Neural mechanisms of sentence production: a volumetric study of primary progressive aphasia

Abstract

Studies on the neural bases of sentence production have yielded mixed results, partly due to differences in tasks and participant types. In this study, 101 individuals with primary progressive aphasia (PPA) were evaluated using a test that required spoken production following an auditory prime (Northwestern Assessment of Verbs and Sentences—Sentence Production Priming Test, NAVS-SPPT), and one that required building a sentence by ordering word cards (Northwestern Anagram Test, NAT).

Voxel-Based Morphometry revealed that gray matter (GM) volume in left inferior/middle frontal gyri (L IFG/MFG) was associated with sentence production accuracy on both tasks, more so for complex sentences, whereas, GM volume in left posterior temporal regions was exclusively associated with NAVS-SPPT performance and predicted by performance on a Digit Span Forward (DSF) task. Verb retrieval deficits partly mediated the relationship between L IFG/MFG and performance on the NAVS-SPPT.

These findings underscore the importance of L IFG/MFG for sentence production and suggest that this relationship is partly accounted for by verb retrieval deficits, but not phonological loop integrity. In contrast, it is possible that the posterior temporal cortex is associated with auditory short-term memory ability, to the extent that DSF performance is a valid measure of this in aphasia.

Introduction

Volumetric magnetic resonance imaging (MRI) studies in individuals with brain damage can help determine which regions of the brain are associated with certain linguistic processes, by establishing associations between lesioned (or atrophied) gray matter (GM) and performance on language measures. Two recent studies (den Ouden et al. 2019; Lukic et al. 2021) used Voxel-based Lesion-Symptom Mapping (VLSM; Bates et al. 2003) in individuals with aphasia resulting from stroke and reported associations between sentence production accuracy and lesions to the posterior superior temporal gyrus (pSTG), with one study finding additional associations with anterior temporal (Lukic et al. 2021) and one with inferior parietal (den Ouden et al. 2019) regions. Despite the longstanding tradition associating lesions to Broca’s area (which includes the left pars triangularis and opercularis of the inferior frontal gyrus, IFG) to agrammatic production (e.g. Caramazza and Hillis 1989; Grodzinsky 2000; Miceli et al. 1983), these two studies did not find any association between damage to the IFG and sentence production.

Research in primary progressive aphasia (PPA) has, however, yielded different results. PPA is a neurodegenerative disorder that affects neural tissue associated with language processing, sparing tissue involved in other cognitive processes (e.g. memory, executive functions and comportment) for at least 2 years after onset (Mesulam 2001, 2003). By employing voxel-based morphometry (VBM) or cortical thickness measurements, studies in PPA have reported associations between sentence production performance and atrophy in all the left IFG parcellations, i.e. pars orbitalis (IFGorb, DeLeon et al. 2012; Rogalski et al. 2011a), triangularis (IFGtri; Canu et al. 2019), and opercularis (IFGoper, DeLeon et al. 2012; Rogalski et al. 2011a). They have also found associations with atrophy in the left middle frontal (MFG, Mesulam et al. 2021b) and superior frontal gyri (SFG, Mesulam et al. 2021b; Rogalski et al. 2011a). In addition, associations between atrophy patterns and selected measures of narrative production (at the sentence level) have been found in the left IFG (Wilson et al. 2010a; Ash et al. 2013; Mesulam et al. 2019; Mesulam et al. 2021b), the supplementary motor area (SMA; Wilson et al. 2010a) or the anterior temporal lobe (Ash et al. 2013).

Together, these studies indicate that sentence production relies on the integrity of inferior/middle frontal, posterior temporal, and inferior parietal regions in the left hemisphere, in line with the assumptions made by current neurocognitive models of sentence production (e.g. Thompson and Meltzer-Asscher 2014; Matchin and Hickok 2020). However, these studies also underscore some major inconsistencies between the literature on stroke and that on PPA. On the one hand, the literature on PPA has highlighted the primary role of left IFG/MFG regions in supporting sentence production and downplayed the role of left pSTG. On the other hand, the literature on stroke aphasia has painted the opposite picture, with only two published studies both showing associations between deficits in sentence production and lesions to the left anterior and posterior temporal, but not left inferior frontal, regions.

Factors affecting sentence production

Among the factors at the root of these inconsistencies, differences in the measure(s) used to quantify sentence production play a major role. For example, some studies employed paradigms that require participants to produce a sentence either based on an auditorily presented prime sentence (e.g. Den Ouden et al. 2019; Lukic et al. 2021) or by completing a short story following a prompt (DeLeon et al. 2012). By requiring overt production, these tasks engage motor planning and articulation, thereby being affected by fluency and/or motor speech deficits, such as apraxia of speech, which may (but do not have to) co-occur with syntactic deficits (i.e. sentence structure building) (see e.g. Thompson et al. 2012a; Josephs et al. 2013). Measures derived from spontaneous speech, such as the percentage of grammatically correct sentences (Mesulam et al. 2021b) or the number of sentence embeddings (Wilson et al. 2010a) suffer from the same limitation. Conversely, some studies have employed an anagram task (the Northwestern Anagram Test, NAT, Thompson et al. 2012b; Weintraub et al. 2009) that bypasses overt production by asking participants to order word cards and build a sentence to describe a given picture (Rogalski et al. 2011a; Canu et al. 2019), thereby testing syntactic abilities even when reduced fluency and/or severe apraxia of speech are present.

In addition to fluency and motor speech deficits, accuracy on sentence production tasks may also be affected by difficulties in lexical retrieval. As a result, impaired performance on measures derived from spontaneous speech (Wilson et al. 2010a; Mesulam et al. 2021b) or sentence completion (DeLeon et al. 2012) could reflect word-level deficits, thereby not exclusively targeting the ability to build sentence structures. In particular, verb production deficits, which are common in aphasia, preclude production of grammatically correct sentences. An attempt to account for verb retrieval deficits was made in the study by Lukic et al. (2021), where performance on verb naming was regressed out from VLSM analyses that investigated the relationship between sentence production and lesioned tissue in stroke aphasia. Because performance on verb, but not sentence production, was associated with lesioned tissue in parts of the IFG, the authors argued that lesions in the IFG resulted in failure to retrieve verbs, which in turn prevented sentence structures to be built (given that verbs guide sentence structure building, see Friederici 2012; Thompson and Meltzer-Asscher 2014). To our knowledge, no study in PPA has attempted to control for verb retrieval deficits.

Furthermore, there is evidence that auditory short-term or working memory capacity can affect sentence production (Martin 2003; Slevc 2011; Sung et al. 2018). This is relevant to the findings of studies that employed overt sentence production following an auditory prime (Den Ouden et al. 2019; Lukic et al. 2021), as individuals with impaired auditory short-term or working memory are unable to benefit from the auditory prime to determine which sentence structure they are required to produce. As a result, lower scores on such task may reflect difficulties in using the auditory prime as a model for sentence production, in addition to, or in the absence of, impaired sentence structure building. The effects of auditory short-term and working memory on the relationship between atrophy (or vascular lesions) and sentence production remains incompletely understood (see Peelle et al. 2008, for effects of working memory in comprehension), in part, because commonly used measures of auditory short-term or working memory (i.e. Digit Span tasks) require language production, which is often impaired in people with aphasia.

Another factor that plays a major role in the investigation of brain areas associated with sentence production is the syntactic complexity of the sentence. In our linguistic framework of reference (i.e. Government and Binding Theory, Chomsky 1981), canonical sentences follow native speakers’ preferred word order (i.e. for English, Agent-Verb-Theme). For example, in the active sentence the boy was kissing the girl, the boy is the Agent (i.e. the person carrying out the action) and the girl is the Theme (i.e. person undergoing the action). Conversely, in the passive sentence the girl was kissed by the boy, the Theme comes before the Agent. Within Chomsky’s framework, non-canonical sentences are derived from canonical forms through syntactic movement: namely, the passive sentence the girl was kissed by the boy is derived from movement of the element the girl from the post-verbal, object, position to the position occupied by the grammatical subject. As a result, a phonologically silent element (trace) is left in the original position and co-referenced (i.e. linked) to the final position occupied by the moved element, to ensure correct assignment of thematic roles (e.g. Agent, Theme).

The effect of syntactic complexity has been investigated in both stroke aphasia and PPA, yet, findings are inconsistent. In the two studies on sentence production in stroke (den Ouden et al. 2019; Lukic et al. 2021), no association was found between production of non-canonical sentences and lesions to the left IFG/MFG; rather, Lukic et al. (2021) reported associations with the left pSTG. Conversely, studies in PPA have mostly reported associations between production of non-canonical sentences and the left IFG and SFG/SMA atrophy (Wilson et al. 2010a; Canu et al. 2019; Mesulam et al. 2021b). Notably, PPA studies either looked at non-canonical sentences alone (Mesulam et al. 2021b) or at canonical and non-canonical sentences separately (Canu et al. 2019), leaving the question of the extent to which syntactic complexity affects sentence production unanswered.

Finally, some studies have suggested that PPA subtype may play a role in determining which areas of the brain are involved in language functions (Peelle et al. 2008). The current consensus criteria (Gorno-Tempini et al. 2011) identify three PPA subtypes: a nonfluent/agrammatic, a semantic, and a logopenic subtype. Among these, sentence production deficits are most frequently observed in the nonfluent/agrammatic subtype, which is typically associated with syndrome-specific atrophy patterns in the IFG, extending to the MFG/SFG as well as the temporo-parietal junction (TPJ; Gorno-Tempini et al. 2004, 2011; Mesulam et al. 2009; Mesulam et al. 2021a; Rogalski et al. 2011a; Wilson et al. 2012). Conversely, sentence production tends to be better preserved in both the semantic and logopenic subtypes (Gorno-Tempini et al. 2011; Mesulam et al. 2021b), which are typically associated with atrophy in the left anterior temporal lobe (ATL) or the TPJ, respectively (Gorno-Tempini et al. 2004, 2011; Rogalski et al. 2011b; Mesulam et al. 2014). Accordingly, associations between sentence production accuracy and atrophy to certain brain regions are expected to differ between PPA subtypes, as reported by Peelle et al. (2008) for sentence comprehension.

Aim of the study

Overall, research on the neural basis of sentence production has been limited to a few studies, which given the coexisting methodological issues, have provided inconsistent results. The present study aims to address the issues raised above by investigating the effects of (i) fluency/motor speech, verb retrieval, and performance on Digit Span tasks, (ii) syntactic complexity, and (iii) PPA subtype on the relationship between atrophy and sentence production performance in PPA.

Materials and methods

Participants

One-hundred-one individuals with PPA and 33 healthy participants were recruited from our PPA Research Program, a large longitudinal study that is funded by the National Institute on Aging. Participants were evaluated at the Mesulam Cognitive Neurology and Alzheimer’s Disease Center in Chicago, IL. The study was approved by the Northwestern University Institutional Review Board and conducted in compliance with the Declaration of Helsinki. All participants provided written informed consent prior to entering the study.

Diagnosis of PPA was performed by an expert neurologist (M.-M.M.) based on the evidence of isolated and progressive language deficits, in line with previously published criteria (Mesulam 2003; Gorno-Tempini et al. 2011). Classification of participants’ language profiles into the three major subtypes (i.e. agrammatic or PPA-G, logopenic or PPA-L, and semantic or PPA-S) was established following neuropsychological and language assessment, as well as qualitative evaluation of spontaneous speech, and based on a slight modification (Mesulam et al. 2014; Mesulam et al. 2021a) of the original consensus criteria (Gorno-Tempini et al. 2011). Namely, a diagnosis of PPA-G was made for individuals who showed evidence of agrammatism in production and preserved single-word comprehension; PPA-S was defined by evidence of single-word comprehension impairment, and PPA-L was defined by evidence of word-finding difficulties in the absence of agrammatism and single-word comprehension deficits (see Mesulam et al. 2021a, for details). Individuals not meeting criteria for any of the three major subtypes, including those presenting with mixed PPA (i.e. a proposed fourth subtype of PPA; see Mesulam et al. 2012; Mesulam et al. 2021a), were excluded from the study. Patients with a history of neurological (other than PPA) and/or psychiatric disorders were also excluded. Of the 101 participants included in the study, 49 were classified as PPA-G, 29 as PPA-L and 23 as PPA-S. All participants were right-handed.

Demographics for all participant groups are reported in Table 1. Groups did not differ in age (Kruskal–Wallis χ² = 4.176, P = 0.243) or gender (χ² = 2.719, P = 0.437). As for education, PPA-L were more educated than individuals with PPA-G (t = 3.556, P <. 001), but all other comparisons were not significant. PPA groups did not differ in the duration of the disease (as measured by the self-reported number of years since symptoms were first noticed; Kruskal–Wallis χ² = 0.061, P = 0.970).

Table 1.

Demographics for all participant groups (M ± SD).

Group	Gender	Age	Education	Symptom Duration
PPA-G	21 M, 28F	65.5 ± 6.5	15.4 ± 2.5	3.7 ± 1.9
PPA-L	16 M, 13F	66.5 ± 5.7	17.0 ± 1.5	4.0 ± 2.5
PPA-S	13 M, 10F	63.3 ± 5.6	16.4 ± 2.7	3.6 ± 1.5
Healthy	13 M, 20F	64.5 ± 6.3	16.4 ± 2.5

Language and neuropsychological measures

Individuals with PPA were administered several standardized language tests, including: the Western Aphasia Battery Revised (WAB-R, Kertesz 2007), the Boston Naming Test (BNT, Kaplan et al. 1983), the Peabody Picture Vocabulary Test (PPVT, Dunn and Dunn 1997), the Northwestern Assessment of Verbs and Sentences (NAVS, Thompson 2012), and the Northwestern Anagram Test (NAT; Thompson et al. 2012b). Individuals with PPA were also tested on several neuropsychological measures, including the Mini-Mental State Examination (MMSE, Folstein et al. 1983), the Digit-Span Task Forward (DSF) and the ratio between the Digit Span Forward and Backward (henceforth DSF/DSB ratio), both derived from the Wechsler Memory Scale—Revised¹ (Wechsler 1997), and the Trail Making Test (TMT; Tombaugh 2004). The DSF/DSB ratio was computed by dividing the difference between total scores on the Digit Span Forward and Backward by their sum (i.e. (DSF−DSB)/(DSF + DSB)), and aimed to provide a measure of how much harder the DSB task was compared to the DSF task. The DSF/DSB ratio can therefore be considered as a proxy of auditory working memory.

The Northwestern Assessment of Verbs and Sentences

The Northwestern Assessment of Verbs and Sentences (NAVS) (Cho-Reyes and Thompson 2012; Thompson 2012) includes five subtests, two of which are relevant to the present investigation. The Verb Naming Test (VNT, Fig. 1a) evaluates verb production by requiring participants to name the action displayed in the picture. Sentence production is assessed by the Sentence Production Priming Test (SPPT, Fig. 1b), in which participants are shown two reversible pictures depicting the same action with the same participants. After listening to the examiner produce a sentence that describes the picture on the left side, participants are asked to make a similar sentence for the picture on the right side. The test assesses production of 30 sentences, 15 canonical (active, subject-relative sentences, subject Wh-questions), and 15 non-canonical (passive, object-relative sentences, object Wh-questions).

Fig. 1.

Example of test items extracted from the NAVS VNT (a) and SPPT (b) assessing production of the verb wash (a) and of the passive sentence the woman was kissed by the man (b), respectively.

The Northwestern Anagram Test (NAT)

The NAT (Thompson et al. 2012b; Weintraub et al. 2009) evaluates production of six sentence types, using the same sentence and picture stimuli as in the SPPT. Contrary to the NAVS SPPT, the NAT does not require any overt production; rather, participants are presented with pictures as the one shown in Fig. 2, where participants and action are labeled, as well as with a set of cards that they have to order correctly to describe the picture using the intended sentence structure. As shown in the example, the examiner provides the beginning of the sentence (i.e. The man) and participants are required to complete the sentence by ordering the word cards. The test assesses production of 30 sentences, 15 canonical (active, subject-cleft sentences, subject Wh-questions), and 15 non-canonical (passive, object-cleft sentences, and object Wh-questions).

Fig. 2.

Example of test item included in the NAT, assessing production of the sentence the man is kissed by the woman.

Structural imaging

Data acquisition

For each participant, structural T1-weighted 3D images were acquired using MP-RAGE sequences on a 3T Trio or Prisma scanner (slice thickness: 1 mm, TR = 2300 ms, TE = 2.91 ms, flip angle = 9°, FOV = 256 × 256mm). The distribution of individuals scanned on 3 T Trio or Prisma scanners did not differ among the 4 groups (i.e. healthy individuals, PPA-G, PPA-L, and PPA-S, χ² = 1.338, P = 0.720). Images were manually re-oriented to the AC-PC line using SPM12 (https://www.fil.ion.ucl.ac.uk/spm/software/spm12/) and then pre-processed using the Computational Anatomy Toolbox 12 toolbox (CAT12, Gaser et al. 2022; https://neuro-jena.github.io/cat/) running in MatLab 2022b.

Pre-processing in CAT12 takes place in two steps: an initial pre-processing that includes denoising, bias correction and affine registration using SPM unified segmentation (Ashburner and Friston 2005), and a refined voxel-based pre-processing including—among others—skull stripping, local intensity transformation of all tissue classes, adaptive maximum a posteriori (AMAP) segmentation, co-registration, and normalization to the a common reference space using Geodesic Shooting (Ashburner and Friston 2011). Modulated, normalized (1.5 × 1.5 × 1.5 mm) GM volume images were ultimately smoothed using an 8 mm kernel. For a more detailed description of the pre-processing steps, the reader is referred to Lukic et al. (2022), and to the CAT12 Manual (https://neuro-jena.github.io/cat12-help/). GM volume images were checked for quality in CAT12, which computes an image quality rating (IQR), i.e. a weighted measure that combines information about noise contrast ratio, inhomogeneity contrast ratio, and resolution. Participants were included in the study only if their IQR values were equal to or greater than 80 (i.e. good or excellent).

Data analyses

Behavioral data

Regression analyses were run in R 4.2.2 (R Core Team 2022, https://www.R-project.org/) to compare performance on language and neuropsychological tests between PPA subtypes. Analyses were run using the (g)lm function included in the stats package (https://stat.ethz.ch/R-manual/R-devel/library/stats/html/stats-package.html): models were fitted with participant-level average accuracy on a given test as the dependent variable, and with Subtype entered as the only predictor. In these types of linear models, statistical significance of predictors (P value) reflects the significance of the model when compared to a null model (i.e. of a model in which the value of the dependent variable is equal to its mean). Post-hoc comparisons were carried out using the glht function included in the package multcomp (Hothorn et al. 2016) and Tukey’s test for pairwise mean comparisons, then corrected by applying False Discovery Rate (FDR) correction (Benjamini and Hochberg 1995). Effect sizes were expressed as the multiple R-squared values provided by the summary function, which is used to call the results of the linear models. R-squared values provide the fraction of the variance that is explained by the model.

Neuroimaging analyses

Neuroimaging data statistical analyses were performed using VBM (Ashburner and Friston 2000) in CAT12 (http://dbm.neuro.uni-jena.de/cat) and Statistical Parametric Mapping (SPM12), both running in Matlab 2022b.

Atrophy maps for the entire group of 101 participants with PPA, and for each of the three PPA subtypes separately, were calculated by performing two-sample t-tests and comparing GM volume images of the PPA group(s) to those of 33 healthy participants. Atrophy maps were thresholded in SPM12 using a voxelwise Family-Wise Error (FWE) corrected P < 0.05, which sets the probability of encountering one or more false positive t-tests (i.e. one or more false positive voxels in the entire search volume) to 5%, based on Random Field Theory (see SPM12 Manual for details, https://www.fil.ion.ucl.ac.uk/spm/doc/spm12_manual.pdf). Furthermore, an arbitrary cluster size threshold of k ≥ 100 was applied to the thresholded maps, with the aim to restrict results to clusters of meaningful size.

VBM regression analyses were used to assess the relationship between performance on sentence production tasks (i.e. the NAVS SPPT or the NAT) and GM volume across participants with PPA. For these analyses, models were specified through the Basic Models function in CAT12, using a full factorial design in which normalized, modulated and smoothed GM volume images from all participants with PPA were used as input, and and either the NAVS SPPT or the NAT proportion total correct responses were used as the main predictor. Participant-related variables (age, education, years from symptoms onset, and scanner) were entered as covariates of no interest, in addition to total intracranial volume (which accounts for differences in brain volume). Models were estimated using the “classical” (i.e. Restricted Maximum Likelihood) method provided by SPM12. Additional VBM analyses reported in the Supplementary Materials employed the same methodology to investigate the relationship between GM volume and verb retrieval (using the VNT total correct responses) or auditory short-term memory (using the Digit Span Forward, DSF total score). All VBM regression statistical maps were thresholded in SPM12 using a voxelwise uncorrected threshold of P < 0.001, and a cluster-level FWE correction (P <. 05). The FWE correction at the cluster level indicates the cluster size that is necessary to ensure that the probability of encountering a false positive cluster per search volume is less than 5%. We note that although this method of statistical thresholding is more liberal than applying a voxelwise FWE (P < 0.05) correction, it is a very widely used method in functional MRI analysis (Woo et al. 2014).

Significant clusters derived from each of the whole-brain VBM regression analyses on sentence production (i.e. the clusters that passed the statistical significance of voxelwise uncorrected P < 0.001 and cluster-level FWE (P < 0.05) correction) were extracted from SPM12, binarized and overlaid onto each PPA participant’s (thresholded at 0.2) GM volume maps, to compute the number of GM voxels in each participant’s map that overlapped with each of the significant clusters. Then, the proportion atrophy for each of the significant clusters was computed for each participant and entered as dependent variable in simple multiple linear regressions, using the same methodology as described in the “Behavioral Data” section. This was done to evaluate the effect of (i) verb retrieval, DSF and DSF/DSB ratio, (ii) syntactic complexity, and (iii) PPA subtype on the relationship between atrophy and sentence production. For the purposes of (i), the proportion total correct responses on the VNT, the total number of correct responses on the DSF, and the DSF/DSB ratio were entered as the only predictors in separate simple linear regression models. For goal (ii), the proportion correct responses on the non-canonical (NC) and canonical (C) items included in the SPPT (or NAT, depending on the cluster) were entered as the only predictors in separate simple regression analyses. For goal (iii), a model including the proportion correct responses on the SPPT (or NAT, depending on the cluster) was first fitted, and PPA subtype was entered in the model first as a main effect and then in interaction with the SPPT (or NAT) score, to assess whether the relationship between atrophy and sentence production was different between subtypes. The contribution of PPA subtype to the model fit was evaluated by comparing models with and without PPA subtype or with and without the PPA subtype*SPPT (or NAT) interaction, using the anova function included in the stats package.

Results

Language and neuropsychological scores

Scores on standardized language and neuropsychological tests are listed in Table 2, for all individuals with PPA and separately for each subtype. Results of simple linear regressions are reported in detail in the (Supplementary Materials Table S1). Briefly, individuals with PPA-L showed higher WAB-AQ (i.e. lower aphasia severity) than both PPA-G and PPA-S. Noun production and comprehension (as indexed by the BNT and the PPVT, respectively) were significantly more impaired in individuals with PPA-S than in the other two subtypes, and the same was true for verb production. Verb production was also significantly more impaired in PPA-G than in PPA-L. Production of canonical and non-canonical sentences on the NAT and the NAVS SPPT was significantly lower in PPA-G than in both PPA-L and PPA-S subtypes, whereas, no differences were observed between PPA-L and PPA-S. Comprehension of canonical and non-canonical sentences on the NAVS SCT was also more impaired in PPA-G than in the other two subtypes, with no additional differences between PPA-L and PPA-S.

Table 2.

Results of the standardized language and neuropsychological tests by PPA subtype, and for the entire group of individuals with PPA. For all language measures, raw scores are reported, and maximum scores are indicated in parentheses. For neuropsychological measures, raw scores are reported for the MMSE, the DSF and the DSB, and time in seconds for the TMT.

Measure	PPA-G		PPA-L		PPA-S			All PPA
	Mean	SD	Mean	SD	Mean	SD		Mean	SD
Language Measures
WAB-AQ (100)	82.8	7.4	91.2	5.2	84.8	9.0		85.7	8.1
BNT (60)	45.0	10.9	49.6	10.5	10.6	7.5		38.5	18.3
PPVT (36)	33.0	3.0	34.1	2.1	20.5	7.4		30.4	6.9
NAT C (15)	12.1	4.0	14.5	0.7	14.8	0.7		13.4	3.1
NAT NC (15)	6.7	3.1	12.1	2.3	13.3	2.5		9.7	4.1
NAT Total (30)	18.8	6.5	26.5	2.6	28	3.0		23.1	6.5
NAVS VNT (22)	18.8	3.2	20.9	1.0	16	4.6		18.8	3.6
NAVS VCT (22)	21.7	0.8	22.0	0.0	21.4	1.1		21.7	0.8
NAVS SPPT C (15)	11.2	4.5	14.2	1.9	14.8	0.7		12.9	3.6
NAVS SPPT NC (15)	7.4	5.3	12.2	2.7	14.3	1.6		10.4	4.9
NAVS SPPT Total (30)	18.6	9.1	26.5	4.1	29.0	2.2		23.4	8.1
NAVS SCT C (15)	13.5	2.3	14.7	0.9	14.8	1.0		14.2	1.8
NAVS SCT NC (15)	12.3	2.3	14.1	1.5	14.3	1.5		13.3	2.2
NAVS SCT Total (30)	25.8	4.1	28.8	2.3	29.1	2.5		27.4	3.6
Neuropsychological Measures
MMSE (30)	25.1	3.4	26.0	2.7	24.9	5.0		25.3	3.6
DSF (14)	4.6	2.2	5.4	2.1	8.67	2.1		5.7	2.7
DSB (14)	3.6	1.4	4.8	1.5	6.6	1.8		4.6	1.9
TMT—PartA	44.7	21.2	38.9	16.2	32.4	10.1		40.4	18.4
TMT—PartB	161.5	76.7	130.2	70.6	80.7	30.6		135.0	74.0

Note. BNT = Boston Naming Test; C = Canonical; DSF = Digit Span Forward; DSB = Digit Span Backward; MMSE = Mini-Mental State Examination; NAT = Northwestern Anagram Test; NAVS = Northwestern Assessment of Verbs and Sentences; NC = Non-canonical; SCT = Sentence Comprehension Test; SPPT = Sentence Production Priming Test; TMT = Trail Making Test; VCT = Verb Comprehension Test; VNT = Verb Naming Test.

While no differences between subtypes were observed on the MMSE, PPA-S performed better than both PPA-G and PPA-L on the Digit Span Task and on Part B of the TMT (Table S1). PPA-G and PPA-L subtypes exhibited similar performances on all neuropsychological measures, with the exception of Digit Span Backward, where participants with PPA-G showed lower scores than individuals with PPA-L.

Whole-brain VBM analyses

Atrophy maps (two-sample t-tests)

Figure 3a shows regions of atrophy in the entire group of individuals with PPA, as compared to a group of healthy individuals. Across all PPA subtypes, atrophy encompassed the entire left temporal lobe, and extended rostrally to encompass the insula and portions of the left IFG, MFG, SFG and frontal pole, caudally to include portions of the inferior and superior divisions of the lateral occipital cortex (iLOC, sLOC) and dorsally to incorporate the angular (AG) and supramarginal gyri (SMG) in the left hemisphere. Significant clusters of atrophy were also found in the right hemisphere, albeit restricted to the temporal lobe and to the frontal pole. Subcortically, atrophy encompassed the amygdala and the anterior parahippocampal gyrus bilaterally, whereas atrophy in other subcortical structures, namely the putamen, caudate, nucleus accumbens, thalamus and hippocampus, was only found in the left hemisphere.

Fig. 3.

Atrophy maps for all individuals with PPA combined (a) and for each of the three PPA subtypes: PPA-G (b), PPA-L (c), and PPA-S (d). Areas in color indicate regions of significant atrophy (i.e. regions in which gray matter volume in PPA was less than in healthy participants) at the voxelwise FWE-corrected threshold of P < 0.05 (k ≥ 100). The color bar reflects the range of t values (from 0 to 15). The t value corresponding to the statistical significance threshold of choice is 4.760. Lighter colors (e.g. light yellow and white) reflect larger t values, i.e. regions of greater atrophy, whereas darker colors (e.g. dark red) reflect smaller t values, i.e. regions of lesser atrophy.

In all three subtypes (Fig. 3b-d), atrophy affected the anterior and posterior portions of the STG and MTG, as well as the fusiform gyrus and the thalamus, in the left hemisphere. In PPA-G (Fig. 4b), atrophy was prominent in the left frontal lobe, where it affected the IFG, as well as portions of the MFG and SFG, and the precentral gyrus. Medially, significant regions of atrophy were observed in the anterior cingulate and in the paracingulate gyri. In the parietal lobe, parts of the AG and posterior SMG were also significantly atrophied. Furthermore, areas of subcortical atrophy were noted in the putamen, caudate, nucleus accumbens, thalamus, amygdala, and hippocampus, all in the left hemisphere. Smaller regions of atrophy were found in the right hemisphere, namely within the posterior MTG and STG, and within the frontal pole. Atrophy in PPA-L (Fig. 3c) was considerably less widespread than in PPA-G, as it reached significance only in left middle/superior temporal, and inferior parietal regions, as well as in a small portion of the left frontal pole and of the right posterior STG. Subcortically, PPA-L exhibited significant atrophy in the left thalamus and amygdala, whereas other structures, including the left putamen and caudate, did not show any significant reduction in GM volume. Finally, PPA-S showed the typical pattern of atrophy centered around the temporal poles, bilaterally although more so in the left than in the right hemisphere. This group exhibited extensive subcortical atrophy in the left hemisphere, including the left putamen, caudate, nucleus accumbens, thalamus, and hippocampus, and was the only group in which the amygdala was atrophied in both left and right hemispheres.

Fig. 4.

Areas of association between gray matter (GM) volume and sentence production accuracy measured using the total score on the NAVS SPPT (a) and the total score on the NAT (b). Both analyses included age, education, duration of disease (i.e. number of years since reported onset of symptoms), scanner, and total intracranial volume, as covariates. Statistical maps were thresholded at an uncorrected voxelwise P < 0.001 and cluster-level FWE correction for multiple comparisons, which corresponded to k ≥ 1086 for the analysis in (a) and to k ≥ 1800 for the analysis in (b). The color bar reflects the range of t values (from 0 to 7). The t value corresponding to the statistical significance threshold of choice is t = 3.180. Lighter colors (e.g. light yellow and white) reflect larger t values, i.e. regions where the association between GM volume and sentence production accuracy was stronger, whereas darker colors (e.g. dark red) reflect smaller t values, i.e. regions of where the association was weaker.

VBM regression analyses

Results of the VBM analyses conducted on sentence production are reported in Fig. 4, for analyses run using the NAVS SPPT total score (Fig. 4a) and the NAT total score (Fig. 4b). Significant clusters for both analyses are reported in Table 3, with labels for all regions derived from the Harvard–Oxford atlas (Desikan et al. 2006). Accuracy on the SPPT was associated with GM volume in two clusters: the largest cluster encompassed portions of the IFG (pars opercularis and triangularis), as well as parts of the MFG and frontal pole in the left hemisphere; the second cluster was centered on the left parietal operculum and extended laterally and caudally to include parts of the posterior STG and the planum temporale. Accuracy on the NAT was associated with GM volume in a large left-hemisphere cluster centered around the IFG (and encompassing all three subdivisions), extending rostrally and dorsally to include portions of the MFG, frontal pole, SFG, as well as the SMA and the most ventral part of the precentral gyrus (coinciding with the ventral premotor cortex) in the left hemisphere; in addition, the cluster encompassed portions of the anterior cingulate and paracingulate gyri. A second, much smaller, cluster was found in the right hemisphere and included the three IFG parcellations (orbitalis, triangularis, and opercularis) and part of the frontal operculum.

Table 3.

Clusters derived from the VBM regression analyses on sentence production measures derived from the NAVS SPPT (a) and the NAT (b). Region labels are derived from the Harvard–Oxford atlas.

a)
Cluster p-value (FWE)	Cluster Size	Peak T-value	Peak Coordinates			Hemi	Peak Region	Extent
			x	y	z
<.001	4439	5.214	−52.5	21	27	L	IFGoper	IFGtri, Frontal Pole, MFG
0.030	1086	4.144	−37.5	−33	16.5	L	Parietal Operculum	pSTG, Planum Temporale, Central Operculum, Heschl’s Gyrus
b)
Cluster p-value (FWE)	Cluster Size	Peak T-value	Peak Coordinates			Hemi	Peak Region	Extent
			x	y	z
<.001	26,971	6.297	−49.5	18	28.5	L/R	MFG	IFGtri, IFGoper, IFGorb, Frontal Pole, Frontal Operculum, PCG, antCing, L/R Paracingulate, L/R SFG, SMA
0.004	1800	4.498	52.5	21	−1.5	R	IFGtri	IFGorb, IFGoper, Frontal Operculum

Note. antCing = Anterior Cingulate; IFGoper = Inferior Frontal Gyrus, pars opercularis; IFGorb = Inferior Frontal Gyrus, pars orbitalis; IFGtri = Inferior Frontal Gyrus, pars triangularis; MFG = Middle Frontal Gyrus; PCG = Precentral Gyrus; SFG = Superior Frontal Gyrus; SMA = Supplementary Motor Area; pSTG = Superior Temporal Gyrus, posterior division.

Relationship between atrophy in clusters derived from VBM regression analyses and factors affecting sentence production

Effects of verb production ability and performance on digit span tasks on sentence production

Results of the regression analyses (Table 4) showed that accuracy on the VNT was predictive of atrophy (i.e. the lower the score, the more atrophy) only in the largest cluster derived from the SPPT analyses, i.e. the one peaking in the left IFG pars opercularis, although statistical significance did not survive correction for multiple comparisons. Atrophy in the SPPT cluster encompassing the left pSTG and planum temporale, as well as in the two clusters associated with performance on the NAT, was not predicted by accuracy on the VNT. In keeping with these findings, VBM regressions performed using the VNT total score as the main predictor, and including all the covariates entered for the analyses reported above, yielded associations between VNT score and GM volume in a large left fronto-temporal cluster that included the entire left MTG (including the temporal pole portion), parts of the inferior temporal gyrus, the insula, planum polare and orbitofrontal cortex, as well as some subcortical structures such as the amygdala, the hippocampus and the thalamus. Accuracy on the VNT was also associated with GM volume in a cluster centered on the right temporal pole and extending to the anterior MTG (Supplementary Materials, Fig. S1 and Table S2). Notably, regions of atrophy associated with verb production accuracy did not overlap with areas associated with sentence production.

Table 4.

Results of simple linear regressions evaluating the relationship between performance on the verb naming test (VNT), digit span forward (DSF) and DSF/DSB ratio, and the extent of atrophy in each of the two clusters yielded by the whole-brain VBM analyses using the SPPT (a) or the NAT (b) as measures of sentence production.

a)
SPPT	Estimate	SE	t	P	FDR-corrected P	R2
Proportion atrophy in cluster 1 SPPT ~ VNT
Intercept	0.229	0.047	4.831	0.000		0.059
VNT total	−0.133	0.055	−2.448	0.016	0.096
Proportion atrophy in cluster 1 SPPT ~ DSF
Intercept	0.144	0.02	7.319	0.000		0.040
DSF	−0.006	0.003	−1.976	0.051	0.153
Proportion atrophy in cluster 1 SPPT ~ DSF/DSB ratio
Intercept	0.107	0.009	11.902	<0.001		0.008
DSF/DSB ratio	0.029	0.034	0.864	0.390	0.486
Proportion atrophy in cluster 2 SPPT ~ VNT
Intercept	0.237	0.055	4.274	4.53E-05		0.026
VNT total	−0.102	0.064	−1.602	0.112	0.224
Proportion atrophy in cluster 2 SPPT ~ DSF
Intercept	0.238	0.023	10.307	<0.001		0.157
DSF	−0.016	0.004	−4.188	0.000	0.001
Proportion atrophy in cluster 2 SPPT ~ DSF/DSB ratio
Intercept	0.156	0.011	13.973	<0.001		0.022
DSF/DSB ratio	−0.060	0.042	−1.441	0.153	0.262
b)
NAT	Estimate	SE	t	P	FDR-corrected P	R2
Proportion atrophy in cluster 1 NAT ~ VNT
Intercept	0.187	0.035	5.395	0.000		0.042
VNT total	−0.082	0.040	−2.046	0.044	0.153
Proportion atrophy in cluster 1 NAT ~ DSF
Intercept	0.138	0.014	9.675	0.000		0.036
DSF	−0.004	0.002	−1.867	0.065	0.156
Proportion atrophy in cluster 1 NAT ~ DSF/DSB ratio
Intercept	0.111	0.006	17.304	<0.001		0.015
DSF/DSB ratio	0.028	0.024	1.179	0.241	0.361
Proportion atrophy in cluster 2 NAT ~ VNT
Intercept	0.119	0.053	2.256	0.026		0.006
VNT total	−0.046	0.061	−0.762	0.448	0.489
Proportion atrophy in cluster 2 NAT ~ DSF
Intercept	0.076	0.011	6.863	0.000		0.007
DSF	−0.001	0.002	−0.836	0.405	0.486
Proportion atrophy in cluster 2 NAT ~ DSF/DSB ratio
Intercept	0.066	0.005	13.439	<0.001		0.004
DSF/DSB ratio	0.012	0.018	0.647	0.519	0.519

Note. Statistics derived from regression models including each variable as the only predictor are provided. The formula used to compute each model is provided in the form of (DV ~ predictor), where DV stands for dependent variable. For each model and predictor, beta estimates, standard error (SE), t values, uncorrected and FDR-corrected P values, as well as the modified R-squared, are reported. Statistically significant (i.e., FDR-corrected P < .05) and marginally significant (i.e., FDR-corrected P < .1) values are marked using bold and italic fonts, respectively.

Performance on the Digit Span Forward task, but not the DSF/DSB ratio, predicted atrophy in the left temporal cluster derived from the VBM-SPPT analysis, where greater atrophy was associated with lower performance on the task. In line with these results, VBM regressions performed using the Digit Span Forward total score as main predictor yielded one significant cluster encompassing the left planum temporale and posterior STG (Supplementary Materials, Fig. S2 and Table S3). The score on the DSF/DSB ratio was not predictive of atrophy in any of the clusters resulting from the whole-brain VBM analyses (Table 4).

Effect of syntactic complexity

All clusters derived from both the SPPT and the NAT VBM analyses were significantly associated with accuracy on both canonical and non-canonical items. However, some differences were noted when looking at the percentage of accounted variance. On the SPPT (Table 5a), atrophy in the larger, left frontal cluster was better predicted by accuracy on non-canonical (vs. canonical) items (% of total variance accounted for by accuracy on canonical: 16.7%; non-canonical: 25.7%), whereas the opposite pattern was noted for the smaller, left superior temporal cluster (canonical: 11.4%; non-canonical: 8.2%). On the NAT (Table 5b), atrophy in the larger, left frontal cluster was equally well predicted by accuracy on canonical (28.1%) and non-canonical (26.7%) items, whereas, atrophy in the right frontal cluster was better predicted by performance on canonical (18.6%) than non-canonical (9.0%) items.

Table 5.

Results of simple linear regressions evaluating the relationship between production of canonical and non-canonical sentences on the SPPT (a) and the NAT (b), and the extent of atrophy in each of the two clusters yielded by the whole-brain VBM analyses using the SPPT (a) or the NAT (b) as measures of sentence production.

a)
SPPT	Estimate	SE	t	P	FDR-corrected P	R2
Proportion of atrophy in cluster 1 SPPT ~ SPPT Canonical (C)
Intercept	0.219	0.026	8.259	0.000		0.167
SPPT C	−0.128	0.03	−4.335	0.000	0.000
Proportion of atrophy in cluster 1 SPPT ~ SPPT Non-canonical (NC)
Intercept	0.190	0.016	11.993	<0.001		0.257
SPPT NC	−0.118	0.021	−5.708	0.000	0.000
Proportion of atrophy in cluster 2 SPPT ~ SPPT Canonical (C)
Intercept	0.277	0.038	7.374	0.000		0.114
SPPT C	−0.146	0.042	−3.471	0.001	0.001
Proportion of atrophy in cluster 2 SPPT ~ SPPT Non-canonical (NC)
Intercept	0.215	0.024	8.866	0.000		0.082
SPPT NC	−0.091	0.032	−2.892	0.005	0.005
b)
NAT	Estimate	SE	t	P	FDR-corrected P	R2
Proportion of atrophy in cluster 1 NAT ~ NAT Canonical (C)
Intercept	0.263	0.024	10.720	<0.001		0.281
NAT C	−0.164	0.027	−6.130	0.000	0.000
Proportion of atrophy in cluster 1 NAT ~ NAT Non-canonical (NC)
Intercept	0.195	0.014	13.553	<0.001		0.267
NAT NC	−0.121	0.021	−5.919	0.000	0.000
Proportion of atrophy in cluster 2 NAT ~ NAT Canonical (C)
Intercept	0.257	0.040	6.514	0.000		0.186
NAT C	−0.202	0.043	−4.679	0.000	0.000
Proportion of atrophy in cluster 2 NAT ~ NAT Non-canonical (NC)
Intercept	0.146	0.024	6.022	0.000		0.090
NAT NC	−0.107	0.035	−3.081	0.003	0.003

Note. Statistics derived from regression models including each variable as the only predictor are provided. The formula used to compute each model is provided in the form of (DV ~ predictor), where DV stands for dependent variable. For each model and predictor, beta estimates, standard error (SE), t-values, uncorrected and FDR-corrected P-values, as well as the modified R-squared, are reported. Statistically significant (i.e., FDR-corrected P < .05) and marginally significant (i.e., FDR-corrected P < .1) values are marked using bold and italic fonts, respectively.

Differences between subtypes

Results of these analyses are reported in the Supplementary Materials. Regression analyses on the SPPT showed that the introduction of Subtype in models including accuracy on the SPPT total as predictor did not improve the model fit when the dependent variable was either atrophy in the larger left frontal cluster (F = 1.627, P = 0.202) or atrophy in the smaller left temporal cluster (F = 0.485, P = 0.617). Although the interaction between SPPT total score and Subtype improved the model fit for atrophy in the left frontal cluster (F = 4.120, P = 0.019), a closer inspection of the data revealed that this interaction was driven by one participant in the PPA-S group (Fig. S3a), and was no longer significant or contributing to the model fit after the outlier was removed (Fig. S3b and Table S4a). For atrophy in the left posterior temporal cluster, the interaction between SPPT total and Subtype was not significant and did not contribute to the model fit (F = 0.534, P = 0.588). Analyses conducted on the NAT (Table S4b) showed a similar picture: the introduction of Subtype in models including accuracy on the NAT total as predictor did not improve the model fit when the dependent variable was either atrophy in the large left frontal cluster (F = 1.339, P = 0.267) or atrophy in the smaller right frontal cluster (F = 0.502, P = 0.607). In addition, in both analyses, the interaction between NAT proportion total correct and Subtype was not significant and did not improve the model fit (left frontal cluster: F = 1.019, P = 0.365; right frontal cluster: F = 0.663, P = 0.518).

Discussion

The present study investigated regions of atrophy associated with sentence production using VBM in a group of 101 individuals with PPA, with the aim to shed light on the effects of (i) fluency/motor speech, verb retrieval, and Digit Span task performance, (ii) syntactic complexity, and (iii) PPA subtype, on the relationship between sentence production and regions of atrophy. Two measures of sentence production were used: one that required participants to produce a spoken sentence to describe a picture following an auditorily presented prime (i.e. NAVS SPPT total score), and one requiring participants to arrange word cards to describe a picture (i.e. NAT total score). Notably, both measures assessed production of the same sentence structures², using identical picture stimuli, thereby offering the opportunity to compare the results of the two and avoid confounds related to differences in e.g. frequency and length of target sentences. In addition, both tests included sentences with simple, high-frequency words that were easy to understand even for individuals with moderate word comprehension deficits, such as those exhibited by individuals with semantic PPA.

Common and distinct areas associated with sentence production in the two tasks

Whole-brain VBM analyses revealed substantial overlap in frontal brain regions associated with performance on both sentence production measures, namely, the ventral portion of the left MFG and the dorsal portion of the left pars triangularis and opercularis of the IFG, and parts of the left frontal pole. Associations between sentence production and GM volume or cortical thickness in these areas have been found in previous studies in PPA (Canu et al. 2019; DeLeon et al. 2012; Mesulam et al. 2021b; Rogalski et al. 2011a; see also Ash et al. 2013; Wilson et al. 2010a, for evidence from connected speech). As the two tasks differ in their task requirements while including the same sentence structures, regions of overlap between these two likely reflect processes engaged by both tasks, such as hierarchical structure building (i.e. building the sentence structure, or syntactic tree), morpho-syntactic linearization (i.e. ordering words in a sentence), and/or the computation of syntactic movement (see “Effect of syntactic complexity”, for a discussion).

Analyses also highlighted several differences in the regions associated with sentence production accuracy in the two tasks. First, GM volume in left posterior temporal regions, including the left planum temporale and pSTG, was associated with sentence production accuracy on the NAVS SPPT, and not the NAT. This finding is in line with previous studies that reported associations between sentence production and GM volume in these regions using the NAVS SPPT (Lukic et al. 2021), and with the lack of associations with these regions in studies employing the NAT (Canu et al. 2019). This result is likely to reflect differences in the task requirements (see “Effects of verb retrieval and digit span task performance on sentence production”, for a discussion). Conversely, sentence production accuracy on the NAT (but not the NAVS SPPT) was associated with GM volume in the left SMA and ventral premotor cortex, regions linked to pre-articulatory motor programming (see Kearney and Guenther 2019 and Price 2012, for reviews), suggesting that the NAT, which requires no overt articulation, nevertheless, engages covert motor processes. Albeit unexpected, this result may reflect subvocal rehearsal processes that are involved during reading comprehension (e.g. Slowiaczek and Clifton Jr 1980) and supported by these areas (Paulesu et al. 1993; Smith et al. 1998; Müller and Knight 2006). In addition, sentence production accuracy on the NAT was associated with a left frontal cluster that—albeit including some of the same regions associated with accuracy on the SPPT—encompassed several areas that were not associated with accuracy on the SPPT, namely the left orbitofrontal cortex and frontal pole, the dorsomedial portion of the left superior frontal gyrus (SFG), and the anterior cingulate and paracingulate. These regions are part of domain-general networks, such as the cingulo-opercular and the fronto-parietal networks (Dosenbach et al. 2007; Sadaghiani and D’Esposito 2015; Seeley et al. 2007; Vincent et al. 2008), and are involved in task initiation, selective attention, top-down control processes and cognitive flexibility (see also “Effect of syntactic complexity”, for a discussion). Furthermore, atrophy in the right IFG and frontal operculum was associated with sentence production accuracy on the NAT, and not the NAVS SPPT. Associations between performance on the NAT and GM volume or cortical thickness in the right hemisphere of individuals with PPA have been reported in a previous study (Canu et al. 2019) and may reflect engagement of domain-general resources (see “Effect of syntactic complexity”, for a discussion).

Effects of verb retrieval and digit span task performance on sentence production

With respect to the role of verb retrieval deficits, analyses showed that for the SPPT, GM volume in the left frontal cluster was predicted by accuracy on the VNT. However, the percentage of variance accounted for by verb naming accuracy was significantly lower (5.9%) than that of sentence production (16.7% for canonical, and 25.7% for non-canonical sentences), and the result was only marginally significant after correction for multiple comparisons. Conversely, verb retrieval difficulties did not affect the relationship between sentence production accuracy on the SPPT and GM volume in left posterior temporal regions. In keeping with these findings, areas associated with accuracy on the VNT on a whole-brain VBM analysis did not overlap with those associated with accuracy on the SPPT; rather, they were mostly centered around the left middle and inferior temporal gyrus, and only included the left insula and left orbitofrontal cortex in the frontal lobe. Together, these results suggest that verb retrieval deficits only marginally account for the relationship between performance on the SPPT and left IFG/MFG atrophy. Notably, performance on the VNT did not affect the relationship between GM volume and sentence production on the NAT. This result is in keeping with the requirements of the task, in which lexical labels for verbs are provided and verb retrieval deficits are bypassed (as long as reading is possible).

Turning to Digit Span test performance, Digit Span Forward scores were predictive of GM volume in the left posterior temporal cluster associated with sentence production accuracy on the SPPT. Notably, the percentage of variance accounted for by this measure (15.7%) was higher than that accounted for by sentence production (11.4% for canonical, and 8.2% for non-canonical sentences); in addition, this cluster exhibited almost complete overlap with the areas associated with performance on the Digit Span Forward task on a whole-brain VBM analysis. To the extent that Digit Span Forward performance reflects auditory short-term memory processing, these findings are in line with previous studies pointing to this area as the phonological or articulatory loop (Buchsbaum et al. 2011; Lukic et al. 2019; Pisoni et al. 2019; Forkel et al. 2020), and with research identifying the Sylvian parietal–temporal (Spt) area—which was part of the cluster—as an auditory-motor interface area (Hickok et al. 2003, 2009; Buchsbaum et al. 2011). Notably, the ratio between the Digit Span Forward and Backward scores did not affect the relationship between SPPT scores and GM volume. This finding suggests that the ability to maintain phonological information in short-term memory, rather than the additional manipulations required by the Digit Span Backward, may affect underlie the relationship between sentence production deficits and left posterior temporal atrophy. Conversely, the relationship between performance on the SPPT and left IFG/MFG atrophy was not affected by Digit Span task performance, a result that is at odds with the idea that the IFG may support working memory processes associated with sentence computation (Chein et al. 2002; Martin 2003; Rogalsky and Hickok 2011; Bornkessel-Schlesewsky and Schlesewsky 2013). In line with our expectations, no effect of Digit Span task performance was found on the relationship between atrophy and sentence production accuracy on the NAT.

These data suggest that integrity of the phonological/articulatory loop may partly account for the association between sentence production accuracy and atrophy in left posterior temporal (but not inferior/middle frontal) areas on the SPPT (but not the NAT). We note that while the Digit Span tasks are widely accepted as measures of auditory short-term and working memory in people with unimpaired language abilities, their validity in the assessment of individuals with aphasia has been questioned (Dede et al. 2014). Additional research examining the relation between auditory short-term/working memory ability and sentence processing using tasks that do not rely on production ability, such as n-back, picture span, and other tasks (Mayer and Murray 2012; Dede et al. 2014), is needed to shed further light on this issue.

Together, these results indicate that the investigation of the neural bases of sentence production may yield different results depending on the task requirements, and that some of the inconsistencies found in the literature can be attributed, among others, to methodological (task) differences.

Effect of syntactic complexity

A second aim of the study was to assess the effect of syntactic complexity, and specifically syntactic movement (see Chomsky 1981, and Introduction), on the relationship between atrophy and sentence production accuracy. This was achieved by evaluating whether the association between atrophy and sentence production accuracy (computed as the total score, i.e. including both canonical and non-canonical sentences) was better predicted by accuracy on canonical or non-canonical sentences.

Within our theoretical framework of reference (Chomsky 1981), non-canonical sentences (which entail movement) require greater processing demands than canonical sentences (which do not) for hierarchical structure building (Friederici 2012; Thompson and Meltzer-Asscher 2014) and/or morpho-syntactic linearization (i.e. the process of ordering words in a sentence, see Matchin and Hickok 2020), two processes that have been linked to the pars opercularis and triangularis of the left IFG. Our results indicate that the relationship between left inferior/middle frontal atrophy and sentence production accuracy on the SPPT was better predicted by accuracy on non-canonical sentences, than on canonical sentences. Our findings are, therefore, in line with the neurocognitive models of sentence processing mentioned earlier (Friederici 2012; Thompson and Meltzer-Asscher 2014; Matchin and Hickok 2020), and support the literature on healthy language processing, which shows greater IFG activation during processing of non-canonical (versus canonical) sentences (e.g. Europa et al. 2019; Friederici et al. 2006; Kinno et al. 2008; Mack et al. 2013; Santi and Grodzinsky 2007; Santi and Grodzinsky 2010; see Grodzinsky et al. 2021, for a review).

Conversely, atrophy in the left posterior temporal cluster associated with sentence production accuracy on the SPPT was equally predicted by canonical and non-canonical sentences, which accounted for 11.4% and 8.2% of the variance, respectively. This finding was unexpected, in view of current neurocognitive models of sentence processing that assume that left posterior temporal regions play a primary role in assigning thematic roles (i.e. in deciding “who is doing what to whom” in the sentence, see Thompson and Meltzer-Asscher 2014) or in the integration of syntactic with semantic information (Friederici 2012), processes that—if disrupted—cause greater difficulties in production (and/or comprehension) of non-canonical (versus canonical) sentences. Along the same lines, neuroimaging studies in healthy participants (see Grodzinsky et al. 2021, for a review) have linked activation of the left pSTG to processing of syntactic movement in non-canonical sentences, and studies in stroke aphasia have found associations between lesions to the pSTG and production of non-canonical sentences (Lukic et al. 2021). Although our results should not be interpreted as challenging the interpretation that the pSTG supports syntactic processes (based on the evidence that performance on non-canonical sentences accounted for part of the variance), they do suggest that syntactic movement affects left inferior/middle frontal regions to a greater extent than left posterior superior temporal regions. Future studies should—by employing other methods, such as functional neuroimaging—clarify the role of pSTG in non-canonical sentence production.

Turning to the NAT, atrophy in the left frontal cluster associated with sentence production on this task was predicted in equal measure by accuracy on canonical and non-canonical sentences (26.7 vs. 28.1%). This finding may find an explanation in the extent of the left frontal cluster associated with performance on the NAT, which—as opposed to the one associated with accuracy on the SPPT—extended well beyond the IFG/MFG, encompassing the left orbitofrontal cortex, the dorsomedial portion of the left superior frontal gyrus, the supplementary motor area, and the anterior cingulate and paracingulate. These regions are part of domain-general networks, such as the cingulo-opercular and the fronto-parietal networks that support—among others—attentional and executive processes (Collette et al. 2006; Dosenbach et al. 2007; Sadaghiani and D’Esposito 2015; Seeley et al. 2007; Vincent et al. 2008). The notion that attention and executive functions play a role in sentence (and speech) production is not new and relates to the knowledge that sentence production requires, e.g. selecting appropriate lexical items and syntactic structures from a pool of available options (e.g. Rogalski et al. 2010; Cannizzaro and Coelho 2013; Engelhardt et al. 2013; Thothathiri et al. 2017; Tetzloff et al. 2018). Although the NAT bypasses lexical selection, attention and executive functions are still largely at play, as participants are required to build a sentence using cards and a prompt (i.e. the beginning of the target sentence) as guide. In doing so, participants are likely to build multiple sentence structures at the same time and then select the appropriate one based on the joint consideration of the prompt and the picture (see Fig. 2).

Furthermore, the NAT requires reading, a process that is not engaged in the SPPT. As reading relies on a network that includes both inferior parietal and inferior frontal regions in the left hemisphere (see Joubert et al. 2004; Mechelli et al. 2005; Wilson et al. 2009; Purcell et al. 2011; Paz-Alonso et al. 2018), the more extensive involvement of the IFG in sentence production measured with the NAT (versus SPPT) could also reflect reading processes.

Role of PPA subtype

An additional aim of the study was to evaluate whether the relationship between regions of atrophy and sentence production deficits varied depending on PPA subtype, and specifically, whether it was stronger in individuals with PPA-G than in other PPA subtypes, based on the extensive literature that points to sentence production deficits as one of the defining features of PPA-G (Gorno-Tempini et al. 2011; Mack et al. 2021; Mesulam et al. 2009; Thompson et al. 2012a; Weintraub et al. 2009; see Thompson and Mack 2014, for a review). Results showed that atrophy did not differ between subtypes in any of the clusters associated with sentence production on the SPPT or the NAT, and, most importantly, there was no evidence that the relationship between atrophy and performance on the SPPT or NAT differed between subtypes.

Overall, this finding suggests that sentence production is supported by a large-scale network that can be variably disrupted in PPA, resulting in graded levels of impairment across participants. In addition, it suggests that when tissue that is critical for sentence production is atrophied, impairment ensues across all language profiles. One aspect that should be considered in future studies is whether areas associated with sentence production accuracy may vary depending on the underlying neuropathology, which can show predilection for specific areas of the brain (see Mesulam et al. 2022; Mesulam et al. 2023a).

Comparison with the literature on stroke aphasia

The present findings diverge from those of the two volumetric studies of sentence production in stroke-induced aphasia (Den Ouden et al. 2019; Lukic et al. 2021). While we found that sentence production deficits on the NAVS SPPT were associated with atrophy in left IFG/MFG and left pSTG/planum temporale, Lukic et al. (2021), using the same measure, only found associations with left anterior (i.e. temporal pole, aSTG) and posterior (planum temporale, pSTG) temporal regions, and not with left frontal areas. As reported in their paper, Lukic et al. (2021) suggested that the lack of association between lesions to left frontal regions and sentence production accuracy on the SPPT could stem from the presence of significant verb production deficits in their group of participants. According to linguistic theory, verbs guide sentence structure building, in that they determine e.g. the number and type of arguments (i.e. required elements) that need to be produced in order for a sentence to be grammatical (see Thompson and Meltzer-Asscher 2014, for details). Therefore, deficits in verb production can be at the root of sentence production deficits in some people with aphasia. In the study by Lukic et al. (2021), lesioned areas associated with sentence production were obtained by regressing out performance on a verb naming task; therefore, the presence of verb production deficits in their stroke group may have masked the association between lesioned tissue in left frontal regions and sentence production deficits. This claim was supported by the evidence that verb retrieval accuracy was associated with left temporal and frontal regions, including the insula and the IFG, in the same study (Lukic et al. 2021, Supplementary Materials). Notably, the PPA group examined in this study scored higher on tests of verb (and sentence) production than the stroke group included in Lukic et al. (Fig. 5), unmasking the contribution of frontal regions to sentence production. However, our regression analyses demonstrated that verb retrieval deficits only marginally accounted for the association between atrophy in left IFG/MFG and sentence production accuracy. Moreover, areas associated with verb production accuracy on a whole-brain VBM analysis did not overlap with those associated with accuracy on the SPPT in the present study. These findings indicate a need for further research examining the relation between verb and sentence production impairments in aphasia and associated neural correlates to fully understand the role of left frontal regions in sentence production.

Fig. 5.

Percent correct accuracy on language measures used in the study by Lukic et al. (2021), conducted on individuals with stroke-induced aphasia, and in the present study (PPA). WAB-AQ = Western Aphasia Battery – Aphasia Quotient (Kertesz 2007); SPPT = Sentence Production Priming Test (included in the Northwestern Assessment of Verbs and Sentences, NAVS, Thompson 2012); C = canonical sentences; NC = non-canonical sentences. Note that verb production scores were derived from different tests in the two studies (i.e. Northwestern Naming Battery, NNB, Thompson and Weintraub (2014) in Lukic et al. (2021), and NAVS for the present study).

Differences in the way the neuropathology underlying PPA and stroke affects the brain should also be mentioned. As discussed elsewhere (seeMesulam et al. 2014; Mesulam et al. 2019; Silveri et al. 2019), stroke results in cell death, completely precluding the functionality of neural tissue within damaged regions, whereas atrophy in PPA affect some, but not all, regional neural tissue (see Sonty et al. 2003, for evidence that atrophied regions are significantly active during language tasks). Further, stroke is a sudden onset disorder, whereas PPA progresses slowly over time. The slow progression in PPA may leave room for the language network to partially re-organize before language functions are completely lost (e.g. Wilson et al. 2010b; Beeson et al. 2011; Thompson et al. 2021; Tao et al. 2022); however, there is extensive evidence for functional re-organization in individuals with chronic aphasia as well (Fridriksson et al. 2012; Kiran et al. 2015; Barbieri et al. 2019; Barbieri et al. 2023). Whether or not these factors may account for the inconsistencies in the literature on the neural bases of sentence production, but also word and sentence comprehension (see Mesulam et al. 2015; Matchin et al. 2022; Mesulam et al. 2023b), is unclear and should be informed by studies that directly compare data derived from the two clinical populations. Although this has rarely been done, a recent study (Tao et al. 2022) has provided evidence for decreased resting-state functional connectivity between homotopic lateral posterior temporal areas in individuals with stroke aphasia and increased homotopic connectivity in individuals with PPA with similar levels of language severity and patterns of damage. Based on these findings, one may suggest that in stroke, damage to left posterior temporal regions disrupts the left fronto-temporal language network to a greater extent that in individuals with PPA, who may be able to compensate by recruiting right hemisphere homotopic regions. As a result, associations between language deficits and damage to left posterior temporal areas may be more pronounced in stroke aphasia than in PPA. As functional connectivity was not measured in either the current or the Lukic et al. (2021) study, this hypothesis remains to be tested.

Another source of difference between PPA and stroke aphasia lies in the extent to which white matter is affected by pathology. White matter damage in PPA is typically less extensive than in stroke (Mesulam et al. 2014; Mesulam et al. 2019; Silveri et al. 2019), whereas stroke lesions in posterior temporal and inferior parietal regions often extend to white matter fibers connecting these regions to anterior (i.e. inferior frontal) regions, thereby resulting in a disconnection between them. Nevertheless, individuals with PPA-G may exhibit lower structural connectivity in the superior longitudinal fasciculus (SLF, Agosta et al. 2013; Galantucci et al. 2011)—a bundle of fibers connecting left temporo-parietal areas with the frontal lobe, and studies suggest that sentence production deficits in PPA are associated with structural integrity of the SLF (Wilson et al. 2011; Marcotte et al. 2017) or with functional connectivity between the IFG and the MTG (Bonakdarpour et al. 2019). Thus, the extent to which white matter damage contributes to the observed differences in the regions associated with sentence production (and other language) deficits in the two clinical populations remains to be determined.

In sum, the results obtained from this study are compatible with previous evidence in PPA, but not with the (limited) available evidence from the literature on stroke aphasia. Further research is needed, especially in stroke aphasia, to clarify whether these discrepancies stem from differences in methodology (e.g. the use of voxel-based morphometry versus voxel-based lesion symptom mapping techniques) or they reflect damage to different parts of the language network in the two clinical populations.

Clinical significance of the study

Results of this study can—at least in part—inform clinical practice. On the one hand, they suggest that the two selected measures of sentence production, the NAVS SPPT and the NAT, are able to capture difficulties in aspects of sentence production that are engaged by both tasks, such as hierarchical structure building and word ordering, which relies on the left IFG and part of the MFG. On the other hand, they underscore that each task also engages some processes that are unique and supported by brain regions other than left IFG/MFG. Thus, these results provide evidence in favor of employing both the NAVS SPPT and the NAT when investigating sentence production deficits in PPA, in line with our current protocols in place at the Mesulam Center for Cognitive Neurology and Alzheimer’s Disease. In addition, results point to the NAT as a tool assessing this relationship in a way that is independent from verb retrieval and short-term memory deficits (Weintraub et al. 2009), although relying on reading and domain-general processes to a greater extent thant the NAVS SPPT.

Conclusions

The present study provides evidence that sentence production deficits in PPA are primarily associated with GM loss in left inferior and middle frontal regions, as determined by the convergence of the results derived from analyses on two different measures of sentence production (Fig. 4). The relationship between left inferior/middle frontal atrophy and sentence production impairment was stronger for syntactically complex than for simple sentences and was not accounted for by performance on a Digit Span task. Conversely, associations with left posterior temporal/inferior parietal regions were only found for a measure of primed sentence production that also required overt articulation and were associated with Digit Span task performance. Relationships found between sentence production impairments and regions of atrophied tissue did not differ between PPA subtypes, suggesting that when tissue that is critical for sentence production is atrophied, impairment ensues across all language profiles, although the consideration of other factors (such as the type of neuropathology underlying PPA) may offer future insights into the underlying source of sentence production deficits.

Author contributions

Elena Barbieri: Conceptualization, Formal Analysis, Funding acquisition, Investigation, Visualization, Writing—original draft. Sladjana Lukic: Conceptualization, Methodology, Writing—review and editing. Emily Rogalski: Funding acquisition, Resources, Writing—review and editing. Sandra Weintraub: Resources, Writing—review and editing. Marek-Marsel Mesulam: Conceptualization, Funding acquisition, Resources, Supervision, Writing—review and editing. Cynthia K. Thompson: Conceptualization, Funding acquisition, Supervision, Writing—review and editing.

Funding

National Institutes of Health (grant numbers 5R01AG077444 to M.-M.M. and E.R., R01 AG056258 to E.R., R01DCO1948-24 to C.K.T, and P30AG072977 to R.J.V.), and Karen Toffler Charitable Trust (scholarship awarded to E.B.).

 

Conflict of interest statement: The authors have no conflicts of interest to disclose.