The impact of phonological versus semantic repetition training on generalisation in chronic stroke aphasia reflects differences in dorsal pathway connectivity Rachel Holland1, Sasha L. Johns2, & Anna M. Woollams2 1Division of Language and Communication Science
City University London
London, EC1R 0JD, England
2Neuroscience and Aphasia Research Unit
School of Psychological Sciences, Zochonis Building
University of Manchester, Brunswick Street
Manchester, M13 9PL, England
Dr Anna M. Woollams
Neuroscience and Aphasia Research Unit
School of Psychological Sciences, Zochonis Building
University of Manchester, Brunswick Street
Manchester, M13 9PL, England
Running title: Repetition training and dorsal connectivity
Acknowledgement: We are very grateful to Rebecca Butler and Matthew Lambon Ralph for assistance with background behavioural and neuroimaging data for the patients reported in this paper.
It has been suggested that neuroimaging can be used to inform therapeutic intervention.
The current study aimed to determine whether an individual would benefit more from training engaging their intact or damaged neural pathway. Two males with chronic stroke aphasia participated, with DM showing milder disruption of connectivity along the dorsal language pathway relative to JS, according to distortion corrected diffusion-weighted MRI. Each patient received two blocks of six repetition training sessions over two weeks, one of which was “phonological” and the other “semantic” in nature. Both phonological and semantic training produced significant gains for both patients for trained items. For the untrained control items, significant gains were specific to training type for each patient. Only phonological training elicited significant generalisation for DM, which was greater than that seen for JS. Conversely, only semantic training elicited significant generalisation for JS, which was greater than that seen for DM. This double dissociation in generalisation effects suggests that a restitutive approach is more effective for patients with milder damage while a compensatory approach may be more effective for those with more severe damage. These results indicate the utility of neuroimaging to optimise relearning strategies and promote generalisation to untrained items.
KEYWORDS: APHASIA, NEUROIMAGING, GENERALISATION, PHONOLOGICAL, SEMANTIC
There is increasing evidence within speech and language literature of two distinct pathways for language within the human brain (Hickok & Poeppel, 2000, 2004, 2007). Two cortico-cortical pathways project from the bilateral superior temporal gyrus, a region engaged in early speech perception, to form a ventral stream and a dorsal stream. The ventral stream maps between acoustics and articulation via meaning and is implicated in tasks involving auditory comprehension. The dorsal stream maps sound onto articulation which supports sub-lexical speech processing tasks and is crucial for auditory-motor integration of both linguistic and non-linguistic processes. Consistent with this framework, Saur et al. (2008) found that auditory comprehension was subserved by a ventral pathway mediated via the extreme capsule, connecting middle and inferior temporal regions to the ventrolateral prefrontal cortex. In contrast, repetition of nonwords relied upon a dorsal pathway connecting the superior temporal lobe and premotor regions via the arcuate and superior longitudinal fasciculi.
In terms of dysfunction, damage to these two different language pathways has been strongly implicated in aphasia (Binder, Medler, Desai, Conant, & Liebenthal, 2005; Ueno, Saito, Rogers, & Lambon Ralph, 2011). Indeed, different aphasic language profiles support the existence of two simultaneous, parallel anatomical pathways involved in language processing (Friederici & Gierhan, 2013). Disturbance of the dorsal pathway may lead to conduction aphasia, which is characterised by a selective impairment of repetition with preserved comprehension and the production of phonological paraphasias. In contrast, disruption of the ventral route may lead to transcortical sensory aphasia, whose predominant feature is preserved repetition and production in the context of poor comprehension (Kummerer et al., 2013; Noonan, Jefferies, Corbett, & Lambon Ralph, 2010). The degree of lateralisation of function in each processing pathway in the dual-stream model has also been informed by neuropsychological data (Hickok & Poeppel, 2007). The ventral stream is organised bilaterally with each hemisphere supporting different, but complementary parallel processing systems. The dorsal stream, on the other hand, is proposed to be strongly left-hemisphere dominant (Berthier, Lambon Ralph, Pujol, & Green, 2012; Catani & Mesulam, 2008)
Although independent, the ventral and dorsal streams are highly interactive. Rolheiser, Stamatakis and Tyler (2011) considered 24 chronic stroke patients’ performance across 10 tests involving key aspects of language production and comprehension and how this related to the results of diffusion weighted imaging. Phonological processing was found to load most heavily on the arcuate fascicle, implicated in the dorsal stream (Hickok & Poeppel, 2007). Conversely, semantic tasks were found to load on the extreme capsule, which was implicated in the ventral stream (Hickok & Poeppel, 2007). Importantly, whole-brain correlations showed that only performance on tasks loading strongly on either phonology or semantics fit into this dual-stream model, whereas complex linguistic functions of syntax and morphology required integrity of both pathways (Rolheiser, et al., 2011).
A number of studies have identified that both phonological therapies that engage the dorsal route and semantic therapies that engage the ventral route can produce appreciable improvements in aphasic individuals (e.g., Barthel, Meinzer, Djundja, & Rockstroh, 2008; Bruce & Howard, 1987; Coelho, McHugh, & Boyle, 2000; Fridriksson et al., 2009; Lorenz & Ziegler, 2009; Nettleton & Lesser, 1991; Raymer, Thompson, Jacobs, & Le Grand, 1993). There are however, mixed opinions as to which types of therapy are superior. Some have suggested that semantic treatments may be more effective, based in part on increased generalisation to untreated items immediately post-therapy (Howard, 1985; but c.f. Howard, 2000 for a re-analysis showing equivalent gains). Others have suggested that phonological treatments show stronger immediate gains, but for some patients semantic treatments have greater longevity (Lorenz & Ziegler, 2009).
Nickels (2002) has proposed that a combination of phonology and semantics may be the most effective treatment, consistent with the synergistic view of speech processing via the dorsal and ventral pathways. However, (Howard, 2000) argued that the difference between semantic and phonological tasks is often overstated. For example, in studies that aim to employ a “semantic” treatment such as word-to-picture matching, or a yes/no decision task (e.g., is a cat an animal?) the spoken or written phonological form of the target is also provided. Likewise, in “phonological” treatment tasks such as phonologically-cued picture naming semantic activation is also elicited via the presence of a pictorial stimulus. Therefore, both intervention types are engaging the language system in a similar way: by strengthening mappings between semantics and phonology.
There is evidence that the treatment efficacy interacts with nature of the patients’ impairment. In a recent study by Best and colleagues (2013), two patient groups were administered an identical phonological cueing treatment for picture naming. One group of patients was classified as having relatively less of a semantic difficulty (as measured by spoken and written word to picture matching) and more of a phonological output deficit (as measured by length effects and phonological errors in picture naming) when compared to the other group. It was the patients with the greater degree of phonological impairment and lesser degree of semantic impairment who then demonstrated generalisation to untreated items. Importantly, outcome did not relate to traditional aphasia classification, but rather was driven by characterisation of retained behavioural skill.
Critically, in previous research considering the relative efficacy of phonological vs semantic therapies, one factor that is rarely considered in determining the patients’ response is the nature of that person’s underlying brain damage (Abel, Weiller, Huber, Willmes, & Specht, 2015). Neuroimaging is being harnessed to predict recovery (Price, Seghier, & Leff, 2010; Seghier et al., 2016), and could also be utilised to inform intervention. The dual stream model clearly suggests that brain damage can differentially affect the dorsal and ventral pathways, and this has been validated in recent lesion-symptom mapping studies (Butler, Lambon Ralph, & Woollams, 2014). This is obviously a factor that will have some impact on the relative effectiveness of phonological versus semantic treatment strategies. Yet if brain damage does affect one pathway more than the other, then the question becomes whether therapy should focus on rebuilding the function associated with that pathway (e.g., a phonological therapy for a patient with dorsal damage) or rather enhance the use of relatively intact pathways (e.g., a semantic therapy for a patient with dorsal damage). The general issue as to whether therapy should be impairment-focused (e.g., Coelho, et al., 2000; Fridriksson, et al., 2009; Louis et al., 2001; Nettleton & Lesser, 1991) or draw upon intact processing abilities (Yampolsky & Wayers, 2002) is a matter of ongoing debate in the literature.
The goal of this study is to explore how the underlying neuropathology interacts with intervention type in terms of both direct therapeutic gains and the potential for generalisation to untrained items. Using the dual-stream model as a basis to inform therapy for word repetition difficulties, the current study aimed to determine whether an individual would benefit more from restitutive training to restore the function of the damaged neural pathway or compensatory training that takes advantage of the function of the intact neural pathway. Word repetition was selected as the target as this is a “degenerate” task (Price & Friston, 2002) that can be accomplished via either the dorsal or ventral pathways. Furthermore, the ability to repeat is often a capacity required in order to effectively engage with most traditional therapeutic interventions for word finding difficulties (e.g., Abel, Schultz, Radermacher, Willmes, & Huber, 2005; Bastiaanse, Bosje, & Franssen, 1996; Nickels, 1992, 2002). We employed a phonological and a semantically oriented relearning protocol to tap the capacity of the damaged dorsal and intact ventral pathways respectively. We compared the effectiveness of phonological vs. semantic therapy for repetition in two individuals, patients DM and JS, with differential degrees of damage to the dorsal language pathways, as determined by Diffusion Tensor Imaging and subsequent probabilistic tractography (Anatomical Connectivity Mapping).
2. Material and Methods
Two native English speaking, right-handed males (DM and JS) with chronic stroke aphasia were recruited from a larger study concerning the role of white matter connectivity in chronic stroke aphasia (Butler, et al., 2014). Both patients had a single left-hemisphere stroke, more than one year previous, resulting in chronic stroke aphasia. Both DM and JS are classified as Broca-type aphasic speakers and were impaired on the Cambridge 64-item picture naming test (Bozeat, Lambon Ralph, Patterson, Garrard, & Hodges, 2000; Hodges, Patterson, Oxbury, & Funnell, 1992) and the Boston Naming Test (Kaplan, Goodglass, & Weintraub, 1983). Table 1 provides demographic information and summarises the performance of the patients on a variety of neuropsychological tests. For those neuropsychological tasks without published normative data, Butler et al. (2014), collected control data from healthy control participants (three females, 10 males): mean age = 68.69 years (SD = 6.55), range = 59–80 years; mean years of education = 12.55(SD = 2.38), range = 10–17 years.
In comparison to control data, both patients showed impairments across assessments that engaged phonological and semantic knowledge. For example, both patients were impaired at repetition, suggesting that the dorsal pathway is most likely damaged. However, repetition of known words may also be achieved via meaning and thus the ventral pathway. DM is impaired at all repetition tasks, JS only at immediate repetition of words and nonwords. Patient JS was impaired at the synonym task and the Camel and Cactus picture association task, which could be ascribed to damage to the ventral pathway, but the synonym task also involves a speech perception component that is mediated by the dorsal pathway. Both patients were impaired on the spoken sentence comprehension task: a task that relies upon both the dorsal pathways for processing of the spoken input and ventral pathway for access to meaning. Given these intricacies of standardised assessments across phonological and semantic boundaries the degree of damage to the dorsal and ventral pathways cannot be easily identified on the basis of behavioural profile alone.
Neither of the patients was receiving any individual or group therapy for the treatment of naming deficits during the course of this study, but both have a history of speech and language treatment. Both patients and their carers gave informed consent to participate in the study in accordance with the NHS approved ethics associated with the study.
Table 1: Demographic and behavioural assessment battery scores for each participant as measured at the time of their brain scans. Scores are given as percentages. Scores marked in bold fall below the cut-off for normal performance. The cut-off was calculated as 2 SD below the mean control performance. a Cut-off based on published norms. b No cut-off available.
Experimental stimuli for the pre-test consisted of a set of 240 words varied by imageability (120 low; 120 high imageability words). Both high and low imageability sets were matched for overall frequency according to the Kucera and Francis (Kucera & Francis, 1967) and CELEX (Baayen, Piepenbrock, & Gulikers, 1995) measures, number of syllables and phonemes (MRC psycholinguistic database, Coltheart, 1981). Training stimuli were individualised for each patient and drawn from words that patients failed to accurately repeat on both of the pre-testing occasions. For each patient, consistently failed items were then split into four matched sets of fifteen words that comprised the relearning conditions (phonological or semantic) with the remaining two sets as untrained control for each relearning condition. For each patient, t-tests revealed no significant differences on any of the psycholinguistic measures across training and control sets used in each therapy, nor across training and control sets within each condition (all t-values < 1 for both patients, p-value ranges from 0.35 – 1). Although the two patients received different items, the training and control stimuli were matched as closely as possible across patients in terms of imageability, frequency and word length (number of syllables, phonemes and letters). Nevertheless, stimuli for DM tended to be somewhat longer than for JS. In particular syllable length of the phonological training (t=2.30, df = 28, p=0.03) and control set (t=2.59, df = 28, p=0.02) were greater for DM than JS. Likewise, the semantic training and control condition were also longer in terms of the number of syllables for DM (training: t=2.14, df = 28, p=0.04; control: t=2.96, df = 28, p=0.006) than JS. Despite these differences in length measures, imageability and frequency measures were closely matched across patients (all t values <1.35). See Table 2 for average measure values for each participant in each condition.
Table 2: Mean stimulus properties for each condition for each patient.
Two pre-training assessments separated by one week were conducted to establish baseline repetition performance of the 240-item experimental stimuli. At each pre-training assessment time points, patients were required to immediately repeat each heard word as quickly and as accurately as they could. A fixation cross appeared on a computer screen to signal the end of the auditory stimuli and prompt a response. Repetition was self-paced and a keypress was required to initiate the next trial. Training sessions then began a week after the second pre-training assessment session. Verbal responses in the pre-training assessments, baseline assessments, training sessions, and post-training assessments were digitally recorded for offline coding of response accuracy.
2.4. Training protocol
Training blocks consisted of three one-hour sessions per week for two weeks with each patient receiving a total of twelve sessions over two training blocks. Training was either “phonological” or “semantic” in nature with both training conditions requiring that patients repeat each heard word. In the phonological condition, each heard word was accompanied by a video of a mouth saying the same word simultaneously. Such audio-visual integration has been shown to not only improve picture naming performance in aphasic speakers on both trained and untrained items but also created an “errorless learning” environment which patients found particularly enjoyable (Fridriksson, et al., 2009). Fridriksson and colleagues (2009) demonstrated that treatment of speech production in non-fluent aphasic patients can make use of motor speech perception, even though this process is also somewhat impaired in aphasia (Schmid & Ziegler, 2006). Perception of audio-visual speech activates left frontal regions also involved in speech production; hence the use of such stimuli aphasia therapy aims to activate these regions to stimulate any residual function. In contrast, in the semantic training condition, each heard word was paired with an associated picture (e.g., high imageability: “beef” with a picture of a sliced roast; “bandage” with a picture of an arm being bandaged; low imageability: “hazard” with a picture of a warning sign, “envy” with a picture of a green eye). We used pictures of a concrete associates in order to allow the same therapy approach for both high and low imageability items (Hoffman & Lambon Ralph, 2011) although of course the word referent was more often present in the picture for the high imageability words than the low imageability words. Stimuli were presented using the DMDX software (Forster & Forster, 2003) on a Dell Laptop. Presentation was self-paced with each therapy session consisting of six repetitions of the 15 item training set, with each block of the set randomised anew.
Each patient received two blocks of training, with one block consisting of six semantic sessions and the other block consisting of six phonological sessions, the order of which was counterbalanced across patients (DM received semantic training first, while JS received phonological training first). Training sessions were equally spaced within each week of the training block. Before beginning the first training block, each patient was retested on the training and control items for that training block condition to provide a current baseline measure. After completion of the first training block, patients’ repetition performance was immediately assessed on their trained and matched control words. Maintenance of any gains was determined by a follow-up assessment one week later. Before beginning the second training block, each patient was retested on the training and control items for that training block condition to provide a revised baseline measure. Immediate gains in word repetition at the second training block and at one-week follow-up were again assessed (see Figure 1 below for protocol).
Post-training Assessment (FU) of block 1 items & Baseline Assessment of Block 2 items
2-week training Block 2
15-item training set 2
Post-training Assessment (FU) of block 2 items
Figure 1: Schematic of study design. Assessment days are highlighted in blue. Training blocks are highlighted in green. If set 1 corresponded to the phonological training items, set 2 would therefore correspond to the semantic items and vice versa.
2.5. Acquisition of Imaging Data
Patients’ imaging data were acquired as part of a larger case series study by Butler, Lambon Ralph, and Woollams (2014). All scans were acquired on a 3T Philips Achieva scanner (Philips Healthcare, Best, The Netherlands) using an 8-element SENSE head coil. High resolution structural MRI scans were acquired using a T1-weighted inversion recovery sequence with 3D acquisition, with the following parameters: TR (repetition time) = 9.0 ms, TE (echo time) = 3.93 ms, flip angle = 8 °, 150 contiguous slices, slice thickness = 1 mm, acquired voxel size 1.0 mm x 1.0 mm x 1.0 mm, matrix size 256 x 256, FOV = 256 mm x 256 mm, TI (inversion time) = 1150 ms, SENSE acceleration factor 2.5. Distortion corrected diffusion-weighted images were acquired using a pulsed gradient spin echo echo-planar imaging sequence implemented with TE = 54 ms, Gmax = 62 mT/m, half scan factor = 0.679, 112 x 112 image matrix reconstructed to 128 x 128 using zero filling, reconstructed resolution 1.875 mm x 1.875 mm, slice thickness 2.1 mm, 60 contiguous slices, 43 non-collinear diffusion sensitization directions at b = 1200 s/mm2 (Δ = 29.8 ms, δ = 13.1 ms), 1 at b = 0, SENSE acceleration factor = 2.5. Artefacts arising from pulsatile brain movements (Jones & Pierpaoli, 2005) were minimised by cardiac gating the diffusion sequence using a peripheral pulse unit placed on the participant’s finger. Acquisition time for the diffusion MRI data was approximately 28 minutes, although this varied slightly based on the participant’s heart rate. For each diffusion gradient direction, phase encoding was performed in right-left and left-right directions, giving two sets of images with the same diffusion gradient directions but opposite polarity k-space traversal, and hence reversed phase and frequency encode direction, allowing correction for geometric distortion (Embleton, Haroon, Morris, Ralph, & Parker, 2010). A co-localised T2 weighted turbo spin echo scan with 0.94 mm x 0.94 mm in-plane resolution and 2.1 mm slice thickness was also obtained for use as a structural reference scan in distortion correction (Embleton, et al., 2010).
2.6. Pre-processing of Imaging Data
Pre-processing of T1-weighted data was conducted in SPM8 (SPM8, Wellcome Trust Centre for Neuroimaging, http://www.fil.ion.ucl.ac.uk/spm/). Patients’ T1-weighted scans were normalised and segmented, together with 19 age- and education-matched healthy control patients’ brains using a modified unified normalisation-segmentation procedure (Seghier, Ramlackhansingh, Crinion, Leff, & Price, 2008). The normalised images were then smoothed using an 8 mm full-width-at-half-maximum (FWHM) Gaussian kernel. Susceptibility- and eddy current-induced distortions in diffusion data were corrected using Embleton’s et al.’s (2010) distortion correction method implemented in MATLAB. Distortion-corrected diffusion weighted images were then processed using the model-based bootstrap (Haroon, Morris, Embleton, & Parker, 2009), applied to constrained spherical deconvolution (CSD) (Tournier, Calamante, & Connelly, 2007; Tournier et al., 2008). The bootstrapped CSD was used to derive probability density functions (PDFs) that were used to produce whole brain probabilistic tractography-derived connection maps called Anatomical Connectivity Maps (ACMs) (Embleton, Morris, Haroon, Lambon Ralph, & Parker, 2007). ACMs quantify the total number of probabilistic paths recorded passing through each voxel of the brain, thereby providing a measure of the degree of tractography-derived anatomical connectivity passing to, from and through each voxel. ACMs were generated using the probabilistic index of connectivity (PICo) tractography algorithm (Parker, Haroon, & Wheeler-Kingshott, 2003), with ten tractography streamlines launched from every brain voxel. Each participant’s T1-weighted image and ACM were co-registered using a rigid-body transformation and normalised to Montreal Neurological Institute (MNI) space in SPM8. All coordinates are reported in MNI space.
3.1. Behavioural data
Raw repetition accuracy scores for each patient at each time point for each intervention type (semantic or phonological) relative to a matched control condition are reported in Table 3. The number of items incorrect at baseline that were then correct after training is provided in parentheses.
Gains relative to baseline: Considering data for those items that were incorrect at baseline, ExactMcNemar tests (one-tailed) revealed that patient DM showed a significant increase in word repetition accuracy for trained items immediately after the phonological (χ2 =5.14, p=.008) and semantic training conditions (χ 2 =5.14, p=.008), both of which were maintained at follow up (χ 2 =5.14, p=.008; χ 2 =5.14, p=.008). The control conditions also showed a significant increase immediately after phonological training (χ 2 =3.20, p=.031), which was marginal at follow up (χ 2 =2.25, p=.062), and contrasted with the lack of significant changes to the control items after semantic training (χ 2 =0.50, p=.250, χ 2 =0.50, p=.250).
Patient JS also showed significant improvement in word repetition accuracy for trained items after both the phonological (χ 2 =8.10, p=.001) and semantic (χ 2 =7.11, p=.002) training conditions, which were again maintained at follow up (χ 2 =4.17, p=.016, χ 2 =5.14, p=.008). JS showed no reliable gains for control items in the phonological training condition with performance remaining at zero correct at both post-training assessment time points, but did show a significant improvement on control items for the semantic training condition (χ 2 =4.17, p=.016), although not maintained at follow-up (χ 2 =1.33, p=.125).
In summary, both phonological and semantic therapy was produced significant gains for both patients for trained items. What is more interesting are the results for the untrained control items, which showed significant increases for the phonological condition only for DM and the semantic condition only for JS. While it could be argued that increases for control items were a simply the result of contact associated with the process of training, the fact that these gains were not general but specific to training type, and indeed to different training types across patients, suggests that these effects may instead be better viewed as effective generalisation.
Table 3: Number of items correct for each training condition for each patient, with the number of items incorrect at baseline but correct after training in parentheses. Pre = baseline performance before training block; Post = immediately post training; FU= one week follow-up.
Comparison of gains over treatments: Calculation of the magnitude of gains (i.e., the proportion of trained and control items incorrect at baseline that were correct after training) are shown in Figure 2.
For DM, semantic and phonological training was equally effective for trained items both immediately and at follow up (χ 2 =0.03, p=.857; χ 2 =0.03, p=.857). For control items, phonological training was more effective than the semantic training immediately (χ 2 =1.58, p=.159) and at follow-up (χ 2 =1.03, p=.311), although this difference did not reach significance. This difference in performance for control items over training type meant that the advantage of trained over untrained items was significant for the semantic condition (immediate: χ 2 = 3.97, p=.046; follow-up: χ 2 = 3.97, p=.046) but not the phonological condition (immediate: χ 2 =0.58, p=0.446; follow-up; χ 2 = 1.35, p=.244). It is important to note that while this latter contrast could be taken as indicating greater treatment gains for the semantic condition for DM, if we take control item performance as indicating generalisation, the opposite is true (net gains over trained and untrained of 12/11 items for phonological vs 9/9 items for semantic immediately and at follow-up, respectively).
For JS, semantic and phonological therapy were also equally effective for trained items both immediately (χ 2 =0.02, p=.885) and at follow up (χ 2 =0.53, p=.464). For control items, semantic training was significantly more effective than the phonological training immediately (χ 2 =8.11, p=.004) and marginally so at follow-up (χ 2 =3.59, p=.058). The advantage of trained over untrained items was clearly significant for the phonological condition (immediate: χ 2 =15.00, p<.001, follow-up: χ 2 =7.50, p=.006) but was only marginally so at follow-up for the semantic condition (immediate: χ 2 =1.90, p=.168; follow-up: χ 2 =3.04, p=.082). Again, this could be taken as indicating greater treatment gains for the phonological condition for JS, but if we take control item performance as indicating generalisation, the opposite is clearly the case (net gains over trained and untrained of 10/6 items for phonological vs 15/10 items for semantic immediately and at follow-up, respectively).
Figure 2: Number of items incorrect at baseline but correct after training for patients DM and JS. Post = immediately post training; FU= one week follow-up.
Comparison of gains over patients: For the trained items, there were no significant differences in the overall extent of gain made by patients after either the phonological (χ 2 = 0.8266, p=.3620) or the semantic (χ 2 = 1.45, p=.229) condition immediately post training, or at follow up (phonological: χ 2 = 0.29, p = 0.584; semantic: χ 2 = 0.14, p = 0.705). However, direct comparison of the overall gains made by each patient after each training condition for control items showed a striking interaction: patient DM repeated significantly more control items following phonological therapy than patient JS both immediately and at follow-up (χ 2 = 6.47, p=.011; χ 2 = 4.97, p=.025;). Conversely, patient JS repeated marginally significantly more control items following semantic therapy than patient DM (χ 2 = 3.16, p=.075), although this difference was not maintained at follow-up due to a decay of JS’s gains (χ 2 = 0.333, p=.564).
The behavioural results therefore suggest an initial double dissociation over patients in the effectiveness of training on performance for untrained items. Given the internal control provided by the within-subject cross-over design and pairwise matching of training and control sets on a variety of psycholinguistic variables, a key factor that may support differing generalisation effects observed in each patient result from variation in the nature of the patient’s brain damage. We consider not only the integrity of grey and white matter, but also the degree of reductions in connectivity along white matter pathways in each patient using ACMs. According to the dual stream model of speech processing, we would expect that differences between patients in damage to the dorsal and/or ventral pathways may have mediated the differential generalisation of training effects to control items.