Neuroanatomical Correlates of Semantic Features of Narrative Speech in Semantic and Logopenic Variants of Primary Progressive Aphasia

The semantic variant of a primary progressive aphasia (svPPA) is characterized by progressive disruption of semantic knowledge. This study aimed to compare the semantic features of words produced during a narrative speech in svPPA and the logopenic variant of PPA (lvPPA) and to explore their neuroanatomical correlates. Six patients with svPPA and sixteen with lvPPA underwent narrative speech tasks. For all the content words, a semantic depth index (SDI) was determined based on the taxonomic structure of a large lexical database. Study participants underwent an MRI examination. Cortical thickness measures were extracted according to the Desikan atlas. Correlations were computed between SDI and the thickness of cortical regions. Mean SDI was lower for svPPA than for lvPPA. Correlation analyses showed a positive association between the SDI and the cortical thickness of the bilateral temporal pole, parahippocampal and entorhinal cortices, and left middle and superior temporal cortices. Disruption of semantic knowledge observed in svPPA leads to the production of generic terms in narrative speech, and the SDI may be useful for quantifying the level of semantic impairment. The measure was associated with the cortical thickness of brain regions associated with semantic memory.


Introduction
Primary progressive aphasia (PPA) is a group of neurodegenerative disorders in which the fundamental disturbance is progressive disruption of language processing. The current clinical classification includes three main variants: semantic variant (svPPA), non-fluent variant (nfvPPA), and logopenic variant (lvPPA) [1]. This classification is based on the nature of language disturbances observed in patients, supported by data obtained using neuroimaging techniques (magnetic resonance imaging [MRI] and positron emission tomography).
Progressive and multimodal loss of semantic knowledge is the main feature of the svPPA. At the disease onset, the principal disturbances are represented by anomia and difficulties in comprehending single words, which are more evident for words and concepts characterized by low familiarity and frequency [2,3]. Confrontation-naming and single-word comprehension tasks are typically used to assess these disturbances. Patients with svPPA produce semantic paraphasias, tending to substitute low-familiarity words number of open-class words were observed [30]. Notwithstanding that a core diagnostic feature of lvPPA is the absence of agrammatism, an impairment at the syntactic level has been reported by the reduced proportion of well-formed sentences [26,30].
In svPPA, a reduced number of nouns has been reported [15], with some studies showing an increased frequency of nouns and verbs produced [21]. Syntactic processing is generally intact; however, some analyses have shown that the proportion of well-formed sentences is reduced compared with that in normal individuals [31]. As is easily predictable, semantic errors are a constant finding in these patients [32,33]. However, the typical pattern of semantic errors described in svPPA, that is, the production of superordinate words and generic terms instead of the requested terms, has generally been described using single-word production (for example, in confrontation naming) and comprehension.
In this study, we aimed to compare the microlinguistic features, and particularly the semantic features, of words produced during a narrative speech in svPPA and lvPPA and to explore their neuroanatomical correlates.

Participants
We recruited 22 participants (16 lvPPA patients; 6 svPPA patients) at the Center for the Medicine of the Aging of the Department of Geriatrics, Neurology, and Orthopedics of the Catholic University of Rome between 2013 and 2014. Aphasic patients fulfilled the criteria for both lvPPA and svPPA [1]. All the patients were native Italian speakers. They underwent a brain MRI scan within a maximum interval of two months from the neuropsychological examination (i.e., either before or after). Patients showed relatively preserved autonomy in activities of daily living and were included only if they had at least five years of formal education, did not express themselves in dialect, and completed the neuropsychological examination. We excluded patients with a history or radiological evidence of cerebrovascular disease, a history of major psychiatric disorders, or alcohol or drug abuse. Moreover, we excluded patients with visual-perceptual difficulties that could interfere with the perception of the stimulus picture.
All participants were right-handed, according to the Edinburgh Handedness Inventory. All participants provided written informed consent to participate in the study after all procedures were fully explained. This study was approved by the Catholic University Institutional Ethics Committee.

Neuropsychological Assessment
The cognitive profiles of all patients were carefully evaluated. They underwent a full neuropsychological examination, which included tasks of verbal episodic memory, shortterm memory, visuospatial analysis, constructive praxis, and intelligence. These cognitive functions are required to complete a narrative task. The complete battery took 45-60 min (in one or two sessions) to administer. The general neuropsychological assessment included memory tasks (Rey's immediate and delayed free recall of 15 words, Rey Osterrieth figure delayed recall), visuospatial tasks (line cancellations), phonological and semantic fluency, Raven's colored matrices, digits, and spatial span.
Language was investigated using Capasso and Miceli's [34] Esame neuropsicologico per l'Afasia-ENPA (Neuropsychological Examination for Aphasia), evaluating both the lexical and the sublexical mechanisms in production and comprehension.

Assessment of Narrative Abilities
Narrative ability was assessed using pictures and story descriptions. Each participant was asked to produce a set of narratives elicited with the help of two pictures, the Cookie Theft Scene (Boston Diagnostic Aphasia Examination) and the Picnic Scene (Western Aphasia Battery) and to talk about his or her illness and typical day. The Cookie Theft Scene is a black-and-white line drawing of an animated kitchen. In the picture, a woman is standing at the kitchen sink, drying a plate while the sink is overflowing with water. Meanwhile, a boy (presumably the woman's son) has climbed a stool that is rocking unsteadily and is cooking from a jar in the cupboard. A girl (presumably the boy's sister) has her hand raised upward to receive the cookies that her brother (the boy) is taking.
In the Picnic Scene, there is a man and a woman having a picnic near a lake. The man is reading a book, and the woman is pouring a drink into a glass. In this scene, another man is flying a kite. A dog is standing next to him.
To avoid poor performance due to short-term memory limitations, the pictures remained visible until the patient had finished his or her description. The examiners minimized prompts to allow uninterrupted speech, although a prompt was offered after 10 s of silence.
To better assess language production, a short informal conversation was reproduced. It starts with an open-ended question (i.e., "Tell me about your family") to introduce a topic and encourage conversation. Participants were free to answer as they liked in order to reproduce a spontaneous conversation about common events in daily life (family, hobbies, career).
Each storytelling and conversation was tape-recorded and subsequently transcribed verbatim by an examiner. The transcription included phonological fillers, pauses, false starts, and extraneous utterances.
The close transcription of the speech samples requires approximately an hour. Speech errors were classified as phonological (phoneme substitution, deletion, or insertion) or semantic paraphasia (when a target word is substituted by a semantically related word). Moreover, we considered the number of well-formed sentences, which are complete sentences free of grammatical errors, regardless of content and the frequency of anomic pauses and conduites d'approche (CDA) in the speech. Each word produced by the patients was transcribed in a database and analyzed for the number of times it was repeated (number of occurrences) and, for open-class words, lexicosemantic features that included frequency of use (FU) [35] and semantic depth index (SDI).
The SDI was derived based on the taxonomic structure of the large lexical database WordNet, an English database developed at Princeton University [36]. WordNet is organized in a hierarchical fashion in which lexical and conceptual links connect words and groups of words. The hierarchy is organized starting from generic terms (the most generic being "entity") towards terms that become more and more specific (see Figure 1 for an example). The fundamental element of WordNet is the synset (set of synonyms), which is defined by gloss, a verbal description of the concept to which it refers. Several conceptual relationships are considered in WordNet. For example, nouns are organized according to hypernyms, hyponyms, meronyms, and holonym relations. Hypernym and hyponym refer to the taxonomic "is-a-kind-of" relationship (if "cookie" is a kind of "food", then "cookie" is a hyponym of "food" and "food" is a hypernym of "cookie"). Meronyms and holonyms are based on the "is-a-part-of" relationship (if "wheel" is a part of "bicycle," then "wheel" is a meronym of "bicycle," and "bicycle" is a holonym of "wheel" wheel'). Each synset was embedded according to an ordered hierarchy constructed on the "is a" relationship [37], resembling the categorical organization of conceptual knowledge. From a general point of view, WordNet can be regarded as a nodes-and-paths network, in which nodes are represented by synsets and paths by the connections between synsets. The SDI was determined, for each open-class word, by computing the number of nodes separating the produced word from the root term (for example, "entity" in the case of nouns). Thus, higher values corresponded to a deeper exploration of the lexical-semantic system to produce a more precise lexical entry, whereas lower values indicated the production of terms closer to the root, thus being more generic. Figure 1 shows an example of SDI computation. We report, as an example, part of a svPPA speech: "l'acqua, non lo so se era chiuso oppure o oppure c'erano delle cose . . . " (translation: the water, I don't know or or..or if it was closed or there were things: SDI water : 6; SDI thing : 1; SDI close : 1). In a case of lvPPA, we recorded the following speech: "c'è il . . . i vari sportelli e l'acqua che esce dal lavandino da sola" (translation: there is the . . . the various doors and the water coming out of the sink on its own; SDI water : 6; SDI door : 7; SDI sink : 8; SDI come out : 3). generic. Figure 1 shows an example of SDI computation. We report, as an example, part of a svPPA speech: "l'acqua, non lo so se era chiuso oppure o oppure c'erano delle cose …" (translation: the water, I don't know or or..or if it was closed or there were things: SDIwater: 6; SDIthing: 1; SDIclose: 1). In a case of lvPPA, we recorded the following speech: "c'è il … i vari sportelli e l'acqua che esce dal lavandino da sola" (translation: there is the … the various doors and the water coming out of the sink on its own; SDIwater: 6; SDIdoor: 7; SDIsink: 8; SDIcome out: 3).

Figure 1.
Complete hierarchical tree of the lexical entry "cat" as derived from WordNet. The Semantic Depth Index (SDI) of the specific term "cat" (in red) corresponds to 10, i.e., the number of nodes from the root term "entity" to the term "cat". The SDI of the generic, superordinate term "animal" (in blue) corresponds to 3.
Since the narrative speech taken into account was in Italian, in the present study, we referred to MultiWordNet [38], a multilingual lexical database whose taxonomy is aligned to that of WordNet, which proved to be reliable in previous studies on semantic features of words produced in fluency tasks [39].

MRI Acquisition and Processing
All participants (except one patient with lvPPA) underwent a 1.5-T MRI (Signa, GE Healthcare) with an eight-channel phased-array neurovascular coil to obtain diffusionweighted imaging, fluid-attenuated inversion recovery, turbo spin echo, and gradient echo images. Cortical thickness analysis was performed for all PPA participants within two months (before or after) the neuropsychological examination using a T1-weighted fast gradient echo in the sagittal plane for multiplanar reconstruction (TR = 2000 ms, TE = 3.42 ms, flip angle = 15°, FOV = 256 mm, 1.0-mm slice thickness for a total of 160 slices per slab, matrix size = 256, NEX = 1). The raw 3D T1 MRI data underwent automated processing for surface-based cortex reconstruction and volumetric segmentation using the FreeSurfer image analysis software (version 6.0.0, Martinos Center for Biomedical Imaging, Charlestown, Massachusetts), which is documented and freely available for download online (http://surfer.nmr.mgh.harvard.edu/). Briefly, after brain extraction, highresolution 3D T1 images were registered to the standard space, and the gray-white and gray-cerebrospinal fluid borders were computed [40][41][42]. An experienced neuroradiologist visually inspected the image outputs from each stage of FreeSurfer processing and edited them to refine the segmentation and correct any software delineation errors. Cortical parcellations were made according to Desikan's atlas [42], and the cortical thickness was extracted for all regions included in the atlas.
A set of cortical areas previously associated with concept neural representation [43,44] was selected as part of the semantic network, including the following regions of Figure 1. Complete hierarchical tree of the lexical entry "cat" as derived from WordNet. The Semantic Depth Index (SDI) of the specific term "cat" (in red) corresponds to 10, i.e., the number of nodes from the root term "entity" to the term "cat". The SDI of the generic, superordinate term "animal" (in blue) corresponds to 3.
Since the narrative speech taken into account was in Italian, in the present study, we referred to MultiWordNet [38], a multilingual lexical database whose taxonomy is aligned to that of WordNet, which proved to be reliable in previous studies on semantic features of words produced in fluency tasks [39].

MRI Acquisition and Processing
All participants (except one patient with lvPPA) underwent a 1.5-T MRI (Signa, GE Healthcare) with an eight-channel phased-array neurovascular coil to obtain diffusionweighted imaging, fluid-attenuated inversion recovery, turbo spin echo, and gradient echo images. Cortical thickness analysis was performed for all PPA participants within two months (before or after) the neuropsychological examination using a T1-weighted fast gradient echo in the sagittal plane for multiplanar reconstruction (TR = 2000 ms, TE = 3.42 ms, flip angle = 15 • , FOV = 256 mm, 1.0-mm slice thickness for a total of 160 slices per slab, matrix size = 256, NEX = 1). The raw 3D T1 MRI data underwent automated processing for surface-based cortex reconstruction and volumetric segmentation using the FreeSurfer image analysis software (version 6.0.0, Martinos Center for Biomedical Imaging, Charlestown, Massachusetts), which is documented and freely available for download online (http://surfer.nmr.mgh.harvard.edu/, accessed on 14 June 2022). Briefly, after brain extraction, high-resolution 3D T1 images were registered to the standard space, and the gray-white and gray-cerebrospinal fluid borders were computed [40][41][42]. An experienced neuroradiologist visually inspected the image outputs from each stage of FreeSurfer processing and edited them to refine the segmentation and correct any software delineation errors. Cortical parcellations were made according to Desikan's atlas [42], and the cortical thickness was extracted for all regions included in the atlas.
A set of cortical areas previously associated with concept neural representation [43,44] was selected as part of the semantic network, including the following regions of interest (ROIs): the anterior temporal lobe, superior temporal gyrus, middle temporal gyrus, inferior temporal gyrus, fusiform gyrus, parahippocampal gyrus, entorhinal cortex, inferior frontal gyrus, caudal middle frontal gyrus, superior frontal gyrus, precuneus, supramarginal gyrus, posterior cingulate gyrus, and rostral anterior cingulate gyrus.

Statistical Analyses
Data were analyzed with Statistical Package for Social Sciences (SPSS, IBM Corporation, Armonk, NY, USA) version 24.0. The two-tailed level of statistical significance was set at p < 0.05. Comparisons between groups were performed using the Mann-Whitney U test. Concerning cortical thickness analyses, the two clinical groups were statistically contrasted Brain Sci. 2022, 12, 910 6 of 13 against those of 30 Healthy Controls (HC) matched for age, sex and education recruited in the neuroimaging laboratory. Pearson's correlation coefficients (r) were computed between the mean of the semantic depth index and the thickness of the ROI of the semantic network from Desikan's atlas [45] in the entire sample of patients with lvPPA and svPPA. The results of correlation with cortical thickness were considered statistically significant at p < 0.05 after the false discovery rate (FDR) correction controlling procedure for multiple comparisons [46].

Microlinguistic Features of Narrative Speech
The sociodemographic and neuropsychological features of the patients are summarized in Table 1. Patients affected by lvPPA produced fewer open-class words (in particular, verbs and adjectives) than those affected by svPPA (p = 0.002, p = 0.006, and p = 0.010, respectively) ( Table 2). Furthermore, the number of corrected sentences was higher for svPPA than for lvPPA (p < 0.001), whereas phonemic paraphasia was more common for lvPPA (p = 0.040). Finally, SDI was higher in the lvPPA group than in the svPPA group (p = 0.033). Individual SDI scores are reported in Supplementary Table S1.

Between-Group Comparisons on Cortical Thickness
The comparison of cortical thickness between subjects affected by svPPA and lvPPA and a group of HC is reported in Table 3. Table 3. Means and standard deviations of the cortical thickness of brain regions extracted using the Desikan's atlas in the three groups (svPPA, lvPPA, and HC) and comparisons performed between each clinical group and healthy controls (HC) with Mann-Whitney U Tests. Abbreviations: SD-standard deviation; svPPA-semantic variant of Primary Progressive Aphasia; lvPPA-logopenic variant of Primary Progressive Aphasia. P-values were adjusted using the False Discovery Rate-(FDR) correction (Benjamini and Hochberg, 1995 SvPPA vs. HC: In more detail, cortical thickness was significantly reduced in patients with svPPA compared with HC in several regions covering the medial aspect of the temporal lobe (bilateral temporal pole, entorhinal cortex, parahippocampal gyrus, fusiform gyrus), Brain Sci. 2022, 12, 910 8 of 13 the lateral aspect of the temporal lobe (bilateral banks of the superior temporal sulcus, inferior, middle, and superior temporal gyri), the frontal lobe (bilateral pars opercularis of the inferior frontal gyrus, left caudal and rostral middle frontal gyrus, superior frontal gyrus, frontal pole, lateral and medial orbital frontal cortex), the parietal lobe (bilateral inferior parietal cortex and left postcentral gyrus, precuneus, superior parietal cortex, and supramarginal gyrus), the occipital lobe (left cuneus), the bilateral insula and the cingulate cortex (left isthmus, posterior and anterior cingulate cortex) (all p FDR < 0.05).
LvPPA vs. HC: Similarly, a reduction of cortical thickness was also observed in patients with lvPPA compared to HC in numerous areas of all lobes including the medial aspect of the temporal lobe (bilateral temporal pole, entorhinal cortex, fusiform gyrus, and left parahippocampal gyrus), the lateral aspect of the temporal lobe (bilateral banks of the superior temporal sulcus, inferior, middle, and superior temporal gyri and left transverse temporal cortex), the frontal lobe (bilateral pars opercularis of the inferior frontal gyrus, rostral and caudal middle frontal gyrus, superior frontal gyrus, precentral gyrus and paracentral lobule, and left pars triangularis of the inferior frontal gyrus), the parietal lobe (bilateral inferior and superior parietal cortex, postcentral gyrus, precuneus and supramarginal gyrus), the occipital lobe (bilateral lingual gyrus, lateral occipital cortex, and left cuneus), the bilateral insula and the cingulate cortex (bilateral isthmus, posterior cingulate and left anterior cingulate cortex) (all p FDR < 0.05).

Discussion
The analysis of connected speech is a knowledgeable source of information for the study of aphasic patients, and its role is becoming more relevant in the field of neurodegenerative disorders in general and PPA in particular [21].
In this study, we explored the main features of connected speech in patients affected lvPPA and svPPA and introduced the SDI, a measure of the integrity of the lexicalsemantic system.

Discussion
The analysis of connected speech is a knowledgeable source of information for the study of aphasic patients, and its role is becoming more relevant in the field of neurodegenerative disorders in general and PPA in particular [21].
In this study, we explored the main features of connected speech in patients affected lvPPA and svPPA and introduced the SDI, a measure of the integrity of the lexicalsemantic system.
These behavioral results are consistent with those of previous studies assessing connected speech in PPA. Compared with svPPA, patients with lvPPA produced a reduced number of open-class words, as previously reported [30], and as expected, they also produced a higher number of phonemic paraphasias [26]. We did not explore the syntactic aspects of connected speech in detail; however, as previously reported, the number of correct sentences produced by lvPPA was lower than that produced by svPPA [26,30]. This finding is in contrast to the notion that the absence of agrammatism is required to diagnose lvPPA [1]. Possibly, such difficulties may reflect a strategy to cope with lexical impairment: if the patient cannot retrieve the target word from the lexicon, he or she simply skips the undergoing proposition introducing a new argument [47].
As for the svPPA group, the only score in which we observed scores lower than those obtained by the lvPPA patients was the SDI. Even if svPPA patients produced a higher number of semantic paraphasias than lvPPA patients, this difference was not statistically significant. However, the production of semantic paraphasias may depend on both the disruption of semantic memory and difficulties in accessing the lexicon [48]. Therefore, it is not surprising that patients with lvPPA may also produce such errors in narrative speech, as previously shown by other authors [29]. However, the SDI was conceived based on the typical behavior of patients affected by svPPA, which often produces semantic errors of the "superordinate" type [3,17]. From a linguistic point of view, this could be translated into a tendency to produce words that are hypernyms [37] of the target words. However, the "superordinate category" concept can be easily applied to a relatively small group of concepts, and, to our knowledge, it is generally based on a priori definitions. Nevertheless, hierarchical relationships may be broader than those that can be easily recognized (for example, the superordinate category "animal" for "cat"). For this reason, the SDI may be helpful in detecting the disruption of semantic memory, even in the absence of typical superordinate-type errors.
The reliability of this measure was confirmed by the results obtained by exploring its correlation with neurostructural measurements. We found evidence of a significant association between the SDI and the cortical thickness of brain regions associated with semantic memory, covering a large area in the temporal cortex (left hemisphere > right hemisphere), including the bilateral temporal pole, parahippocampal cortex, entorhinal cortex, and left middle and superior temporal cortex.
Overall, our correlational results suggest an association between SDI and the structural integrity of temporal regions implicated in the perceptual system for processing concrete concepts rather than of the verbal system implicated in the processing of abstract concepts. This is not surprising, considering that our semantic index was derived from the connected speech collected during the picture description. The lateralization of most correlations between SDI and the thickness of the left regions is consistent with left-hemispheric functional specialization for speech production.
More specifically, the correlation with the bilateral anterior temporal lobes (ATL) is in line with the hub-and-spoke model, which recognizes this area as central to semantic cognition [49][50][51][52], serving as a transmodal hub integrating information from the surrounding sensorimotor to form amodal semantic representations [53,54]. ATLs have been hypothesized to map between different modality-specific representations and to make proper generalizations based on crucial semantic relationships rather than superficial similarities in each particular domain [55]. This region is highly interconnected via interactions with other nodes or 'spokes' in the network [56], and its hub-like functions allow the mediation of learned associations distributed throughout the human cortex [57]. Damage to the ATL severely disrupts semantic processes, as supported by the microlinguistic analysis of our svPPA patients.
With respect to the correlation with the parahippocampal gyrus, it is important to note that this region contributes to the processing of visual and imageable spatial property knowledge during semantic tasks [58,59].
No association emerged between SDI and regions involved in semantic control system, comprising the inferior frontal gyrus (IFG), which is engaged alongside the semantic system in directing and constraining the retrieval of relevant semantic knowledge [51,[60][61][62]. The lack of correlation with this region might be explained by the involvement of this area, especially in retrieving information with weak semantic relations and selecting taskrelevant information [63]. On a methodological note, we acknowledge that the number of participants in this study was relatively small. Nevertheless, we acknowledge that this is an exploratory analysis, and further quantitative studies applying voxel-wise approaches are also required to evaluate the findings of the present work.

Conclusions
LvPPA displayed reduced ability to perform an adequate lexical retrieval and also some difficulties in sentence correctness that may deserve further exploration in future studies. Errors due to semantic disruption in the svPPA can be detected and quantified in narrative using the SDI, which showed a robust association with brain regions devoted to semantic memory.
Supplementary Materials: The following are available online at https://www.mdpi.com/article/10 .3390/brainsci12070910/s1, Table S1: Means and standard deviations on cortical thickness of the three groups (svPPA, lvPPA, HC) and comparisons performed between each clinical group and healthy controls (HC) with Mann-Whitney U Tests. Institutional Review Board Statement: Ethical review and approval were waived for this study due to the fact that the procedures described are part of the standard clinical care of the Institution.
Informed Consent Statement: Informed consent was obtained from all subjects involved in the study to undergo standard clinical procedures.

Data Availability Statement:
The data presented in this study are available on request from the corresponding author.

Conflicts of Interest:
The authors declare no conflict of interest.