task-based functional connectivity studies (e.g, Fair et al., 2007; Johnson et al., 2020; Sandberg et al., 2015; Tao & Rapp, 2019) and intended to limit the extent to which simple task-induced coactivation in pairs of regions that are not otherwise functionally correlated might impact or bias the connectivity results. Functional data were also high-pass filtered at .01 Hz, consistent with other recent investigations of task-based/background functional connectivity (Browndyke et al., 2018; Johnson et al., 2020; Tao & Rapp, 2019).

2.5.4. Regions of interest

Thirty-eight anatomical regions of interest (ROIs) were selected based on literature demonstrating their involvement in picture naming (see Supplementary Table 1). ROIs were defined according to the AAL2 atlas (Rolls et al., 2015). Patients’ left hemisphere ROIs were modified using MarsBaR (Brett et al., 2002) by intersecting subject-specific lesion maps with the atlas-based ROIs and deleting any overlap between the two, leaving only the spared portions of each ROI (Johnson et al., 2019, 2020; Meier et al., 2016, 2018, 2019a; Sims et al., 2016). Thus, while the same ROIs were utilized in all participants, each patient’s ROIs were uniquely specified to reflect their own lesion. Importantly, none of the ROIs were completely destroyed in any of the patients.

2.5.5. Functional connectivity

For every subject and each time point at which they were scanned, CONN was used to calculate bivariate Pearson correlations (which were subsequently Fisher-transformed to z values) in the denoised, HRF-convolved BOLD time series between each pair of ROIs for the condition of interest (i.e., trained items). Thus, for each time point (i.e., pre and post) and participant, we obtained a 38×38-ROI connectivity matrix comprising the Fisher-transformed correlation coefficient between each pair of regions. Because we were not interested in self-connections and our functional connectivity data did not reflect directionality (that is, the connection from ROI 1 to ROI 2 was the same as the connection from ROI 2 to ROI 1), each matrix represented a total of 703 unique, pairwise connections.

2.6. Statistical analysis

Data management, statistical analyses and visualizations were conducted with the CONN Toolbox (Whitfield-Gabrieli & Nieto-Castanon, 2012), R 3.6.1 (R Core Team, 2019), RStudio (RStudio Team, 2019), and the following packages for R: tidyverse (Wickham, 2017), lme4 (Bates et al., 2015), optimx (Nash, 2014; Nash & Varadhan, 2011), lmerTest (Kuznetsova et al., 2017), and emmeans (Lenth, 2019), purrr (Henry & Wickham, 2019), and ggplot2 (Wickham, 2016).

To address our first aim of determining the extent to which functional connectivity in the picture-naming network differed between healthy controls and patients with chronic aphasia prior to the introduction of naming therapy, we conducted 703 two-sample t tests (i.e., one test per connection in the network) to compare functional connectivity (i.e., Fisher-transformed correlation coefficients) in all 30 patients at pre-treatment/hold to that of healthy controls. Connections in which a significant difference was identified after an FDR correction for multiple comparisons (i.e., q < .05) were characterized as “abnormal.” This set of connections, which we refer to as the “abnormal connectome,” were utilized in our subsequent analyses.

Our second aim was to investigate whether naming treatment was associated with changes in functional connectivity. This aim was addressed by constructing a linear mixed effects regression (LMER) model in which functional connectivity in the abnormal connectome identified under aim 1 was the dependent variable. Fixed effects included group (i.e., treated or untreated patients), time (i.e., pre or post), and the group-by-time interaction⁵. To account for repeated measurement of patients over time, subjects were entered as a random effect. Additionally, because we had no strong hypotheses regarding the relative functional connectivity of individual connections within the abnormal connectome, connections were also entered as a random effect. Several random effects structures were compared (see Supplementary Methods and Analysis for details), and the final best-fit model included by-subjects random intercepts and slopes for time and by-connections random intercepts. The fixed effects in the regression model were dummy coded so that reference levels for group and time were the treatment group and pre, respectively. Thus, the coefficient for the overall model intercept represented average functional connectivity in the reference group at the reference time point (i.e., connectivity in the treatment group at pre-treatment). The effect of group represented the difference in average functional connectivity between treated and untreated patients at pre; the effect of time represented the change in functional connectivity in treated patients from pre to post; and, of greatest interest for our purposes, the group-by-time interaction represented the difference between treated and untreated patients in the extent to which their connectivity changed from pre to post.

Due to differences in PPT performance between treated and untreated patients, we also constructed an alternative model that included the PPT and its interactions with group and time as fixed effects to determine if differences in semantic processing might explain potential group or time-based differences in functional connectivity. The structure and reference coding of the alternative model was consistent with the original model described above.

The third aim of the study was to determine if changes in functional connectivity over time in the abnormal connectome differed depending on treatment outcome. Thus, an LMER model was constructed with group as a 3-level factor comprising treatment responders, nonresponders, and untreated patients. The dependent variable and fixed and random effects were otherwise the same as those in the analysis for aim 2. In this analysis, the fixed effects were dummy coded with responders as the reference level for group and pre as the reference level for time.

To determine if any behavioral or demographic variables should also be included in the Aim 3 regression, one-way ANOVAs were performed to compare responders, nonresponders, and untreated patients on the WAB-AQ, BNT, PPT, age, education, months post-onset of aphasia, and lesion volume. Only the ANOVA for PPT scores showed a significant effect of group (F_2,27 = 6.12, p < .01), and a post-hoc Tukey test indicated nonresponders had lower scores than responders (difference = 5.40, p = .03) and untreated patients (difference = 7.13, p = .01). Thus, an alternative regression model that included the PPT and its interactions with group and time as fixed effects was constructed and compared to the initial model described above.

3. Results

3.1. Behavioral results

The focus of this investigation is on the impact of treatment on functional connectivity in the picture-naming network. However, it is important to first establish the overall efficacy of the treatment. The original treatment results, which were based on overlapping participant samples as the present study, demonstrated positive group-level outcomes after treatment (Gilmore et al., 2018; Johnson et al., 2019). Furthermore, these beneficial effects were verified in the present sample, as the average effect size was significantly larger in the treated group than the untreated group, and treated patients improved significantly more in naming accuracy during the fMRI task than did untreated patients (Table 1). Based on criteria established in our previous studies of the behavioral effects of the treatment (Gilmore et al., 2018) and benchmarks in the anomia literature (Beeson and Robey, 2006), 14 treated patients were classified as responders and six as nonresponders. Individual treatment effect sizes and response classifications are presented in Table 2.

3.2. Functional connectivity

3.2.1. Aim 1: Patients with chronic aphasia at baseline (i.e., pre-treatment/hold) vs. healthy controls

Two-sample t tests comparing functional connectivity in 30 patients and 17 healthy controls in the 703-connection picture-naming network indicated that patients had significantly lower functional connectivity in 31 connections (Figure 2). There were no connections in which patients had significantly higher functional connectivity than controls. All 31 connections and associated t and q (i.e., FDR-corrected p) values are reported in Supplementary Table 2. Ten connections in this abnormal connectome were between two left hemisphere regions and 21 were between one left and one right hemisphere region; thus, every connection involved at least one left hemisphere region. A total of 16 regions (out of 38 possible), including nine in the left hemisphere and seven in the right hemisphere, were represented in the abnormal connectome (Figure 2).

Figure 2.

The abnormal connectome revealed by comparisons between patients and healthy controls. On the left, functional connectivity was significantly lower in patients than controls in all 31 connections shown; the color bar represents t values, with darker shades indicative of a greater difference between groups. On the right is a list of all regions in the abnormal connectome, their abbreviations, and the number of connections in which they were involved.

3.2.2. Aim 2: Functional connectivity in the abnormal connectome in treated versus untreated patients

Results of the regression model predicting functional connectivity from group (treated and untreated) and time (pre- and post-treatment/hold) are reported in Table 3 (part A). The effect of group was not significant, indicating that functional connectivity in the abnormal connectome did not significantly differ between treated and untreated patients at the pre- time point. The effect of time was significant (β = 0.038, p = .032), indicating that functional connectivity increased from pre to post in treated patients (Figure 3A). However, the group-by-time interaction was not significant; thus, there was insufficient evidence to conclude that treated patients experienced a greater change in connectivity from pre to post than did untreated patients.

Table 3.

Results of regression models for functional connectivity at pre- and post-treatment in (A) the abnormal connectome in treated and untreated patients (i.e., Aim 2), (B) the abnormal connectome in responders, nonresponders, and untreated patients (i.e., Aim 3), and (C) the normal picture-naming connectome in responders, nonresponders, and untreated patients (i.e., Follow-up Analysis 1). Beta coefficients for model intercepts reflect mean functional connectivity in the treatment group (A) or responders (B and C) at pre-treatment. For each fixed effect, the factor levels being evaluated are indicated in parentheses. Bold text indicates p < .05.

A) Aim 2 model
Fixed Effects	β	SE	DF	t	p
Intercept	0.054	0.041	42.437	1.310	.197
Group (untreated)	−0.021	0.031	28.000	−0.676	.505
Time (post)	0.038	0.017	27.999	2.253	.032
Group (untreated) x Time (post)	−0.036	0.029	27.999	−1.226	.230
Random Effects	Variance	SD
Connections (Intercept)	0.042	0.204
Subjects (Intercept)	0.005	0.074
Subjects x Time (Slope)	0.003	0.062
Residual	0.031	0.176
B) Aim 3 model
Fixed Effects	β	SE	DF	t	p
Intercept	0.065	0.043	46.822	1.529	.133
Group (untreated)	−0.033	0.033	27.000	−0.976	.338
Group (nonresponders)	−0.038	0.039	27.000	−0.975	.338
Time (post)	0.072	0.017	27.000	4.182	<.001
Group (untreated) x Time (post)	−0.069	0.027	27.000	−2.617	.014
Group (nonresponders) x Time (post)	−0.111	0.031	27.000	−3.545	.001
Random Effects	Variance	SD
Connections (Intercept)	0.042	0.204
Subjects (Intercept)	0.005	0.074
Subjects x Time (Slope)	0.002	0.046
Residual	0.031	0.176
C) Follow-up 1 model
Fixed Effects	β	SE	DF	t	p
Intercept	0.092	0.012	275.700	7.414	<.001
Group (untreated)	−0.020	0.010	27.000	−1.930	.064
Group (nonresponders)	−0.015	.012	27.000	−1.220	.233
Time (post)	0.005	0.007	27.000	0.741	.465
Group (untreated) x Time (post)	0.001	0.010	27.000	0.064	.949
Group (nonresponders) x Time (post)	−0.012	0.012	27.000	−0.957	.347
Random Effects	Variance	SD
Connections (Intercept)	0.075	0.274
Subjects (Intercept)	0.001	0.023
Subjects x Time (Slope)	0.001	.023
Residual	0.037	0.191

Figure 3.

Mean functional connectivity in the abnormal connectome at pre- and post-treatment/hold in (A) treated (TX) and untreated (UN) patients, and (B) responders (R), nonresponders (N), and untreated patients (U). Error bars represent standard error derived from the respective regression models. Dashed horizontal significance lines refer to the effect of time (i.e., in A, change in treated patients from pre to post and in B, change in responders from pre to post); the curved significance lines in B refer to the group-by-time interactions (i.e., changes from pre to post in responders vs nonresponders and responders vs. untreated patients). *p < .05; **p = .01; ***p < .001.

(C-E) Histograms of t statistics from the abnormal- and random-connectome models for effects associated with changes in functional connectivity from pre to post in (C) responders, (D) responders vs. untreated patients, and (E) responders vs. nonresponders (i.e., follow-up analysis 2). The t statistics from random-connectome models are shown in gray; t statistics from the abnormal-connectome model are shown in black and surrounded by a red box.

Including the PPT and its interactions as fixed effects did not improve the fit of the model, and the results of the PPT-inclusive model were largely consistent with those of the original model described above. Thus, only the original model is addressed in detail in the main text; however, details and results of the PPT-inclusive model are provided in the Supplementary Methods and Analysis section.

3.2.3. Aim 3: Functional connectivity in the abnormal connectome in responders, nonresponders, and untreated patients

Results of the regression model predicting functional connectivity from group (responders, nonresponders, untreated) and time (pre- and post-treatment/hold) are reported in Table 3 (part B). There was no significant difference in functional connectivity in the abnormal connectome between responders and untreated patients or nonresponders at pre, as indicated by the effects of group. However, the effect of time (β = 0.072, p < .001) and the group-by-time interactions with both untreated patients (β = −0.069, p = .014) and nonresponders (β = −0.111, p = .001) were significant. These results suggest that functional connectivity increased in responders from pre- to post-treatment, and that this increase was greater than whatever changes occurred in the subsets of patients who either did not receive treatment (i.e., the untreated group) or whose naming did not improve in response to treatment (i.e., nonresponders). Indeed, as shown in Figure 3B, functional connectivity in the abnormal connectome increased in responders, but remained virtually unchanged in untreated patients and decreased in nonresponders. Post hoc comparisons of estimated marginal means confirmed there were no significant changes in functional connectivity in untreated patients (p = .915) or nonresponders (p = .145). Furthermore, these two groups did not significantly differ from one another in terms of functional connectivity at pre, post, or change over time (all p > .221).

Including the PPT and its interactions as fixed effects did not improve the fit of the model, and results of the models with and without the PPT were generally consistent with one another. Thus, the model without the PPT is addressed in the main text, but complete details and results of the PPT-inclusive model are provided in the Supplementary Methods and Analysis section.

3.2.4. Follow-up analyses: Specificity of treatment-related changes in functional connectivity

The analysis for aim 3 indicated that functional connectivity in the abnormal connectome changed over time (i.e., after treatment) in responders but not in nonresponders or untreated patients. However, because the analysis was intentionally constrained to the abnormal connectome, it did not indicate whether these changes were specific to the abnormal connections, which might be indicative of a focal treatment effect, or if they were representative of wider-spread changes throughout the larger picture-naming network. Therefore, we investigated this question via two analyses, as described below.

3.2.4.1. Follow-up analysis 1

In the first follow-up analysis, we investigated whether the larger picture-naming network changed in a similar way as the abnormal connectome in responders or any other sub-group. Thus, the 31 connections that formed the abnormal connectome were excluded from the initial 703-connection picture-naming network. This resulted in a “normal” picture-naming connectome that included 672 connections in which functional connectivity did not differ between controls and patients at baseline. Next, an LMER model with the same structure as that used to address aim 3 was constructed (i.e., functional connectivity predicted by group, time, and their interaction, with random intercepts for connections and subjects and random slopes for subjects over time).

As shown in Table 3 (part C), none of the effects of interest was significant in this model, indicating that functional connectivity in the “normal” portion of the picture-naming network did not change as a function of time/treatment in any group, and, thus, the changes observed under aim 3 were specific to the abnormal connectome. However, the normal connectome was substantially larger than the abnormal connectome (i.e., 672 vs. 31 connections, respectively); thus, it was plausible that noise associated with the sheer size of the normal connectome contributed to the null finding in the first follow-up analysis. We therefore conducted a second follow-up analysis (described in section 3.2.4.2, below) to address the potential confound of the size of the network being analyzed.

3.2.4.2. Follow-up analysis 2

Like the first follow-up analysis, the goal of the second follow-up analysis was to assess the specificity of the effects identified in the primary analysis for aim 3, while also taking network size into consideration. As such, follow-up analysis 2 was conducted using principles derived from permutation testing, as follows.

Once again, the 31 connections in the abnormal connectome were excluded from the initial 703-connection picture-naming network. From the remaining 672 connections, we generated 1,000 unique subsets, each comprising 31 randomly selected connections (i.e., 1,000 “random connectomes”). This step ensured that the random connectomes were the same size as the abnormal connectome but did not include the same connections.

Next, for each random connectome, we constructed an LMER model identical to that used to address aim 3. Of the 1,000 random-connectome models, 401 failed to converge and were therefore excluded from further analysis. The t statistics for the three fixed effects of interest (i.e., functional connectivity change in responders, differences in functional connectivity changes between responders and untreated patients and responders and nonresponders) were extracted from the 599 models that successfully converged. The t statistic for each of these effects from the model for the abnormal connectome (i.e., the t values reported in rows 4–6 of Table 3, part B) was compared to the distribution of t statistics for the corresponding effect from the random-connectome models in order to test the null hypothesis that the effects of interest were not specific to the abnormal connectome. A p value for each of these comparisons was computed as the number of t statistics from random-connectome models that were at least as extreme as the values derived from the abnormal-connectome model plus 1 (to account for the t statistic from the abnormal-connectome model), divided by the total number of connectomes analyzed (i.e., 1 abnormal connectome and 599 random connectomes = 600 total connectomes). This p value was compared to an alpha of .05 to determine whether to accept or reject the null hypothesis.

As shown in Figure 3C–E, none of the random-connectome models produced t statistics for an effect of interest that was as extreme as those from the abnormal-connectome model. Thus, the probability of randomly selecting a set of 31 connections from the larger picture-naming network and obtaining an effect of interest with equal or greater magnitude as those from the abnormal connectome was .002 (1/600) per effect. Given that this value was considerably lower than the a priori alpha of .05, this result was taken as evidence that the effect of treatment on functional connectivity was specific to the abnormal connectome. In contrast, the rest of the picture-naming network, whether viewed as a whole (per follow-up analysis 1) or in smaller sub-networks equal in size to the abnormal connectome (per follow-up analysis 2), remained relatively static.

4. Discussion

In this study, we investigated functional connectivity in patients with chronic stroke-induced aphasia relative to healthy older adults and compared connectivity in patients who did and did not receive language therapy for naming deficits. Patients’ naming ability and task-based functional connectivity were assessed before and after a period of approximately 12 weeks, during which 20 patients received the naming treatment and 10 patients received no intervention. Key findings and their implications are discussed below.

4.1. Abnormally low functional connectivity in the picture-naming network in chronic aphasia

In our first analysis—a comparison of functional connectivity in a 38-region, bilateral picture-naming network in healthy controls and all 30 patients prior to intervention—we found that patients had abnormally low functional connectivity in 31 connections. In general, this tendency toward lower-than-normal functional connectivity is consistent with previous studies of stroke and aphasia (e.g., Klingbeil et al., 2017; Meier et al., 2019a; Sandberg, 2017; Siegel et al., 2016; Tao & Rapp, 2020). Interestingly, although the abnormal connectome included just 4.41% of the 703 connections that were analyzed, it encompassed 42.11% (i.e., 16) of the 38 regions in the initial picture-naming network. Moreover, several of these regions, particularly those that were involved in numerous abnormal connections (i.e., the left superior and middle temporal gyri and bilateral inferior frontal gyri), had lower connectivity in patients with chronic aphasia than controls in a previous naming treatment study (van Hees et al., 2014). Furthermore, while all of the regions in the initial analysis were selected because of their presumed contributions (or capacity to contribute to) to naming, it is of little surprise that regions in the left temporal and inferior frontal cortex constituted a substantial portion of the abnormal connectome in patients with naming deficits, as these areas have repeatedly been implicated in lexical-semantic and/or phonological processes that support successful word retrieval and production (Binder et al., 2009; Indefrey, 2011; Indefrey & Levelt, 2004; Price, 2012; Vigneau et al., 2006; Whitney et al., 2012)

All of the abnormal connections identified in the patients involved at least one left hemisphere region. Thus, left hemisphere stroke damage resulting in chronic aphasia was associated with reduced functional synchrony of language regions ipsilesionally and interhemispherically, whereas contralesional connectivity was relatively preserved, perhaps because the underlying tissue was wholly intact. This general pattern of connectivity is consistent with the phenotype of stroke injury for language identified by Siegel et al. (2016). However, unlike Siegel et al. (2016), who found abnormally high connectivity involving the dorsal attention and default mode networks, we did not identify any connections in which patients had higher functional connectivity than controls. Importantly, however, we investigated a picture-naming network that was highly relevant to our aims but did not explicitly examine the same networks as Siegel et al. (2016). Additionally, participants in the present study were all in the chronic stage of recovery. It is, therefore, possible that our participants may have had even more abnormal connectivity—including, higher-than-normal connectivity in some areas—early after stroke, and that these abnormalities diminished (but did not resolve entirely) by the chronic stage. Consistent with this hypothesis is a finding from another study by Siegel et al. (2018), in which patients were followed from two weeks to one year post-stroke. In that study, modularity, a functional connectivity-based measure of network segregation, was abnormal at the sub-acute time point, but tended to normalize by one-year post-onset, though some patients with persistent cognitive-linguistic impairments still exhibited abnormal modularity.

4.2. Effects of treatment and treatment outcome on functional connectivity in the abnormal connectome

The initial comparison of functional connectivity in treated and untreated patients established that the two groups did not differ significantly in terms of functional connectivity prior to treatment. In conjunction with comparisons of demographics and standardized assessment scores—which revealed somewhat better non-verbal semantic processing skills (per the PPT) in the untreated group—this analysis confirmed that the treated and untreated groups were well-matched comparators.

The same analysis also showed that functional connectivity increased significantly in the treatment group from pre- to post-treatment, indicating that these patients experienced enhanced synchrony between connections that were abnormally weak prior to treatment. However, longitudinal changes in connectivity in the treated patients did not significantly differ from those of untreated patients. Viewed on its own, this result suggests that naming treatment was not responsible for connectivity changes in the treated group. However, we hypothesized that the comparison of treated and untreated patients may have been confounded by differential outcomes among patients in the treatment group. Indeed, in a prior study involving an overlapping participant sample and a subset of the regions analyzed here, we found that responders experienced a significant increase in functional activation, while nonresponders experienced virtually no change in activation (Johnson et al., 2019). The present aim 3 analysis allowed us to investigate the potential confound of treatment outcome by re-examining the effects of time/treatment while accounting for treatment responsiveness. Crucially, this analysis confirmed our hypothesis, as functional connectivity in the abnormal connectome increased significantly in responders from pre- to post-treatment and did so to a significantly greater extent than it did in both nonresponders and untreated patients. While relatively few treatment studies have examined functional connectivity in those with post-stroke language deficits, normalization of and/or increases in connectivity or related markers have previously been reported in conjunction with spelling (Tao & Rapp, 2019, 2020), imitation (Duncan & Small, 2016; Santhanam et al., 2018), and naming (Sandberg et al., 2015; van Hees et al., 2014) interventions. In contrast to responders, nonresponders experienced no significant change in functional connectivity; in fact, they showed a non-significant trend of decreasing connectivity over time (Figure 3B). This pattern of results may indicate that a fundamental difference between patients who benefit most and least from treatment is the capacity (or lack thereof) to modulate connections that are functionally abnormal in response to treatment. Moreover, in light of the stability of functional connectivity in the untreated group, the downward trend in nonresponders may, perhaps, represent a maladaptive neural response to treatment that underlies or is at least associated with limited behavioral improvement. This interpretation is speculative, however, given the null results of post hoc comparisons between nonresponders and untreated patients.

An additional result that bears addressing was that there was no difference in functional connectivity in the abnormal connectome between responders and nonresponders before treatment. This finding bolsters our assertion that differences in the two groups’ responsiveness to treatment were responsible for their different patterns of changing connectivity. Nevertheless, the lack of pre-treatment differences was somewhat surprising in light of our previous study of an overlapping sample, in which we identified several functional connectivity-based network properties that differed between responders and nonresponders at pre-treatment and which were associated with patients’ eventual treatment outcome (Johnson et al., 2020). However, in that study, comparisons were based on a 946-connection semantic processing network, whereas the central comparison in this study was based on the 31-connection abnormal connectome. Furthermore, Johnson et al. (2020) used methods based on graph theory and therefore utilized additional data processing steps, such as thresholding out negative correlations, that were not applied in the present study. Considered together, the results of these two studies suggest that metrics summarizing large-scale network characteristics or relationships between a given node and an entire network are more useful predictors of treatment outcome than pairwise relationships between nodes in a much more circumscribed network.

4.3. Specificity of changes in functional connectivity

A unique aspect of the present study is that the results pertain specifically to a subset of connections within the broader picture-naming network that were initially found to be functionally abnormal in the patients prior to treatment. Indeed, the follow-up analyses indicated that the pattern of results identified in the abnormal connectome (i.e., significantly greater increases in connectivity in responders than nonresponders and untreated patients) were not representative of sweeping changes throughout the entire picture-naming network. This suggests that the observed treatment-related changes in connectivity were more representative of what might be considered normalization (i.e, changes in the direction expected based on healthy controls), rather than compensatory mechanisms, which would instead have been indicated by changes in connections that were already functioning “normally” before treatment. In fact, it may be the case that we only observed changes in abnormal connections precisely because they had been functioning abnormally and were, therefore, primed to be modulated by treatment. On the other hand, the vast majority of connections in the larger network were functioning normally before treatment; as such, there may have been little need to modulate them. However, this is a speculative explanation, as we cannot state definitively that there were no compensatory mechanisms at work in our patients, only that our analyses of the picture-naming network did not provide evidence of such mechanisms.

An additional notable finding was that changes in connectivity over time were fairly consistent across the entire abnormal connectome. This was demonstrated through the model selection process (described in the Supplementary Methods), which showed that the inclusion of by-connections random slopes for time did not improve the fit of the model. Specifically, a favorable treatment response had a consistent impact on the abnormal connectome as a cohesive unit, rather than differential effects on individual connections.

One limitation of the present study is that our analyses focused solely on the picture-naming network. This approach allowed us to examine relationships among regions believed to be highly relevant to our population of interest (i.e., individuals with naming impairments), treatment (i.e., semantically focused naming therapy), and fMRI task (i.e., overt picture naming), but our results and associated implications are therefore strictly limited to the network of connections we examined. While our picture-naming network encompassed much of the cortex, it did not include some connections or subsystems whose function may have been altered by stroke (e.g., components of the dorsal attention network, as noted previously). Thus, it is possible that an alternative or more comprehensive brain parcellation consisting of established functional networks (e.g., Yeo et al., 2011) would provide additional or alternative insights to those reported here. However, the extent to which such parcellations—which have generally been derived using data from healthy, younger adults—are appropriate for those with chronic stroke-induced aphasia and adequately reflect their functional organization is unknown; thus, we utilized the more conservative approach of defining ROIs based on anatomical boundaries as outlined in the methods.

Another limitation of the study is the difference in sample sizes across the various groups/subgroups, as well as the relatively small size of the groups, which may have resulted in varying levels of precision in terms of the connectivity estimates for each group. While we had sufficient power to detect statistically significant group differences in the Aim 3 analysis, which involved the smallest subgroup (i.e., nonresponders), it is possible that some groups (e.g., nonresponders) are less representative of their respective populations than others (e.g., responders) and that larger and more equal samples would have produced somewhat different and/or more robust results. In fact, sample size differences (and associated differences in the precision and/or variability around the within-group estimates of functional connectivity) between the treated and untreated groups may have contributed to the lack of a significant groupby-time interaction in the Aim 2 analysis, despite its coefficient (−0.036) being nearly equal in magnitude to that of the effect of time (0.038), which was significant.

A final issue in this study was that there was some discrepancy between several patients’ treatment outcome classifications and changes in their accuracy on the in-scanner naming task from pre- to post-treatment. Specifically, in-scanner scores of two responders (i.e., BU15 and BU22) decreased after treatment, while scores of three nonresponders (i.e., BU11, BU21, and BU26) increased. These counterintuitive results were likely driven at least in part by within-subject variability in naming performance, a documented phenomenon in individuals with aphasia (Freed et al., 1996). In fact, this variability informed our decision to measure naming accuracy multiple times outside the scanner before and after treatment (i.e., to obtain reliable estimates of naming performance) and to use as a primary outcome an effect size measure that accounts for variability by incorporating multiple measurements at each time point. It is also possible that repeated assessment on the same items or increased familiarity with the task and/or imaging procedures could have impacted patients’ performance, which might explain the increased scores of nonresponders, as well as some of the untreated patients. For these reasons, we would argue that the effect size estimates (and associated outcome classifications) based on out-of-scanner naming scores are likely a better index of patients’ naming ability than changes in in-scanner task accuracy.

5. Conclusions

We investigated the functional connectivity of a picture-naming network in patients with chronic aphasia as compared to healthy controls and over time in the presence and absence of treatment. Patients had significantly lower connectivity than controls in 31 connections involving several regions that contribute to critical aspects of naming, including semantic processing, lexical selection and control, and phonological processing. Functional connectivity in this abnormal connectome increased significantly in patients who benefited most from treatment, but not in those who either did not benefit from or did not receive treatment. Follow-up analyses suggested that connectivity changes were specific to the abnormal connectome and not the broader naming network. These results highlight the importance of accounting for treatment outcome, in addition to the presence/absence of treatment, in studies of language therapy and neural function in aphasia. They also suggest that the potential for increased connectivity among certain connections that are abnormally weak before treatment may represent an important prerequisite for treatment success.

Highlights.

• At study entry, patients had abnormally low naming-network functional connectivity.

• Connectivity increased in patients who responded best to treatment.

• Changes in responders were specific to abnormal connections.

• Connectivity did not change in nonresponders or untreated patients.