Online sentence processing impairments in agrammatic and logopenic primary progressive aphasia: Evidence from ERP

Abstract

Evidence from psycholinguistic research indicates that sentence processing is impaired in Primary Progressive Aphasia (PPA), and more so in individuals with agrammatic (PPA-G) than logopenic (PPA-L) subtypes. Studies have mostly focused on offline sentence production ability, reporting impaired production of verb morphology (e.g., tense, agreement) and verb-argument structure (VAS) in PPA-G, and mixed findings in PPA-L. However, little is known about real-time sentence comprehension in PPA. The present study is the first to compare real-time semantic, morphosyntactic and VAS processing in individuals with PPA (10 with PPA-G and 9 with PPA-L), and in two groups of healthy (22 young and 19 older) individuals, using event-related potentials (ERP). Participants were instructed to listen to sentences that were either well-formed (n = 150) or contained a violation of semantics (e.g., *Owen was mentoring pumpkins at the party, n = 50), morphosyntax (e.g., *The actors was singing in the theatre, n = 50) or VAS (*Ryan was devouring on the couch, n = 50), and were required to perform a sentence acceptability judgment task while EEG was recorded. Results indicated that in the semantic task both healthy and PPA groups showed an N400 response to semantic violations, which was delayed in PPA and older (vs. younger) groups. Morphosyntactic violations elicited a P600 in both groups of healthy individuals and in PPA-L, but not in PPA-G. A similar P600 response was also found only in healthy individuals for VAS violations; whereas, abnormal ERP responses were observed in both PPA groups, with PPA-G showing no evidence of VAS violation detection and PPA-L showing a delayed and abnormally-distributed positive component that was negatively associated with offline sentence comprehension scores. These findings support characterizations of sentence processing impairments in PPA-G, by providing online evidence that VAS and morphosyntactic processing are impaired, in the face of substantially preserved semantic processing. In addition, the results indicate that on-line processing of VAS information may also be impaired in PPA-L, despite their near-normal accuracy on standardized language tests of argument structure production.

1. Introduction

Primary progressive aphasia (PPA), an acquired neurodegenerative disease, preferentially affects the neural networks for language, relatively sparing networks that subserve other aspects of cognition (Gorno-Tempini et al., 2011; Mesulam, 1982; 2003: Mesulam et al., 2009). Current diagnostic criteria identify three main types of PPA: semantic (PPA-S) associated with bilateral (left > right) anterior temporal lobe (ATL) atrophy; agrammatic (PPA-G) linked with peak atrophy in the left inferior frontal gyrus (IFG), and logopenic (PPA-L) correlated with atrophy within the left temporo-parietal junction (TPJ) (Gorno-Tempini et al., 2011; Mesulam et al., 2009).

Whereas patients across subtypes present with naming impairment, PPA-S is associated with difficulties in single-word comprehension, but relatively preserved object knowledge (Mesulam et al., 2013; Riley et al., 2018; Seckin et al., 2016). Conversely, comprehension of single words and object knowledge are preserved in both PPA-G and PPA-L subtypes. Individuals with PPA-G also exhibit agrammatic sentence production, i.e., difficulties in producing verbs, function words, and bound morphemes (see Thompson et al., 2013; Thompson and Mack, 2014), and PPA-L patients show word-finding pauses and phonological paraphasias in speech production (Gorno-Tempini et al., 2004; Mack et al., 2015; Mesulam et al., 2009; Teichmann et al., 2013).

1.1. Sentence processing patterns in PPA

Most sentence processing studies have elucidated offline production impairment patterns, finding reduced speech rate and sentence length, as well as reduced proportion of grammatically correct sentences, verb inflections and verbs produced with correct argument structure in the spontaneous speech of individuals with PPA-G and occasionally PPA-L, but not in PPA-S (Thompson et al., 2012a; Wilson et al., 2010b). In addition, accuracy on a sentence production priming task was impaired in both PPA-G and PPA-L (compared to healthy controls), and for PPA-G more than PPA-L, for sentences with non-canonical word order (Thompson et al., 2013). Fewer studies have investigated sentence comprehension. These too are primarily offline studies (Amici et al., 2007; Thompson et al., 2013; Wilson et al., 2010a), indicating impaired comprehension of non-canonical sentences in both PPA-G and PPA-L, with few studies examining real-time sentence processing in PPA.

Online paradigms (e.g., eye-tracking or event-related potentials, ERP) allow for the investigation of the time course of language comprehension and production, without the need for overt responses which may be impacted by decision-making processes, thereby providing measures that are more sensitive to language deficits in PPA, compared to offline measures such as standardized tests. Online paradigms may therefore help characterize the language profile of subtypes of PPA for which behavioral tests provide a mixed picture, as for sentence processing in PPA-G and PPA-L.

The current study investigates online sentence processing in PPA using EEG (ERPs), examining semantic, morphosyntactic (subject-verb agreement) and lexical-syntactic (VAS) processing in individuals with PPA-G and PPA-L.

1.2. ERP components related to sentence processing

1.2.1. Semantics

Many ERP studies with healthy participants have investigated semantic processing, using sentences containing semantic incongruences, including violations of the verb thematic requirements (e.g., *The cloud was buried; Friederici et al., 1998¹). Results consistently show an N400 (Kutas and Hillyard, 1980; see Kutas and Federmeier, 2011 for a review), which may be delayed and attenuated in healthy older adults (Kutas and Iragui, 1998; Federmeier et al., 2003; see also Wlotko et al., 2010 for a review). Studies in PPA (Hurley et al., 2009, 2012) have reported no N400 to pairs of semantically-related words in PPA-S and normal-like performance in individuals with PPA-G or PPA-L. In line with these findings, eye-tracking studies (Faria et al., 2018; Seckin et al., 2016) have reported abnormal fixation patterns to semantically-related, non-target, objects in PPA-S, but not PPA-G or PPA-L. A notable exception is a study by Kielar and colleagues (Kielar et al., 2018), in which delayed and attenuated electrophysiological responses to semantic violations (e.g., *She will go to the bakery for a loaf of books) were found - using an auditory grammaticality judgment task - in a mixed group of individuals with PPA-G and PPA-L. Evidence of delayed lexical-semantic access in PPA-G (and not PPA-L) was also reported in a recent eye-tracking study (Mack et al., 2019). Overall, research suggests that subtle abnormalities in online semantic processing, which are not captured by the administration of standardized language tests, may be found in individuals with PPA-G and PPA-L. In keeping with these findings, research in stroke-induced agrammatic aphasia (Kielar et al., 2012) showed that N400 responses to semantic violations were attenuated.

1.2.2. Morphosyntax

Studies of morphosyntactic (for example, number-agreement) processing with healthy listeners (e.g., Friederici et al., 1998; Hagoort et al., 2003; Wassenaar et al., 2004) have found, for sentences containing violations of subject-verb agreement, a positive-going, centro-parietally distributed, wave peaking around 600 ms after the onset of the critical word (P600), which is generally not affected by aging (King and Kutas, 1995; Kemmer et al., 2004). In PPA, online morphosyntactic processing has been investigated in only a few studies, with inconsistent findings that point to either a delayed sensitivity to subject-verb agreement violations or to a complete lack of sensitivity to tense violations in PPA-G (Grossman et al., 2005; Peelle et al., 2007), in spite of the robust evidence from psycholinguistic research showing deficits in production of verb morphology in PPA-G (and not PPA-L; Graham et al., 2004; Knibb et al., 2009; Thompson et al., 2012a; Wilson et al., 2010b; see Thompson and Mack, 2014, for a review). Only one study (Kielar et al., 2018) investigated electrophysiological responses to morphosyntactic violations (e.g. *She will going to the bakery for a loaf of bread) in PPA, reporting delayed and attenuated responses (compared to healthy individuals) in a combined group of PPA-G and PPA-L. In line with these findings, ERP studies conducted on individuals with stroke-induced agrammatic aphasia have shown lack of online sensitivity (i.e., no P600) to morphosyntactic violations (Hagoort et al., 2003; Wassenaar et al., 2004).

1.2.3. Verb-argument structure

A smaller number of studies have investigated the ERP correlates of VAS processing, with inconsistent findings. In healthy participants, violations of the number of arguments (e.g., *Anne sneezed the doctor and the nurse) were associated with a N400–P600 pattern (Friederici and Frisch, 2000; Friederici and Meyer, 2004; Kielar et al., 2012), while violations of the verb subcategorization frame (e.g., *Anna knows that the inspector helped to the banker²) and violations of structural preference (*The doctor charged the patient was lying) elicited a P600, but not an N400 (Friederici and Frisch, 2000; Osterhout and Holcomb, 1993; Osterhout et al., 1994). In PPA-G, impaired online processing of VAS information, including violations of the number of arguments (e.g., *The child crawls grass through mud) has been reported in a few studies (Price and Grossman, 2005; Peelle et al., 2007). In a recent eye-tracking study (Mack et al., 2019), individuals with PPA-L and PPA-G were asked to listen to sentences that lacked the verb direct object (e.g., Tomorrow Susan will open the …) while looking at a 4-picture array that included a plausible verb object (target, e.g., a jar) and 3 implausible verb objects (e.g., a pencil, a plate, and a flute). Results indicated impaired verb-argument integration (i.e., difficulty in predicting the plausible direct object based on the verb thematic requirements) in PPA-G and not PPA-L. These findings are in line with previous psycholinguistic research reporting impaired VAS production in PPA-G (and not PPA-L; Thompson et al., 1997; Thompson et al., 2012a; Thompson et al., 2012b; Wilson et al., 2010b), and with data obtained from individuals with stroke-induced agrammatic aphasia that show impaired sensitivity to violations of the number of verb arguments (Kielar et al., 2012) or to a mismatch in thematic mapping (Wassenaar and Hagoort, 2007).

1.3. Current study

The present study examined online semantic, morphosyntactic and VAS processing in individuals with PPA-L and PPA-G, by measuring ERPs as participants performed an auditory sentence acceptability judgment task. We expected semantic violations to elicit an N400 in both young (YA) and older (OA) healthy adults. We also anticipated that an N400 of similar amplitude and onset as in the OA group would be found in both PPA-G and PPA-L individuals. For morphosyntactic violations, we expected a P600 in both YA and OA groups; compared to OA, we expected a P600 of similar amplitude and onset in the PPA-L group, and abnormal (i.e, reduced in amplitude or absent) P600 in the PPA-G group, based on the evidence from ERP studies in stroke-induced agrammatic aphasia. Finally, for VAS violations we expected, based on previous studies, either an N400–P600 pattern or a P600 alone in both YA and OA groups. We anticipated, based on both ERP studies in stroke-induced agrammatic aphasia as well as on online research in PPA, abnormal P600 responses to such violations in individuals with PPA-G. For PPA-L, predictions were less straightforward: on the one hand, we expected normal-like ERP responses to VAS violations, based on evidence that VAS processing is primarily intact in PPA-L (Thompson et al., 2012a, 2012b; Mack et al., 2019; also see Thompson and Mack, 2014, for a review); on the other hand, given that PPA-L often display cortical atrophy in the TPJ, a region associated with processing of VAS information in healthy individuals (e.g., Den Ouden et al., 2009; Thompson et al., 2007; Thompson et al., 2010; see also Thompson and Meltzer-Asscher, 2014, for a review), we anticipated that ERP responses to VAS violations may be delayed and reduced in amplitude compared to those in the OA group.

With respect to the relation between ERP components and behavioral performance in the PPA group, we anticipated that N400 amplitude would correlate with measures of lexical-semantic processing. Furthermore, P600 amplitude in the morphosyntactic and VAS conditions was expected to be correlated with the severity of agrammatism, as indexed by tests of verb and sentence production. Finally, we predicted a positive correlation between sensitivity to violations and amplitude of ERP components across all conditions.

2. Materials and methods

2.1. Participants

Nineteen individuals with PPA (PPA-G³: n = 10 and PPA-L: n = 9; see Gorno-Tempini et al., 2011; Mesulam et al., 2009, for diagnostic criteria and subtyping), and 41 healthy individuals (YA: n = 22; OA (>35 years old): n = 19) participated in the study. Group sample sizes were chosen on the basis of standards in the field and were comparable to those used in prior studies of auditory sentence comprehension in aphasia (7–15 patients; 12–15 controls; Hagoort et al., 2003; Kielar et al., 2012; Swaab et al., 1997; Wassenaar and Hagoort, 2007). Demographic information for all participants is provided in Table 1. All participants were right-handed (Oldfield, 1971) monolingual native speakers of English with normal or corrected-to-normal vision and hearing, had no prior history of neurological, language, learning, or psychiatric disorders, and were not being treated with antidepressants (see Feige et al., 2002; Alhaj et al., 2011). As anticipated, age significantly differed between YA and OA groups based on a Wilcoxon rank sum test (W = 396, p < .0001), while both individuals with PPA-G and with PPA-L were matched in age to OA (W = 103.5, ns; W = 66.5, ns, respectively) and to each other (W = 21.5, ns). The OA group had a greater number of years of education than YA (W = 291.5, p = .0092), but no differences were found in years of education between individuals with PPA and OA (PPA-G vs. OA: W = 107, ns; PPA-L vs. OA: W = 63, ns) or between the PPA-L and PPA-G groups (W = 24, ns). Duration of the disease - indexed by the self-reported number of years since the appearance of the first symptoms of PPA - did not differ between the PPA-L and the PPA-G groups (W = 40, ns, Table 2). In addition, none of the participants in the OA group showed evidence of cognitive impairment on a brief neuropsychological test battery including the following tests: The Montréal Cognitive Assessment (MoCA, Nasreddine et al., 2005); the Wechsler Memory Scale-Revised (Wechsler, 1987; selected subtests: Digit Span Forward and Backward, Logical Memory I and II, Visual Reproduction I and II); the Shape Cancellation Test (Weintraub and Mesulam, 1985); the Trail Making Test (Reitan, 1971); the Wisconsin Card Sorting Test (WCST, Heaton et al., 1993) and the Boston Naming Test (Kaplan et al., 2001).

Table 1.

Demographic info for each group. Mean (SD).

ID	Age	Gender	Duration of Symptoms (years)	Handedness	Education (years)
PPA-L1	65	M	3.5	Right	19
PPA-L2	71	M	10.0	Right	18
PPA-L3	70	M	2.0	Right	18
PPA-L4	62	M	3.5	Right	18
PPA-L5	75	F	4.0	Right	18
PPA-L6	61	M	4.0	Right	18
PPA-L7	70	M	1.5	Right	16
PPA-L8	69	M	9.0	Right	18
PPA-L9	66	M	2.5	Right	16
PPA-L Mean (SD)	67.7 (4.5)	1F	4.4 (3.0)		17.7 (1.0)
PPA-G1	63	M	7.0	Right	20
PPA-G2	60	M	2.5	Right	12
PPA-G3	65	F	1.5	Right	18
PPA-G4	52	F	6.0	Right	16
PPA-G5	57	M	6.0	Right	18
PPA-G6	76	M	3.0	Right	14
PPA-G7	54	M	1.5	Right	16
PPA-G8	63	F	2.5	Right	16
PPA-G9	70	F	5.5	Right	16
PPA-G10	56	F	3.0	Right	16
PPA-G Mean (SD)	61.6 (7.4)	5F	3.8 (2.1)		16.2 (2.2)
Healthy age-matched adults (n= 19) Mean (SD)	61.5 (14.5)	10F	N/A	all right-handed	16.9 (2.4)
Healthy younger adults (n= 22) Mean (SD)	21.9 (3.2)	12F	N/A	all right-handed	15.1 (1.5)

Table 2.

Scores obtained from participants with PPA on measures of cognitive and language functions. Measurement units for each measure are provided, and total maximum scores (for measures computing the number correct responses (N)) are indicated in parenthesis. P-values (determined using a Wilcoxon rank sum test) for statistically significant between-group differences are provided.

	Cognitive functions						Language
ID	MMSE (30)	DSF	DSB	Shape Canc	TMT-A	TMT-B	BNT (60)	WABAQ (100)	PPVT (36)	NNB AC (verb)	NNB AC (noun)	NNB naming (verb)	NNB naming (noun)	NAVS VNT (trans)	NAVS ASPT (trans)	NAVS SCT (can)	NAVS SCT (noncan)	NAVS SPPT (can)	NAVS SPPT (noncan)	NAT (can)	NAT (noncan)	NAVI (finite)	NAVI (nonfinite)	WPM	synt corr sent
unit	N	span	span	time	time	time	N	N	N	% corr	% corr	% corr	% corr	% corr	% corr	% corr	% corr	% corr	% corr	% corr	% corr	% corr	% corr	N	% corr
PPA-L1	28	6	3	145	44	153	56	88	36	100	100	100	100	100	100	100	100	100	80	100	60	83	100	61	71
PPA-L2	24	5	4	159	60	300	36	85	34	100	100	88	88	93	92	100	87	93	47	87	67	78	85	103	78
PPA-L3	29	6	3	96	30	73	59	94	36	100	100	100	100	100	96	100	100	100	100	93	93	95	100	73	81
PPA-L4	25	3	3	90	22	56	56	95.6	36	92	100	94	100	93	96	87	80	100	73	100	93	100	100	61	90
PPA-L5	26	7	5	84	39	66	26	96.8	32	100	100	100	94	95	92	100	100	100	100	100	100	98	100	118	86
PPA-L6	24	4	3	77	24	73	5	69.8	27	100	100	63	56	78	92	100	93	80	73	100	67	50	100	164	82
PPA-L7	23	5	3	173	41	174	41	91.4	34	100	100	100	100	93	100	100	100	100	93	93	93	65	100	119	87
PPA-L8	29	7	4	180	52	144	36	93	35	100	100	100	100	86	92	100	100	93	87	60	60	88	100	112	85
PPA-L9	23	4	3	114	28	102	27	92.9	31	100	100	94	88	88	100	100	100	100	100	47	60	58	100	88	73
PPA-L (mean)	25.67	5.22	3.44	124.22	37.78	126.78	38.00	89.61	33.44	99.11	100.00	93.22	91.78	91.78	95.56	98.56	95.56	96.22	83.67	86.67	77.00	79.44	98.33	99.96	81.49
PPA-L (SD)	2.31	1.31	0.68	38.08	12.21	73.27	16.53	7.81	2.83	2.51	0.00	11.41	13.55	6.56	3.50	4.09	6.99	6.41	16.61	18.49	16.21	17.10	4.71	31.46	6.07
PPA-G1	14	3	2	180	45	263	39	68.3	35	90	100	81	94	76	88	93	100	33	7	27	13	50	95	58	63
PPA-G2	22	4	3	130	29	123	36	78.9	35	100	100	88	81	83	92	93	100	73	53	93	47	95	100	70	74
PPA-G3	27	3	3	116	29	187	40	91.5	31	100	92	81	100	95	100	93	93	100	27	93	47	55	100	86	90
PPA-G4	29	8	4	99	31	63	58	93	36	100	100	88	100	100	96	100	93	100	100	87	73	95	95	65	86
PPA-G5	21	4	3	106	24	96	30	71.7	34	100	100	69	75	32	48	87	73	20	13	87	33	68	100	114	67
PPA-G6	23	4	4	180	74	156	34	84.3	34	100	100	94	100	100	88	100	93	100	73	87	60	68	90	41	47
PPA-G7	26	5	5	180	17	66	49	93.6	36	100	100	88	100	74	96	100	100	100	60	100	33	83	100	101	68
PPA-G8	28	5	3	180	36	145	60	93.5	35	100	100	88	100	86	100	100	93	93	33	100	60	43	95	77	83
PPA-G9	26	4	0	107	33	172	47	74.5	35	100	92	81	94	100	76	73	87	47	0	7	7	18	30	51	24
PPA-G10	28	5	3	NA	48	91	44	79.3	34	100	100	88	100	86	88	87	53	67	27	7	7	50	85	30	13
PPA-G (mean)	24.40	4.50	3.00	142.00	36.60	136.20	43.70	82.86	34.50	99.00	98.40	84.60	94.40	83.20	87.20	92.60	88.50	73.30	39.30	68.80	38.00	62.50	89.00	69.14	61.47
PPA-G (SD)	4.32	1.36	1.26	34.91	15.17	58.76	9.41	9.18	1.36	3.00	3.20	6.55	8.63	19.39	14.73	8.16	14.07	29.04	30.10	36.74	22.25	23.07	20.22	24.66	24.51
sig.	n.s.	n.s.	n.s.	n.s.	n.s.	n.s.	n.s.	n.s.	n.s.	n.s.	n.s.	0.012	n.s.	n.s.	n.s.	n.s.	n.s.	n.s.	0.007	n.s.	0.002	n.s.	n.s.	0.045	n.s.

Note. AC = auditory comprehension; ASPT = Argument Structure Production Test; BNT = Boston Naming Test; DSF = Digit Span Forward; DSB = Digit Span Backward; MMSE = Mini-Mental State Examination; NAT = Northwestern Anagram Test; NAVI = Northwestern Assessment of Verb Inflections; NAVS = Northwestern Assessment of Verbs and Sentences; NNB = Northwestern Naming Battery; PPVT = Peabody Picture Vocabulary Test; TMT = Trail Making Test; WAB = Western Aphasia Battery; WPM = Word Per Minute. Can = canonical; noncan = non-canonical, synt corr sent = syntactically correct sentences produced during the re-telling of Cinderella’s story; trans = transitive verbs.

Individuals with PPA showed evidence of mild to moderate aphasia on the Western Aphasia Battery - Revised Version (WAB-R, Kertesz, 2006), as indicated by the WAB Aphasia Quotient, which did not differ between PPA-G and the PPA-L groups (Table 2). Single-word comprehension on the Northwestern Naming Battery (NNB, Thompson and Weintraub, 2014) was at ceiling for both nouns and verbs in both PPA groups, as also indicated by the Peabody Picture Vocabulary Test (PPVT; Dunn and Dunn, 2006). Noun naming accuracy on the NNB and the BNT (Kaplan et al., 2001) ranged from mild to moderate, with PPA-G and PPA-L showing similar performance on both tests. Verb production accuracy on the NNB was lower in the PPA-G than in the PPA-L group (p = .012), with similar trends observed on the Northwestern Assessment of Verbs and Sentences (NAVS, Thompson, 2011), in which production of transitive verbs in isolation (Verb Naming Test, VNT) and within a sentence context (Argument Structure Production Test, ASPT) was numerically lower in PPA-G than in PPA-L. Production of tense morphology on the Northwestern Assessment of Verb Inflections (NAVI, Thompson and Lee, 2017) was also numerically (albeit not statistically) lower in PPA-G than PPA-L for both finite and non-finite verbs. As expected, based on the classification criteria (see Mesulam et al., 2009), individuals with PPA-G were significantly more impaired than the PPA-L group in production of non-canonical sentences on the NAVS (Sentence Production Priming, SPPT, p = .007) and on the Northwestern Anagram Test (NAT, Thompson, Weintraub, & Mesulam 2012 (Thompson et al., 2012c); p = .002). Conversely, production of canonical sentences on both tests and comprehension of both canonical and non-canonical sentences on the NAVS Sentence Comprehension Test (SCT) were similar between groups. Analyses of spontaneous speech samples collected through the re-telling of Cinderella’s story, conducted using the Northwestern Narrative Language Analysis (Hsu and Thompson, 2018; Fromm et al., 2020), indicated that fluency ranged between 30 and 164 words per minute (WPM) and was lower in PPA-G than in PPA-L groups (p = .045); in addition, individuals with PPA-G produced a numerically (albeit not statistically) lower percentage of syntactically accurate sentences.

Individuals with PPA were also screened for non-language cognitive deficits: performance on the Mini-Mental State Examination (Folstein et al., 1975) did not differ between PPA-G and PPA-L groups and was either within normal limits (>26) or mildly impaired in all participants but one (PPA-G1, Table 2), who evinced moderate impairment. No differences between the two PPA groups were found on tests of verbal short-term and working memory (Forward and Backward Digit Span, Wechsler Memory Scale - Version III, Wechsler, 1997), visuo-spatial attention (Shape Cancellation Test, Weintraub & Mesulam, 1985), or executive attention (Trail Making Test, Reitan, 1971).

All participants passed a pure-tone audiometric screening prior to taking part in the experiment, to ensure adequate perception of the auditory stimuli, and all provided written informed consent prior to the study according to Northwestern University Institutional Review Board policies.

2.2. Experimental stimuli

The auditory sentence acceptability judgment task included three conditions: a semantic condition, a morphosyntactic condition, and a verb argument structure (VAS) condition. Each included an equal number of acceptable and violated sentences (n = 62/each for the semantic and VAS conditions, n = 50/each for the morphosyntactic condition, for a total of 348 sentences). All sentences were presented in the active, present progressive form, with was/were as auxiliaries. Sentences in the semantic condition began with a proper name (noun phrase, NP), followed by the verb, an NP direct object and a prepositional phrase (PP), that served as an adjunct (e.g., Owen was carving pumpkins at the party). Sentences in the morphosyntactic and VAS conditions had a similar structure but no direct object (NP-V-PP, e.g., Ryan was eating on the couch, see Table 3 for examples of the experimental items included in each condition). All sentences were recorded by a female native English speaker at a normal rate of speech (speech rate: 4.1 syll/sec) and digitized at 44.1 Hz; the onset and duration of each word was extracted from Praat (Boersma, 2001). An additional 12 sentences (4 per condition) were created and used for practice.

Table 3.

Examples of the stimuli used in the sentence acceptability task: bolded words mark the critical regions for ERP analysis.

Violation Condition	Grammatical Sentences	Violated Sentences
Semantic	Owen was carving pumpkins at the party.	Owen was mentoring pumpkins at the party.
	Fiona was mentoring students at the conference.	Fiona was carving students at the conference.
Verb-argument structure	Ryan was eating on the couch.	Ryan was devouring on the couch.
Morpho-syntactic	The hiker was camping on the mountain.	The hikers was camping on the mountain.
	The actors were singing in the theater.	The actor were singing in the theater.

For the semantic condition, 62 transitive verbs were paired to form 31 unique verb pairs (e.g., mentoring - carving). Semantic violations were created by replacing the verb in each acceptable sentence (e.g., Fiona was mentoring students at the conference) with the other verb in the pair (e. g., *Fiona was carving students at the conference), resulting in a semantic incongruity between the verb and its direct object, which also coincided with a violation of the verb thematic requirements.⁴ Within each pair, each verb appeared once in a semantically appropriate sentence context and once in a sentence containing a semantic violation (e.g., Fiona was mentoring students at the conference; *Owen was mentoring pumpkins at the party).

For the morphosyntactic condition, 50 simple active sentences, all containing an intransitive unergative verb and beginning with a singular (N = 25) or a plural NP (N = 25), were generated; violations (N = 50) were obtained by replacing the auxiliary in each acceptable sentence (e. g., The actors were singing in the theatre) with one of different number (e. g., *The actors was singing in the theatre).

For the VAS condition, 31 unique verb pairs were created so that one verb in each pair was an optionally transitive verb (i.e., a verb that selects for an internal argument (i.e., direct object) that may be legally omitted, e.g., eat), and one was an obligatory transitive verb (i.e., a verb that selects for an internal argument (i.e., direct object) that must be present in the sentence, e.g., devour) that could be used in a similar semantic context. VAS violations were realized by replacing the optionally transitive verb (e.g., Ryan was eating on the couch) in each acceptable sentence with the obligatory transitive verb in the pair (e.g., *Ryan was devouring on the couch), so that the VAS requirements for obligatory transitive verbs (e.g., devour) were violated. Notably, due to the experimental manipulation performed to create VAS violations, acceptable (n = 62) and violated (n = 62) sentences in this condition contained different verbs.

Prior to the study, sentence acceptability judgements on all the 348 sentences were collected from 13 healthy young adults (11 F; mean age: 22.7 ± 3). Sentences were presented in a pseudorandomized order on a Dell Desktop computer running OpenSesame version 2.9 (Mathôt et al., 2012). Participants were instructed to listen to each sentence and decide if the sentence was acceptable or unacceptable in standard English by pressing a pre-specified key on a QWERTY keyboard (“N” key = acceptable; “C” key = unacceptable). Sentence pairs with accuracy <80% (combined across acceptable and violated sentences) were excluded (n = 24; 12 from the semantic and 12 from the VAS condition; none from the morphosyntactic condition).

The final version of the experiment included 300 sentences (n = 100 per condition); acceptable (n = 50) and violated (n = 50) sentences did not differ in sentence length (in milliseconds (ms), semantic: t (98) = −0.579, p = .564; morphosyntactic: t (98) = −1.579, p = .117; VAS: t (98) = −0.841, p = .402) or speech rate (i.e., syll/sec, semantic: t (98) = 0.657, p = .513; morphosyntactic: t (98) = 1.758, p = .082; VAS: t (98) = −1.190, p = .237). Verbs used in acceptable and violated sentences within the VAS condition were matched for the log-transformed frequency of usage (acceptable: 4.56 ± 0.58, violated: 4.41 ± 0.6; t (98) = 1.258, p = .211) based on the Corpus of Contemporary American English (COCA, https://corpus.byu.edu/coca/).

2.3. Procedure

2.3.1. Experimental task

A sentence acceptability judgment task was performed by all participants on a Dell Desktop computer running OpenSesame 2.9. Participants were seated in a comfortable chair in a dimly lit sound-proof chamber, facing a computer monitor, and were instructed to listen to sentences and indicate if they were acceptable in standard English. A short practice block (n = 10 sentences, with examples from each condition) followed. During the experiment, sentences were presented in a pseudorandomized order with the following constraints: 1) sentences requiring the same response (e.g., acceptable) never appeared more than three times in a row; 2) acceptable and violated versions of the same sentence were separated by at least nine items, and 3) sentences belonging to the same condition never appeared more than three times in a row. To control for order effects, two different pseudorandomized orders (each fulfilling the conditions stated above) were created, and each participant was randomly assigned to one order. Each trial began with the presentation of a fixation cross in the center of the screen for 500 ms followed by a sentence presented auditorily over speakers (Fig. 1). The cross remained on the screen for the duration of the sentence (Mean = 2497 ms). Next, a cartoon face with a question mark (duration = 2000 ms) prompted participants to provide their response either using a keyboard (YA group only, see Section 2.2) or using a 5-key button box (OA and PPA groups). Participants were instructed to respond only when the cartoon face was present on the screen, in order to avoid motion artifacts in EEG recording during sentence presentation. Trials were separated by an interstimulus interval (ISI) with jittered duration (Mean = 1500 ms). Breaks were inserted into each list after every 50 items.

Fig. 1.

Schematic representation of an experimental item for the auditory grammaticality judgment task.

2.3.2. EEG procedures

2.3.2.1. Recording.

Data recording took place in a single session, lasting about 1.5 h, including set-up time. EEG data were recorded using a 32-electrode cap with active electrodes (Brainvision Acti-cap), amplified with a 100 Hz low-pass filter and digitized at 500 Hz. Scalp electrodes were placed according to the International 10–20 system at the following locations (FP1, FP2, F7, F3, Fz, F4, F8, FC5, FC1, FC2, FC6, T7, C3, Cz, C4, T8, CP5, CP1, CP2, CP6, P7, P3, Pz, P4, P8, O1, Oz, O2). Additional electrodes were positioned on the outer canthi of each eye (1 cm above or below midline, following Spriggs (2009)) to monitor for eye movements and blinks. Electrode impedances were kept below 15 KΩ (Mean = 8.5 KΩ). EEG was referenced online to the left mastoid, and re-referenced offline to the average of the left and right mastoids. All electrical equipment (computers, monitors) were connected through a power conditioner (Furman PL-8C) to reduce 60 Hz line noise. Prior to starting the experiment, participants were familiarized with the sensitivity of EEG to artifacts and were encouraged to find a comfortable position in the chair that allowed them to feel as relaxed as possible. Throughout the duration of the experiment, individuals were instructed to keep their eyes on the fixation cross and avoid blinking while listening to sentences in an attempt to reduce eye movement-related artifacts.

2.3.2.2. Data pre-processing.

EEG data were processed with EEGLAB (14.1.1) and ERPLAB (6.1.3) running in MATLAB R2017b. After removing 60 Hz noise (using Cleanline plugin 1.03), raw EEG data were high-pass filtered at 1 Hz (Winkler et al., 2015), manually screened for large artifacts and subjected to independent component analysis (ICA) using AMICA v.1.5. ICA weights were then applied to an unfiltered copy of the data and components reflecting eye-movements or other artifacts were removed. Next, a band-pass (0.1–30 Hz) IIR Butterworth filter (4th order roll-off) was applied. For the semantic and morphosyntactic conditions, participants’ EEG were time-locked to the onset of the word signaling a violation; target epochs extended from 0 to 1000 ms post-stimulus onset, with a 200 ms baseline. For the VAS violation condition, the EEG was time-locked to the onset of the verb; target epochs extended from 0 to 1800 ms post-stimulus onset (with a 200 ms baseline) to account for potential lexical differences between the obligatory versus optionally transitive verbs. After removing epochs with amplitude greater than ±75 μV, ERPs were computed by averaging the remaining data separately for each sentence type (grammatical, violation) for each condition at each electrode, for each participant.

2.3.3. Data analysis

2.3.3.1. Behavioral data analysis.

Demographic, neuropsychological, and language measures were compared across groups using Wilcoxon rank sum test (with p < .05) running in R 3.5.2. (R Core Team, 2018). Principal components analyses (PCAs) were conducted on the correlation matrices for all of the language measures in Table 2, using the prcomp function available in the R stats package. Only components that altogether accounted for at least 80% (with each accounting for at least 5%) of the total variance were interpreted and entered in further analyses. The contribution of each language measure to each PCA component was calculated by using the following formula, where L is the loading for the language measure i, SD is the standard deviation, and n indicates the total number of language measures:

$$\frac{{\left(L_i*S{D}_i\right)}^2}{{\sum }_{k=i}^n{\left(L_k*S{D}_k\right)}^2}$$

Components were interpreted by identifying the top language score (s) that individually accounted for at least 10% (and altogether accounted for at least 40%) of the component (Table 4).

Table 4.

Principal Component Analysis variance decomposition and loadings for each of the tests entered in the analysis. Values indicate the percentage of the variance in each component that is accounted for by performance on each test. Bolded values indicate the tests that loaded most on each component and that were used for interpretation of each component.

		PC1	PC2	PC3	PC4
SD		2.69	1.84	1.31	0.92
Prop_variance		0.45	0.21	0.11	0.05
Cum_variance		0.45	0.66	0.77	0.82
Loadings
BNT		0.35	19.61	8.78	4.86
PPVT		0.02	16.03	22.05	3.20
NNB	verb production	5.59	10.76	0.01	2.39
NAVS	VNT (transitive verbs)	2.70	11.67	18.17	0.16
	ASPT (transitive verbs)	6.72	2.84	11.87	4.75
	SCT (canonical sentences)	8.79	1.98	0.50	3.10
	SCT (non-canonical sentences)	5.05	0.29	2.02	45.89
	SPPT (canonical sentences)	10.53	1.39	4.92	1.67
	SPPT (non-canonical sentences)	10.27	0.02	0.81	9.95
NAT	canonical sentences	7.13	3.66	5.48	1.34
	non-canonical sentences	10.41	0.44	0.05	5.94
NAVI	finite verbs	6.88	0.01	15.31	7.17
	non-finite verbs	5.66	5.46	6.71	0.00
Cinderella	prop syntactically correct sentences	8.46	3.48	3.04	7.23
	WPM	2.34	18.75	0.19	0.17
WAB	Aphasia Quotient (WABAQ)	9.08	3.61	0.09	2.19

Note. ASPT = Argument Structure Production Test; BNT = Boston Naming Test; NAT = Northwestern Anagram Test; NAVI = Northwestern Assessment of Verb Inflections; NAVS = Northwestern Assessment of Verbs and Sentences; NNB = Northwestern Naming Battery; PPVT = Peabody Picture Vocabulary Test; SCT = Sentence Comprehension Test; SPPT = Sentence Production Priming Test; VNT = Verb Naming Test; WAB = Western Aphasia Battery.

Performance on the experimental task was analyzed by computing a measure of sensitivity (d-prime) developed by the field of signal detection theory (Macmillan and Creelman, 2004), which can be estimated from the observed rate of hits (i.e., acceptable sentences judged as acceptable) and false alarms (i.e., violated sentences judged as acceptable). For each condition and each participant, d-prime scores (computed using the psycho R package) were entered into a regression analysis with group, condition, and the group*condition interaction introduced as predictors. Statistical significance of each predictor was determined by comparing models with and without a given predictor, using the anova function. Significant effects were followed with planned post-hoc comparisons conducted using the mcp function in the multcomp package, and by applying FDR (False Discovery Rate) correction for multiple comparisons.

2.3.3.2. ERP data analysis.

For healthy individuals, time windows for data analysis were selected based on the existing literature on the N400 and P600 components and on visual inspection of the data: 300 ms–800 ms and 500 ms–1000 ms following the onset of the critical word (i.e., the word signaling a violation) for the semantic and the morphosyntactic condition, respectively, and 1000 ms–1500 ms following onset of the main verb for the VAS condition. Notably, the late time window selected for the VAS condition accounted for the time elapsed (Mean = 530 ms) between the onset of the main verb and the onset of the critical word (i.e., the preposition), and corresponded to an average of 470 ms–970 ms following the onset of the preposition. Time windows for the analysis of ERP data obtained from the PPA group were selected based on the latency and duration of each component in the OA group. A subset of 21 electrodes were selected for statistical analyses and grouped into five regions: left anterior (F3, F7, FC1, and FC5), right anterior (F4, F8, FC2, and FC6), left posterior (CP1, CP5, P3 and P7), right posterior (CP2, CP6, P4 and P8), and midline (Fz, Cz, Pz, C3 and C4). Within the pre-selected time windows, mean amplitude at each electrode was extracted, from consecutive 100 ms time bins for each participant and sentence type (acceptable, violated), and entered into mixed-effect regression analyses, with time bin, electrode region, and sentence type as fixed effects and participant and electrode as random effects. Separate regression analyses were run for each condition and each participant group. First, the model containing all the fixed effects and their interactions was implemented, and the best-fit model was determined by using the automated backward elimination procedure included in the lmerTest package (Kuznetsova et al., 2017). In the presence of significant interactions, planned comparisons were run and p-values were corrected using FDR correction for multiple comparisons. Only components that were sustained for at least 200 ms (i.e., two consecutive time bins) were considered as significant, entered in further analyses, and interpreted. Within time windows for which significant ERP components were identified and within the electrode region where the effects were largest, difference waves were computed for each participant by subtracting the mean amplitude for acceptable sentences from the mean amplitude for violated sentences, for the electrode showing the largest effect in each individual. Difference waves for each participant were then entered into simple regressions with group (YA, OA, PPA-G and PPA-L) as the only fixed effect, and time bin as the only random effect, to investigate between-group differences in mean amplitude of the ERP components. For individuals with PPA, additional regression analyses were run - after combining individuals in a single group - to investigate the relation between the amplitude of the identified ERP components and accuracy on the experimental task (d-prime score), the reported duration of their disease, and the PCA components derived as in 2.3.3.1.

3. Results

Across all groups and all conditions, an average of 3.85% ± 5.04 trials were rejected (3.04% ± 4.11 from correct and 3.2% ± 3.95 from violated sentences). More trials were removed from the VAS (M = 5.37% ± 6.1) than from the other conditions (semantic: M = 3.14% ± 4.25, p = .045; morphosyntactic: M = 2.87% ± 4.11, p = .03), and more trials were removed from the group of young adults (M = 4.95% ± 5.6) than from the PPA group (M = 2.58 ± 3.73, p = .018).

Three participants from the YA and three from the OA group were excluded from analysis because more than 20% of their EEG data were affected by artifacts. The total number of participants included in the analyses was 54 (YA: N = 19; OA: N = 16; PPA: N = 19). Due to time constraints, some individuals with PPA (4 PPA-G and 5 PPA-L) were run on a short version of the experiment that omitted the morphosyntactic violation condition.

3.1. Principal component analysis (PCA)

Results of the PCA, namely, the variance accounted for by the first four principal components (summing up to 82.5% of the total variance) and the contribution of each language measure to each PCA component, are reported in Table 4.

Performance on the NAVS SPPT (canonical and non-canonical sentences) and NAT (non-canonical sentences only) were the major contributors to the first component (PC1). Given that poor performance on the non-canonical sentences included in the NAVS SPPT and NAT is the primary criteria for a PPA-G diagnosis, this component was considered as an index of agrammatic production. However, it should be noted that these factors only accounted for 31% of PC1, and that WAB-AQ also accounted for 9%, suggesting that PC1 may also reflect aphasia severity. The variables that most contributed to PC2 were the number of words per minute (WPM), which loaded positively, and performance on the BNT, which loaded negatively, on the component. Given that fluent speech (i.e., higher WPM) and poor performance on the BNT are frequently found in PPA-S, PC2 was considered as an indicator of lexical-semantic impairment. Performance on the NAVS VNT (transitive verbs, loading positively) and on the PPVT (loading negatively) were the major contributors to PC3, which was interpreted as reflecting impaired verb retrieval in the absence of word comprehension deficits. Finally, performance on the NAVS SCT (non-canonical sentences only) contributed alone to more than 40% of PC4, which was therefore considered as an index of syntactic comprehension deficits.

3.2. Experimental results

3.2.1. Behavioral data

Following inspection of accuracy and d-prime data, one participant in the YA group was considered as an outlier (d-prime < 0 in all three conditions) and therefore excluded from further analyses. Mean accuracy, d-prime scores and SD for each participant group are provided in Table 5. Regression analyses on d-prime scores revealed a significant group*condition interaction (F = 3.633, p = .003). Post-hoc analyses showed, for all conditions, decreased sensitivity in both PPA groups compared to the OA group (semantic condition: PPA-G: z = −5.534, p < .0001; PPA-L: z = −4.163, p < .0001; morphosyntactic condition: PPA-G: z = −2.499, p = .025; PPA-L: z = −2.347, p = .028; VAS condition: PPA-G: z = −3.623, p = .0006; PPA-L: z = −3.007, p = .004). No difference was found between individuals with PPA-L and individuals with PPA-G in any condition (all ps>.2; see Table 5). In addition, YA and OA groups did not significantly differ on either the semantic or the VAS condition (all ps > .2), but a marginally significant difference was found on the morphosyntactic condition, where OA performed more poorly than YA (p= .053; Table 5).

Table 5.

Percentage correct responses (and corresponding d-prime scores) on the experimental task, listed by condition. Statistical significance indicates: for older adults, the comparison with young adults; for the PPA-L group, the comparison with older adults; for the PPA-G group, the comparison with older adults and with the PPA-L group, respectively. Symbols indicate p-values ranging between 0.05 and 0.1 (~), from 0.01 to 0.05 (*), from 0.001 to 0.01 (**) and smaller than 0.001 (***).

	Semantic condition		Morphosyntactic condition		Verb-argument structure condition
	Accuracy (%correct)	D-prime	Accuracy (%correct)	D-prime	Accuracy (%correct)	D-prime
Young adults	95.8 ± 20.0	3.40 ± 0.52	95.4 ± 21.0	3.33 ± 0.69	79.3 ± 40.6	2.10 ± 0.83
Older adults	92.3 ± 26.7	2.97 ± 0.68 (ns)	83.6 ± 37	2.46 ± 1.49 (~)	77.9 ± 41.5	1.81 ± 0.83 (ns)
PPA-L	68.4 ± 21.9	1.1 ± 1.22 (***)	64.3 ± 14.9	0.81 ± 0.95 (~)	61.1 ± 13.4	0.70 ± 0.73 (*)
PPA-G	58.7 ± 29.1	0.56 ± 1.93 (***, ns)	64.2 ± 29.8	0.96 ± 1.94 (~, ns)	58.4 ± 20.5	0.52 ± 1.18 (**, ns)

3.2.2. Event-related potentials

3.2.2.1. Semantic condition

3.2.2.1.1. Healthy participants.

Individuals in both the YA and the OA groups showed a significant effect of sentence type, with mean amplitude for sentences containing a semantic violation more negative than for semantically acceptable sentences (YA: t = −20.7, p < .0001; OA: t = −16.3, p < .0001). The negativity (N400) elicited by semantic violations was significant in all time bins within the 400–800 ms time window for the YA group and within the 500–800 ms time window for the OA group. The N400 effect reached its maximum amplitude in the 500–600 ms window in the YA, and in the 600–700 ms window in the OA, group. A significant sentence type*region interaction emerged in both groups (YA: F = 4.048, p = .003; OA: F = 2.27, p = .059). Follow-up analyses indicated that the effect was larger along the midline (t = −12.147, p < .0001) and on right posterior electrodes (t = −10.206, p < .0001) for the YA group, and along the midline (t = −8.749, p < .0001) and on left posterior electrodes (t = −8.232, p < .0001) in the OA group (see Fig. 2(a and b)). Between-group comparisons performed on the difference waves (semantic violation - acceptable) indicated no significant differences in mean amplitude between the two groups (t = −0.749, ns).

Fig. 2.

Grand mean waveforms for correct sentences (black dashed line) and sentences containing a semantic violation (red), for each participant group ((a) young adults; (b) older adults; (c) PPA-G; (d) PPA-L). Examples of correct and violated sentences are provided, with bolded words indicating the word to which ERPs were time-locked (critical word). In each plot, onset of the critical word is marked by the vertical line, and time from onset (in milliseconds) is plotted on the x-axis, while voltage (in μV) is plotted on the y-axis. For each participant group, only the electrode with the largest effect is displayed. For visualization purposes, a 15 Hz low-pass filter was applied to the data.

3.2.2.1.2. Individuals with PPA.

Participants in both PPA-G and PPA-L groups showed a significant effect of sentence type within the 500–800 ms window, again indicating a significant N400 effect for semantic violations (PPA-G: t = −3.83, p = .0001; PPA-L: t = −4.797, p < .0001). The N400 effect was largest in the 600–700 ms window for the PPA-G group, and in the 500–600 ms window for the PPA-L group. Although the sentence type*region interaction was only marginally significant in the PPA-L (F = 2.31, p = .056) and not significant in the PPA-G (F = 1.385, ns) group, post-hoc comparisons indicated, for both groups, a significant N400 only on the electrodes along the midline (PPA-G: t = 2.382, p = .043; PPA-L: t = 3.666, p = .0006) and in the left posterior region (PPA-G: t = −2.838, p = .023; PPA-L: t = −3.691, p = .0006), as also shown in Fig. 2(c and d). Between-group comparisons indicated that the mean amplitude was not significantly different across groups (OA vs. PPA-G: t = 0.812, ns; OA vs. PPA-L: t = 0.463, ns; PPA-G vs PPA-L: t = 0.293, ns). N400 amplitude was larger in individuals with higher d-prime scores (t = −4.114, p = .0001) and in individuals who reported a shorter duration of the disease (t = 3.241, p = .002).

3.2.2.2. Morphosyntactic condition

3.2.2.2.1. Healthy participants.

Both YA and OA individuals showed a significant effect of sentence type within the pre-selected time window, with mean amplitude for sentences containing a subject-verb agreement violation more positive than for acceptable sentences (YA: t = 22.974, p < .0001; OA: 36.907, p < .0001). For both groups, the positivity (P600) elicited by morphosyntactic violations was significant in all time bins in the 500–1000 ms window, and largest in the 600–700 ms window. A significant sentence type*region interaction was found in both groups (YA: F = 35.385, p < .0001; OA: F = 19.999, p < .0001). Post-hoc comparisons revealed that the effect was larger along the midline and on posterior electrodes in both groups (Fig. 3(a and b)). Between-group comparisons on the difference waves (morphosyntactic violation - acceptable) indicated greater amplitude of the P600 in the OA than in the YA group (t = 2.536, p = .012).

Fig. 3.

Grand mean waveforms for correct sentences (black dashed line) and sentences containing a morphosyntactic violation (red), for each participant group ((a) young adults; (b) older adults; (c) PPA-G; (d) PPA-L). Examples of correct and violated sentences are provided, with bolded words indicating the word to which ERPs were time-locked (critical word). In each plot, onset of the critical word is marked by the vertical line, and time from onset (in milliseconds) is plotted on the x-axis, while voltage (in μV) is plotted on the y-axis. For each participant group, only the electrode with the largest effect is displayed. For visualization purposes, a 15 Hz low-pass filter was applied to the data.

3.2.2.2.2. Individuals with PPA.

For the PPA-G group, no main effect of sentence type was found (t = 1.573, ns; Fig. 3(c)). Although the interaction sentence type*time bin was significant (F = 4.489, p = .001), post-hoc comparisons revealed only a marginally significant difference within the 800–900 ms time bin, in the direction of a positivity for violated (compared to acceptable) sentences (t = 2.068, p = .097). This positivity, however, did not continue in the following time window (900–1000 ms), where morphosyntactic violations were associated with a negativity for grammatically correct sentences (t = −3.039, p = .012). A significant effect of sentence type was found in the PPA-L group within the pre-defined time window (t = 6.063, p < .0001), in the same direction as for the healthy participants, i.e., mean amplitude was more positive for violated than for acceptable sentences (Fig. 3(d)). A significant sentence type*time bin interaction was found (F = 7.72, p < .0001), with post-hoc comparisons indicating that the positivity (P600) was found only within the 600–1000 ms time window, and was largest in the 800–900 ms time window. Post-hoc comparisons run following a significant sentence type*region interaction (F = 8.864, p < .0001) indicated that the effect was restricted to the midline and posterior electrodes. A significant difference in P600 amplitude was found between the OA and both PPA groups (PPA-G: t = −3.234, p = .0037; PPA-L: t = −2.423, p = .0237). Analyses within the PPA group showed a positive relation between ability to detect morphosyntactic violations and P600 amplitude (t = 6.599, p < .0001).

3.2.2.3. Verb-argument structure (VAS) condition

3.2.2.3.1. Healthy participants.

Individuals in the YA and OA groups showed a significant effect of sentence type within the pre-selected time window, i.e. 1000–1500 ms from verb onset (corresponding to an average of 470–970 ms from onset of the post-verbal preposition that marked the violation). In both groups, mean amplitude was significantly more positive for violations than for acceptable sentences (YA: t = 7.527, p < .0001; OA: t = 4.362, p < .0001, Fig. 4(a and b)). A significant sentence type*time interaction emerged in both groups (YA: F = 4.934, p = .0006; OA: F = 4.139, p = .0024), with post-hoc comparisons indicating that the P600 was significant only within the 470–770 ms window from violation onset in both groups and largest in the 570–670 ms window for YA (t = 6.072, p < .0001) and in the 470–570 ms window for OA (t = 4.237, p < .0001). Sentence type significantly interacted with region only in the OA group (F = 2.693, p = .029), revealing a significant P600 effect on the midline and posterior electrodes (midline: t = 3.901, p = .0002; left posterior: t = 3.545, p = .0007; right posterior: t = 3.925, p = .0002). Although no significant sentence type*region interaction was found in the YA group, planned post-hoc comparisons indicated that the P600 was larger on midline and posterior electrodes (midline: t = 5.552, p < .0001; left posterior: t = 4.859, p < .0001; right posterior: t = 3.922, p = .0001) than on anterior electrodes (left anterior: t = 2.533, p = .012; right anterior: t = 2.512, p = .012). Between-group analyses showed no difference in amplitude between YA and OA groups (t = 1.506, ns).

Fig. 4.

Grand mean waveforms for correct sentences (black dashed line) and sentences containing a verb-argument structure (VAS) violation (red), for each participant group ((a) young adults; (b) older adults; (c) PPA-G; (d) PPA-L). Examples of correct and violated sentences are provided, with bolded words indicating the word to which ERPs were time-locked (critical word). In each plot, onset of the critical word is marked by the vertical line, and time from onset (in milliseconds) is plotted on the x-axis, while voltage (in μV) is plotted on the y-axis. For each participant group, only the electrode with the largest effect is displayed. For visualization purposes, a 15 Hz low-pass filter was applied to the data.

3.2.2.3.2. Individuals with PPA.

Within the 1000–1300 ms time window from verb onset (470–770 ms from violation onset), selected based on the significant time window in the OA group, the PPA-G group showed a significant effect of sentence type with opposite polarity than healthy controls, (i.e., a negativity for VAS violations, t = −2.923, p = .0155, Fig. 4(c)). Significant sentence type*time (F = 5.196, p = .0057) and sentence type*region (F = 3.102, p = .0149) interactions indicated, however, that this negativity was brief (i.e., significant only in the 470–570 ms time window from violation onset: t = −3.845, p = .0004; no significant difference in other time windows) and localized on the anterior electrodes (left anterior: t = −2.365, p = .0185; right anterior: t = −3.040, p = .0025; all other ps > .15). Turning to the PPA-L group, no significant effect of sentence type was found (t = 0.990, ns, Fig. 4(d)). Although the sentence type*time interaction was marginally significant (F = 2.867, p = .0573), post-hoc comparisons conducted in each time window failed to reach significance (all ps > .05). Despite the finding of a significant P600 in the healthy, but not in the PPA, groups, between-group comparisons showed no significant differences in mean P600 amplitude between the OA and the PPA groups (OA vs. PPA-G: t = 1.644, ns; OA vs. PPA-L: t = 1.061, ns). Given that no P600 was found in either PPA group in the same time window as healthy individuals, an additional set of analyses was run in the following time window (1300–1700 ms from verb onset, equivalent to an average of 770–1170 ms from violation onset). In this time window, as in the previous, individuals with PPA-G did not show any significant effect of grammaticality (t = 0.490, ns, Fig. 4(e)). Despite the finding of a significant sentence type*time interaction (F = 3.406, p = .017), post-hoc comparisons did not show any significant differences in mean amplitude between acceptable and violated sentences in any of the time windows (all ps > .1). For the PPA-L group, analyses revealed a significant effect of sentence type, in the same direction as the healthy control groups, i.e., more positive amplitude for violated than for acceptable sentences (t = 4.088, p < .0001). Following a significant sentence type*time interaction (F = 3.979, p = .0078), post-hoc comparisons indicated that the positivity was marginally significant in the 870–970 ms time window (t = 1.848, p = .0864) and significant in the 970–1170 ms time windows (t = 3.863, p = .0005 and t = 3.140, p = .0034, respectively), with a more widespread distribution than in healthy individuals (see Fig. 4(f), all ps < .05 except for posterior right electrodes). Unlike the semantic and morphosyntactic conditions, the amplitude of this VAS positivity was not related to behavioral performance (d-prime score) on the task (t = −0.998, ns). Furthermore, P600 amplitude was larger for individuals who reported longer duration of the disease (t = 3.19, p = .0029).

3.2.3. Relation between ERP components and language deficits in PPA

Analyses showed a significant relation between N400 amplitude to semantic violations and PC1, in the direction of greater N400 for individuals with a greater syntactic production impairment (i.e., worse agrammatism, t = 4.922, p < .0001). No relation was found between N400 amplitude and PC2. P600 amplitude to morphosyntactic violations was also significantly predicted by the PC1 component, with smaller amplitudes found in individuals with greater syntactic production deficits (t = −4.228, p = .0001). A significant relation was also found between P600 amplitude to morphosyntactic violations and PC3 (t = 4.046, p = .0002), which indicated smaller P600 amplitude in individuals with greater impairment in verb retrieval. Turning to VAS violations, P600 amplitude in the earlier time window (470–770 ms from violation onset) - albeit not reaching significance at the group level - was positively related to PC3 (t = 3.440, p = .0011), suggesting that individuals with better preserved verb retrieval displayed trending positivities. Interestingly, this relation was not found in the following time window (770–1170 ms from violation onset; t = 1.523, ns). Rather, the amplitude of the late positivity was significantly related to PC4 (t = 2.062, p = .0465), with greater syntactic comprehension impairments being associated with larger positivities.

4. Discussion

The present study aimed to investigate online semantic and syntactic processing in two groups of individuals with Primary Progressive Aphasia (PPA), a group with the agrammatic (PPA-G) and a group with the logopenic (PPA-L) subtype. For this purpose, event-related potentials (ERPs) were recorded while individuals with PPA and groups of young and older healthy participants listened to sentences that contained semantic, morphosyntactic (i.e., subject-verb agreement), or verb-argument structure (VAS) violations.

4.1. Healthy individuals

Both young and older adults showed, as expected, an N400 to semantic violations, in line with previous studies (see Kutas and Federmeier, 2011, for a review). The N400 has been suggested to reflect ease of access to semantic information from memory (Kutas and Federmeier, 2000; DeLong et al., 2005), or integration of the meaning of a word with the preceding context (see Osterhout and Holcomb, 1995). More recently, studies have proposed dynamic accounts in which the N400 reflects, in a sequential, cascaded fashion, context-dependent pre-activation of semantic features in memory (i.e., prediction) and compositional semantic processes (semantic unification/integration; see Baggio and Hagoort, 2011; Fleur et al., 2020; Nieuwland et al., 2020). In our study, semantic violations (e.g., *Owen was mentoring pumpkins at the party) were obtained by replacing the verb with another that was semantically incongruent with the remainder of the sentence context (i.e., violated the verb thematic requirements). Following recent dynamic accounts (Baggio and Hagoort, 2011; Fleur et al., 2020; Nieuwland et al., 2020), the N400 found in healthy participants may index both the (un-predictability) of the verb direct object (i.e., pumpkins) and the semantic unification/integration processes that are triggered by it in the attempt to integrate the meaning of a possible argument with the thematic requirements of the verb. This latter interpretation is in line with previous studies investigating animacy violations (Friederici et al., 1998; Nieuwland and Van Berkum, 2006; see also Nieuwland et al., 2013).

While the N400 is typically observed between 300 ms and 600 ms and peaks around 400 ms following the onset of the critical word (see Friederici et al., 1998; Kutas and Federmeier, 2011), the N400 in this study had a later onset and peaked later (around 500 ms) than in previous studies. Although auditory N400s tend to last longer than visual N400s (Kutas and Federmeier, 2011), the delayed onset and peak latency found in this study may stem from the experimental stimuli that were employed: in many of the classical N400 studies semantic incongruities occurred on the sentence-final word or in the presence of a preceding discourse context (e.g., Nieuwland and Van Berkum, 2006; see reviews in Kutas and Van Petten, 1994; Osterhout and Holcomb, 1995); whereas, in our study, the context preceding the critical word was provided solely by the verb. Given that lexical access is delayed for verbs compared to nouns (Malyutina & den Ouden, 2015; Vigliocco et al., 2011; see Crepaldi et al., 2011 for a review), a delayed and weaker prediction made upon encountering the verb may result in an N400 with later onset and longer latency (compared to N400 responses resulting from predictions based on a semantically rich context). In support of this interpretation, longer N400 latencies have been reported for verbs (compared to nouns) in a lexical decision task (Gomes et al., 1997); in addition, late-onset N400s were found for actions performed on implausible objects and for violations of verb thematic restrictions (Proverbio and Riva, 2009; Rösler et al., 1993). The seemingly long duration of the N400 in our study (up to 800 ms post-onset) is in line with recent findings (Nieuwland et al., 2020) that show extended duration of N400 components reflecting plausibility (integration) effects. Importantly, in spite of its atypical latency and duration, the N400 observed in healthy participants in our study showed the typical central-parietal distribution.

Violations of subject-verb agreement elicited, as expected, a P600 response that was significant starting from 500 ms until 1000 ms following the onset of the critical word (auxiliary) and largest in the 600–700 ms time window. Results are in line with numerous studies that reported a P600 in response to morphosyntactic violations (e.g., Coulson et al., 1998; Hagoort et al., 1993; Hagoort et al., 2003; Kaan and Swaab, 2003), which can be interpreted as reflecting processes of re-analysis and repair (Friederici et al., 2002; Hahne and Friederici, 1999; Kaan and Swaab, 2003).

VAS violations, which were obtained through a manipulation of the verb resulting in a missing direct object following an obligatory transitive verb (e.g., *Ryan was devouring on the couch), also elicited a P600 in both younger and older adults. This component was significant in the 1000–1300 ms window following verb onset, which corresponded to an average of 470–770 ms from the onset of the critical word (preposition) and was largest on midline and posterior electrodes, in line with the typical P600 distribution (Kaan and Swaab, 2003). The absence of an N400 to VAS violations is seemingly at odds with previous studies (e.g., Friederici and Frisch, 2000; Kielar et al., 2012); however, it should be noted that the experimental stimuli in the present study differed from previous research because violations consisted of a missing (vs. an additional) argument and because the word signaling a VAS violation was a preposition (vs. a noun phrase). Since N400s reflect either the (un-)predictability of upcoming words or the effort to integrate them with the preceding semantic context (see Discussion above), they may be observed when an additional argument is presented following an intransitive verb (e.g., in *Anne sneezed the doctor and the nurse, see Kielar et al., 2012) and when an attempt to integrate the meaning of the noun (recognized as a possible argument) with the meaning of the preceding verb is made. However, such an attempt may not be made upon encountering a post-verbal preposition, as this may signal the presence of a probable adjunct (rather than an argument). Alternatively, the findings may indicate that VAS violations realized by omitting an obligatory argument are processed the same way as violations of structural preference, which typically elicit only a P600 (Osterhout and Holcomb, 1993; Osterhout et al., 1994).

4.1.1. Age-related changes in sentence processing

The comparison between younger and older adults showed a delayed onset (but no reduced amplitude) of N400s to semantic violations in older (vs. younger) participants that is in line with previous studies (Kutas and Iragui, 1998; Wlotko et al., 2010). As suggested by some authors (Federmeier and Kutas, 2005; Federmeier et al., 2003; Kutas and Iragui, 1998; see Wlotko et al., 2010, for a review), this delayed response may reflect an age-related decline in the ability to use sentential context to facilitate word processing. Specifically, it may be indicative of less efficient prediction of the post-verbal argument or of difficulties in its integration with the verb thematic requirements, which however did not affect offline accuracy. Differences between young and older adults were also found in N400 scalp topographies: while the former showed a more right-lateralized response, N400 amplitude in the latter peaked on left posterior regions. Although N400 scalp topographies are often unstable (see Kutas and Federmeier, 2011), these findings - together with the evidence of delayed N400 onset in the older group - may indicate age-related changes in the neural networks supporting thematic integration.

Conversely, P600 responses to both morphosyntactic and VAS violations had similar onset, peak latency, and spatial distribution in both younger and older healthy adults, as seen in other studies (King and Kutas, 1995; Kemmer et al., 2004). However, P600 amplitude in response to morphosyntactic violations was larger in older (vs. younger) adults. Although unexpected, this finding is in line with a recent study suggesting that P600 amplitude may be larger in individuals with intermediate levels of cognitive control, compared to individuals with low or high cognitive control skills (Brothers et al., 2019; see also Kropotov et al., 2016, for evidence on other components). Enhanced P600 amplitude in aging cohorts may reflect increased engagement of neural resources during re-analysis/resolution of conflicting representations, as suggested by Brothers et al. (2019). This hypothesis is further supported by the (marginally significant) lower sensitivity to morphosyntactic violations found in older (vs. younger) healthy adults, which suggests age-related changes in morphosyntactic processing.

4.2. Individuals with PPA

In both PPA groups, semantic violations elicited an N400 with similar timing and spatial distribution as in older adults, but showed numerically (albeit not statistically) reduced amplitude. These findings suggest that the ability to predict verb arguments and/or integrate their meaning with the verb thematic requirements in PPA may be relatively, but not entirely, preserved, as also indicated by the reduced offline sensitivity to violations of the verb thematic requirements, as well as by the positive association between N400 amplitude and offline sensitivity to these violations. Contrary to our predictions, no relation was found between N400 amplitude and measures of lexical-semantic processing (e.g., PC2); rather, N400 amplitude was associated with the severity of sentence production deficits (as indicated by the first PCA component, PC1), i.e., larger ERP responses to semantic violations were observed in individuals with more severe agrammatic production. Although seemingly counterintuitive, this result may be interpreted as suggesting that individuals who are more clearly classifiable as PPA-G (in light of their performance on non-canonical sentence production on the NAT and NAVS SPPT, two factors that loaded heavily on PC1) also tend to show more preserved semantic processing, i.e., to show a more “pure” PPA-G profile (compared to individuals who present with a more ‘mixed’ profile). In keeping with these findings, N400 amplitude in PPA was larger in participants with longer (vs. shorter) duration of the disease, suggesting that, as the disease progresses, semantic processing deficits may appear in both PPA-G and PPA-L individuals.

These results are in line with previous ERP studies showing a normal-like N400 during semantic processing in individuals with PPA-G and PPA-L (Hurley et al., 2009, 2012), with studies showing intact sensitivity to thematic violations using online behavioral paradigms (Peelle et al., 2007; but see Price and Grossman, 2005, for opposite results), and with evidence coming from behavioral studies that reported largely intact semantic processing in both groups (e.g., Mesulam et al., 2009; Riley et al., 2018; Rogalski et al., 2011; Seckin et al., 2016). Furthermore, the materials employed in the semantic condition of this study were similar to the experiment which tested argument access in the eye-tracking study by Mack et al. (2019), where the authors found that both PPA-G and PPA-L were facilitated in accessing the semantic content of a noun (e.g., jar) based on the thematic restrictions posited by a verb (e.g., open), in a similar way as for healthy individuals. Altogether, these results suggest that both PPA-G and PPA-L have access to the verb thematic requirements and attempt to predict and/or integrate the post-verbal noun phrase with the verb, which results in an N400 when the verb thematic requirements are not met. However, based on the data reported in one of the experiments performed by Mack et al. (2019), for PPA-G, the ability to access verb thematic requirements and use that information to facilitate sentence processing may be dependent on the task demands, and may be impaired when the verb direct object needs to be retrieved.

Turning to morphosyntactic processing, in line with our predictions, subject-verb agreement violations elicited a P600 in the PPA-L but not in the PPA-G group. The absence of a P600 in PPA-G is consistent with studies that reported difficulties in production of verb morphology on standardized language tests and in connected speech (Thompson et al., 1997, 2012a, 2013; Wilson et al., 2010b; see Thompson and Mack, 2014, for a review), as well as with studies showing decreased online sensitivity to violations of tense and agreement (Grossman et al., 2005; Peelle et al., 2007). In keeping with these findings, and in line with our predictions, P600 amplitude was positively associated with offline sensitivity to morphosyntactic violations, and was smaller in individuals with more severe agrammatic production (as indexed by PC1) and with more pronounced deficits in verb retrieval (as indicated by the third PCA component, PC3), two of the most common symptoms of PPA-G. In individuals with PPA-L, the P600 elicited by subject-verb agreement violations had delayed onset and peak latency, reduced amplitude, but similar spatial distribution, compared to older adults. Taken together, these findings suggest that processes of re-analysis and repair in PPA-L may rely on the same neural sources as in older adults, but may be delayed and less efficient. Evidence for impaired production of verb inflections in PPA-L has been reported in a few studies (see Meteyard and Patterson, 2009; Wilson et al., 2010b, 2014), and ascribed to a ‘core’ phonological deficit (Wilson et al., 2014).

Similarly to morphosyntactic violations, VAS violations did not elicit a P600 in individuals with PPA-G. This finding is in line with our predictions and with several studies showing impaired VAS production in PPA-G both in standardized language tests and connected speech (Thompson et al., 2012a, 2012b, 2013). The PPA-L group also lacked a P600 to such violations in the same time window in which the effect was found in healthy individuals. Interestingly, P600 amplitude in this window was related to PC3, the PCA component reflecting impairment in verb retrieval, suggesting that - although no evidence of P600 was found in either group - individuals with better preserved verb retrieval showed trending positivities to these violations. The PPA-L group showed, however, a late positive component in the 770–1070 ms window from preposition onset, which displayed (compared to older adults) an atypical spatial distribution that included anterior regions. Amplitude of this late component was larger in individuals with a reportedly longer duration of the disease, as well as more impaired comprehension of syntactically complex sentences (indexed by PC4), and was not associated with verb retrieval as in the earlier window. Taken together, these findings suggest that this component may reflect inefficient processing of VAS violations, as also indicated by the lack of association with the d-prime score. Although the large majority of offline psycholinguistic studies has reported largely intact VAS production in PPA-L (Thompson et al., 2012a, 2012b, 2013), neuroimaging research has shown that cortical atrophy in PPA-L often affects the temporo-parietal junction (TPJ, Mesulam et al., 2009; Teichmann et., 2013), a region that is associated with processing of VAS information (e.g., the number of verb arguments (Den Ouden et al., 2009; Meltzer-Asscher et al., 2015; Thompson et al., 2007, 2010) in healthy individuals. Taken together, results obtained from this condition indicate, as predicted, that processing of VAS violations was impaired in both groups, but more so in PPA-G than in PPA-L. Namely, while both groups showed reduced offline sensitivity (vs. healthy adults) to VAS violations, individuals with PPA-L exhibited significant (albeit delayed and abnormally-distributed) ERP responses, suggesting that they may retain an online sensitivity to such violations, in spite of being unable to carry out the syntactic re-analysis and repair processes indexed by the P600 in healthy individuals. Conversely, in the PPA-G group, mean amplitude for acceptable and VAS-violated sentences did not differ throughout the entire epoch, a result that - in conjunction with the offline results - suggests a failure to detect violations of verb-argument structure.

Although the present study has the limitation of reporting findings from a small number of participants with PPA, we note that the statistical approach employed for the analysis of both behavioral and ERP data (i.e., mixed-effects regression) accounts for inter-individual variability and ensures that group-level results are not driven by single individuals. In addition, regression analyses conducted on the entire sample of individuals with PPA (N = 19) indicate strong relationships between ERP components and behavioral measures that distinguish PPA-G from PPA-L, thereby supporting the distinctive ERP signatures of sentence processing found in the two groups.

5. Conclusion

To our knowledge, this is the first study to investigate semantic, morphosyntactic, and verb-argument structure processing in primary progressive aphasia (PPA) using online, electrophysiological measures (ERP). Our findings indicate that sentence comprehension is substantially impaired in individuals with PPA-G. While the current diagnostic criteria for PPA-G are mostly focused on sentence production deficits, the results of this study elucidate the sentence comprehension deficits in this patient group by showing a dissociation between (relatively) spared semantic and impaired lexical-syntactic (argument structure) and morphosyntactic (subject-verb agreement) processing in comprehension. This pattern is consistent with the verb-argument structure and morphosyntactic production impairment found on standardized language tests and connected speech. In addition, this pattern replicates what was previously found in an ERP study conducted on individuals with stroke-induced aphasia, which showed relatively preserved semantic, but impaired verb-argument structure, processing (Kielar et al., 2012).

Evidence of impaired sentence comprehension was also found in PPA-L, who showed abnormal processing of lexical-syntactic (argument structure) information, in the face of largely spared semantic and morphosyntactic processing. While the finding of impaired verb-argument structure processing is seemingly at odds with the available literature in production, it is consistent with the hypothesis that the temporo-parietal junction - the primary region displaying cortical atrophy in PPA-L - supports argument structure processing.

Lastly, the study points to the importance of online (including electrophysiological) studies for the investigation of language deficits in comprehension, as they may provide insight into difficulties in real-time sentence processing that may not be captured by offline language measures.