Neural plasticity and treatment-induced recovery of sentence processing in agrammatism

Healthy participants

Five male and seven female non-brain-damaged volunteers ranging in age from 32 to 79 years (M = 54) were recruited from the greater Chicago area. All were strongly right handed (based on scores derived from the Edinburgh Handedness Inventory; Oldfield, 1971), monolingual, native speakers of English, and were without history of neurological, speech, language or learning problems. The mean number of years of education was 17.3 (range 15–20). All participants passed a hearing screening at 500, 1000, 2000, and 4000 Hz at 30 dB in both ears and had normal or corrected-to-normal vision (self-reported).

Agrammatic aphasic participants

The aphasic participants included 5 males and 1 female, age-matched with the healthy participants (range = 38–66 years; M = 54). All presented with aphasia secondary to a single left hemisphere stroke involving the middle cerebral artery (MCA), with the exception of one participant (A1) who had a crossed aphasia with a right hemisphere lesion (see Fig. 1 and Table 1 for lesion information for all participants). Participants were between 6 and 146 months post stroke at the time of the study. Years of education ranged from 12 to 20 and all were employed at the time of their stroke, apart from A5, who was close to obtaining his degree from medical school (see Table 2). The Western Aphasia Battery (WAB; Kertesz, 1982) and the Northwestern Assessment of Verbs and Sentences (NAVS; Thompson, experimental version) were administered to all participants. In addition, narrative language samples were collected by asking participants to tell the story of Cinderella after viewing a picture book of the story. These samples were analyzed for a variety of linguistic variables, viz. mean length of utterance (MLU), percentage of grammatical sentences, the ratio of open to closed class words, ratio of noun to verb production, and percentage of verbs used with correct arguments (after Thompson et al., 1995). A reading test designed to include the words and sentence types used in the study also was administered, which all participants passed. All participants also passed screenings for more than mild dysarthria and/or apraxia of speech.

Results derived from the WAB, NAVS, and narrative language samples are shown in Table 3. WAB Aphasia Quotients (AQs) ranged from 66.8 to 85, indicating aphasia in the mild to moderate severity range. For all participants, production was more impaired than comprehension. Spontaneous speech was characterized by reduced phrase length (MLU ranged from 4.78 to 7.61), grammatical sentences (the percent of grammatical sentences produced ranged from 29% to 73%), and use of closed class words (mean open:closed class ratio = 1.78). In addition, the patients showed better production of nouns as compared to verbs (mean noun:verb ratio = 1.65). Performance on the NAVS also showed better sentence comprehension than production, but all participants showed superior performance in both modalities for canonical (subject-verb-object (S-V-O), in English) compared to noncanonical structures. None of the participants were able to produce object relative (OR) structures prior to treatment.

Sentence structures and stimuli

Participants were trained to produce object relative (OR) sentences, as in 1a, using TUF (see Thompson 2008, for details). Throughout treatment, generalization was tested to untrained object cleft (OC) structures as in 1b and object extracted wh-questions (OWH) as in 1c.

• 1. a.

  Pete saw the groom that the bride carried _____. (OR)

•  b.

  It was the groom that the bride carried _____. (OC)

•  c.

  Who did the bride carry _____? (OWH)

Based on derivational syntax, these structures are linguistically similar in that all involve Wh-movement of a wh-element (i.e., who, that) from a position following the verb (marked by blank spaces in the examples above) to a position earlier in the sentence. This movement introduces a level of complexity that is not inherent in, for example, simple active sentences. However, among the sentences, 1a is more complex than 1b (due to the additional presence of subject raising in 1a, but not in 1b), which is more complex than 1c (because 1b, but not 1c, involves clause embedding) (Chomsky, 1977, 1981, 1995; Cheng & Corver, 2006; also see Thompson & Shapiro, 2007). Notably, previous treatment research has shown that training 1a improves production and comprehension of all three sentence types (Thompson et al., 2003; Thompson et al., 2010).

Ten active sentences with semantically reversible participant roles were developed (i.e., The groom carried the bride) and for each sentence two corresponding black and white line-drawn pictures were prepared (e.g., a groom carrying a bride and a bride carrying a groom), for a total of 20 picture stimuli. All sentences contained transitive verbs, controlled for length, frequency, and other variables known to influence lexical processing and production.

Experimental design

A modified multiple baseline design across participants, combined with a multiple probe technique, was used to evaluate the effects of treatment (Thompson, 2006). Prior to treatment, two baseline probes, which tested both comprehension and production of all sentence types, were administered (see below). For participants A2, A4, and A6 these probes were administered on two separate days during week 1 of the study, following which treatment was begun during week 2. For Participants A1, A3, and A5, baseline probes were extended across 4 weeks, with treatment commencing during week 5, with the longer baseline period for the latter subjects serving as experimental control (see Fig. 2).

Following baseline testing, treatment was begun, with participants seen for two 2-h sessions per week. During each session, the 20 OR items were trained in random order. Daily, prior to each training session, probes identical to those used for baseline testing were administered so that acquisition of OR structures and generalization to OC and OWH structures could be observed. A randomly selected half of the items were administered each day (30 production and 30 comprehension items), resulting in a full probe weekly. For each participant, treatment was provided until 80% correct OR production was achieved on two successive probes or for a maximum of 20 sessions. For a subset of production and comprehension probes (15%), responses were scored by two independent raters, revealing 99.1% inter-rater reliability for production probes and 100% reliability for comprehension probes. After treatment was completed, the NAVS was readministered.

Comprehension and production probe tasks

The 20 picture stimuli were presented three times each, once for elicitation of each sentence type, for a total of 60 comprehension and 60 production items. Probes were computer generated using Superlab Pro (version 2.0.4; http://www.cedrus.com), such that the probe tasks would be as similar as possible to the task used in our fMRI conditions (see below).

Comprehension probes used a sentence picture matching task: presented with picture pairs on the computer screen. Participants were instructed to “point to the picture that goes with the sentence”, which was then auditorily presented by a pre-recorded female voice. A 5-s response time was provided, after which the computer advanced to the next item. For production, a sentence production priming procedure was used: a picture pair appeared on the computer screen and a pre-recorded female voice modeled the target sentence type using the picture on the left side of the screen as it was highlighted (“For this picture, I could say: <example sentence>”). Instruction to make a similar sentence for the target picture was then given as the picture on the right side of the screen was highlighted (“For this picture, YOU could say: …”). A maximum of 10 s was provided for responding, after which the computer proceeded to the next item. Responses were considered correct if the target sentence type was produced with a verb and both the agent and theme arguments were represented. Semantically appropriate noun and verb substitutions were accepted (e.g., man for boy), but only substituted verbs with the same number and type of arguments engendered by the target verb were accepted (e.g., chase for follow). Functional category errors (e.g., verb tense substitution/omission, omission of determiners) also were accepted, but overt production of that (or who) was required for correct object relative and object cleft structures. That is, reduced relative clauses, such as Pete saw the groom the bride carried., were not accepted even though they are grammatically correct.

Treatment data analysis

The primary dependent variable, reflecting responsiveness to treatment, was the proportion of correct OR responses noted on production and comprehension of probes. Generalization to untrained sentence types was a secondary variable of interest. Results of the study were examined, primarily, on an individual basis, in keeping with single-subject experimental design conventions. In addition to visual inspection of learning and generalization curves, effect sizes for the production probes were calculated for each participant, for each sentence type, using Busk & Serlin’s (1992) variation on Cohen’s d statistic (Cohen, 1988), as advocated by Beeson and Robey (2006). We subtracted the mean of the first three probes, which included the two baseline probes, from the mean of the two final probe scores and divided the result by the standard deviation of the first three probe scores. When the first three probes all yielded scores of 0, and thus no variation, we used the standard deviation of the respective participant’s pooled baseline scores for the other two sentence types. With respect to interpretation of effect-size magnitude, we compared d values to the effect sizes reported by Robey, Schultz, Crawford, and Sinner (1999), based on their review of treatment research in aphasia: 2.6 (small), 3.9 (medium) and 5.8 (large).

We also evaluated treatment effects by computing performance change for all sentence types between baseline probe 1 and probe 2 for participants A1, A3, and A5 and comparing that to performance change for Participants A2, A4, and A6 from baseline probe 2 to the final treatment probe for production and comprehension, using independent-samples Mann-Whitney U tests. NAVS performance from pre- to post-treatment also was analyzed non-parametrically for the group using a Wilcoxon signed ranks test.

Primary neuroimaging task

An auditory verification task was selected to obtain the functional neuroimaging data. Participants, presented with an auditory sentence and a visual scene, indicated by button-press (yes/no) whether or not the two matched (see Fig. 3). Sentence types were object clefts (OC), which involve Wh-movement, subject clefts (SC), which are not derived through Wh-movement and have a canonical word order, and simple past tense actives (ACT), matched for sentence length in syllables (see 2a–2c). Half of the sentences presented in each condition matched the pictures, while the other half consisted of role-reversed mismatches. Sixty trials per condition were pseudorandomly distributed over 4 runs, in an event-related design. Picture stimuli remained on screen for 8000 ms, with ISI jittered between 300 and 22,000 ms, based on optimal design calculations using a customized version of OPTSEQ (http://surfer.nmr.mgh.harvard.edu/optseq/).

• 2. a.

  It was the groom that the bride carried. (OC)

•  b.

  It was the bride that carried the groom. (SC)

•  c.

  Yesterday, the bride carried the groom. (ACT)

Healthy control participants were scanned once, whereas pre-post treatment scans were obtained for the aphasic participants. With the exception of aphasic participant A1, who had a left-sided hemiparesis, all subjects used their left hand for the button-press response.

FMRI data acquisition and preprocessing

Scanning was carried out on a Siemens 3T TIM Trio scanner. A T1-weighted anatomical scan was obtained at the start of each protocol with the following parameters: TR = 2300 ms; TE = 3.36 ms; flip angle = 9°; image matrix = 224 × 256; FOV = 256; voxel size = 1 mm × 1 mm × 1 mm. During the experimental runs, functional volumes with BOLD contrast were obtained using gradient echo-planar imaging sequences (TR = 2000 ms; TE = 30 ms; flip angle = 80°; matrix size = 64 × 64; FOV = 220 mm; voxel size = 3.44 mm × 3.44 mm × 3 mm; 32 slices). To allow for image saturation, the first 6 volumes of each experimental run were discarded.

Data preprocessing and statistical analysis was performed with SPM5 (http://www.fil.ion.ucl.ac.uk/spm). Functional scans were corrected for slice-acquisition timing and realigned to a mean functional volume, with which the anatomical volume was then coregistered. Spatial normalization of the anatomical image to the Montreal Neurological Institute (MNI) 152-subject template brain (ICBM, NIH P-20 project) took place through ‘unified normalization’ in SPM5, which combines segmentation, bias correction and spatial normalization in one model (Crinion et al., 2007). For the lesioned brains, medium regularization was used, whereas non-lesioned brains where normalized with light regularization. Functional volumes were then normalized using transformation parameters obtained from normalization of the coregistered T1s, and resliced at a resolution of 3 mm × 3 mm × 3 mm. Spatial smoothing was performed with a 6 mm (full-width, half-maximum) isotropic Gaussian kernel. Effects of global signal were removed from the functional time-series using the method described by Macey, Macey, Kumar, and Harper (2004). In the aphasic participants’ data, signal intensity was scaled to allow comparison of pre- and post-treatment scans.

Long-trials task

The aphasic individuals also participated in a long-trials fMRI task, which was the same as the experimental task, but used a fixed SOA of 30 s and only 40 trials. We included this task to examine participant-specific hemodynamic response functions (HRFs), which are known to be delayed in some patients with stroke-induced aphasia and can, therefore, influence detection of fMRI activation in some regions of the brain (Bonakdarpour et al., 2007).

Perfusion imaging

For identification of hypoperfused tissue, which might affect cortical activation levels as reflected by the BOLD response, the aphasic participants underwent perfusion imaging by arterial spin labeling (ASL) with a modified PICORE/Q2TIPS sequence (Chen, 2008; Luh, Wong, Bandettini, & Hyde, 1999; Wong, Buxton, & Frank, 1997). Pulse sequence parameters were TI1 = 600 ms and TI2 = 1600 ms. Twenty Pairs of interleaved control and tag images were collected (TR = 2500 ms, TE = 2300 ms (minimum), slice thickness = 5 mm, with 1 mm gap; inplane resolution 3.44 mm × 3.44 mm). Resting state images were acquired using 50 averages.

Flow images were calculated from the interleaved control and tag time series using a surround subtraction method and averaged over the entire scan to provide an adequate signal-to-noise ratio (Chen, 2008; Wong et al., 1997). Results were quantified in ml/100 g/min and the images were then coregistered to the anatomical images and spatially normalized, using nearest neighbor interpolation for strict maintenance of flow values. Using each participant’s normalized grey matter mask that formed the output of the T1 segmentation process in SPM5, we calculated the mean flow value for the grey matter in each ROI. We also calculated the mean flow value per hemisphere, again using the grey matter mask.

Correlations between perfusion and region-specific time-to-peak of the HRF were calculated using a Pearson test. In addition, relations between region-specific upregulation of fMRI activation after treatment (i.e., regions showing greater post- compared to pre-treatment activation) and perfusion and time-to-peak also were examined with independent-samples t-tests, dividing ROIs into two groups, viz. those that showed upregulation of activation after treatment and those that did not. The same analyses were performed using downregulation (i.e., regions showing lesser activation post-treatment) as the grouping factor and using ‘differential activation’ (up- or downregulation versus no change) as the grouping factor.

Control participants’ fMRI data analysis

In the first-level analysis, a high-pass filter of 256 s was used to eliminate scanner drift. Response times and accuracy were modeled parametrically for each trial, to covary out their effects on the main conditions of interest. Individual participants’ summary activation maps for main effects were entered into a second-level within-subjects 2 × 3 analysis of variance (ANOVA) with sentence-picture matching (2 levels) and sentence type (3 levels) as variables. All second-level statistics were evaluated at a voxelwise significance threshold of p<.05, corrected for multiple comparisons per false discovery rate (FDR: Benjamini & Hochberg, 1995; Genovese, Lazar, & Nichols, 2002). A minimum cluster size threshold of 3 contiguous voxels (81 mm³) was used.

Aphasic participants’ fMRI data analysis

The fMRI data of our aphasic participants were analyzed individually to avoid problems associated with small sample sizes and grouping data from patients with heterogeneous lesions (Price, Crinion, & Friston, 2006). Within-session and between-session effects were computed and estimated in one fixed-effects model for each aphasic participant, with the four runs within each session modeled as separate regressors. As Smith et al. (2005) point out, drawing conclusions about longitudinally differential activation based solely on visual comparison of thresholded activation maps or on counts of total numbers of activated voxels has considerable drawbacks. Session-specific levels of baseline intensity affect which activated voxels ‘survive’ a particular threshold. This may result in Type I errors, in that the illusion of differences between sessions is created, while these differences may not be statistically significant when directly compared in terms of normalized activation strength.

We performed ROI analyses using participant-specific HRF functions, derived from analysis of the long-trials fMRI task, to examine activation in regions of significant activation in the control group (see below): Brodmann’s Areas 7, 9, 13, 21, 22, 39, 40, 44, 45, bilaterally. The participants’ true HRF in spared tissue was calculated in these 18 regions of interest, using Brain Voyager software (QX 1.4, Maastricht, The Netherlands). HRF latency maps were formed using linear correlation lag analysis of the stimulation onsets and the time series on a voxel by voxel basis. The map was thresholded at r = .11–.20, depending on signal to noise. Within a suprathresholded region of interest, a stimulus-locked average formed the HRF curve for that particular cluster (see Bonakdarpour et al., 2007, for further details). The obtained hemodynamic response patterns were fed as weights to the relevant FIR-modeled post-event time bins in the statistical model. Specific contrasts were calculated by giving positive and negative values to these weights, thus allowing t-tests to compare the different conditions. Brodmann’s Area ROIs were based on the WFU-Pickatlas toolbox, version 2.0 (Maldjian, Laurienti, Kraft, & Burdette, 2003).

Behavioral Data

Repeated measures ANOVAs of scan task performance showed main effects of sentence type for both reaction time (F(1.1,11.8) = 14.45, p<.05) and accuracy (F(1.2,13.7) = 8.85, p<.05). Bonferroni corrected post-hoc testing revealed reaction times for SC sentences (M = 2641 ms) to be significantly faster than for OC and ACT sentences (M = 3101 and M = 2861 ms, respectively). With respect to accuracy, OC sentences yielded more erroneous responses (M = 96% correct) than SC and ACT sentences (both at M = 99% correct). There was also a significant correlation between reaction time and accuracy, with accurate responses being faster than inaccurate responses (Pearson r(34) = −.47, p<.01).

FMRI Activation Patterns

Factorial analysis of the healthy control participants’ neuroimaging data revealed significant main effects of sentence type and trial type (i.e., matched versus no-match), but no interaction. The main effect of trial type was present in a small cluster (108 mm³) in the left-hemisphere lingual gyrus and a slightly larger cluster (324 mm³) in the left posterior insula. Further investigation of this effect, by comparing matched versus mismatched trials in directional t-tests, revealed that the effect in the left insula was carried by the contrast matched>mismatched. The lingual gyrus effect was not significant when measured with directional t-tests. The absence of an interaction between trial type and sentence type, as well as the minimal trial type effect, prompted us to collapse the matched and mismatched trials in further analyses of sentence type effects to increase power.

Analysis of sentence-type effects revealed a main effect of sentence type, driven by the contrast of OC>SC, which showed activated clusters in a large left-lateralized network, including the middle and inferior frontal gyri, precentral gyrus, anterior middle and superior temporal gyri, angular gyrus and anterior insula (see Table 6 and Fig. 6). In all clusters, data plotting revealed that the effects were driven by increases over baseline signal level for the OC condition, as opposed to decreases for the SC condition. A similar network was found when subject-cleft and simple active sentences were collapsed and compared to OC structures (OC>(SC + ACT)) (which contrasted sentences with Wh-movement to those without Wh-movement), with the addition of left caudate (head) and bilateral thalamic activation. The contrast SC>OC revealed a small cluster of activation in the left posterior insula, which turned out to be driven by a decrease in the BOLD response for OC trials, relative to the baseline signal level (as determined visually by plotting responses in the activated cluster). No other contrasts showed significant activation.

Behavioral Data

Table 7 shows scan task accuracy and reaction time data for the aphasic participants. With regard to accuracy, four participants showed improved performance from pre- to post-treatment scans across sentence type: A1, A2, A3 and A6. However, when subjected to statistical analysis (Chi² (α = .05)), significant effects were found only for Participant A6 (for all three sentence types) and Participant 5 (for object clefts only). A Pearson analysis of the correlation between changes in accuracy and changes in response time revealed no correlation (r(16) = .09, p = .7).

The Reaction time data were analyzed with subject-specific 3 × 2 ANOVAs, with the factors sentence type (SC, OC, ACT) and time (pre, post). We only report significant main effects and interactions here, with main effects of sentence type tested post hoc with Bonferroni correction (α = .05). Participants A1, A2, and A5 showed significantly slower RTs post treatment (A1: F(1,354) = 27.28, p<.05, A2: F(1,354) = 13.08, p<.05, A5: F(1,354) = 8.74, p<.05) as compared to pre-treatment, as well as main effects of sentence type (A1: F(2,354) = 12.08, p<.05, A2: F(2,354) = 5.94, p<.05, A5: F(2,354) = 5.60, p<.05), reflecting slower RTs for OCs compared to the other sentence types. Conversely, participants A3 and A4 showed significantly faster RTs post treatment (A3: F(1,354) = 8.57, p<.05; A4: F(1,344) = 10.16, p<.05) and a main effect of sentence type was found for A3 (F(2,354) = 4.33, p<.05). In keeping with the other participants, this latter effect derived from longer RTs for OC compared to SC sentences. Participant A6 did not show a significant main effect for sentence type (F(2,351) = 2.91, p = .056).

Long-trials (HRF) and perfusion imaging results

Results of the whole-brain perfusion imaging analyses are graphically presented in Fig. 7, with region-specific blood flow values given in Table 8. Table 8 also shows the aphasic participants’ time-to-peak (TTP) of the HRF in each ROI, based on the long-trials data. TTP cells are empty if no fluctuating HRF correlated with trial onset was observed in a particular region, usually due to lack of sufficient sparing of functional grey matter. The hemispheric asymmetries shown in Table 8 and the pattern visible in Fig. 7 indicate that perfusion was generally lower in the hemisphere directly affected by the stroke, with all participants showing higher perfusion in the contralesional hemisphere (i.e., in the right hemisphere for A2–A6 and the left hemisphere for A1, whose stroke affected the right hemisphere). Relatively low perfusion also extended beyond the immediate perilesional areas.

Correlational analysis examining the relation between blood flow values (i.e., perfusion) and TTP revealed a significant negative correlation (Pearson r(103) = −.20, p<.05). That is, lower perfusion was correlated with longer TTP for the hemodynamic response associated with neural activation (see Fig. 8). In addition, independent samples t-tests revealed significantly shorter TTPs (t(101.9) = −4.793, p<.001)) and higher perfusion levels (t(106) = 2.089, p<.05)) in regions showing post-treatment upregulation of activation (see below) compared to regions that did not (see Fig. 9). Conversely, there were no differences between regions showing or not showing downregulation for either TTP or perfusion levels. Finally, regions with pre-post changes in activation in either direction showed shorter TTPs than regions that showed no change (t(91.0) = 3.024, p< 0.05), and there were no differences in perfusion levels across these regions.

FMRI Activation Patterns

Analyses of the aphasic participants’ pre- versus post-treatment fMRI data examined several ROIs, including frontal regions (BA 44, 45 and 9), the insula (BA 13), and temporoparietal areas (BA 21, 22, 39, 40, and 7). As noted above, these regions were selected because of their role in language processing noted in the literature and also based on the network identified for processing complex sentences in our healthy control participants. Because we anticipated tissue recruitment both ipsilateral and contralateral to the lesion, we examined for activation in these ROIs bilaterally. We report main-effect differences between the pre- and post-treatment scans, collapsing the three sentence types (although sentence types were separately modeled).

The aphasic participants with significant activation differences in the targeted ROIs for pre-treatment>post-treatment (pre>post) and post-treatment>pre-treatment (post>pre) contrasts are shown in Table 9. Fig. 10 also graphically displays these data, depicting the proportion of participants showing differential pre/post activation for main effects within each ROI. Note that some participants show both up- and downregulation within the same ROI, reflecting activation shifts within a particular region. Finally, in Fig. 11 we present cortical and axial maps for individual aphasic participants showing pre>post (blue) and post>pre (red) activation differences projected onto their skull-stripped and normalized brains.

Despite individual variation across patients, some common changes in activation were noted from pre- to post-treatment. In particular, in the left hemisphere, downregulation (that is, less activation post- as compared to pre-treatment) stood out in the insula (BA 13) and the superior temporal gyrus (BA 22), with minor downregulation in the triangular part of the inferior frontal gyrus (BA 45), while upregulation (that is, greater activation post- as compared to pre-treatment) was prominent in the middle temporal and angular gyri (BAs 21 and 39), as well as in superior parietal cortex (BA 7), with minor upregulation in the opercular part of the inferior frontal gyrus (BA 44). Pre to post-treatment changes were also noted in the right hemisphere: there was somewhat more downregulation than upregulation in the insula (BA 13) and the middle temporal gyrus (BA 21) and concomitant minor upregulation in frontal and prefrontal regions (BAs 44 and 9), with more prominent upregulation, compared to downregulation, in the supramarginal gyrus (BA 40) and superior parietal cortex (BA 7). The left supramarginal gyrus (BA 40) and prefrontal cortex (BA 9) and the right insula (BA 13), showed mixed patterns of both up- and downregulation, in some cases driven by activation shifts within the same regions by individual subjects, particularly patients A2 and A4.

The most consistent activation shifts from pre- to post-treatment noted in the main effects analysis for the patients showed up in the left angular gyrus (BA 39) and the superior parietal region bilaterally (BA 7). Only one patient (A4) showed pre>post activation in BA 39, whereas four of the five participants showed post>pre activation in this region (the lesion of A6 encompassed all of the angular gyrus). As for superior parietal area activation, A4 again showed bilateral pre>post activation, whereas A1, A2, and A4 showed post>pre activation in the left BA 7 and A1, A2, A4, and A6 showed post>pre activation in the right BA 7.