Research Participants

Two patients (G01, J02) with chronic nonfluent aphasia received syntactic mapping treatment under two different conditions as part of a larger study, and underwent fMRI pre- and posttreatment. Pretreatment assessment suggested that both G01 and J02 demonstrated aphasia profiles most consistent with transcortical motor aphasia, with repetition, naming, and auditory comprehension scores relatively higher than fluency scores (see Table 1).

Pretreatment language assessment

Patient G01

G01 was a 72-year-old right-handed woman treated 53 months postonset of a left hemisphere cerebral vascular accident (CVA). She obtained 12 years of education and had been a homemaker prior to her stroke. She reportedly had received 6 months of speech and language therapy following her stroke. An MRI scan revealed a lesion consistent with a proximal occlusion of the left middle cerebral artery, with damage to the left temporal and parietal lobes, basal ganglia, thalamus, sensorimotor area, premotor cortex, and centrum semiovale (Figure 1).

T1-weighted axial magnetic resonance images in radiologic space (left is on the right) showing the extent of left cerebral vascular infarction for G01 and J02.

In addition to deficits in fluency and word retrieval on the Western Aphasia Battery (WAB; Kertesz, 1982), results of the pretreatment Action Naming Test (ANT; Obler & Albert, unpublished) and Boston Naming Test (BNT; Kaplan et al., 2000) indicated that G01 had greater difficulty with verb than noun retrieval (see Table 1). Verb retrieval errors tended to be the use of nonspecific or “light” verbs (such as “using a saw” for “sawing”). Performance on the Circles and Squares Test (Schwartz et al., 1980) revealed auditory comprehension of reversible active sentences was quite good (23/24 correct), however, comprehension of passive sentences, particularly when the action was depicted from left to right, was impaired (16/24 correct).

Sentence production in response to reversible action pictures highlighted the difficulty in generating accurate active and passive sentences. A 90-sentence subset of the pretreatment responses was coded for type of error, and half (45) of these sentences were coded by a second examiner. Interjudge agreement for presence of error was 100%, and for type of errors was 83%. Disagreements in error coding were resolved by consensus. Errors were predominantly word order errors (e.g., beginning a passive sentence with the agent or “actor” noun), use of a light verb in place of the target verb (e.g., “went after” for the target “follow”), and incorrect verb selection (e.g., “teasing” for “biting”). Pretreatment performance on the baseline and probe measures revealed a stable deficit in sentence generation (z = 1.04; p > .05). Analysis of narrative discourse using the Quantitative Production Analysis (Berndt et al., 2000) and additional linguistic measures revealed persistent deficits with word order and morphology, decreased representation of verbs, and reduced syntactic accuracy and complexity.

Syntactic Mapping Treatment

The syntactic mapping treatment differed from what has been most often reported in the literature in that the intervention focused on sentence production rather than comprehension of thematic roles. Sentence production was paired with a color and spatially coded mapping template focusing on thematic role positions (see Figure 2). The patients were encouraged to use the illustration in the template to guide the placement of the lexical items in the sentence to accurately reflect the thematic roles. Errors in morphological endings and article omissions were ignored for scoring purposes, with the exception of the passive morphology for passive sentence construction (e.g., “was + verb + by”).

Illustration of color and spatially coded mapping template used to train active and passive sentences in both phases (Errorless Mapping Treatment, Errorful Mapping Treatment) of the intervention.

The treatment was administered in two phases. In the first phase, Errorless Learning Mapping (ELM) utilized constant time delay as an errorless training procedure for sixteen 90-minute sessions, occurring 3–4 times per week (Wolery et al., 1992; Maher et al., 2002). This was followed by a second phase, Errorful Mapping (EFM), which utilized the more traditional approach for the same number and frequency of sessions. Active sentences were trained first (8 sessions), followed by passive sentence training (8 sessions) in each of the phases. In ELM, sentence production errors were prevented by providing an immediate model for repetition of the target sentence. The patient was instructed to only produce the sentence if he/she was certain it would be correct. Otherwise, the therapist would provide the model and the patient would repeat the target sentence. After 80% of the sentences were produced correctly at this stage, then a time-delay between the presentation of the picture and the model was introduced, allowing the patient increasing independence in producing the sentence. However, if an error was either produced or anticipated, the model for the correct response was provided yielding a near errorless learning environment. In contrast, in EFM, the patients first generated their independent responses to the target, and errors were corrected through a query approach of the lexical items to fill the respective thematic roles of agent (actor) and patient (recipient of action) in the target sentence (Schwartz et al., 1994b). In both the ELM and EFM phases, the color-coded template of the sentence constituents (and when appropriate the passive morphology) was present. In the ELM phase, written words corresponding to the agent, recipient, and verb were placed in the appropriate locations to “fill in” the map prior to the response (or after a specified time delay). In EFM, the map was filled in as the patient produced the lexical items, either independently or with cues for the therapist. The comparative impact of ELM and EFM has been reported elsewhere (Maher et al., 2002).

For each participant, a corpus of reversible sentences was probed to yield three different sets of transitive verbs that were balanced for word frequency (Kucera & Francis, 1967) for each patient. All of the verbs required only an agent (actor) and a patient or goal (recipient) to be syntactically correct. Within each set, no two verbs began with the same initial phoneme. These verbs were paired with high frequency animate nouns to yield reversible sentences that were pictured in simple line drawings (see Figure 2). Multiple baselines of active and passive sentence production in response to the three sets of reversible action pictures were obtained prior to the initiation of treatment. These same sentences were probed throughout the two phases of treatment to determine treatment and generalization effects. Again, errors in morphologic endings and article omissions were ignored for scoring purposes, with the exception of the passive morphology for passive sentence construction (e.g., “was + verb + by”). Synonyms for the lexical items (e.g., “lady” for woman; “bashing” for hitting) were scored as correct. Each response was transcribed and scored offline by the therapist and by a second examiner who was unfamiliar with the patients for reliability. Interjudge agreement exceeded 90%; disagreements were resolved by a third examiner.

Following the determination of a stable baseline as indicated by the C-statistic (Tryon, 1982), ELM treatment was initiated using the first set of pictures, while the other two sets were not treated but were probed with the same frequency as the treated set. The second set of pictures was used for training during EFM, and the third set of pictures was never treated, to allow for a generalization comparison for both phases of treatment.

Functional Magnetic Imaging Task

G01 and J02 underwent fMRI scanning both pre- and postmapping treatment. A silent sentence generation task alternated with a passive viewing task in a block paradigm. For the baseline task, patients passively viewed nonsense objects thought to have minimal semantic representation (Kroll & Potter, 1984). A similar paradigm has been used for detecting the neural substrates of simple sentence production in healthy adults (Peck et al., 2004b). The use of perceptual tasks during baseline states is recommended to interrupt ongoing neural activity that may involve the same brain regions activated during linguistic retrieval (Binder et al., 1999). For the sentence production task, patients were instructed to silently produce sentences describing events depicted in black-and-white line drawings of an action occurring between two nouns (one actor and one recipient). Because stimuli presented in the scanner were similar to the items presented during treatment probes, performance of patients in the scanner could be estimated based on their performance during the treatment-related sentence generation tasks. Although often erroneous in lexical selection and/or thematic role assignment, even prior to treatment, G01 produced active sentences containing an argument structure (characterized by the presence of noun/verb combinations) 99% of the time, and for J02, 94% of the active sentences he produced contained an argument structure, indicating generation of a syntactic frame during each attempt. Therefore, we are confident that both patients engaged in syntactic processing during the active blocks of the paradigm.

Silent sentence production was chosen, despite its drawbacks, to minimize imaging motion artifacts brought about by extended speech. Although artifacts from overt language production can be averaged away across participants in group studies of healthy adults, Barch et al. (1999) advised against individual participant analyses using overt production. Given the variability in lesion size, shape, and location, studies of stroke patients are constrained to within-subject designs to avoid grossly underestimating perilesional activity. Such underestimation can result with group averages when perilesional activity for some patients falls within the lesion territory of other patients. Methods to reduce motion-related signal change by discarding motion-corrupted images obtained while speaking (Birn et al., 1999) cannot easily be applied to overt sentence production in aphasic patients because of the variable onsets and variable durations of their overt responses. Furthermore, those early images are crucial for capturing processes related to ordering sentence elements because those must occur prior to spoken sentence production (Garrett, 1980, 1984; Berndt, 1998). Lastly, comparisons of overt and covert language production reveal similar patterns of activity regardless of response modality, with the addition of regions associated with motor aspects of speech production in overt responding (Palmer et al., 2001; Ackermann & Riecker, 2004; Kan & Thompson-Schill, 2004).

Each 26.4-second active (silent sentence production) block contained 3 line drawings, each presented for 8.8 seconds to allow time for patients to respond. A run comprised 6 active blocks interleaved with 7 passive viewing intervals that varied in length. Each experimental run began and ended with a passive viewing block. Passive viewing intervals were pseudorandomly varied between 17.6 sec (4 images), 22 sec (5 images), and 26.4 sec (6 images) to mitigate low frequency artifacts (Zarahn et al., 1997). The minimum interval was long enough to allow hemodynamic responses to return to baseline before the subsequent active block began. Each run lasted 316.8 sec during which 72 whole-brain functional images were acquired. Each patient completed 3 runs for 216 images total. Stimuli were projected onto a translucent screen at the patients' feet via a Solo 2500 LS Laptop Computer (Gateway) and an XG-NV7XU LCD Projector (Sharp Electronics) using E-Prime Version 1 software.

Image Acquisition

Images were acquired on a 1.5T GE Signa MR scanner using a dome-shaped quadrature radio frequency head coil (MRI Devices). Head motion was minimized using foam padding. For functional images, 24 contiguous sagittal slices covering the whole brain were acquired using a gradient echo 2-shot spiral scan sequence (TE = 40 ms; TR = 2200 ms; FA = 90°; FOV = 180 mm) (Macovski, 1985; King et al., 1995; Noll et al., 1995). Slice thickness differed slightly (less than 4%) between sessions to accommodate slight differences in head angle within the field of view between sessions (6.0 mm pre- and 6.2 mm posttreatment for G01; 6.6 mm pre- and 6.4 mm posttreatment for J02). In a 2-shot spiral sequence, each spiral is acquired during a separate TR; thus, the total time for acquisition of a single whole-brain volume of functional images was 4.4 seconds. For purposes of analysis, data were interpolated to a 128 × 128 matrix. Subsequent to functional images, structural T1-weighted images were acquired for 124 × 1.3 mm thick sagittal slices, using a 3D spoiled GRASS volume acquisition (TE = 7 ms; TR = 27 ms; NEX = 1; FOV = 240 mm; FA = 45°; matrix size = 256 × 192). Dual-echo T2-weighted images were also acquired (TE_eff = 34 ms and 85 ms; TR = 3000 ms; NEX = 1; FOV = 240 mm; matrix size = 256 × 128) using a fast spin echo sequence.

Image Analysis

Functional images were analyzed and overlaid onto structural images with the Analysis of Functional Neuroimaging (AFNI) program from the Medical College of Wisconsin (Cox, 1996). To minimize the effects of head motion, time series images were spatially registered in 3-dimensional space. Images were visually inspected for gross artifacts and viewed in a cine loop to detect residual motion. Neither patient's data contained enough images with gross artifacts or uncorrected head motion in pre- or posttreatment scans to warrant exclusion from the study. To mitigate large vessel effects and other sources of error, voxels where the standard deviation of the signal change exceeded 5% of the mean signal intensity were set to a value of zero. This threshold was developed in our laboratory for use with the current instrument and spiral pulse sequences as a convenient way to eliminate nonbrain voxels and voxels in large veins identified on MR angiograms, while at the same time preserving brain activity outside of major veins. Linear trends were removed from individual imaging runs by orthogonalization (i.e., detrending) (Birn et al., 2001).

Imaging runs were concatenated into a time series of 216 serial images. Hemodynamic responses for the active task were deconvolved from the time series on a voxel-by-voxel basis. Deconvolution was based on a 44-sec or 10-image lag comprising the active block length (6 images) plus 4 images (the shortest baseline interval). Deconvolution estimates the hemodynamic response by empirically determining the optimum shape of waveform matching the data and finding its goodness-of-fit or R² (Ward, 1998). Because the shape of the hemodynamic response in a given area for a particular task can vary considerably from one individual to another (Aguirre et al., 1998), deconvolution analysis should offer greater sensitivity between patients than methods making a priori assumptions about the shape of the hemodynamic response.

Signal-to-noise ratios (SNRs) were calculated based on the method of Weisskoff (1996), using the formula SNR = (0.655 × Sbrain)/Sbackground, with Sbrain being the mean of the signal in a 5 × 5 region of the periventricular white matter in a medial slice and Sbackground being the standard deviation in a 5 × 5 region of the signal in the background of the image (i.e., outside the head) in a region free from motion-related signal change. Although for G01 the SNRs were similar between sessions (pre: 45.2; post: 46.7), the SNRs for J02 were 59.7 pretreatment and 48.1 posttreatment. Since the noise structure varied between these scans, prior to analysis of pre- and posttreatment functional MR images, images were equated for sensitivity to BOLD response across sessions on a voxel-by-voxel basis using a procedure developed in our laboratory. In brief, this procedure uses residual variances from the deconvolution analyses to estimate the noise structures of pre- and posttreatment images. Different known amounts of signals were added to those noise structures. Probability density curves were generated for each noise structure at each signal level, showing the fraction of voxels as a function of R². Detection probability was then used to equate the pretreatment and posttreatment R²s for sensitivity. This procedure was applied to both patients' images.

After equating for sensitivity, each patient's pre- and posttreatment images were spatially coregistered in 3-dimensional space and converted to 1-mm cubic voxels having the same spatial coordinates for both scans. Activation was defined by statistical thresholds of R² ≥ .30 for G01 and R² ≥ .40 for J02. To compare amplitude of response in areas activated in pre- or posttreatment images, a composite image was created such that a voxel would be declared activated if statistical threshold was met either pre- or posttreatment (or both). A contiguity threshold of volume ≥ 200 microliters was used to select areas of activity for this union of pre- and posttreatment scans, so that the possibility of interpreting false positive activity was minimized. Hemodynamic response waveforms for each voxel in an area so identified were extracted from pre- and posttreatment deconvolution analyses. Conversion to Talairach coordinate space (Talairach & Tournoux, 1988) was performed to assist in finding the general regions in which these clusters of activity were located. Because locations of structures vary between individuals and mislocalization could result from using atlas coordinates alone, clusters were localized by reference to anatomic landmarks, using the atlas as a guide when necessary. A smaller contiguity threshold (volume > 100 microliters) was adopted for interpretation of the results of individual scans.

Finally, differences in volume of the active clusters could be observed by comparing pre- to posttreatment scans. The likelihood that these differences occurred by chance was evaluated by testing with binomial statistics whether there were different proportions of active voxels within a larger volume containing the pre- and posttreatment active clusters. That is, an active voxel was counted as a success, and the probability of success was estimated from the average of pre- and posttreatment counts. There were two caveats for these calculations. First, voxel counts were based on acquisition voxels (not 1 mm cubic voxels), because the former are more nearly independent, an important assumption about binomial trials. Second, there was no ideal way to determine the size of the larger volume within which the active clusters were embedded, that is, by analogy, the total number of binomial trials within which proportions of successes could be evaluated. Therefore, a wide range of volumes were used in the calculations, from the smallest possible for embedding the clusters (min) to larger than the whole brain, in steps of 2 × min, 4 × min, 10 × min, 100 × min, 1,000 × min, 10,000 × min, and 1,000,000 × min. We accepted as reliably different only those clusters for which, over the whole range of embedding volumes, p < .01 that the observed number of active voxels either pre- or posttreatment resulted from the probability of being active, estimated from their average.