Word production and comprehension in frontotemporal degeneration: A neurocognitive computational Pickian account

Over a century ago, Arnold Pick reported deterioration of word production and comprehension in frontotemporal degeneration, now a common ﬁnding. Individuals with semantic dementia (SD) and behavioral variant frontotemporal dementia (bvFTD) present with word retrieval difﬁculty, while their comprehension is less affected. Computational models have illuminated naming and comprehension in poststroke and progressive aphasias, including SD, but there are no simulations for bvFTD. Here, the WEAVER þþ /ARC model, previously applied to poststroke and progressive aphasias, is extended to bvFTD. Simulations tested the hypothesis of a loss of activation capacity in semantic memory in SD and bvFTD, caused by network atrophy (Pick, 1908a). The outcomes revealed that capacity loss explains 97% of the variance in naming and comprehension of 100 individual patients. Moreover, capacity loss correlates with individual ratings of atrophy in the left anterior temporal lobe. These results support a uniﬁed account of word production and comprehension in SD and bvFTD. ©


Introduction
It has been known since Pick (1892Pick ( /1977Pick ( , 1904Pick ( /1997) that word production and comprehension problems may not only arise from vascular stroke but also from neurodegenerative disease.One of his patients was 75-year-old Anna Jirinec, presenting with fluent speech but severe word retrieval difficulty, impaired word comprehension, and disrupted conceptual knowledge, alongside seemingly spared episodic memory.
Autopsy revealed that atrophy was most pronounced in her left temporal lobe, with photographs of Pick (1904) showing that especially the anterior part was affected.There, cell loss had thinned the cortex by 75% (Fischer, 1911).Fischer (1910) reported that her brain did not contain plaques and tangles (Alzheimer, 1907;Fischer, 1907), thus she was not afflicted with the disease named after Alzheimer.Onari and Spatz (1926) further delineated the temporal lobe atrophy, showing that it is most prominent in the left temporal pole and fusiform gyrus, relatively sparing Wernicke's area and the hippocampi, and with cell loss predominantly in the upper three cortical layers.Later research has shown that frontotemporal lobar degeneration yields behavioral variant frontotemporal dementia (bvFTD) and semantic dementia (SD) as major clinical syndromes (Brun, 1987;Hodges et al., 1992;Neary et al., 1988;Snowden et al., 1989).
Individuals with bvFTD present with personality changes and behavioral problems (i.e., disinhibition, apathy/inertia, loss of sympathy/empathy, hyperorality, repetitive/compulsive behavior, and executive deficits), atrophy of the frontal and/or anterior temporal lobes and corresponding hypoperfusion or hypometabolism, and tau-positive or TDP-43positive pathology at the molecular level (e.g., Rascovsky et al., 2011).Individuals with SD present with word production and comprehension deficits, loss of conceptual knowledge and largely spared episodic memory, atrophy of the anterior temporal lobes (ATL) and corresponding hypoperfusion or hypometabolism, and mostly TDP-43-positive pathology (e.g., Gorno-Tempini et al., 2011;Landin-Romero et al., 2016), predominantly present in the upper three cortical layers (Ohm et al., 2022).The ATL comprises the anterior part of the superior, middle, and inferior temporal gyri, the temporal pole, rhinal cortex, and the anterior fusiform and parahippocampal gyri (e.g., Bonner & Price, 2013).The atrophy of the ATL in SD is typically most prominent in the fusiform gyrus, middle and inferior temporal gyri, and temporal pole of the left hemisphere (Collins et al., 2017;Snowden et al., 2019;Yang et al., 2012).For detailed case descriptions of SD and bvFTD, I refer to Kertesz (2007).
The aphasic symptoms of Anna Jirinec and her circumscribed left temporal lobe atrophy reported by Pick (1904Pick ( /1997) ) resemble the characteristics of SD, which is one of the variants of primary progressive aphasia (PPA), the other variants being nonfluent/agrammatic and logopenic PPA (Gorno-Tempini et al., 2011).Mesulam (2007) argued that "PPA offers a unique experiment of nature for exploring the molecular fingerprints that make the language network a primary disease target and for probing the cognitive architecture of human language as it undergoes a slow but relentless dissolution" (p.11).Although not among the clinical criteria of bvFTD (Rascovsky et al., 2011), it has become increasingly clear that word production and comprehension difficulties are also frequently present in bvFTD (e.g., Geraudie et al., 2021).As Kertesz (2007) stated: "Almost all FTD/Pick's patients have some aphasic speech disturbance of the semantic or the agrammatic nonfluent type earlier or later in their illness and mutism is a mid to end stage symptom in all" (p.185).Thus, the evidence from bvFTD may provide insights that are complementary to those from PPA. Language performance in poststroke and progressive aphasias, including SD, has been illuminated using computational models (e.g., Dell et al., 1997Dell et al., , 2013;;Ueno et al., 2011), but simulations are lacking for bvFTD.Computer simulation is an important tool in testing whether models can account for the data, with the requirement to precisely define the nature of representation and processing, and in case of deterioration, also the nature of the deficit.In the present article, the WEAVERþþ/ARC model, which has provided neurocognitive computational accounts of poststroke and progressive aphasias (Janssen et al., 2020;Roelofs, 2014Roelofs, , 2022)), is extended to language performance in bvFTD.
To explain why the atrophy is circumscribed and leads to focal symptoms, Pick (1908a) put forward an account in terms of a functional network, assuming that "occasionally such a systematically similar group of neurons, i.e., a system in the older sense, succumbs to atrophy earlier than the others, and as a result the function of this system fails in a completely isolated manner" 1 (p.24).He used a drawing of Ram on y Cajal to illustrate how laminar-specific atrophy may occur.Modern research has shown that neurodegenerative diseases target specific functional networks, starting in regions with heavy network traffic and propagating along strong functional and anatomical connections (e.g., Mandelli et al., 2016;Seeley et al., 2009;Zhou et al., 2012).Employing network-sensitive neuroimaging methods, Seeley et al. found that five different neurodegenerative syndromes (i.e., bvFTD, SD, nonfluent/ agrammatic PPA, Alzheimer's dementia, and corticobasal syndrome) are associated with circumscribed atrophy of five distinct functional and structural networks, as determined in healthy human brains.In frontotemporal degeneration, either the upper or the lower three cortical layers tend to be atrophied (Ohm et al., 2022).
In the remainder, I first discuss the main hypothesis in this article, namely that the word production and comprehension difficulties in SD and bvFTD have a common functional ground, which is argued to be a disruption of semantic memory (cf.Grossman et al., 2004;Hardy et al., 2016).Semantic memory comprises our general conceptual knowledge of the world, such as what a cat is, as opposed to knowledge of personal events (e.g., that our cat Guna caught frogs in the garden during summer), called episodic memory (e.g., Eichenbaum, 2012).Moreover, the disruption is assumed to be due to circumscribed atrophy of the brain, in particular in the left ATL, which is also how Pick (1892Pick ( /1977Pick ( , 1904Pick ( /1997Pick ( , 1908aPick ( , 1908b) ) explained the language impairment of his patients.Although this view was ultimately accepted by Wernicke, he initially opposed it, stating: "Simple atrophy, which, as part of general atrophy, has affected a single gyral region, never causes a loss of its functions, causes no focal symptoms" 2 (Wernicke, 1874, p. 46; also cited by Pick, 1901).Next, I describe the assessment of word production, comprehension, and corresponding atrophy of brain regions in a large-scale study of SD and bvFTD by Snowden et al. (2019).Then, the WEAVERþþ/ARC model and its extension to bvFTD are outlined.In subsequent sections, I use the model to fit the naming and comprehension performance of 100 individual patients with SD or bvFTD from the Snowden et al. study.Also, the estimated severities of impairment in the model are tested against the empirical ratings of atrophy in various brain areas of the individual patients.Thus, it is examined whether 1 "dass gelegentlich eine solche systematisch gleichgeartete Neurongruppe, also ein System im € alteren Sinne, fru ¨her als die u ¨brigen der Atrophie verf€ allt, und dadurch ganz isoliert die Funktion dieses Systems ausf€ allt" (Pick, 1908a, p. 24).Pick proposed the account in a plenary address at a large conference in Amsterdam in 1907, and the paper he read was also included in the proceedings (Pick, 1908b).
dissolution of the functional network constituting the core of semantic memory correctly predicts the empirically observed pattern of aphasic symptoms and the spatial distribution of the atrophy of the individual patients, testing a Pickian (1908aPickian ( , 1908b) ) account.

Localized modality-general deterioration of semantic memory
In examining word production and comprehension abilities in SD and bvFTD, Snowden et al. (2019) observed that all individuals with SD (N ¼ 32) showed word production problems, assessed by picture naming, and 53% exhibited word comprehension problems, assessed by word-to-picture matching.Moreover, also 48% of the patients with bvFTD (N ¼ 71) showed word production problems and 17% exhibited word comprehension problems.Staffaroni et al. (2021) found that naming and word-to-picture matching were worse in semantic dementia (N ¼ 185) than in bvFTD (N ¼ 612), and performance in both patient groups was worse than in healthy controls (N ¼ 581).In a systematic literature review of speech and language impairments in bvFTD including 181 studies, Geraudie et al. (2021) observed that picture naming was impaired in 63% and word comprehension in 67% of the studies.Thus, word production and comprehension are not only disrupted in SD but often in bvFTD as well.
Whereas individuals with SD and bvFTD present, to different degrees, with word retrieval and comprehension difficulties, word repetition is preserved in SD (e.g., Leyton et al., 2014) as well as in bvFTD (e.g., Geraudie et al., 2021).This pattern of impaired and spared performance across word production, comprehension, and repetition tasks differs from what is observed in the nonfluent/agrammatic and logopenic variants of PPA, where repetition is commonly worse than comprehension (Janssen et al., 2022;Savage et al., 2013).The similarity in pattern of performance between SD and bvFTD suggests that the underlying language impairment, although differing in degree, is functionally the same.Although atrophy in bvFTD tends to be most prominent in the frontal lobes, the ATLs are often affected too (e.g., Rascovsky et al., 2011).Thus, the word production and comprehension problems in the two patient groups may have a common source in the ATLs.According to a prominent account, the ATLs underpin a hub of modality-general concepts, which are taken to integrate modality-specific features (e.g., shape, color, sound, touch, taste, smell, movement, action) that are stored in widely-distributed perceptual and motor regions of the brain, together making up semantic memory (e.g., Lambon Ralph, 2014;Patterson et al., 2007;Pick, 1931).This invites the hypothesis that a modality-general semantic hub is disrupted by left ATL atrophy in both patient groups and that differences in performance between the groups reflect differences in severity of atrophy (cf.Grossman et al., 2004;Hardy et al., 2016).This hypothesis was opposed by Snowden et al. (2019), who argued that "there may be no domain-general hub in which concepts are represented and which can be disrupted by selective damage.The semantic loss may be a product of widespread loss of connections across the semantic network" (p.32).
To examine the hypothesis of semantic memory disruption by circumscribed left ATL atrophy, computer simulations with the WEAVERþþ/ARC model (Roelofs, 2014(Roelofs, , 2022) ) were run.The importance of simulations was highlighted by Mirman and Britt (2014), who stated that "intuitions and verbal theories are not enough because they can be claimed to predict (or not predict) just about anything.Computational models provide a concrete implementation of a proposed theory that can then be tested empirically to evaluate whether it truly accounts for the observed data" (p.11).Computational models have elucidated word production, comprehension, and repetition deficits in poststroke aphasia and in SD (e.g., Dell et al., 1997Dell et al., , 2013;;Ueno et al., 2011), or have specifically illuminated the disruption of semantic memory in SD (Rogers et al., 2004).The WEAVERþþ/ARC model3 has been applied to word production, comprehension, and repetition in poststroke aphasia (Roelofs, 2014(Roelofs, , 2021) ) and in the three variants of PPA (Roelofs, 2022).The model integrates evidence from behavioral psycholinguistic studies, functional neuroimaging, tractography, and aphasiology.Building on Pick's (1892Pick's ( /1977Pick's ( , 1908a) ) seminal account and modern insights (e.g., Mandelli et al., 2016;Seeley et al., 2009;Zhou et al., 2012), the impairments were assumed to arise from a loss of activation capacity in parts of the language network that are specific to each PPA variant.

1.2.
Word production, comprehension, and atrophy ratings in SD and bvFTD Snowden et al. (2019) assessed word production and comprehension abilities in relatively large cohorts of patients with SD and bvFTD.Word production was assessed using two picture naming tests, one demanding and the other easier.The demanding test consisted of 30 items of increasing difficulty, eliciting floor level performance in patients with SD.The easier Manchester naming test consisted of 40 items, appropriately pitched for SD.Patients spoke the name of the object shown in each picture (e.g., a rabbit, say "rabbit"), one at a time.Word comprehension was assessed using a word-topicture matching test involving the same 40 items as the Manchester naming test.Patients had to match a printed word (e.g., rabbit) with one of four semantically related pictures (the rabbit, not the mouse or another animal).Patients with SD had, on average, only one item of the 30 correct on the demanding test, much reducing the usefulness of that test.The easier Manchester naming and word-to-picture matching tests used the same items, permitting direct comparison of the scores.Therefore, I concentrate on the findings from the Manchester tests in the remainder.
In line with another large-scale study (Staffaroni et al., 2021) and a systematic literature review (Geraudie et al., 2021), Snowden et al. (2019) observed that performance was worse in SD than bvFTD.Patients with SD performed, on average, 40% correct on the naming task and 81% correct on the word-to-picture matching task, whereas patients with bvFTD performed 83% and 96% correct, respectively.Similar patterns of findings have been obtained with the Sydney Language Battery (SYDBAT), which is a screening test designed to distinguish the three variants of PPA (Janssen et al., 2022;Savage et al., 2013).The SYDBAT assesses word production, comprehension, repetition, and conceptual knowledge using, respectively, picture naming, spoken wordto-picture matching, word repetition, and picture-to-picture matching tasks.Using the SYDBAT, Goldberg et al. (2021) observed levels of performance for SD (N ¼ 11) and bvFTD (N ¼ 26) that were similar to those observed by Snowden et al.Patients with SD had 42% correct in picture naming and 71% correct in word-to-picture matching, and patients with bvFTD had, respectively, 78% and 93% correct on these tasks.Thus, the patterns of performance are the same across studies: Word production is worse than comprehension, and both are worse in SD than in bvFTD.The decline is not just proportional.Predicting individual comprehension scores by naming scores assuming a linear relationship explains only 53% of the variance of the patients with SD and 54% of those with bvFTD in the study of Snowden et al.
In addition to assessing behavioral performance, Snowden et al. ( 2019) also rated atrophy.Magnetic resonance imaging (MRI) scans of the patients were visually rated for severity of atrophy using a five-point scale, ranging from no atrophy to severe atrophy (Davies et al., 2009).Voxel-based morphometry analysis was not viable, because the MRI scans had been obtained in different clinical centers using different acquisition protocols.However, earlier research had shown that the visual rating scale has good inter-rater reliability and its scores are strongly correlated with atrophy measures obtained with voxel-based morphometry.Moreover, the atrophy ratings using the scale of Davies et al. agreed with those obtained using another simpler rating scale for MRI scans in SD and bvFTD.Snowden et al. observed that atrophy scores were higher in SD than in bvFTD for several temporal regions, including the left and right temporal poles (approximate Brodmann designations BA 36/38) and anterior fusiform gyri (approximately BA 36/ 20).Conversely, atrophy scores were numerically, but not statistically, higher in bvFTD than in SD in left and right lateral frontal gyri (approximately BA 8/9/46, not including Broca's area BA 44/45) and the basal ganglia.These findings correspond with previous evidence that atrophy is greatest in the ATLs (comprising the temporal pole and anterior fusiform gyrus) in SD, while being greatest in the frontal lobes in bvFTD (e.g., Seeley et al., 2009;Zhou et al., 2012).
Moreover, atrophy correlated with behavioral performance.Snowden et al. (2019) observed that the naming performance of individual patients correlated negatively with the ratings of atrophy in several temporal regions, including the temporal pole and anterior fusiform gyrus of the left hemisphere (i.e., naming performance decreased as atrophy increased), for both SD and bvFTD.Atrophy in the left anterior fusiform gyrus also correlated negatively with word-topicture matching in SD, but not in bvFTD.There were no correlations with atrophy in the right hemisphere counterparts of these regions, and also not with left or right posterior temporal cortex (approximately BA 37) and left or right frontal regions.These correlation patterns agree with the assumption that a semantic memory hub, underpinned by the ATL, particularly in the left hemisphere, is disrupted in both SD and bvFTD.Snowden et al. (2019) examined word production and comprehension, but not repetition, in SD and bvFTD.It is important to note, as already indicated, that whereas word production and comprehension are typically impaired, word repetition is spared in both SD and bvFTD (e.g., Geraudie et al., 2021;Goldberg et al., 2021).This pattern is different from what is observed in the nonfluent/agrammatic and logopenic variants of PPA, where repetition is commonly worse than comprehension (Janssen et al., 2022;Savage et al., 2013).Therefore, I describe word production, comprehension, as well as repetition in WEAVERþþ/ARC for both SD and bvFTD.

The WEAVERþþ/ARC model
The WEAVERþþ/ARC model makes explicit how three major memory systems of the human brain, namely declarative, procedural, and working memory, interact during word production, comprehension, and repetition (Roelofs, 2014(Roelofs, , 2018(Roelofs, , 2021(Roelofs, , 2022; for a review, see Roelofs & Ferreira, 2019).When a goal (e.g., to name a picture, to perform word-topicture matching, to repeat a word) is temporarily kept in working memory, information about concepts and words needed to achieve the goal is retrieved by spreading activation through an associative network stored in long-term declarative memory.Selection of relevant information is achieved by the application of condition-action rules stored in long-term procedural memory.The rules also exercise top-down control.For example, in picture naming, a condition-action rule enhances the activation of the target concept in the associative network.Declarative memory is thought to be underpinned by temporal and inferior frontal regions, procedural memory by frontal regions, basal ganglia, and thalamus, and working memory by dorsolateral prefrontal cortex.Eichenbaum (2012) provides a review of the brain systems for declarative, procedural, and working memory.The structure of WEAVERþþ/ARC and its mapping onto the human brain is illustrated in Fig. 1 (see Kemmerer, 2019a, for a review of the empirical evidence).
The model assumes that concepts are part of a network of modality-general conceptual representations stored in the ATL bilaterally.The concepts are connected to modalityspecific features, assumed to be stored in widely-distributed perceptual and motor regions of the brain (e.g., Lambon Ralph, 2014;Patterson et al., 2007).For example, concepts are connected to higher-level representations of the visual shape of objects, thought to be represented in the posterior fusiform gyrus (e.g., Weiner & Zilles, 2016).Lexical concepts are connected to lemmas that specify the grammatical properties of words, such as that cat is a noun.Lemmas are thought to be stored in the middle part of the left middle temporal gyrus (MTG).They are connected to lexical output forms (e.g., <cat>) in left posterior superior temporal gyrus (STG) and MTG (Wernicke's area).The lexical output forms are connected to output phonemes (e.g., /k/, /ae/, and /t/) in left posterior inferior frontal gyrus (IFG; i.e., Broca's area), which are connected to motor programs (e.g., [kaet]) in ventral precentral gyrus.Input phonemes (e.g., /k/, /ae/, and /t/), and lexical input forms (e.g., <cat>) are thought to be stored in middle to posterior STG and superior temporal sulcus (STS) bilaterally (see Kemmerer, 2022, for a review).In picture naming, activation spreads from lexical concepts to motor programs via lemmas, lexical output forms, and output phonemes; in spoken word comprehension, activation spreads from input phonemes to lexical concepts via lexical input forms and lemmas; and in repetition, activation spreads from input phonemes to motor programs via output phonemes, both directly and indirectly via lexical forms and lemmas.

Assumptions about SD and bvFTD
As indicated, the main hypothesis in the present article is that a deterioration of semantic memory is common to SD and bvFTD.It is assumed that cell loss due to disease reduces the capacity of the conceptual network to transmit activation (cf.Burke et al., 1991;Dell et al., 2013) or diminishes its capacity to maintain activation over time (cf.Martin & Dell, 2019), reflected in the spatial distribution of atrophy (i.e., circumscribed atrophy of the ATLs), corresponding hypoperfusion or hypometabolism, and aphasic symptoms (Pick, 1908a).A loss of activation transmission in WEAVERþþ/ARC corresponds to a reduction of connection weights and a loss of activation maintenance to an increased decay rate (Roelofs, 2014(Roelofs, , 2022)).
The weight decrease concerns all connections to, within, and from the conceptual network, and the decay increase concerns all concept nodes.

Methods
I report all data exclusions, all inclusion/exclusion criteria, whether inclusion/exclusion criteria were established prior to data analysis, all manipulations, and all measures in the study.No part of the study procedures or analysis plans was preregistered prior to the research being conducted.Word-to-picture matching scores were missing for two patients with SD and for one patient with bvFTD.Therefore, the WEAVERþþ/ARC simulations were run for word production and comprehension of 30 patients with SD and 70 patients with bvFTD.Thus, in total, the behavioral performance of 100 individual patients was simulated.

General simulation protocol
The simulation method, including the network structure and set of parameter values, was the same as in earlier simulations (e.g., Roelofs, 2014Roelofs, , 2022)).The target word was cat and the other items were dog and fish (both semantically related), fog (phonologically related to a semantic alternative, namely dog), and mat (phonologically related to cat).The earlier simulations concerned performance on the SYDBAT (Janssen et al., 2022;Savage et al., 2013) and related tests, where word comprehension was assessed by spoken word-to-picture matching rather than written word-to-picture matching.The latter, however, was the task version used by Snowden et al. (2019).The network in Fig. 1 contains input phonemes and lexical input forms for spoken word comprehension (involved in spoken word-to-picture matching) and repetition, thought to be represented in middle to posterior STG and STS bilaterally.This input network can also be used to simulate written word comprehension (involved in written word-to-picture matching), except that the input phonemes and lexical forms are then taken to be input letters and lexical forms.
Whereas spoken words activate their phonemes serially, written words activate their letters in parallel.The letters and orthographic lexical forms are thought to be represented in the left occipitotemporal sulcus (e.g., Kemmerer, 2022;Yeatman & White, 2021).This assumption about written word perception in the present simulations allowed a comparison with other studies of SD and bvFTD that used the spoken word-to-picture matching task of the SYDBAT (e.g., Goldberg et al., 2021).
The simulations were also run with a larger network that additionally included all animal names of the SYDBAT (i.e., butterfly, elephant, caterpillar, dinosaur, rhinoceros, hippopotamus, and orangutan), as had been done in the previous simulation studies of PPA.The Pearson correlation between the simulated patterns of naming and comprehension of the individual patients (i.e., 200 data points) of Snowden et al. (2019) for the small and larger networks was r ¼ .99 for a weight lesion and r ¼ .98 for a decay lesion.Thus, the simulation outcomes do not vary with the size of the network.
Information was retrieved from the network by spreading activation according to Here, a(m, t) is the activation level of node m at point in time t, d is the decay rate, and Dt is the duration of a time step in msec.The sum indicates the amount of activation that m receives between t and t þ Dt, where a(n, t) is the output of neighbor n and r is a weight indicating the strength of the connection between nodes m and n.Capacity loss was simulated by reducing connection weights (r) or increasing the decay rate (d) in the conceptual network.

Simulating behavioral performance
External stimulus activation was provided to lexical concepts for naming, to input phonemes or letters for comprehension, and to input phonemes for repetition.Activation was then allowed to spread for 2 sec in Dt ¼ 25 msec steps, and the mean activation of the relevant nodes was computed.For each of several degrees of capacity loss, the difference in mean activation between target and closest alternative was expressed as a percentage of the normal activation difference.When the activation difference is smaller, selection takes longer or fails and errors are more likely to occur.Thus, lower percentages correspond to poorer performance.For naming and repetition, the activation difference concerned motor program nodes, and for comprehension, the difference concerned concept nodes.An exhaustive search through the parameter space, varying between minimal and maximal capacity loss, was performed to obtain the value of the weight decrease and decay increase that provided the best fit between model and group averages (from Goldberg et al., 2021) or individual patient data (from Snowden et al., 2019).The best fit was defined as the lowest mean absolute difference (i.e., mean absolute error, MAE) between simulated and empirical performance for naming, comprehension, and repetition (Goldberg et al.) or naming and comprehension (Snowden et al.).The weight decrease varied between .99 and 0 (Â r) and decay increase between 1.0 and 1.66 (Â d), both in steps of .01.To provide a measure of goodness of fit between model and data, MAEs and Pearson correlation coefficients are reported.

Correlations with atrophy ratings
For each of the 100 patients of Snowden et al. (2019), the parameter search yielded a value for the capacity loss parameter (i.e., weight decrease or decay increase) that provided the best fit between model and individual performance on the naming and comprehension tasks.According to the model, weight decrease and decay increase are a function of brain atrophy.To test this, I computed correlations between weight decrease or decay increase and atrophy ratings for various brain regions of the individual patients.Given that the atrophy ratings were on a five-point scale, involving an ordinal measure, Kendall's tau correlations were calculated.The assumption of a disruption of a semantic memory hub in both SD and bvFTD, underpinned by the left ATL, implies correlations for some regions but not for others.In particular, correlations should be found for the ratings of atrophy in the left temporal pole and left anterior fusiform gyrus, but correlations should be absent or less strong for atrophy in the right hemisphere counterparts of these regions, in bilateral posterior temporal cortex, and in bilateral frontal regions.Lower weight values (i.e., a higher weight decrease) should correspond to higher atrophy ratings, thus a negative correlation is expected for the weight lesion (and higher decay values should correspond to higher atrophy ratings, thus a positive correlation is expected for the decay lesion).

Bayesian analyses
To quantify the strength of the statistical evidence for the presence or absence of correlations, I performed Bayesian statistical analyses and report Bayes factors (e.g., Wagenmakers et al., 2018).A Bayes factor quantifies the evidence that the data provide for one hypothesis versus another.For example, when the Bayes factor BF À0 (i.e., subscript À0) equals 5, the data are 5 times more likely under the H 1 that a negative correlation is present (BF þ0 for a positive correlation) than under the H 0 of no such correlation.The Bayesian analyses were done using JASP (Love et al., 2019) using Cauchy priors with default parameter settings (for a defense, see Wagenmakers et al.).Under a standard interpretation, a BF of 3e10 indicates "moderate evidence", 10e30 "strong evidence", 30e100 "very strong evidence", and >100 "extreme evidence" for one hypothesis relative to the other.

Availability of the source codes
The simulations with the WEAVERþþ/ARC model were computationally implemented using the C programming language and the programming environment of Microsoft Visual Cþþ 2022.The source codes of the simulation programs are available from the Open Science Framework at https://osf.io/52x3w/ or from the author. 3.

Performance accuracy as a function of activation capacity
In the simulations, performance accuracy of naming and comprehension, but not of repetition, varied as a function of weight decrease and decay increase in semantic memory (i.e., the conceptual network).Fig. 2 shows that weight decrease (left panel) and decay increase (right panel) tend to have similar effects on naming and comprehension.However, there is also an important difference.Whereas word production and comprehension may fully deteriorate with a weight lesion, this cannot happen with a decay lesion.
Some aspects of the patterns of performance that are observed empirically correspond more closely to the decreased performance that results from a weight decrease (left panel of Fig. 2) than from a decay increase (right panel).Snowden et al. (2019) and others have observed for some patients an accuracy of (close to) 0% on naming and comprehension tasks, which may occur under a weight lesion but not a decay lesion in the model.Therefore, I discuss the simulation results for a weight lesion in some detail, whereas the results for a decay lesion are only briefly mentioned.All results are reported numerically in the Supplementary material.

Patient group averages for naming, comprehension, and repetition
In Roelofs (2022), I presented simulations of the pattern of performance for word production, comprehension, and repetition in SD, both for English (Savage et al., 2013) and for Dutch (Janssen et al., 2022).Performance on the SYDBAT exhibits the canonical pattern of severely impaired word production, less affected comprehension, and preserved repetition.For example, for the patients with SD (N ¼ 13), Janssen et al. observed 29% correct for naming, 78% for comprehension, and 96% for repetition.The model captured these findings: 29% correct in naming, 87% in comprehension, and 99% in repetition, with an MAE of 4.0%.Goldberg et al. (2021) used the SYDBAT to assess performance in SD and bvFTD.Patients with SD had 42% correct in picture naming, 71% in comprehension, and 98% in repetition.The model yielded 41% correct in naming, 92% in comprehension, and 100% in repetition, with an MAE of 8.0%.Different from the earlier simulations of performance in SD, the model clearly overstates comprehension performance here, although the ordinal pattern in the data is preserved.For patients with bvFTD, Goldberg et al. observed 78% correct in picture naming, 93% correct in comprehension, and 97% correct in repetition.The model captured these findings: 79% correct in naming, 98% in comprehension, and 100% in repetition, with an MAE of 3.3%.Thus, overall, the model does a reasonable job in capturing the patterns of naming, comprehension, and repetition performance at the group level in SD and bvFTD.

Naming and comprehension of individuals with SD
The left panel of Fig. 3 shows the patterns of performance on the naming and comprehension tasks of the 30 persons with SD in the study of Snowden et al. (2019) together with the group averages.Using box plots to determine outlying observations in the patient scores for each task, 3 of the 60 data points were deemed to be outliers, denoted by numbers (e.g., #6).The figure reveals that patients #1, #6, and #16 had extremely disrupted comprehension.
The right panel of Fig. 3 shows the WEAVERþþ/ARC simulation results for all 30 individual patients in the study of Snowden et al. (2019), assuming a weight lesion.For each patient, denoted by dot and number with the tasks color coded, the predicted performance scores with the lowest MAE are plotted against the observed scores (for details, see Supplementary Table 1).Overall, the model succeeds reasonably well at simulating the performance patterns of the individual cases.The average MAE across patients is 3.1% and the correlation between model and individual patient data is r ¼ .99,p < .001,BF þ0 ¼ 3.0 Â 10 49 (i.e., "extreme evidence" for a positive correlation).Similar outcomes were obtained when assuming a decay lesion (Supplementary Table 2), with an average MAE across patients of 6.8% and correlation of r ¼ .92,p < .001,BF þ0 ¼ 5.7 Â 10 22 .
As can be seen in the right panel of Fig. 3, the predicted comprehension scores for patients #1, #6, and #16 are at the diagonal line, meaning that the model fits the outlying empirical data points well.The good fit is obtained when the weight lesion for these patients is severe.
A similar goodness of fit was obtained between the model and the individual patient data of Janssen et al. ( 2022) concerning the naming, comprehension, and repetition tasks of the Dutch version of the SYDBAT.Restricting the analysis to naming and comprehension, the average MAE across patients was 3.6% and the correlation between model and individual patient data was r ¼ .99,p < .001,assuming a weight lesion.Thus, the model succeeds reasonably well at simulating the individual performance patterns of the persons with SD, regardless of language (English, Dutch), type of pictures (black-and-white drawings, colored photos), and modality of word-to-picture matching (i.e., written or spoken words).

Naming and comprehension of individuals with bvFTD
The left panel of Fig. 4 shows the patterns of performance on the naming and comprehension tasks of the 70 persons with bvFTD in the study of Snowden et al. (2019) together with the group averages.Using box plots to determine outlying observations in the patient scores for each task, 6 of the 70 data points for naming and 11 of the 70 data point for comprehension were deemed to be outliers.This concerned the naming performance of patient #1 or a lower score and the comprehension performance of patient #54 or a lower score.
The right panel of Fig. 4 shows the WEAVERþþ/ARC simulation results for all 70 individual patients in the study of Snowden et al. (2019), assuming a weight lesion.For each patient, denoted by dot and number with the tasks color coded, the predicted performance scores with the lowest MAE are plotted against the observed scores (for details, see Supplementary Table 3).Again, the model succeeds reasonably well at simulating the performance patterns of the individual patients.The average MAE across patients is 2.3% and the correlation between model and individual patient data is r ¼ .98,p < .001,BF þ0 ¼ 1.4 Â 10 89 .Similar outcomes were obtained when assuming a decay lesion (Supplementary Table 4), with an average MAE across patients of 6.8% and correlation of r ¼ .92,p < .001,BF þ0 ¼ 7.4 Â 10 65 .
As can be seen in the right panel of Fig. 4, the predicted naming score for patient #1 lies at the diagonal line, meaning that the model fits the outlying empirical data point for this patient well.Moreover, the predicted comprehension score for patient #18 lies at the diagonal line too, meaning that the model also fits the outlying empirical data point for this patient.However, the model cannot fit the pattern of performance of patient #9, for which the MAE is 17.3%.Snowden et al. (2019) also noticed the deviant performance of this patient.They stated: "Yet in one bvFTD patient in particular there were a relatively large number of instances of correct naming with impaired comprehension performance.It is likely that executive failures accounted for the incorrect response selections in that patient.A feature of bvFTD is lack of adherence to task rules.In a forced-choice condition patients may sometimes base responses on personal preference or idiosyncratic criteria or else respond randomly due to inattention.It is noteworthy that the bvFTD patient showed severely impaired performance on the Weigls test, achieving a score of 0" (p.31).
The Weigls block sorting test requires grouping of 12 colored blocks according to color, shape, or motif, which is thought to measure executive control ability.In line with the extremely poor score on this test, the highest atrophy rating scores of patient #9 concerned bilateral frontal cortex and basal ganglia rather than temporal cortex.

Atrophy ratings of individuals with SD and bvFTD
The model maintains that weight decrease and decay increase are a function of brain atrophy.This predicts correlations between weight decrease or decay increase values in the model and atrophy ratings for particular brain regions of the individual patients.In particular, correlations should be found for the ratings of atrophy in the left temporal pole and anterior fusiform gyrus, but they should be absent or less strong for the atrophy in the right hemisphere counterparts, bilateral posterior temporal cortex, and bilateral frontal regions.Lower weight values should correspond to higher atrophy ratings (and higher decay values should correspond to higher atrophy ratings).Table 1 lists the Kendall's tau correlations between weight decrease in WEAVERþþ/ARC and empirical atrophy ratings for relevant brain regions of the individual patients.The correlations were assessed for the temporal pole, anterior fusiform gyrus, posterior temporal cortex, and lateral frontal cortex, all bilaterally.Lower weight values should correspond to higher atrophy ratings, so the correlations are predicted to be negative.The table shows that negative correlations were obtained for the left temporal pole and anterior fusiform gyrus, but they were absent or less strong for the right hemisphere counterparts, bilateral posterior temporal cortex, and bilateral frontal regions.The Bayes factors for the left temporal pole and left anterior fusiform gyrus are larger than 100, indicating that the evidence for the presence of the negative correlations is "extreme".A Bayes factor between .10 and .33indicates "moderate evidence" for an absence of a correlation (i.e., the absence is 3e10 times more likely than the presence), which holds for the right temporal pole and lateral frontal cortex in both SD and bvFTD.For other regions, the evidence for the  c o r t e x 1 6 3 ( 2 0 2 3 ) 4 2 e5 6 presence or absence of a correlation is less strong or less consistent (e.g., stronger for SD than bvFTD).Similar results were obtained for decay values (see Supplementary Table 5).
To summarize, the estimated weight lesions (and decay lesions) in the model correlate with ratings of atrophy in specific brain regions of the individual patients.In particular, correlations were present for atrophy in the left temporal pole and left anterior fusiform gyrus, whereas they were absent or less strong for the atrophy in the right hemisphere counterparts, bilateral posterior temporal cortex, and bilateral frontal regions.
Given the correlations between model capacity and empirical atrophy (Table 1), and the goodness of fit between model and behavioral data for each patient group (Figs. 3 and  4), the correlations between model performance and atrophy rates in the different regions for each group and for each task are expected to follow the corresponding empirically observed correlations.Supplementary Tables 6e9 give the correlations that were empirically observed and in the model for a weight lesion, separately for each task, group, and brain region.As was to be expected, there is a close correspondence between the correlations in the empirical study and in the model.In particular, correlations between behavioral performances and atrophy ratings are observed for the left ATL and anterior fusiform gyrus but not for other brain regions, both empirically and in the model.There is even a close match between the strength of the correlations that were empirically observed and in the model.However, there is also a salient deviation between model and data.Whereas Snowden et al. (2019) observed no correlations between comprehension performance in bvFTD and atrophy in left ATL and left anterior fusiform gyrus, such correlations are present in the model.It should be noted, however, that the Bayes factor associated with the empirical correlation between comprehension performance in bvFTD and left ATL atrophy was around 1.0, meaning that evidence is equally strong for the absence and the presence of a negative correlation.Moreover, the correlation in the model (i.e., À.28) falls within the 95% CI for the empirically observed correlation (i.e., [À.015, À.282]).The Bayes factor associated with the empirical correlation between comprehension performance in bvFTD and left fusiform gyrus atrophy was 1.82, meaning that there is about two times more evidence for the presence than for the absence of a negative correlation.The correlation in the model (i.e., À.27) falls within the 95% CI for the empirically observed correlation (i.e., [À.023, À.306]).Given the indeterminacy of the present empirical evidence, the predicted correlations should be further investigated in future research.The power of a future empirical test may be increased by using a more demanding word-to-picture matching task and by measuring atrophy using voxel-based morphometry rather than visual rating.

General discussion
Pick (1892/1977, 1904/1997) argued that circumscribed atrophy of the left temporal lobe causes word production and comprehension difficulty.To examine whether a loss of activation capacity in a semantic memory hub of modality-general concepts, underpinned by the left ATL, accounts for the naming and comprehension performance of persons with SD or bvFTD, WEAVERþþ/ARC simulations were run.The simulation outcomes revealed that activation capacity loss explains 97% of the variance in naming and comprehension at the individual patient level.Moreover, the degree of capacity loss in the model correlates with the ratings of atrophy in the left ATL of the individual patients.These results support a unified account of the deterioration of word production and comprehension in SD andbvFTD, along Pickian (1908a, 1908b) lines.
In the simulations with WEAVERþþ/ARC, the affected network was assumed to constitute the central core of semantic memory, which was conceived of as a hub of modalitygeneral concepts in the ATL linked to widely distributed modality-specific features (e.g., Lambon Ralph, 2014;Patterson et al., 2007;Pick, 1931), such as visual forms and spoken names.However, according to an alternative view, defended by Snowden et al. (2019), there exists no modalitygeneral semantic hub in the ATL, but semantic impairment is the result of a loss of connections across a widely distributed network of modality-specific sensory and motor features making up conceptual knowledge (cf.Wernicke, 1874).They argued that the atrophy of the left anterior fusiform gyrus observed for both SD and bvFTD may hamper visual access to the rest of the widely distributed network.Snowden et al. put forward two arguments, based on their own findings, against the notion of the ATL as semantic hub.
The first argument of Snowden et al. (2019) against the ATL as semantic hub, with the temporal pole being key, involved their finding that the strongest correlation between naming and comprehension performance and atrophy ratings in SD concerned the anterior fusiform gyrus.This suggests, according to them, that the temporal polar region does not have a privileged role in semantic memory.However, although correlations were strongest for the anterior fusiform gyrus, they were also clearly present for the left temporal pole.For the pole, Kendall's tau for naming was À.49, p < .002,BF 10 ¼ 461.5, and tau for comprehension was À.40, p < .013,BF 10 ¼ 52.9.Thus, the Bayesian analysis suggests that the evidence for a correlation with the left temporal pole is "extreme" for naming and "very strong" for comprehension.Differential correlations between behavioral performance and the left temporal pole and anterior fusiform gyrus would only be an argument against the ATL as semantic hub if correlations would be present for the fusiform gyrus and absent for the temporal pole, which is clearly not the case.
The second argument of Snowden et al. (2019) against the ATL as semantic hub concerned their observation that associative errors in naming (such as "when it rains" in response to a picture of an umbrella) correlated with left temporal pole atrophy, whereas omission errors (such as "I don't know what that is") correlated with atrophy in left anterior fusiform gyrus.According to Snowden et al., associative errors imply that a patient has some conceptual knowledge, whereas an omission error suggests a loss of knowledge.The differential correlations between type of error and brain regions were taken "to challenge the notion of the temporal polar region as a semantic hub" (p.32).However, although correlations for associative errors were stronger for the temporal pole than for the anterior fusiform gyrus (BF 10 ¼ 30.7), correlations were clearly present for both regions.For the temporal pole, c o r t e x 1 6 3 ( 2 0 2 3 ) 4 2 e5 6 Kendall's tau was .45,p < .001,BF 10 ¼ 8.0 Â 10 8 , and for the fusiform gyrus, tau was .42,p < .001,BF 10 ¼ 2.6 Â 10 7 .Also, while correlations for omission errors were stronger for the anterior fusiform gyrus than for the temporal pole (BF 10 ¼ 11919.0),correlations were also clearly present for both regions.For the fusiform gyrus, Kendall's tau was .50,p < .001,BF 10 ¼ 5.6 Â 10 10 , and for the temporal pole, tau was .40,p < .001,BF 10 ¼ 4.7 Â 10 6 .Thus, associative and omission errors are not uniquely associated with atrophy in one region rather than another, but both types of error are associated with both regions of the ATL.The difference in strength of the correlations suggests different functional roles of the regions in picture naming (e.g., the fusiform gyrus providing visual access to the conceptual network), but it does not challenge the notion of the temporal pole as a key semantic hub region.
In sum, whereas Snowden et al. (2019) held that "[t]he correlative data … challenge the status of the anterior temporal lobes as a semantic hub" (p.33), I do not take the precise ATL location to be a fundamental issue for the hub-and-spoke view.The temporal pole and the anterior fusiform gyrus are both part of the ATL, and atrophy in both regions correlates with behavioral performance.Proponents of the hub-andspoke view do not consider the ATL to be functionally homogeneous.Instead, the ATL semantic hub exhibits graded specialization, reflecting the patterns of structural and functional connections with modality-specific spoke areas (e.g., Binney et al., 2012).Whereas early research suggested that the temporal poles play a central role in the semantic hub, the heart of the hub is now taken to be in the anterior fusiform gyrus, in line with the evidence of Snowden et al. (for discussion, see Patterson & Lambon Ralph, 2016).
Whereas the hub-and-spoke view readily accounts for the evidence that naming difficulty in SD occurs across input modalities, including vision, touch, and audition (e.g., Coccia et al., 2004;Lambon Ralph, 2014;Patterson et al., 2007), this does not hold for the modality-specific distributed view on concepts favored by Snowden et al. (2019).For naming problems to arise across input modalities, atrophy has to be present in each of the corresponding modality-specific distributed brain regions or corresponding connections.Instead, there is circumscribed atrophy of the ATLs, especially in the left hemisphere, which was also the seminal observation of Pick (1892Pick ( /1977Pick ( , 1904Pick ( /1997)).Although Wernicke initially opposed Pick's claim that word production and comprehension difficulties may result from circumscribed atrophy of the brain (i.e., the left ATL), he later endorsed the Pickian account.Wernicke (1906) stated that "Pick has shown that within the framework of a general brain atrophy, there are more pronounced localized atrophies, which reveal themselves through focal symptoms that correspond with the locus" 4 (p.553).
The WEAVERþþ/ARC model does not simulate the different error types, but simulates reduced accuracy in a general way.Given the make-up of the conceptual network that was used in the simulations, errors in the model will concern coordinate concepts or omissions.Snowden et al. (2019) categorized errors as coordinates (e.g., "dog" for "cat"), superordinates (e.g., "animal" for "cat"), associations (e.g., "when it rains" in response to an umbrella), visual misidentifications (e.g., "hat" in response to a mushroom), omissions (e.g., "don't know"), or acceptable alternative responses (e.g., "coat" for "jacket").In bvFTD and SD, the vast majority of errors concerned related concepts (i.e., coordinates, superordinates, and associates), omissions, or acceptable alternative responses.To be able to simulate the relative frequency of these error types, more sophisticated conceptual networks would be required in the model, allowing for errors in all the categories that Snowden et al. distinguished.In future research, the model may be extended further to take account of the various error types.
The naming and comprehension tasks of the Manchester and SYDBAT test batteries (used by Goldberg et al., 2021;Janssen et al., 2022;Savage et al., 2013;Snowden et al., 2019) involve concrete noun concepts only, which constitutes a limitation of these tests.A more comprehensive assessment of semantic memory would include action verbs and abstract concepts as well as other languages than English and Dutch (e.g., Kemmerer, 2019b).Meta-analyses of functional neuroimaging studies have shown that the left dorsolateral ATL and left IFG are activated more by abstract than concrete concepts, whereas the reverse holds true for left inferior temporal gyrus, fusiform gyrus, and medial temporal cortex (e.g., Del Maschio et al., 2022;Wang et al., 2010).The higher activation of left IFG for abstract than concrete concepts has been attributed to a greater need for top-down control in processing abstract concepts (cf.Fig. 1), whereas the higher activation of inferiormedial temporal cortex for concrete than abstract concepts has been attributed to the activation of perceptual features by concrete concepts (see Kemmerer, 2022, for discussion).Atrophy of left IFG in bvFTD leads to worse performance on abstract than concrete concepts (Cousins et al., 2016).However, atrophy of temporal cortex in SD may result in better or worse performance on abstract than concrete concepts, depending on the distribution of atrophy in the temporal lobes, among other factors (e.g., Hoffman et al., 2012).It has been argued that better performance on abstract than concrete concepts is observed in only a minority of patients with SD, whereas the reverse pattern holds for the majority of cases (e.g., Hoffman & Lambon Ralph, 2011).Cousins et al. observed better performance on abstract than concrete concepts in SD, arguing that atrophy predominantly affected perceptual features.Conversely, Jefferies et al. (2009) observed worse performance on abstract than concrete concepts in SD, arguing that atrophy did not much affect perceptual features.
Picture naming is worse than word comprehension in both SD and bvFTD (see Figs. 3 and 4).WEAVERþþ/ARC captures this finding, but this does not hold for other computational models.The Lichtheim 2 model of Ueno et al. (2011) has been applied to SD (but not to bvFTD).In Lichtheim 2, meaning (thought to be stored in ventrolateral ATL) is mapped via two layers of hidden nodes (in anterior STG/STS and in posterior IFG) onto articulatory output (in insular-motor cortex) in picture naming, while auditory word input (in primary auditory areas and surroundings) is mapped via two layers of hidden nodes (in middle STG/STS and in anterior STG/STS) onto meaning (in ventrolateral ATL) in comprehension.
Simulations with Lichtheim 2 conducted by Ueno et al. themselves showed how performance accuracy in SD varies as a function of lesion severity, which concerned the amount of noise over the output of the layer representing meaning and the proportion of damaged links into the layer (in ventrolateral ATL).These simulations revealed that increasing severity reduces naming and comprehension performance to the same extent, contrary to the empirical findings.Thus, WEAVERþþ/ ARC, but not Lichtheim 2, captures the finding that naming is worse than comprehension in SD and bvFTD.
In the WEAVERþþ/ARC model, weight and decay lesions influence the activation of nodes in similar ways, both reducing activation levels.A weight lesion affects the transmission of activation through the network, both locally (e.g., between concept nodes, the hub) and more remotely (e.g., between visual percepts and concepts or between concepts and lemmas, the spokes).A decay lesion affects the maintenance of activation of nodes in the network (here, the concept nodes), which works locally.Studies indicate that atrophy in SD and bvFTD reduces functional and structural connectivity in the brain (e.g., Seeley et al., 2009;Zhou et al., 2012), both locally (e.g., within the ATL) and more remotely (e.g., between the ATL and visual areas).Loss of network connectivity would correspond to a weight lesion rather than a decay lesion.However, weight and decay lesions are not mutually exclusive.For example, Dell et al. (1997) advanced a model of poststroke aphasia in which both lesions were present.In the present simulations of SD and bvFTD, a weight lesion only, and to a slightly lesser extent, a decay lesion only, suffices to explain most of the variance in naming and comprehension performance.Thus, the simplest account would be in terms of either a weight lesion or a decay lesion, but not both.Given that a weight lesion explains more variance than a decay lesion, and is supported by the loss of network connectivity that is empirically observed, a weight lesion is favored over a decay lesion.Future studies may examine the neural basis of both lesion types in more depth (e.g., Benhamou et al., 2020).

Conclusions
Assuming activation capacity loss in a semantic memory hub underpinned by the left ATL, WEAVERþþ/ARC simulations of word production and comprehension in SD and bvFTD revealed that the model succeeds reasonably well in capturing behavioral performance and atrophy ratings.In particular, the model explains 97% of the variance in naming and comprehension of 100 individual patients, and the degree of capacity loss in the model correlates with ratings of atrophy in the left ATL of the patients.Thus, activation capacity loss in a semantic memory hub provides a unified Pickian account of impaired word production and comprehension in SD and bvFTD as well as the cortical distribution of atrophy.

Credit author statement
2.1.Patients Demographic and clinical characteristics of the patients can be found in Goldberg et al. (2021) and Snowden et al. (2019).Goldberg et al. reported group averages for SD (11 patients) and bvFTD (26 patients) tested on the naming, comprehension, and repetition tasks of the SYDBAT.Snowden et al. also reported group averages for SD (32 patients) and bvFTD (71 patients) tested on naming and comprehension tasks, and additionally provided individual patient scores as supplementary data.

Fig. 1 e
Fig. 1 e Illustration of the functional neuroanatomy assumed by the WEAVERþþ/ARC model, superimposed onto the atrophied brain of seminal patient Anna Jirinec (adapted from Pick, 1904).N ¼ noun.

Fig. 2 e
Fig. 2 e Performance accuracy as a function of weight decrease (left panel) and decay increase (right panel) in the conceptual network in WEAVERþþ/ARC simulations of single word naming, comprehension, and repetition.

Fig. 3 e
Fig. 3 e Performance accuracy in semantic dementia for naming and comprehension: The left panel shows the performance of individual patients from Snowden et al. (2019) denoted by different colored lines and the group averages by black lines and squares.Numbered patients (e.g., #6) are discussed in the text.The right panel shows for each patient, denoted by dot and number with the tasks color coded, the predicted performance accuracy with the lowest mean absolute error (MAE) plotted against the observed accuracy.N ¼ number of patients.

Fig. 4 e
Fig. 4 e Performance accuracy in behavioral variant frontotemporal dementia for naming and comprehension: The left panel shows the performance of individual patients from Snowden et al. (2019) denoted by different colored lines and the group averages by black lines and squares.Numbered patients (e.g., #9) are discussed in the text.The right panel shows for each patient, denoted by dot and number with the tasks color coded, the predicted performance accuracy with the lowest mean absolute error (MAE) plotted against the observed accuracy.N ¼ number of patients.

Table 1 e
Kendall's tau correlations between weight decrease in WEAVERþþ/ARC and empirical atrophy ratings for brain regions of individual patients, p-value, and Bayes factor (BF).