|Home | About | Journals | Submit | Contact Us | Français|
Many patients with primary progressive aphasia (PPA) are impaired in syntactic production. Because most previous studies of expressive syntax in PPA have relied on quantitative analysis of connected speech samples, which is a relatively unconstrained task, it is not well understood which specific syntactic structures are most challenging for these patients. We used an elicited syntactic production task to identify which syntactic structures pose difficulties for 31 patients with three variants of PPA: non-fluent/agrammatic, semantic and logopenic. Neurodegenerative and healthy age-matched participants were included as controls. As expected, non-fluent/agrammatic patients made the most syntactic errors. The structures that resulted in the most errors were constructions involving third person singular present agreement, and constructions involving embedded clauses. Deficits on this elicited production task were associated with atrophy of the left posterior inferior frontal gyrus.
Primary progressive aphasia (PPA) is a neurodegenerative syndrome in which focal degeneration of language areas leads to progressive language deficits, with other cognitive domains relatively spared (Mesulam, 1982, 2001; Gorno-Tempini et al., 2011). There are three widely recognized variants of PPA. Non-fluent/agrammatic PPA is characterized by agrammatism and/or apraxia of speech (Grossman et al., 1996; Hodges & Patterson, 1996); semantic PPA (also known as semantic dementia) involves deficits in lexical and semantic knowledge (Hodges et al., 1992; Snowden et al., 1989; Warrington, 1975); and logopenic PPA is associated with phonological and word-finding deficits (Gorno-Tempini et al., 2004, 2008). The three variants differ in terms of distribution of atrophy (Gorno-Tempini et al., 2004) and underlying pathologies (Grossman et al., 2010; Snowden et al., 2011).
Syntactic production and comprehension are impaired in non-fluent/agrammatic PPA and to some extent in logopenic PPA, but are relatively spared in semantic PPA (Gorno-Tempini et al., 2004; Grossman et al., 1996; Hodges & Patterson, 1996; Thompson et al., 1997; Wilson et al., 2010a, 2010b, 2011; for review see Wilson et al., 2012). Assessment of syntactic production is not always straightforward. Most studies that have investigated syntactic production in PPA have done so by quantitative analysis of samples of connected speech (Ash et al., 2006, 2009; Bird et al., 2000; Graham et al., 2004; Gunawardena et al., 2010; Knibb et al., 2009; Meteyard and Patterson, 2009; Orange et al., 1998; Patterson et al., 2006; Patterson and MacDonald, 2006; Rogers and Alarcon, 1998; Thompson et al., 1997, 2012a; Wilson et al., 2010b). While this approach provides rich and comprehensive data, the unconstrained nature of elicited narratives or picture descriptions poses several challenges. Individuals differ in terms of which syntactic structures they will select to tell a narrative or describe a scene. Therefore it is difficult to determine which particular syntactic structures are difficult for patients, because some patients may attempt challenging structures, resulting in errors, whereas others may produce simplified structures in order to avoid errors (Wilson et al., 2010b). Furthermore, sometimes when patients make errors, it is not possible to determine the intended structure with certainty.
An alternative approach, which has been employed in just a few studies, is to use elicited production tasks (Thompson et al., 2012b; Weintraub et al., 2009). Weintraub et al. (2009) proposed the Northwestern Anagram Test (NAT), which requires patients to assemble words on printed cards to produce sentences describing pictures. The words that are provided (the first few of which are placed for the patient) constrain the sentence that can be produced. Using the NAT, the authors showed that PPA patients perform more poorly on non-canonical syntactic structures—passives, object wh-questions, and object relatives—than they do on canonical structures. However PPA patients were not divided according to variants in that study. In a subsequent study, impairments on the NAT were linked to reduced cortical thickness in the left inferior frontal gyrus, ventral sensorimotor cortex, and supramarginal gyrus (Rogalski et al., 2011).
In another study from the same group, Thompson et al. (2012b) investigated syntactic production in non-fluent/agrammatic and logopenic patients using two elicitation procedures. In one, the Sentence Production Priming Test (SPPT), the experimenter would describe a picture using a particular sentence structure, and the patient was required to describe another picture using the same structure. In the other, a sentence completion task was used to elicit verbs in various finite or non-finite forms. The authors showed that non-fluent/agrammatic patients have specific difficulties with non-canonical structures such as passives, object wh-questions, and object relatives, and with production of finite verb forms. Their performance was better when they produced canonical structures and non-finite verb forms. In contrast, logopenic patients made comparatively few errors, and did not show the same decrement in performance on non-canonical structures and non-finite verb forms (Thompson et al., 2012b).
These elicited production studies have provided valuable data about production of syntactic structures in PPA. However, only a limited range of structures have been investigated so far. Furthermore, the NAT and the SPPT likely make significant demands on executive processes and verbal working memory, which may complicate interpretation. Finally, while Thompson et al. (2012b) compared non-fluent/agrammatic and logopenic patients, no study has examined syntactic production using an elicited production procedure in all three PPA variants.
In this study, we investigated syntactic production in the three variants of PPA, using an elicited production task (Goodglass et al., 1972) to probe production of eleven specific syntactic structures varying in complexity. This simple story completion task is easily understood by patients and appears to make limited demands on other processes. The primary aim of the study was to determine which structures are difficult for patients with PPA. A secondary aim was to identify brain regions where atrophy was predictive of syntactic production deficits as quantified by this elicited production task.
Six groups of participants were recruited through the UCSF Memory and Aging Center: three variants of PPA; patients with behavioral variant fronto-temporal dementia (bvFTD); patients with other neurodegenerative diseases (“mixed neurodegenerative”); and healthy age-matched controls. The bvFTD and mixed neurodegenerative groups were included as neurodegenerative control groups. All participants gave written informed consent, and the study was approved by institutional review boards at UCSF and the University of Arizona.
Participants received a comprehensive evaluation including a neurological history and examination, neuropsychological testing, and neuroimaging. Patients were diagnosed with PPA based on recently published criteria (Gorno-Tempini et al., 2011), with bvFTD according to established criteria (Neary et al., 1998), or with other neurodegenerative diseases (see below). The mixed neurodegenerative group were patients whose language was evaluated because they had some language symptoms, but for whom language was not the primary complaint. Additionally, participants were required to be fluent in English, and to have sufficiently preserved language abilities to be able to complete the task.
A total of 58 individuals took part in the study. There were 16 patients with non-fluent/agrammatic PPA, 7 with semantic PPA, 8 with logopenic PPA, 6 with bvFTD, 9 with other neurodegenerative diseases, and 12 healthy age-matched controls. The mixed neurodegenerative group comprised patients who were diagnosed with Alzheimer’s disease (N = 4), corticobasal syndrome with suspected Alzheimer’s pathology (N = 3), mixed bvFTD and Alzheimer’s disease (N = 1), and mixed bvFTD with motor neuron disease (N = 1). Demographic information and neuropsychological data for each group is presented in Table 1. The three PPA variant groups did not differ from one another in terms of age, sex, handedness, education, MMSE, CDR, age of disease onset, or years from first symptom. Because patients who could not complete the task at all were not included, our samples were composed of mild to moderate patients, as reflected in the MMSE and CDR scores.
We used an elicited production task described by Goodglass et al. (1972) to determine which common syntactic constructions are spared or impaired in the three variants of PPA. The examiner began the task by informing the patient ‘I will begin a story and ask you to finish it in the most logical and most simple way possible’. A prompt was then read, such as the first item: ‘My friend comes in. I want him to sit down. So I say to him… what?’ The patient then typically responded ‘Sit down’ or similar. This item targets an intransitive imperative. The examiner repeated the prompt once if requested by the patient, but no other directions or prompts were given.
There were 14 targeted structures, each with 2 items, for a total of 28 items. However the last 3 structures (the last 6 items) rarely yielded the intended response, so we did not include those in our analysis. The complete list of prompts for the 11 structures analyzed, along with the intended responses and targeted structures are shown in Table 2.
Participants’ responses were recorded on a Sony camcorder and digitized with VirtualDub, except for one of the patients with non-fluent/agrammatic PPA who was mute and completed the task by writing. Responses were transcribed and coded by two raters (JDL and MB), both of whom were blind to patient diagnosis.
The raters coded: (1) whether the targeted syntactic construction was attempted; (2) if attempted, whether the targeted syntactic structure was produced correctly; (3) presence of any syntactic errors, e.g. missing determiners or inflections (in the target structure or in other parts of the response, regardless of whether the target structure was attempted); (4) presence of any semantic errors, defined as use of words or phrases that were inappropriate for the intended meaning or context (in the target structure or in other parts of the response, regardless of whether the target structure was attempted).
We scored a response as an attempt at the target syntactic structure if it contained all of the required elements for the particular item. For example, for item 10a to be scored as attempted, the response had to be a declarative sentence including a passive in the past tense. If the target syntactic structure was attempted, we recorded it as a correct attempt if it was free of syntactic errors. The response could still be recorded as a correct attempt if it contained semantic errors, phonological paraphasias or distortions.
We also counted the number of words produced by each subject in total. We excluded non-narrative words such as coordinating conjunctions and comments that did not directly address the prompt. We excluded filled pauses, i.e. words such as ‘ah’ or ‘um’. We also excluded false starts, which included partial words that were either followed by production of the word in completed form (e.g. ‘s- sofa’) or were abandoned without completion of the word (e.g. ‘He sm- well, he laughs’). Contractions such as ‘she’ll’ were counted as one word.
Statistical analysis was performed with R version 2.14.0 (http://www.r-project.org). The six groups were compared using ANOVAs for normally distributed variables, or the Kruskal-Wallis non-parametric test for measures with significant floor or ceiling effects. If the omnibus test was significant, we conducted planned contrasts between each patient group and controls, and between each pair of PPA variants. For ANOVAs, follow-up tests were corrected for multiple comparisons with the default single step procedure implemented in the R program glht, whereas non-parametric follow-up tests were Wilcoxon tests performed with wilcox.exact and corrected for multiple comparisons using p.adjust with Holm’s procedure. Performance on specific items was compared using χ2 tests, with Yates’ continuity correction where appropriate.
Structural T1-weighted images were acquired on 1.5T, 3T or 4T Siemens scanners as described previously (Wilson et al., 2010a, 2010b). The 12 normal controls were not included. Three patients’ scans were not of sufficient quality so were excluded (one patient was diagnosed with non-fluent/agrammatic PPA and two with mixed neurodegenerative disease). Therefore there were 43 participants included in this analysis. Images were registered to each other and to Montreal Neurological Institute (MNI) space using SPM5 (Ashburner & Friston, 2005) and DARTEL (Ashburner, 2007). Modulated gray matter and white matter probability maps were scaled by Jacobians, smoothed with a Gaussian kernel of 8mm full-width at half maximum, then summed together to obtain a map of brain parenchyma.
We correlated percent correct on target structures that were attempted with brain parenchyma probability maps. Covariates of age, sex, total intracranial volume, and scanner type were included in the analysis. The resulting statistical map was thresholded at voxelwise p < 0.01, then corrected for multiple comparisons based on cluster size using a permutation method. Specifically, 1000 randomly permuted maps were created, and the largest cluster in each was used to determine the null distribution of maximum cluster size. Permuted maps were masked to include only left hemisphere perisylvian language areas: the left inferior frontal gyrus, pars opercularis and triangularis, the Rolandic operculum, the superior temporal gyrus, and the supramarginal gyrus, based on an anatomical atlas (Tzourio-Mazoyer et al., 2002). This mask was used to increase statistical power, however it should be noted that in the non-permuted analysis of the real data, no regions outside the mask were significantly associated with the syntactic measure. Three additional analyses were also performed including measures of executive function and working memory (digit span backwards, modified trails, and calculation) as covariates.
The groups produced similar total numbers of words across their responses (F(5, 51) = 1.64, p = 0.17) (Fig. 1a, Table 3). Non-fluent/agrammatic PPA patients produced somewhat fewer words than controls, and semantic PPA patients produced somewhat more, but these differences were not significant.
The groups differed significantly in the frequency with which they attempted the targeted structures (F(5, 52) = 3.05, p = 0.017) (Fig. 1b, Table 3). Follow-up comparisons revealed that all neurodegenerative groups except for the bvFTD group attempted targeted structures less frequently than healthy controls (non-fluent/agrammatic: t = 3.39, p = 0.0094; semantic: t = 2.73, p = 0.055; logopenic: t = 2.87, p = 0.039; mixed: t = 2.76; p = 0.051). The three PPA variants did not differ from one another (all t ≤ 0.035) in how often they attempted the targeted structures.
The 22 items differed in the frequency with which participants attempted the targeted structure (χ2(21) = 186.38, p < 0.001) (Fig. 2). While most targeted structures were obtained more than half the time, a few items were particularly unsuccessful: the two items designed to elicit ditransitives (5a and 5b), one of the items intended to elicit an embedded clause (9a), and one intended to elicit a comparative structure (11a).
For each item, we used a χ2 test to determine whether the three PPA and two neurodegenerative control groups differed in the frequency with which they attempted the intended structure (we omitted controls because we have already shown that they attempt the targeted structures more frequently in general). We set an alpha criterion of p < 0.01 to informally correct for multiple comparisons. No items met this threshold. At an uncorrected threshold of p < 0.05, three items showed different distributions: 3a (p = 0.050), 9a (p = 0.030) and 11b (p = 0.020).
The groups differed significantly in the frequency with which they produced targeted structures correctly, when they did attempt them (Kruskal-Wallis χ2 = 29.41, df = 5, p < 0.001) (Fig. 1c, Table 3). Patients with non-fluent/agrammatic PPA were less accurate than controls (p < 0.001), logopenic PPA patients were marginally less accurate (p = 0.068), and mixed neurodegenerative patients were less accurate (p = 0.028). Non-fluent/agrammatic PPA patients were less accurate than semantic (p = 0.015) or logopenic (p = 0.0067) patients, but semantic and logopenic patients did not differ from one another (p = 0.79).
The 22 items differed in the frequency with which participants produced attempted targeted structures correctly (χ2(21) = 193.76, p < 0.001) (Fig. 2). The most challenging structures were the declarative transitive with 3sg agreement (3a), and the two embedded clauses (9a and 9b).
We used sets of chi square tests to determine which items patients with each PPA variant had most difficulty with, using an alpha criterion of p < 0.01 to informally correct for multiple comparisons. Non-fluent patients performed worse than controls on items 3a, 9a and 9b. Inspection of Fig. 2 suggests that 3sg present tense agreement and embeddings posed the most problems for non-fluent patients, followed by wh-questions. There were no items on which semantic or logopenic PPA patients performed significantly worse than controls, though it should be noted that for both groups most errors occurred on item 9b, an embedded clause.
Examples of targeted structures that were attempted but resulted in syntactic errors are shown in Table 4.
Using voxel-based morphometry, we found that the only region where atrophy was significantly predictive of reduced accuracy on targeted structures was the left posterior inferior frontal gyrus, pars opercularis (center of mass: MNI coordinates –53, 12, 13; maximum t = 3.47; cluster extent = 3192 mm3; corrected p = 0.025; Fig. 3).
The same region was found to predict reduced accuracy when covariates of executive function and/or working memory were included in the model, however its volume was large enough to survive correction for multiple comparisons only for the calculation covariate (extent = 3648 mm3); the cluster extent for syntactic accuracy was reduced to 456 mm3 when digit span backwards was included as a covariate, and 512 mm3 when modified trails was included as a covariate.
We counted syntactic errors irrespective of whether or not the targeted structure was attempted, and divided by the total number of words each participant produced. The groups differed significantly in syntactic errors per word (Kruskal-Wallis χ2 = 34.26, df = 5, p < 0.001) (Table 3). Non-fluent/agrammatic PPA patients produced the most errors.
We also counted semantic errors. The groups differed marginally in semantic errors per word (Kruskal-Wallis χ2 = 10.91, df = 5, p = 0.053) (Table 3). All patient groups produced some semantic errors, but controls did not produce any.
Using an elicited production task, we found that all PPA variants, as well as other neurodegenerative patients, produced targeted syntactic structures less frequently than controls. However, the three PPA variants did not differ from one another in the frequency with which they attempted targeted syntactic constructions. When targeted structures were attempted, patients with the non-fluent variant of PPA made more syntactic errors compared to controls and compared to the other PPA variants. Reduced accuracy on production of targeted syntactic structures was associated with atrophy of the left posterior inferior frontal gyrus.
The results of this study are largely consistent with studies that have investigated syntactic production in PPA using quantitative analysis of connected speech (Ash et al., 2006, 2009; Bird et al., 2000; Graham et al., 2004; Gunawardena et al., 2010; Knibb et al., 2009; Meteyard and Patterson, 2009; Orange et al., 1998; Patterson et al., 2006; Patterson and MacDonald, 2006; Rogers and Alarcon, 1998; Thompson et al., 1997, 2012a; Wilson et al., 2010b) and those using constrained production tasks (Thompson et al., 2012b; Weintraub et al., 2009). This literature has shown that patients with non-fluent/agrammatic PPA are impaired in syntactic production, whereas only moderate syntactic deficits are found in semantic or logopenic PPA (Metayard & Patterson, 2009; Thompson et al., 2012a; Wilson et al., 2010b). Previous studies have shown that atrophy of left inferior frontal cortex is associated with deficits in the production of syntax (Gunawardena et al., 2010; Rogalski et al., 2011; Wilson et al., 2010b, 2011). The left inferior frontal cortex is also functionally abnormal in non-fluent/agrammatic PPA: this region is not modulated by syntactic complexity in these patients as it is in controls (Wilson et al., 2010a). The importance of left inferior frontal cortex for syntactic processing may be associated in part with its role in executive function and/or working memory, though some studies have suggested dissociations between frontal regions important for syntactic and working memory functions (Amici et al., 2007; Makuuchi et al., 2009). In our study, the extent of the region associated with syntactic production was reduced when measures of executive function and/or working memory were included as covariates, especially the widely used measures of digit span backwards and (modified) trails.
The specific syntactic constructions that we investigated differed considerably in the extent to which they posed difficulties to patients with non-fluent/agrammatic PPA. Some structures were produced accurately by most patients: intransitive and transitive imperatives, yes/no questions, declarative past tense passives, and comparatives. Of these, the accurate production of passives is most surprising, since previous studies using elicitation tasks have shown poor performance on passive constructions in PPA patients in general (Weintraub et al., 2009) and in non-fluent/agrammatic patients in particular (Thompson et al., 2012b). It is noteworthy that the passives elicited in the present study—‘(the man) was hit (by the train)’ and ‘(she) was bitten (by the dog)’—are not readily reversible, unlike the passives elicited in these prior studies. Furthermore, participants were not required to produce the subject (since it was already part of the prompt), nor were they required to produce the ‘by’ phrase, and both of the verbs used have past participles that are homophonous with the past tense (optionally in the case of bite).
The structures that proved most difficult were the 3sg present tense marker, and embedded clauses. These results are consistent with Thompson et al. (2012b), who found that non-fluent/agrammatic patients are impaired in using 3sg present agreement (61% correct in an elicitation task), and in producing relative clauses. The embedded clauses in the present study were ‘small clauses’ with infinitive verbs, and these proved difficult not only for non-fluent/agrammatic patients, but for other PPA variants and even other neurodegenerative patients.
In sum, we found that non-fluent/agrammatic patients attempt targeted syntactic structures just as frequently as other PPA variants, but make many more syntactic errors. Constructions differ greatly in the extent to which they are prone to errors, with complex embedded structures and verbal inflection proving the most vulnerable. This information could be useful clinically, since elicitation of just these challenging structures may provide a very quick initial indication as to whether a patient may be agrammatic. However intended structures are not always attempted, and not every agrammatic patient fails on every challenging structure, so it is still important to follow up with a careful assessment of connected speech to confirm the presence or absence of agrammatism.
Supported by NIH (NIDCD R03 DC010878 to SMW, NINDS R01 NS050915 to MLGT, NIA P50 AG03006, NIA P01 AG019724). We thank our colleagues, patients, caregivers and volunteers for their contributions to our research.
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.