Capturing multidimensionality in stroke aphasia: mapping principal behavioural components to neural structures

Assessments

In addition to the BDAE (Goodglass and Kaplan, 1983; Goodglass et al., 2000), a battery of language tests was administered to assess the participants’ language and cognitive abilities in a comprehensive fashion. The assessments involved input and output phonological processing, semantic processing and sentence comprehension, as well as more general cognitive function. Assessments were conducted with participants over several testing sessions, with the pace and number per session determined by the participant.

The language assessments included a variety of subtests from the Psycholinguistic Assessments of Language Processing in Aphasia (PALPA) battery (Kay et al., 1992), including: same-different auditory discrimination using non-word minimal pairs (PALPA 1); same-different auditory discrimination using word minimal pairs (PALPA 2); immediate repetition of non-words (PALPA 8); delayed repetition of non-words (PALPA 8); immediate repetition of words (PALPA 9); and delayed repetition of words (PALPA 9). A number of tests from the 64-item Cambridge Semantic Battery (Bozeat et al., 2000) were also included: the spoken word-to-picture matching task; a written word-to-picture matching version of the same task; the picture version of the Camel and Cactus Test; and the picture naming test. To increase sensitivity to mild naming deficits, the 60-item Boston Naming Test (BNT) (Kaplan et al., 1983) was also used. Similarly, to increase sensitivity to subtle semantic deficits, a 96-trial synonym judgement test with words presented in spoken and written form (Jefferies et al., 2009) was also used. To capture syntax level deficits, the spoken sentence comprehension task from the Comprehensive Aphasia Test (CAT) (Swinburn et al., 2005) was administered. Although we included this and subtests from the BDAE (including the Cookie Theft description) as assessments of discourse level processing, the focus of our analysis in this study was on deficits at the single word processing level, as these are the building blocks of language and involve tasks that are sensitive to residual abilities even in severe cases (Henseler et al., 2014). The additional cognitive tests included forward and backward digit span (Wechsler, 1987), the Brixton Spatial Rule Anticipation Task (Burgess and Shallice, 1997), and Raven’s Coloured Progressive Matrices (Raven, 1962).

On language assessments, apart from the Comprehensive Aphasia Test sentence comprehension test (Swinburn et al., 2005), participants were scored on their first response. For the Comprehensive Aphasia Test test, two points are given for a correct response and one point is given for delayed correct responses or self-corrections. For the two naming assessments, participants’ responses were marked correct if they were given within 5 s of presentation. Minor articulatory dysfluencies, but not phonological errors, in responses were accepted as correct. Repetition of auditory stimuli was provided if requested by participants.

Principal components analysis

Participants’ scores on all assessments were entered into a PCA with varimax rotation (conducted with SPSS 16.0). There is no clear guide on the number of cases needed for PCA, but good results have been obtained with a subject to variable ratio of 1.2 (Barrett and Kline, 1981). We had 17 variables and 31 cases, making the ratio in the current study 1.8. In addition, Preacher and MacCallum (2002) suggest factor recovery is good beyond a sample size of 20. Our same size therefore seems adequate for the purposes of PCA.

Factors with an eigenvalue ≥1.0 were extracted and then rotated. After orthogonal rotation, the factor loadings of each test allowed interpretation of what cognitive-language primary process was represented by that factor. Individual participants’ scores on each extracted factor were then used as behavioural covariates in the neuroimaging analysis.

Acquisition and processing

High resolution structural T₁-weighted MRI scans were acquired on a 3.0 T Philips Achieva scanner (Philips Healthcare) using an 8-element SENSE head coil. A T₁-weighted inversion recovery sequence with 3D acquisition was used, with the following parameters: repetition time = 9.0 ms, echo time = 3.93 ms, flip angle = 8°, 150 contiguous slices, slice thickness = 1 mm, acquired voxel size 1.0 × 1.0 × 1.0 mm³, matrix size 256 × 256, field of view = 256 mm, inversion time = 1150 ms, SENSE acceleration factor 2.5, total scan acquisition time = 575 s.

Participants’ MRI scans were normalized and segmented using a modified unified segmentation-normalization procedure optimized for lesioned brains (Seghier et al., 2008) implemented in Statistical Parametric Mapping (SPM) 8 (Wellcome Trust Centre for Neuroimaging, http://www.fil.ion.ucl.ac.uk/spm/) running under Matlab 2009a. Images were then smoothed with an 8 mm full-width at half-maximum Gaussian kernel and used in the lesion analyses described below. As suggested by Geva et al. (2012), the analyses were conducted on the normalized images incorporating both grey and white matter to allow detection of both cortical and subcortical correlates of deficits.

Automated lesion identification procedure

Automated outlines of brain areas classified as ‘abnormal’ were generated using Seghier et al.’s (2008) modified segmentation-normalization procedure. Data from all participants with stroke aphasia and all healthy controls were entered into the segmentation-normalization. Segmented images were smoothed with an 8 mm full-width at half-maximum Gaussian kernel as recommended by Seghier et al. (2008) and submitted to the automated routine’s lesion identification and definition modules using the default parameters apart from the lesion definition ‘U-threshold’, which was set to 0.5. We modified the U-threshold from 0.3 to 0.5 after comparing the results obtained for a sample of patients to what would be nominated as lesioned tissue by an expert neurologist. The generated images were used to create the ‘lesion’ overlap map in Fig. 2 and the individual ‘lesion’ outlines in Fig. 7.

Although it has been demonstrated that cost-function masking with a hand-traced lesion mask is the optimal method for spatial normalization of lesioned brains (Andersen et al., 2010; Wilke et al., 2011), this technique is both labour intensive and somewhat subjective as to what abnormalities fall within the lesion boundaries. We were interested in adopting an efficient and objective method of ‘lesion’ identification for use with large sample of patients, therefore we selected the fully automated method developed by Seghier et al. (2008). This method has been shown to perform at an acceptable level relative to hand tracing (Wilke et al., 2011), particularly in the case of large lesions, as was true of the majority of patients in our sample. The automated method involves initial segmentation and normalization into tissue classes of grey and white matter, CSF and an ‘extra’ tissue class, which allows for the presence of the ‘lesion’. After smoothing, voxels that emerge as outliers relative to normal participants are identified and the union of these outliers provides the ‘fuzzy lesion map’, from which is derived the lesion outline. It should be emphasized that this method essentially identifies areas of neural abnormality rather than ‘lesion’ per se. It is therefore likely to be affected by the abnormal shape of the ventricles in patients with large lesions, and hence is used with the caveat that periventricular results are treated with caution (Geva et al., 2012). On the other hand, this procedure has the potential to be sensitive to indirect lesion effects that would be missed using hand tracing (Wilke et al., 2011). Ultimately, our decision to adopt an automated lesion identification procedure in this study was driven by a desire to use a method that was easily replicable and that would effectively scale up to larger patient samples.

Syndromes and symptoms

The VBM analyses of BDAE subtypes and symptom groups (Figs 3 and 4) were conducted in SPM8 running on Matlab 2009a and 2012a, respectively. Smoothed and normalized T₁-weighted images from each patient in the relevant group and from the group of 19 healthy older control participants were entered into the analysis. Statistical comparisons were then carried out between the subtype or symptom group and the control group for every brain voxel. The resulting images show clusters of voxels in which the subtype or symptom group had a significantly lower concentration of tissue than the control group.

The syndrome analysis considered the nine patients with a BDAE classification of anomic aphasia, the eight patients with a classification of Broca’s aphasia and the six patients with a classification of mixed non-fluent aphasia. The symptom analysis considered those patients with the nine lowest scores on the Cambridge picture naming test, the eight lowest scores on delayed non-word repetition, and the six lowest scores spoken word-to-picture matching (numbers in each group were selected to match those falling into various syndromes).

Principle component analysis factors and test scores

The VBCM analyses of PCA factors and individual test scores were conducted in SPM8 running on Matlab 2009a and 2012a, respectively, with sets of factors or scores entered simultaneously as continuous behavioural covariates. The outcome of the analyses therefore denote which voxels’ variation in tissue concentration corresponds to the unique variance in a given principle component or test, while controlling for variation in the other components or tests included in that analysis.

The first analysis used the three continuous multidimensional predictors of the PCA factor scores, which are necessarily uncorrelated (orthogonal) with one another. We then contrasted these results with those obtained on the basis of a non-PCA selection of individual tests that seem to tap the same underlying abilities. Lastly, we contrasted these results with those obtained using individual tests, selected on the basis that they had the highest loadings on each PCA factor.

Identifying principal language-cognitive factors

The rotated PCA produced a three factor solution which accounted for 82% of variance in participants’ performance (F1 = 61%; F2 = 14%, F3 = 7%). The factor loadings of each of the different behavioural assessments are given in Table 3, with individual participants’ scores on each factor provided in Table 2. Tasks which tapped input and/or output phonology (e.g. non-word repetition, minimal pairs, picture naming and also digit span, which involves repetition of strings of numbers) loaded heavily on Factor 1, hence we refer to this factor as ‘Phonology’. Factor 2 was interpreted as ‘Semantics’, as the assessments that loaded heavily on it were those involving processing of meaning, whether receptive or expressive (e.g. spoken word-to-picture matching, synonym judgement and picture naming). Note that the two naming assessments loaded heavily on both of these factors, as they clearly require intact phonological and semantic processing to be performed successfully, consistent with previous results (Lambon Ralph et al., 2002; Schwartz et al., 2006). The assessments that loaded heavily on Factor 3 were more diverse. Both Raven’s progressive matrices and Brixton spatial anticipation taps pattern detection and prediction abilities. Although the Camel and Cactus Test does require semantic knowledge as shown by its disruption in semantic dementia (Bozeat et al., 2000), in this sample, performance seems to be affected more by the ability to reason out the basis for association, which is consistent with the semantic control deficits reported in stroke aphasia (Head, 1926; Jefferies and Lambon Ralph, 2006). We presume the loading for minimal pairs is due to the need for patients with phonological processing deficits to adopt an explicit comparative strategy and problem-solving for this task, which will be most apparent for non-words where semantics provides no support and thus the task becomes very challenging for aphasic patients. Overall, the tests loading on the third factor involve modality-independent choice, discrimination or reasoning, hence it was interpreted as the ‘executive-cognition’ factor.

Capturing global severity

When the behavioural data were entered into an unrotated PCA, all tests in the battery loaded heavily on the first unrotated factor, a factor that can be interpreted as reflecting each participant’s overall aphasic severity. This unrotated ‘severity’ factor correlated highly with the phonological factor from the rotated PCA (r = 0.766, P < 0.0005), and to a lesser extent with the semantic and cognitive factors (r = 0.500, P = 0.004 and r = 0.405, P = 0.024, respectively). This suggests that in this group of individuals with chronic stroke aphasia, severity maps quite closely onto the level of phonological processing impairment.

Relationship to aphasia subtypes

Figure 1 shows the relationship between the three factor scores and the BDAE classifications. As noted in the ‘Introduction’ section, each aphasia classification sits within a specific subregion of the 3D PCA space (as colours do within an RGB hue space). Thus, for example, patients with global aphasia are situated in the lower left quadrant of Fig. 1A, indicating poor phonological, semantic and cognitive performance. In contrast, the only participant in the cohort with transcortical sensory aphasia is found in the lower right quadrant of Fig. 1B, reflecting a combination of good phonological and cognitive skills yet impaired semantic performance.

In contrast to a pure categorical organization, it is well known that: (i) within each BDAE subtype, there is considerable variability between individual patients; (ii) there seems to be graded relationships between some subtypes (e.g. Wernicke-to-conduction aphasia); and (iii) many patients do not fit into a specific category (and are given a ‘mixed aphasia’ label). Again, this is like the colour analogy set out above: there are a variety of pinks, oranges and violets; some colours seem to border with each other (e.g. pink-to-red; yellow-to-orange, etc.); and some colours are hard to categorize (e.g. grey or khaki). The three extracted factor scores capture these same graded patterns. For example, the participants with Broca’s aphasia show varying phonological performance yet little variation in their semantic performance. Likewise, the ‘mixed’ non-fluent cases sit in the middle of the three factor space (i.e. with a moderate level of all three impairments). Consequently, by shifting away from a categorical model of aphasia towards a continuous multi-dimensional characterization, these more graded aspects of aphasia are captured while preserving the core differences between prototypical examples of each aphasia type. A key hypothesis for this study was that these continuous and independent factor scores would map more precisely onto key underlying neural regions than alternative categorical and/or unidimensional approaches, which was tested in the next analyses.

Lesion overlap

A lesion overlap map for stroke aphasic participants is provided in Fig. 2, and primarily covers the large left hemisphere area supplied by the middle cerebral artery (Phan et al., 2005). All neuroimaging results are shown overlaid on the Ch2better template in MRIcron (Rorden et al., 2007). The maximum number of participants who had a lesion in any one voxel was 26, in the region of the left rolandic operculum.

Voxel-based morphometry of syndromes and symptoms

Results are thresholded at P < 0.001 voxel-level, P ≤ 0.001 family-wise error (FWE) corrected cluster-level. For the syndrome level analysis, Fig. 3 shows areas of significantly lower tissue concentration in subsets of our participants, as defined by BDAE aphasia classification. Although this approach is multidimensional in the sense that each classification represents a profile of performance over a number of tests, it is nevertheless categorical. Despite some variation across categories, the principal finding was that there was a large lesion overlap between subtypes. This is consistent with previous studies that have used similar approaches (Kümmerer et al., 2013) and presumably reflects the fact that certain brain regions are more likely than others to be affected by a middle cerebral artery stroke (Phan et al., 2005). A symptom level analysis took comparable samples of patients scoring the lowest on individual tests of Cambridge picture naming, delayed nonword repetition and spoken word-to-picture matching (Fig. 4). The results showed even larger overlaps between groups, demonstrating the limitations of a unidimensional categorical approach. Indeed, the VBM analysis for each aphasia subtype and behavioural symptom closely mirrors the lesion overlap map (Fig. 2).

Voxel-based morphometry of principle component analysisfactors and test scores

For the PCA factor analyses, the VBCM results are shown in Fig. 5. Each map shows where tissue concentration covaries uniquely with a given factor score, which are necessarily uncorrelated with each other. Results are thresholded at P < 0.001 voxel-level, P ≤ 0.001 FWE corrected cluster-level. Performance on the phonological factor was uniquely correlated with voxels across a number of left hemisphere regions, principally a cluster containing: primary auditory cortex (Brodmann areas 41 and 42); mid to posterior middle and superior temporal gyri; superior temporal sulcus; and posterior portions of the insula, Heschl’s gyrus and the planum temporale. A second cluster in the left inferior prefrontal region was also identified at a slightly lower statistical threshold (Fig. 5A). The phonological cluster also overlapped with white matter regions, the location which encompasses part of the arcuate fasciculus, a key aspect of the dorsal language pathway (Wise, 2003; Catani and ffytche, 2005; Catani et al., 2005; Duffau et al., 2005; Parker et al., 2005; Saur et al., 2008).

Performance on the semantic factor was uniquely related to a cluster of voxels in the left hemisphere anterior temporal lobe (Fig. 5B). The cluster overlapped with the anterior middle temporal gyrus and the temporal stem (including the dorsal edges of the inferior temporal gyrus and fusiform gyrus). Thus, with regards to white matter, the cluster included an area corresponding to part of the ventral language route, overlapping parts of the inferior longitudinal fasciculus, inferior fronto-occipital fasciculus and uncinate fasciculus (e.g. Wise, 2003; Parker et al., 2005; Catani and Mesulam, 2008; Saur et al., 2008; Schmahmann and Pandya, 2008; Duffau et al., 2009). In contrast to the phonological and semantic factors, there were no clusters that correlated uniquely with the cognitive factor score and survived correction for multiple comparisons.

To highlight the advantages of the continuous-variable, multidimensional approach offered by PCA, we contrasted the results obtained with the extracted PCA factor scores to those found with raw test scores, conducting the SPM8 multiple-regression analysis in the same way. Of course, this necessitates selecting tests from the battery that capture the dimensions of phonology, semantics and executive-cognition. In the first ‘non-PCA selection’ analysis, we chose non-word minimal pairs, synonym judgement and Brixton spatial anticipation, as these represent widely-used ‘direct’ measures of each construct. The intercorrelations between these tests were: minimal pairs–synonyms r = 0.619, P < 0.0005; minimal pairs–Brixton r = 0.590, P < 0.0005; synonyms–Brixton r = 0.514, P = 0.003. At P < 0.001 voxel-level, P ≤ 0.001 FWE corrected cluster-level, non-word minimal pairs showed no significant clusters. Synonym judgement was uniquely associated with a significant cluster with subcortical peaks in lentiform nucleus/putamen and regions underlying left inferior frontal areas (Supplementary Fig. 1). Raven’s progressive matrices showed no significant clusters.

One virtue of the rotated PCA approach is that the rotation attempts to binarize the loading of each test across the extracted factors, which helps cognitive interpretation of each factor. This also has the consequence that we can use the results of the PCA to select the individual tests that best capture the key underlying dimensions, and that are also least correlated with one another. Indeed, when we considered the individual tests with the highest loadings on each PCA factor (delayed non-word repetition, spoken word-to-picture matching, and Raven’s coloured progressive matrices) then the intercorrelations between these tests were lower, albeit still significant in some cases: repetition-matching r = 0.371 P = 0.040; repetition-matrices r = 0.206, P = 0.266; matching-matrices r = 0.328, P = 0.071. As can be seen in Fig. 6, at P < 0.001 voxel-level, P ≤ 0.001 FWE corrected cluster-level, delayed non-word repetition showed a significant cluster centred on the superior temporal gyrus, middle temporal gyrus and posterior insula, which is similar to, although more constrained than, the results seen for the phonology PCA factor. The results for spoken word-to-picture matching showed a significant cluster in the left anterior temporal lobe, although this was more extensive than that seen for the semantics PCA factor. Performance on the Raven’s was not associated with any significant clusters, similar to the results seen for the cognition PCA factor. Hence, the results using individual test scores are much stronger when PCA factor loadings have been used to guide their selection. This highlights the utility of the PCA technique to isolate tests that best capture underlying functional dimensions.

Lesion size

To ensure that our results were not merely attributable to lesion size, each participant’s volume was calculated from the lesion identified by the modified segmentation-normalization procedure (see ‘Materials and methods’ section). When lesion volume alone was regressed against participants’ T₁-weighted scans a large voxel cluster in left hemisphere middle cerebral artery territory emerged as significant (Supplementary Fig. 2), representing the outer belt of the lesion overlap map (Fig. 2). The correlation between lesion volume and unrotated PCA factor score, or overall ‘severity’, was r = −0.545, P = 0.002.

For the PCA factors, lesion volume correlated relatively weakly with the phonology factor (r = −0.325, P = 0.075) and the semantic factor (r = −0.260, P = 0.159), and slightly more strongly the with executive-cognition factor (r = −0.411, P = 0.022). Crucially, including lesion volume in the VBCM model with the independent PCA factor scores did not alter the pattern of results obtained (Supplementary Fig. 3), indicating these continuous multidimensional factors are largely independent of global severity.

For the individual test analyses (cf. Supplementary Fig. 3 and Fig. 6), lesion volume correlated significantly with all the single tests considered in the imaging analyses: non-word minimal pairs (r = −0.462, P = 0.009), synonym judgement (r = −0.602, P < 0.0005), Brixton spatial anticipation (r = −0.419, P = 0.019), delayed non-word repetition (r = −0.389, P = 0.030), spoken word-to-picture matching (r = −0.410, P = 0.022), and identically for Raven’s progressive matrices (r = −0.411, P = 0.022). Lesion volume was included in each VBCM model with results thresholded at P < 0.005 voxel-level, P ≤ 0.01 FWE corrected cluster-level. For the non-PCA selected tests, including lesion volume removed the cluster associated with synonym judgement (i.e. no unique clusters were extracted for any of the three measures). For the tests selected on the basis of the PCA loadings, the delayed non-word repetition still showed the significant clusters for superior temporal gyrus/middle temporal gyrus and insula, spoken word-to-picture matching still showed a significant cluster centred in the left anterior temporal lobe, and Raven’s progressive matrices did not show any significant clusters (Supplementary Fig. 4). These additional test-score analyses demonstrate that more robust results emerge when individual assessments are selected according to the PCA.

Individual cases

The relationship of individual patients to the group-level analyses was explored for two reasons. First, if this form of neuroscience investigation is going to have clinical utility, then it is important to explore how clearly individual behavioural and neuroimaging results relate to the maps for each language-cognitive factor. Secondly and relatedly, exemplar cases can help interpretation of the behavioural factors and their neural correlates in terms of real individual patients [given that PCA, by design, generates scores that are at least one step removed from raw clinical measures: see Lambon Ralph et al. (2003)]. Four exemplar participants were selected to provide contrasting pairs who scored above (‘high’) or below (‘low’) the median for the group for each language factor (Fig. 7).

In the first exemplar pair, Patient AL (high-level anomic: A6) was the patient who on average scored highest above the median on both phonology and semantics whereas Patient LM (global aphasic: G2) was the patient who on average scored in lowest below the median on both factors. This contrast provides an illustration of aphasia severity. As shown in Fig. 7, Patient AL’s lesion falls outside the key areas identified as uniquely supporting phonological and semantic processing, whereas Patient LM’s lesion largely encompasses both areas.

The second exemplar pair illustrates the specificity of impairments. Patient DM (Broca’s aphasic; B3) was the patient with the largest discrepancy between factor scores amongst those who scored ‘low’ on phonology but ‘high’ on semantics whereas Patient KS (transcortical sensory aphasia) was the patient with the largest discrepancy amongst those scoring ‘low’ on semantics but ‘high’ on phonology. As expected, Patient DM’s lesion encompassed brain regions identified as uniquely correlating with phonology, whereas those areas shown as uniquely related to semantics fell largely outside the boundary of his lesion. Conversely, Patient KS’s lesion had the opposite distribution in keeping with his transcortical sensory aphasia profile.