Neural Network Bases of Thematic Semantic Processing in Language Production

Abstract

Semantic processing is a central component of language and cognition. The anterior temporal lobe is postulated to be a key hub for semantic processing, but the posterior temporoparietal cortex is also involved in thematic associations during language. It is possible that these regions act in concert and depend on an anteroposterior network linking the temporal pole with posterior structures to support thematic semantic processing during language production. We employed connectome-based lesion-symptom mapping to examine the causal relationship between lesioned white matter pathways and thematic processing language deficits among individuals with post-stroke aphasia. Seventy-nine adults with chronic aphasia completed the Philadelphia Naming Test, and semantic errors were coded as either thematic or taxonomic to control for taxonomic errors. Controlling for nonverbal conceptual-semantic knowledge as measured by the Pyramids and Palm Trees Test, lesion size, and the taxonomic error rate, thematic error rate was associated with loss of white matter connections from the temporal pole traversing in peri-Sylvian regions to the posterior cingulate and the insula. These findings support the existence of a distributed network underlying thematic relationship processing in language as opposed to discrete cortical areas.

1. Introduction

Semantic processing is a central aspect of language and cognition. The dual-stream language model suggests that semantic processing during language depends on ventral stream structures, which are focused on mapping sounds to meaning (Hickok & Poeppel, 2004). However, a more fine-grained anatomical localization of the components of semantic processing in the brain networks in general, and in the ventral stream more specifically, remains unclear. Different aspects of semantic processing may be mapped in the brain according to a specific topography. Thematic associations are a key component of semantic processing and involve the relationship between words that are associated with an event or scenario (Estes et al., 2011) or in space or time (Lin & Murphy, 2001), e.g., balloons and cake are related by the event birthday party. Thematic knowledge can be represented verbally (in language) and non-verbally (e.g., visually). In this study, we are specifically interested in verbal aspects of thematic knowledge. Thematic associations have been linked to temporoparietal regions (Kalénine & Buxbaum, 2016; Kalénine et al., 2009; Mirman et al., 2017; Schwartz et al., 2011). However, this localization is debatable, and other studies have suggested that thematic relationships are processed within a “thematic network” that may involve regions in the anterior temporal lobe (Jackson et al., 2015; Maguire et al., 2010; Sachs et al., 2008; Semenza et al., 1992). Therefore, it is possible that thematic associations may also be represented within a specific ventral stream network rather than by discrete regions.

Interestingly, individuals with semantic dementia have striking deficits in semantic processing (both thematic and taxonomic; Merck et al., 2020), and typically present with peaked grey matter atrophy well-localized to the anterior temporal lobes (ATLs) (Brambati et al., 2009; Neary et al., 1998; Nestor et al., 2006). Taxonomic errors, which relate to target words by common categorical properties (e.g., cat and dog) have been more exclusively linked to the ATL (Lewis et al., 2015; Mirman et al., 2017; Schwartz et al., 2011), partially due to evidence that individuals with semantic dementia produce more taxonomic than thematic errors (Merck et al., 2020). However, brain atrophy in semantic dementia also involves other fronto-temporal regions (Brambati et al., 2009; Hodges & Patterson, 2007), albeit to lesser extent, thus preventing a more precise anatomical dissociation between taxonomic and thematic associations. In addition, stroke patients with semantic aphasia (aphasia that is predominantly characterized by multimodal semantic deficits) demonstrate both taxonomic and thematic deficits (Thompson et al., 2017) despite lesions most commonly residing in the left inferior frontal and/or posterior temporoparietal cortex (i.e., not the anterior temporal lobes). Thematic deficits in those with semantic aphasia may be more pronounced than taxonomic deficits, which could be explained by differences in semantic control processes required for taxonomic versus thematic associations (Jefferies et al., 2020; Thompson et al., 2017) rather than a specialization of the ATLs for taxonomic associations and temporoparietal regions for thematic associations. Overall, these pieces of evidence call into question the existence of an independent hub for thematic associations in the posterior temporoparietal region.

Indeed, one possible reason for the seemingly divergent findings in the literature may be attributed to methodological challenges in identifying regions or networks that are crucial, and not simply implicated, with a given function. For instance, brain areas showing fMRI activation can be associated with but not necessarily causally related to a function. Thus, the observation of anterior and posterior functional regions being linked to thematic associations may be a reflection of shared cognitive processes non-specific to thematic processing. In this context, lesion-based studies can provide a more robust inference of functional localization, since the loss of a function associated with regional brain damage provides a causal relationship. However, lesion-based studies of semantic associations have thus far been limited for two reasons: first, the areas typically associated with semantic processing (e.g., the anterior portion of the temporal lobe or the parietal lobe) are in vascular perfusion territories shared between cerebral arteries and less likely to exhibit damage in the typical strokes of the middle cerebral artery associated with aphasia. Therefore, the statistical power to identify functional-lesional associations is reduced and prone to both type I and type II statistical errors. Second, complex cognitive functions usually depend on broader white matter networks that involve both grey matter structures and the white matter pathways that connect them, yet classical voxel-based lesion approaches have focused solely on the former. In fact, it is well established that the same behavioral symptoms may be caused by a damaged cortex or by disconnection between critical and intact cortical regions (Bonilha et al., 2014). Current advances in connectome-based lesion-symptom mapping (CLSM) have allowed for investigations of the relationship between lesioned white matter connectivity and behavior (Gleichgerrcht et al., 2017). Therefore, CLSM permits the evaluation of behavioral deficits as the result of disconnection from cerebrovascular injury to white matter pathways. As such, CLSM is an ideal neurobiological framework to study thematic semantic processing, since the areas postulated to be critical for this function are located in the anterior-most and posterior aspects of the ventral stream and are hence more likely disconnected than directly damaged by strokes. In this sense, grey matter lesion approaches are thus less likely to disclose the relationship between structural damage and cognitive impairment. Conversely, lesioned network based approaches are optimally suited for this investigation. To our knowledge, no studies have examined white matter connectivity subserving thematic semantic relationships.

Literature examining white matter connectivity underlying the semantic network, in general, has suggested that the inferior longitudinal, inferior frontal-occipital (IFOF), and uncinate fasciculi support semantic processing (Sierpowska et al., 2019). Indeed, these white matter pathways converge in temporal regions that are known to be important for semantic processing, such as the middle temporal and inferior temporal cortex (Sierpowska et al., 2019). One study showed that semantic abilities are linked with white matter pathways connecting the fusiform cortex to lateral temporal, medial temporal, and occipital regions (Chen et al., 2020). Likewise, Fang and colleagues (2015) proposed three white matter networks that are involved in semantic processing: the medial temporal network, the left frontal-subcortical network, and the frontal-temporal/occipital network. Interestingly, these three networks may serve different semantic functions; Ding et al. (2020) found the medial temporal network to be related to the semantic knowledge deficits in semantic dementia patients and the other two left frontal-subcortical and frontal-temporal/occipital networks to be related to semantic control deficits in semantic aphasia. This converges with findings that white matter projections between the frontal and temporal lobes and the IFOF, in particular, may be involved in lexical access rather than storage (Sierpowska et al., 2019). Overall, the evidence implicates specific white matter pathways that may be involved in semantics with early evidence that medial-temporal pathways relate to semantic knowledge and more lateral frontal-temporal pathways relate to semantic access or control. This suggests that evidence of a differentiated semantic network is supported by studies of white matter as well.

In the current study, we investigated the subnetworks associated with thematic processing in language by examining the relationship between thematic errors during confrontation naming and the integrity of white matter connections after stroke, controlling for taxonomic errors and for nonverbal conceptual-semantic knowledge. We used probabilistic tractography to estimate the connection strength between each pair of grey matter brain regions based on the number of probabilistic streamlines connecting two regions, corrected by length of each streamline and the size (total volume) of the two connected regions. We then examined the association between connection strength and thematic error rate. We hypothesized that thematic errors would be related to damage in white matter networks linking posterior and anterior ventral stream structures.

2. Material & Methods

We report how we determined our sample size, all data exclusions (if any), all inclusion/exclusion criteria, whether inclusion/exclusion criteria were established prior to data analysis, all manipulations, and all measures in the study.

2.1. Participants

Seventy-nine adults with chronic aphasia (47 females, 32 males) who participated in a multi-site aphasia treatment study conducted at the University of South Carolina and the Medical University of South Carolina were included in this retrospective study.³ Institutional Review Boards at both universities approved all study procedures, and informed consent was obtained from participants before study enrollment. Neuroimaging and behavioral data for this study were collected at baseline. Neuroimaging data were available for 92 participants; however, 13 were excluded from the current study because greater than 10% of semantic processing behavioral data were not available. Therefore, we included 79 participants in our analyses. Participants had a left-hemisphere ischemic or hemorrhagic stroke at least 12 months before their participation and had an aphasia diagnosis per the Western Aphasia Battery-Revised (WAB-R; Kertesz, 2007), primarily spoke English for ≥ 20 years before their stroke, and were between the ages of 21 and 80 years. Participants were excluded if they had profoundly limited verbal output (i.e., WAB-R Spontaneous Speech score of 0–1), profoundly limited auditory comprehension (i.e., WAB-R Auditory Verbal Comprehension score of 0–1), bilateral or right hemispheric stroke, or other brain injuries. These inclusion/exclusion criteria were determined prior to data analysis. On average, participants were 49.2 (SD = 51.3) months post-stroke and 55.8 (SD = 11.8) years of age at the time of stroke. For detailed participant demographics, see Table 1.

Table 1.

Participant Demographics

Variable	Mean (SD)	Range
Age (years)	60.0 (11.1)	29–80
Age at Stroke (years)	55.8 (11.8)	27–79
Months Post Stroke	49.2 (51.3)	12–241
Years of Education	15.5 (2.3)	12–20

Gender	Frequency (% of sample, n = 79)
Female	47 (59.5)
Male	32 (40.5)
Race
White	61 (77.2)
African American	17 (21.5)
Asian	1 (1.3)
Handedness
Right	69 (87.3)
Left	9 (11.4)
Ambidextrous	1 (1.3)
Aphasia Type (WAB-R)
Broca’s	37 (46.8)
Anomic	25 (31.6)
Conduction	12 (15.2)
Wernicke’s	3 (3.8)
Transcortical Motor	1 (1.3)
Global	1 (1.3)
Etiology
Ischemic	52 (65.8)
Hemorrhagic	19 (24.1)
Other	8 (10.1)

Note. “Other” etiology refers to cases in which it was unclear whether the stroke was ischemic or hemorrhagic.

2.2. Language Assessment

The Philadelphia Naming Test (PNT; Roach et al., 1996; https://mrri.org/philadelphia-naming-test/), a 175-item standardized object picture naming test, was administered to measure confrontation naming abilities. The PNT is a widely used naming test and is particularly suited for this investigation due to the high number of items and the rigorous scoring procedures, which includes the identification of different types of semantic errors. The PNT was administered twice at baseline (within the same week) to account for day-to-day variability in naming performance commonly associated with aphasia (Small et al., 1995). Participants were also administered the WAB-R to confirm the presence of aphasia and assess overall language abilities and the Pyramids and Palm Trees Test (Howard & Patterson, 1992) to assess nonverbal conceptual-semantic knowledge. Legal copyright restrictions prevent public archiving of the WAB-R and Pyramid and Palm Trees Test, which can be obtained from the copyright holders in the cited references.

During naming, word retrieval errors, or paraphasias, can be broadly categorized as semantic or phonological. Semantic errors have a relationship with the target word based on meaning (e.g., “banana” for “apple”), and phonological errors have a relationship with the target based on speech sounds (“mat” for “cat”). Mixed paraphasias are both semantically and phonologically related to the target (e.g., “rat” for “cat”). Semantic and mixed paraphasias can be further subcategorized based on the type of semantic relationship with the target word. For example, “banana” can be subtyped as a category coordinate of the target “apple” because it is a member of the same semantic category as the target. Similarly, other semantic paraphasia subtypes are superordinate (e.g., “fruit” for “apple”) or subordinate (e.g., “granny smith” for “apple”). Superordinate, category coordinate, and subordinate semantic errors can be grouped as taxonomic errors, as they all relate to the target in their taxonomic/hierarchical relationship. In contrast, thematic errors relate to the target by an event or action and may or may not be taxonomically related to the target (Estes et al., 2011). For example, “peeler” for “apple” (Roach et al., 1996).

Trained graduate research assistants scored accuracy on the PNT and transcribed and coded errors under the supervision of certified speech-language pathologists (SLPs), according to guidelines set by Roach et al. (1996) (see Spell et al., 2020 for details related to training and assessment fidelity). PNT errors were first coded as phonological (real words or nonwords related phonologically), semantically related, semantically unrelated, mixed (related semantically and phonologically), abstruse neologism (phonologically unrelated nonwords), or no response. Next, additional scoring criteria were developed for the purpose of the current analysis. Specifically, semantic and mixed errors were combined and subsequently coded by two separate trained research assistants (again under the supervision of an SLP) according to their semantic relationship to the target. These errors were coded into the following categories outlined in the PNT scoring guidelines: synonym, category coordinate, superordinate, subordinate, associated, and diminutive. Additionally, the two research assistants independently scored 15% of the data for reliability. Inter-rater reliability was 94% agreement, and a third rater was consulted to resolve discrepancies. Category coordinate, superordinate, and subordinate errors were grouped as taxonomic errors, and associated errors were considered thematic errors. Synonyms and diminutives were not used in the analysis. Finally, taxonomic and thematic errors were counted for each participant for each of the two administrations of the PNT. As done by others (Kalénine & Buxbaum, 2016; Schwartz et al., 2011), errors were first coded according to their taxonomic relationship. If one did not exist, it was evaluated for a thematic relationship. Therefore, in cases where an error was both similar to the target (taxonomic) and could also be related thematically by event (e.g., “cat” for “dog” – both are members of taxonomic category “animals” but are also associated by event “cat chases dog”), the error was coded as taxonomic to arrive at a more pure set of thematic errors.

Taxonomic and thematic error scores were measured by calculating the proportion of taxonomic or thematic errors out of total errors to account for other error types (e.g., phonological and other semantic error types). Taxonomic and thematic error proportions were averaged across the two PNT administrations for each participant to obtain a single data point for each error type. There were 10 participants for whom we had data from only one PNT administration for various reasons (e.g., fatigue, technical malfunctions, no semantic errors, >10% missing data). For these participants, the scores from the single PNT administration were used instead of an average.

Prior to the CLSM analysis, we removed the shared variance between taxonomic and thematic error scores by regressing one error type onto the other. This was to ensure observed effects were related to the unique variance of each error type (Schwartz et al., 2011). In addition, non-verbal semantic comprehension abilities as measured by the Pyramids and Palm Trees Test were regressed out from the two error types to isolate the word retrieval stages of naming from the conceptual stage of naming (Schwartz et al., 2011). Residuals from these two regressions were then used as the behavioral scores in the CLSM analysis.

2.3. MRI Acquisition and Processing

High-resolution T1-weighted, T2-weighted, and diffusion echo-planar imaging (EPI) images were obtained from each participant typically within two days of behavioral testing at the University of South Carolina or the Medical University of South Carolina using a Siemens Prisma 3T scanner and 20-element head/neck (16/4) coil. T1-weighted images were acquired using an MP-RAGE sequence with a 256 × 256 matrix size (1 mm isotropic pixels), 192 slice-sequence, a 9-degree flip angle, with repetition time (TR) = 2250 ms, inversion time = 925 ms, and echo time (TE) = 4.11 ms with parallel imaging (GRAPPA = 2, 80 reference lines). T2-weighted images were acquired using sampling perfection with application-optimized contrasts and a different flip angle evolution (3D-SPACE) sequence. This 3D turbo spin-echo scan uses a 256 × 256 matrix size, 160 sagittal slices (1 mm thick), a variable flip angle, TR = 3200 ms, TE = 212 ms, and parallel imaging (GRAPPA = 80 reference lines). Finally, diffusion mono-polar EPI scans were acquired twice using a 210 × 210 mm matrix size, 43 volumes sampling 36 directions with b = 1000 s/mm2 (with seven volumes b = 0), a 90-degree flip angle, TR = 5250 ms, and TE = 80 ms with parallel imaging (GRAPPA = 2, 80 contiguous 1.5 mm slices). In the second series, phase encoding polarity was reversed.

Post-stroke lesions were manually drawn by experienced researchers who were blind to behavioral scores on the T2-weighted images using the software MRIcron. The manually drawn lesions were then mapped into standard space by first coregistering the individual T1 and T2 scans using the transforms in order to reslice the lesion into native T1 space. Then, to remove jagged edges associated with manual drawing, we smoothed the resliced lesion maps with a 3 mm full-width half-maximum Gaussian kernel. Third, to normalize the T1 images into standard space, we applied an enantiomorphic segmentation-normalization approach using SPM12’s unified segmentation-normalization (Ashburner and Friston, 2005) and a series of custom MATLAB scripts (Rorden et al., 2020). This approach creates a mirrored image of the right hemisphere, which was then co-registered to the native T1 image. A chimeric image was then created based on the native T1 scan with the lesioned tissue replaced by tissue from the mirrored hemisphere (Nachev et al., 2008). SPM12’s unified segmentation-normalization warped this chimeric image to standard space, the resulting spatial transform was applied to the T1 scan, and the lesion map and T2/DWI image. Finally, we binarized the lesion mask such that only voxels with a probability of > 50% were kept in the final normalized lesion mask.

Each participant’s individual whole-brain connectome was reconstructed from the neuroimaging data as described in Fridriksson et al. (2018). First, using SPM12’s unified segmentation-normalization, we segmented T1-weighted images into probabilistic grey and white matter maps. Second, each individual’s grey matter map was divided into 189 regions using the Johns Hopkins University (JHU) neuroanatomical atlas (Faria et al., 2012). Third, we registered the grey matter parcellation maps non-linearly into DTI space. Fourth, we computed pairwise probabilistic DTI fiber tracking for all pairs of grey matter regions. Diffusion images were undistorted using TOPUP (Andersson et al., 2003) and Eddy (Andersson and Sotiropoulos, 2016). Tractography was estimated using FSL’s FMRIB’s Diffusion Toolbox (FDT) probabilistic method (Behrens et al., 2007). We used FDT’s accelerated BEDPOST (Hernandez et al., 2013) to measure default distributions of diffusion parameters at each voxel. Probabilistic tractography was then performed using FDT’s probtrackX (parameters: 5000 individual pathways drawn through the probability distributions on principal fiber direction, 0.2 curvature threshold, 200 maximum steps, 0.5 mm step length, and distance correction). Fifth, we determined the strength of each pairwise connection based on the number of probabilistic streamlines connecting each grey matter region pair. We corrected the pair-wise connectivity strengths by distance travelled by each streamline and the total volume of the connected regions. Finally, we constructed a weighted adjacency matrix M of size 189 × 189 for each participant with M_i,j representing the connectivity strength between regions-of-interest I and j, (averaged with the connectivity strength between regions of interest j and I, since diffusion tensor imaging is not directional).

2.4. Statistical Analysis

We performed CLSM (Gleichgerrcht et al., 2017) to analyze the relationship between white matter structural connectivity and residualized thematic error proportions. We also analyzed this relationship for residualized taxonomic error proportions as a comparison. The analysis was performed using NiiStat toolbox (https://www.nitrc.org/projects/niistat/) on MATLAB software (R2021a, version 9.10.0.1739362). We restricted the analysis to a set of 20 language-specific and domain-general brain regions in the perisylvian and pericentral areas that have been linked to language or support language (see Figure 1 for a depiction of regions included and the Appendix for a list of regions). The selection of these regions was primarily based on a body of work by Fedorenko and colleagues (Fedorenko, 2014; Fedorenko & Thompson, 2014; Fedorenko et al., 2010, 2011, 2012). These regions were also selected because they are commonly affected by middle cerebral artery strokes. Because univariate statistics provide independent tests of each brain region, and multivariate methods are able to leverage the synergistic contribution of multiple regions, we conducted both types of analyses. The multivariate approach is able to consider all connections in one model and can thus account for connections that are not fully independent of one another (DeMarco & Turkeltaub, 2018). All multivariate statistics are based on leave-one-out training and testing. In the univariate approach, connection strength between each set of ROIs, a continuous measure based on probabilistic tractography as described above, is included in a regression model as a predictor of the behavioral variable (residualized thematic or taxonomic error rate), controlling for total lesion volume. We used a false discovery rate of .05 to correct for multiple comparison. In the multivariate approach, we used NiiStat to perform a support vector regression in which the connection strength for each pair of ROIs is included in a single regression model as a predictor until the optimal model to predict the behavioral variable is selected. We controlled for total lesion volume in the model as well.

Figure 1.

Regions of Interest

Note. These 20 ROIs were chosen based on work by Fedorenko and colleagues (Fedorenko, 2014; Fedorenko & Thompson, 2014; Fedorenko et al., 2010, 2011, 2012) demonstrating that these regions play a role in language-specific or domain-general functions related to language. See the Appendix for a full list of regions from the JHU atlas.

2.5. Data Availability

Connectome and behavioral data to replicate the analysis are available at https://osf.io/xbhz8/. The conditions of our ethics approval do not permit public archiving of anonymised raw data. Data will be made available upon request to the corresponding author and in accordance with ethical procedures governing the reuse of sensitive data, which includes completion of a data sharing agreement and approval by the local ethics committee.

3. Results

3.1. Behavioral Data

On a single PNT administration, participants produced an average of 5.3 (3.0%) semantic errors and 2.9 (1.7%) mixed errors, totaling an average of 8.2 (4.7%) combined semantic errors per person. Participants produced 427 taxonomic and 210 thematic errors across all 79 participants. Participants produced an average of 5.4 taxonomic errors (SD: 4.9, Range: 0–24; 3.1%) and 2.7 thematic errors (SD: 3.0, Range: 0–16; 1.5%) on the PNT. Most participants (78.5%) produced more taxonomic than thematic errors, 8.9% produced the same number of taxonomic and thematic errors, and 12.7% produced more thematic errors than taxonomic. A Wilcoxon signed-rank test showed that participants produced significantly more taxonomic than thematic errors (z = −6.12, p < .001). Additionally, the number of taxonomic and thematic errors were significantly correlated (r = .63, p < .001). Taxonomic errors were not significantly correlated with lesion volume (Spearman’s correlation: .218, p = .053), but thematic errors were significantly correlated with lesion volume (Spearman’s correlation: .317, p = .004). After converting raw error counts to error proportions, neither taxonomic nor thematic error proportions were significantly correlated with lesion volume (Spearman’s correlation: −.096, p = .402, Spearman’s correlation: .110, p = .335, respectively). See Table 3 for language scores of included participants. See Figure 2 for visualizations of the behavioral data.

Table 3.

Participant Scores on Language Tests

Variable	M(SD)	Range
WAB_AQa	62.4 (21.8)	22.8–95.4
Percent PNT Correct	50.0 (33.7)	0.0–98.3
PPTTb	45.3 (6.1)	14–52

Note.

^aWAB_AQ = Western Aphasia Battery-Revised Aphasia Quotient. Max score is 100.

^bPPTT = Pyramids and Palm Trees Test. Raw score out of 52.

Figure 2.

Behavioral Data

Note. A. The distributions of the number of taxonomic and thematic errors produced. B. A scatterplot of thematic versus taxonomic error proportions with a regression line. The size of each data point is proportional to the participant’s lesion volume. C. A scatterplot of taxonomic (gold) and thematic (blue) error proportions versus PPTT scores with regression lines. The size of each data point is proportional to the participant’s lesion volume. One participant who scored as an outlier on the PPTT (raw score = 14) was removed from this plot to aid in visualization.

Taxonomic error proportions were also significantly higher than thematic error proportions (Wilcoxon signed-rank test: z = −5.74, p < .001). Taxonomic and thematic error ratios were significantly correlated (r = .42, p < .001), justifying the residualization of the error proportions to account for shared variance before lesion-symptom mapping. Both error proportions were also significantly correlated with the Pyramids and Palm Trees Test (PPTT) (Taxonomic: r = .26, p = .02; Thematic: r = .27, p = .02), supporting our decision to regress out the variance associated with this test of semantic knowledge before performing lesion-symptom mapping.

3.2. Connectome-based Lesion Symptom Mapping

Figure 3 shows the lesion overlay of all included participants. Using regions from the JHU atlas, the top three regions of peak overlap were in the left 1) superior longitudinal fasciculus, 2) postcentral gyrus, and 3) posterior insula. CLSM using the univariate approach showed two white matter connections among the 20 ROIs in the left hemisphere that were significantly associated with the residualized thematic error rate after accounting for lesion volume: connections between the 1) pole of the middle temporal gyrus (MTG) and posterior cingulate gyrus (PCG; z = −3.71, p = .0001), and 2) inferior temporal gyrus (ITG) and insula (z = −3.58, p = .0002; See Figure 4). There were no white matter connections that were significantly associated with residualized taxonomic error rate, which was included as a control. Figure 5 is a scatterplot of the connection strengths between each pair of significant regions and thematic error proportions.

Figure 3.

Lesion Overlay

Note. Lesion overlay of all 79 participants included in the study. Warmer colors denote brain regions where more participants had a stroke lesion in that area. The scale bar indicates the number of participants.

Figure 4.

CLSM Results

Note. A. A depiction of the strongest connections between the 20 left-hemisphere ROIs, averaged across all participants. Colored bars represent average percent lesion volume in that region across participants. The presence of a strong connection was binarized based on a set threshold of connection strength for visualization purposes. B. A connectome matrix of the 20 left-hemisphere ROIs using the JHU atlas. Brighter/warmer colors indicate stronger white matter connections between regions, depicting the connection strength on a continuous scale. See region abbreviations in the Appendix. C. Significant connections in left hemisphere associated with residualized thematic error proportions: 1) connection between the ITG and insula and 2) connection between the pole of the MTG and posterior cingulate gyrus. Breakdown of the main tracts within each connection is shown in legend. Of note, the pathways shown here were reconstructed through deterministic tractography using data from 842 neurotypical participants in the Human Connectome Project (Yeh et al., 2013) for visualization purposes. The CLSM analyses employed probabilistic tractography as described in the text.

Figure 5.

Plot of Connection Strength in Relation to Thematic Error Proportions

The multivariate SVR analysis revealed a significant association between structural connectivity and residualized thematic error proportions (C = .10, r = .219, p = .03), with the top 10 most influential connections listed below. The two connections that were revealed in the univariate analyses using stringent multiple comparisons procedures are in bold. As with the univariate analyses, there were no SVR models that significantly predicted taxonomic error rates.

Henceforth, the asterisk (*) denotes a connection.

1. STGpole*posterior MTG (z = −2.42, p = .008)

2. STGpole*posterior cingulate gyrus (z = −2.37, p = .009)

3. insula* ITG (z = −2.36, p = .009)

4. insula*MTGpole (z = −2.25, p = .012)

5. precentral gyrus*angular gyrus (z = −2.11, p = .017)

6. insula*MTG (z = −2.07, p = .019)

7. insula*posterior MTG (z = −1.82, p = .034)

8. insula*superior temporal gyrus (STG) (z = −1.82, p = .034)

9. MTGpole*posterior cingulate gyrus (z = −1.77, p = .038)

10. insula*superior parietal gyrus (z = −1.76, p = .039)

3.3. Outliers

There was one extreme outlier among participants who fell outside three times the quartile range of taxonomic error proportions. Therefore, we conducted a sensitivity analyses to determine whether exclusion of this datapoint affected results. After excluding this participant and recalculating residuals, an additional white matter connection was significantly associated with residualized thematic error proportions. The connection between the STGpole and the PCG (z = −3.41, p = .0003) and the STGpole and the posterior MTG (z = −2.97, p = .001) were significantly associated with residualized thematic error proportions in addition to the MTGpole*PCG connection. Of note, these two connections were also present in the top 10 influential connections within the SVR results reported when including all participants (the top 2 most influential connections). If we remove all outliers among taxonomic and thematic error proportions (outside 1.5x the quartile range) as identified by boxplots in SPSS, the insula*ITG connection is no longer significantly associated with residualized thematic error proportions and the MTGpole*PCG (z = −3.09, p = .0003), and STGpole*PCG (z = −3.27, p = .0003) connections remain significant. Across all these analyses, the most consistent connection associated with residualized thematic error proportions is the MTGpole*PCG connection. In every analysis for which we removed at least one outlier, the STGpole*PCG connection is also significantly associated with residualized error proportions. If we redo the multivariate SVR analysis without the extreme outlier, we get another model that significantly predicts thematic error proportions above chance (C = .025, r = 0.261, p = 0.01).

1. STGpole*PCG (z = −2.75, p = .003)

2. insula*ITG (z = −2.63, p = .004)

3. insula*MTG (z = −2.59, p = .005)

4. MTGpole*PCG (z = −2.39, p = .008)

5. insula*STG (z = −2.17, p = .015)

6. STGpole*posterior MTG (z = −2.16, p = .015)

7. STG*PCG (z = −2.08, p = .019)

8. insula*posterior MTG (z = −1.93, p = .027)

9. insula*superior frontal gyrus (z = −1.78, p = .038)

10. MTG*PCG (z = −1.78, p = .038)

The three new connections are in bold after removing the outlier, which replaced these three connections from the top 10 connections in the original SVR model: insula*MTGpole, insula*superior parietal gyrus, and precentral gyrus*AG. Of note, the STGpole*PCG connection along with the original insula*ITG and MTGpole*PCG connections are in the top four connections of this model without the extreme outlier.

3.4. Post-hoc Analyses

Due to the fact that there are a variety of methods possible to measure taxonomic and thematic error rates, we conducted a series of post-hoc analyses to further validate our findings by measuring error rates in relation to a) semantic errors only, b) total items (excluding missing data), and c) total items, accounting for other error types [number of taxonomic or thematic errors/(total completed items – number of other non-taxonomic or -thematic errors)]. We followed the same residualization procedures described above wherein we regressed taxonomic and thematic error rates against the other and regressed out variance related to conceptual semantic abilities via the PPTT scores. We then conducted three additional CLSM procedures for each of these different behavioral variables.

For the univariate CLSM analysis examining connections between regions associated with residualized thematic error rates in relation to total semantic errors, we found that residualized thematic error rates were significantly associated with connections between the left i) insula and posterior STG (z = −4.06, p < .0001), ii) ITG and posterior MTG (z = −3.90, p < .0001), and iii) insula and posterior MTG (z = −4.01, p < .0001). For the univariate CLSM analysis in which we calculated error rates in relation to total items (excluding missing data), we did not find any significant associations between behavior and connection strength. Finally, for the univariate CLSM analysis in which we calculated error rates in relation to total items (excluding missing data) and accounting for other (non-taxonomic or -thematic error types), we, again, did not find any significant associations between behavior and connection strength. As with our original analysis, we did not find any significant associations between taxonomic error rates and connection strength for any of the above post-hoc analyses. We report details of multivariate CLSM analyses using SVR for each of these different outcome variables in Supplementary Material as well, but in brief, results using the error type proportion out of total semantic errors differed from the univariate analysis, and we found no significant models associated with error types in proportion to total items or total items accounting for other error types.

To further examine how these two connections might fit into the semantic system, we examined data from the Human Connectome Project and the fMRI meta-analytic tool, Neurosynth (neurosynth.org). Neurosynth synthesizes data from research studies that use fMRI to examine various neuropsychological processes. Users are able to examine meta-analytic data from many studies that are grouped by key words included in the abstract of the article. We chose to use the term “associative” to synthesize studies investigating associative memory because there was no “thematic” term available for searching and we considered the term “association(s)” to be potentially too broad. We used the term “semantic” to synthesize studies investigating semantic memory as this term resulted in far more meta-analytic data than other terms such as “semantically”, “semantic memory”, “semantics”. Though this synthesis approach is likely imperfect, according the Neurosynth validation study, this term approach to searching for related studies has been shown to be valid (Yarkoni et al., 2011). Looking at the activation map of associative semantic memory from Neurosynth (term: “associative”), associative memory is most represented in the medial temporal lobe, including the hippocampus, parahippocampus, and the perirhinal and entorhinal cortices (Achim et al., 2007; Ford et al., 2010; Hales et al., 2009). We visualized these associative memory cortical regions along with the white matter connections we found to be important for thematic associations, using data from 842 neurotypical participants in the Human Connectome Project using DSI Studio (Yeh et al., 2013; See Figure 6). Based on the proximity of these white matter connections to the associative memory system, it is possible that, over time, thematic associations encoded by the medial temporal lobe are gradually mapped onto regions in the lateral temporal lobe important for multimodal processing and language. Indeed, a similar map for semantic processing (term: “semantic”) demonstrates regions that are projected from the connections observed in our study onto the neocortical structures. Therefore, when these more lateral temporal regions are damaged, thematic errors in naming may occur, at least in part, due to disconnection of this network. Although this hypothesis was not directly tested in this study, it is a possible interpretation of the current results that aligns with findings that thematic processing is linked to neural substrates of memory (Maguire et al., 2010), and that neural substrates of associative memory shift from medial to lateral regions within hours of encoding (Nieuwenhuis et al., 2012). Therefore, this hypothesis lends itself to further testing.

Figure 6.

Associative Memory and White Matter Connections Related to Thematic Error Production

Note. The top mosaic demonstrates the spatial relationship between meta-analytical brain regions involved with associative tasks (pooled from data from 295 studies using Neurosynth). The bottom mosaic demonstrates the areas associated with semantic processing (based on data from 1031 studies using Neurosynth). The “hot” colormap indicates the voxel-wise Z score (both corrected for multiple comparisons using a false discovery rate of p = 0.01). On both mosaics, the tract density images representing the pathways found in this study are shown in blue (the blue color bar indicates the number of streamlines per voxel). Note the anatomical proximity between the white matter pathways and associative memory and semantic processing.

In addition, though not the focus of this paper, we attempted to replicate the VLSM analysis reported by Schwartz et al. (2011), which found a neuroanatomical dissociation between taxonomic and thematic errors. In short, we were unable to replicate their findings and did not find any significant voxels associated with residualized taxonomic and thematic error rates despite following their methods as closely as possible and with a comparable sample size. Additionally, we were unable to replicate Schwartz et al. (2011) using our primary outcome measure. Details of this analysis are reported in Supplementary Material.

4. Discussion

To better understand the neural network underlying thematic associations, we examined whether the integrity of white matter connections within the ventral stream of language processing was associated with the production of thematic semantic errors during picture naming in participants with chronic post-stroke aphasia. Behavioral results showed that participants produced more taxonomic errors than thematic errors, which has been reported previously in the aphasia literature (Schwartz et al., 2011). In addition, univariate CLSM results showed two white matter connections associated with the production of thematic errors during picture naming: the connection between the pole of the MTG and the PCG and the connection between the ITG and insula. These are pathways traversing the temporal lobes linking anterior and inferior temporal regions to peri-Sylvian structures. The multivariate SVR CLSM results showed 10 significant white matter connections that were significantly associated with residualized thematic error rates, two of which are the connections revealed in the univariate analysis. Additional connections between the insula and MTG, MTG pole, posterior MTG, STG, and superior parietal gyrus were revealed as well as three other connections linking anterior and posterior regions (STGpole*posterior MTG, STGpole*PCG, precentral gyrus*angular gyrus).

Post-hoc univariate analyses revealed that when measuring thematic error rates in relation to semantic errors, the connections between the left i) insula and the posterior MTG, ii) insula and posterior STG, and iii) ITG and posterior MTG were significantly associated with thematic error rates. Although these three connections are different than those previously found in relation to total errors, all four of these regions (insula, posterior MTG, posterior STG, and ITG) were identified in the connections found in relation to total errors, and they traverse the temporal lobe, connecting anterior to posterior regions.

The ATLs have previously been hypothesized as a semantic “hub” wherein semantic information from modality-specific sensorimotor brain areas is integrated, and an amodal/transmodal semantic representation is stored (Lambon Ralph et al., 2010; Lambon Ralph, 2014; Schwartz et al., 2009). The ATL “hub-and-spoke” model asserts a bilateral semantic system with evidence of left-hemisphere specialization for accessing semantic representations from a written word stimulus (versus auditory or pictorial stimuli, for example) or for naming tasks (versus other semantic tasks) (Rice et al., 2015; Rice et al., 2018). However, the exact anatomic location within the ATL that functions as the hub is still under debate. For example, the temporal pole (Lambon Ralph et al., 2009; Pobric et al., 2007) and fusiform gyrus (Binney et al., 2010; Chen et al., 2020) have each been suggested as possible focal points. Recent studies appear to garner support for the ventral ATL (which includes the fusiform cortex) as the semantic hub, however, it has been proposed that both the fusiform and temporal pole may be hubs and play similar or distinct roles (Chadwick et al., 2016; Chen et al., 2020). The current findings are consistent with the view that thematic information from the posterior cingulate is transferred to the ATL for integration. These results are also consistent with the alternative possibility that the ATL is used to access the thematic network supported by the posterior cingulate in verbal tasks. Some have postulated the existence of a thematic hub in the temporoparietal junction (Mirman et al., 2017; Schwartz et al., 2011). Indeed, one of the two significant connections associated with residualized thematic errors links a region close to the temporoparietal junction (PCG) with the anterior temporal region (temporal pole) and two additional multivariate connections associated with residualized thematic errors also link regions in or surrounding the temporoparietal region with the temporal pole (STGpole*posterior MTG, STGpole*PCG). Though these findings do not support a fully independent thematic hub, they may support the role of the posterior temporal and parietal lobes in thematic processing as part of the larger semantic network involving the ATL. The MTG, more generally, and the posterior MTG, specifically, have also been linked with thematic (Davey et al., 2016; Henseler et al., 2014; Thompson et al., 2017) and action/tool semantic processing (Binder et al., 2009; Kalénine et al., 2010), making it possible that thematic relationships are partially represented in the temporoparietal area, where action representations are also thought to be stored and integrated with the more anterior amodal/transmodal semantic processing hub. The posterior MTG has also been linked with anterior regions of the brain (the anterior inferior frontal gyrus) in the semantic control of thematic associations and has been proposed as a region that integrates automatic semantic spreading activation controlled by the default mode network and executive multiple demand network to facilitate activation of weaker semantic associations (Davey et al., 2016). Our findings support such claims surrounding the role of the MTG in thematic associations.

The posterior cingulate cortex (PCC) has been shown to play a role in a range of functions, including attention, retrieval of autobiographical memory, planning for the future, and the default mode network (Leech & Sharp, 2014). The default mode network and semantic network engage similar brain regions (Binder & Desai, 2011; Binder et al., 2009; Humphreys et al., 2015) and may share some functions. It has been suggested that semantic representations are likely active during the conscious resting state as part of internal thought (Binder et al., 2009). The PCC also appears to be involved in spatial processing and action in space (Barrett et al., 2019; Rolls, 2019). The PCC’s exact role in language is not yet clear, but it has been implicated in semantic processing (Binder et al., 2009) and theorized to be involved in integrating memory, emotion, and language (Maddock et al., 2002). The role of the PCC in spatial and action processing aligns with thematic relationships, which are primarily characterized by actions and events (e.g., dog chews a bone). The PCC is also strongly connected to the hippocampus (Beckman et al., 2009), which supports its role in event-based or episodic representations. These results appear in line with research that has found a link between action (verb) retrieval and the temporoparietal junction (Bedney et al., 2014; Kalénine et al., 2010; Martin et al., 1995; Pisoni et al., 2018). These thematic relationships may be stored alongside the action/verb representations and then integrated with the semantic representations in the ATL, or accessed through the ATL.

Similarly, the insula appears to be involved in various functions, including sensorimotor processing, emotional processing, and cognition (Ibañez et al., 2010; Uddin et al., 2017), and its exact role in language processing has been debated. The insula has previously been implicated in the dorsal stream of language as part of the articulatory network (Hickok and Poeppel, 2007), and damage to the insula has been associated with conduction aphasia (Ardila, 2018), which is characterized by phonemic paraphasias. These previous reports raise questions about why several connections between regions in the temporal lobe and the insula were associated with thematic semantic error production. Indeed, damage to the insula has also been linked to Broca’s aphasia (especially the anterior insula), Wernicke’s aphasia (posterior insula), and aphasia without phonological impairments (Ardila, 2018; Baratelli et al., 2015). Furthermore, two meta-analyses examining the role of the insula in speech and language processing found the insula to be involved in a range of language functions, in addition to articulation (Ardila et al., 2014; Oh et al., 2014). These studies suggest that the insula is connected to several language regions, and it is thus not surprising that it may serve a function in semantic processing, though its exact function is still unclear. The insula is a region most commonly affected by MCA strokes and insular damage is correlated with larger stroke lesions (Kodumuri et al., 2016). Although we controlled for lesion volume, this prevalence of insular damage among stroke survivors has led to its association with a wide range of functions (Kodumuri et al., 2016). Therefore, further research is needed to clarify the insula’s role in semantic processing.

The ITG has been included in the semantic ventral stream of language (Hickok and Poeppel, 2004), particularly the posterior portion of the ITG (Binder et al., 2009). Previously, semantic errors in aphasia have been associated with damage to and/or hypoperfusion in Brodmann’s areas (BA) 22 (posterior superior temporal gyrus/Wernicke’s area) and 37 (posterior inferior temporal and fusiform gyri). Cloutman and colleagues found that BA 22 was involved in semantic processing for both naming and comprehension, but BA 37 was only implicated during naming (Cloutman et al., 2009). The authors concluded that BA 37 is likely involved in semantic access rather than semantic knowledge. Given that participants from our study completed a naming task, our finding regarding the involvement of the ITG may also point to a function of semantic access as it pertains to thematic relationships. This interpretation is bolstered by the fact that we controlled for semantic knowledge via the Pyramids and Palm Trees Test in our behavioral measure, which suggests that the significant connections in this study are relevant to processes of lexical-semantic access during retrieval rather than semantic representations themselves; however, future research is required to examine this claim further. In sum, there is evidence that both the insula and the ITG are involved in language processing, and our data suggest these regions may be involved in thematic associations. It is worth noting, however, that when excluding an extreme outlier, this insula*ITG connection did not come up in the univariate analysis (though it did in the SVR analysis). As such, the specific roles of these structures and their interaction are still unclear and require further inquiry.

Connections involving the STG were also present in the multivariate SVR analyses and the outlier/post-hoc analyses and suggest a role of the STG in verbal thematic processing. Specifically, the connections between the STGpole and the a) PCG and c) posterior MTG, which came up in univariate analyses when excluding outliers and in the multivariate SVR analysis, is consistent with the hypothesis discussed previously that action processing in the PCG and posterior MTG is integrated with transmodal semantic representations in the temporal pole. Historically, the STG has been closely linked with language comprehension abilities (Wernicke, 1874; Geschwind, 1971; Hillis et al., 2001); however, more recent evidence suggests that its primary role is in phonological processing rather than semantic processing (Binder et al., 2009). Therefore, its appearance in these results as related to thematic processing is somewhat surprising. Nevertheless, these data suggest the STG may play a role in thematic processing.

In addition, the multivariate analyses also revealed that the connection between precentral gyrus and angular gyrus may be associated with verbal thematic processing. The angular gyrus has been regularly documented as related to semantic processing, and thematic processing, in particular (Jefferies et al., 2020; Schwartz et al., 2011). Damage to the precentral gyrus has been noted to be a significant predictor of PNT naming performance, especially due to its role in speech production (Fridriksson et al., 2018). The significance of the connection between the angular gyrus and precentral gyrus may represent the transfer of thematic information to the motor cortex prior to production. Indeed, additional research to clarify the roles of these regions and their connections is needed.

Of note, data from the Human Connectome Project showed that connections between the ITG and insula contain tracts within the uncinate fasciculus (57% of insula to ITG tract), arcuate fasciculus (28% of insula to ITG tract), and parahippocampal cingulum (5% of insula to ITG tract) (see Figure 4C). Indeed, the uncinate fasciculus and parahippocampal cingulum have been associated with semantic processing (Hoenig & Scheef, 2005; Pisoni et al., 2018). The arcuate fasciculus is more regularly linked with phonological processing in the dorsal stream (Dick & Tremblay, 2012); however, Glasser and Rilling (2008) point out that the medial portion of the arcuate may be involved in lexical-semantics. White matter tracts between the pole of the MTG and PCC contain fibers within the inferior thalamic radiation (67% of the MTG to PCC tract) and the parahippocampal cingulum (30.3% of the MTG to PCC tract). The inferior thalamic radiation has been suggested to play a role in integrating lexical-semantic information between the thalamus and cortical areas (Crosson, 2013), and the parahippocampal cingulum may play a role in processing of lexical-semantic ambiguity (Hoenig & Scheef, 2005). We have put the full breakdown of the tracts included in these two connections in the Supplementary Material. Therefore, the white matter connections we found to be associated with thematic errors have previously been shown to be involved in semantic processing.

The finding that damage to these white matter connections was associated with a higher rate of thematic errors could suggest that these connections are involved in either inhibiting or activating thematic competitors during naming. One avenue for future research is to determine the extent to which type of thematic association, such as thematic associations that are causal (e.g., rain and flood), action-based (e.g., knife and steak), or spatially related (e.g., balloons and cake), might be differentially represented in the thematic semantic errors produced by people with post-stroke aphasia depending on the lesion. For example, the strength of thematic associations has been related to whether there is an action-based thematic relationship (as with manipulable objects) or a non-action-based thematic relationship (Tsagkaridis et al., 2014). Moreover, integrity of structures and functional connectivity among regions that are important for movement have been uniquely linked to comprehension of manipulable compared to nonmanipulable objects (Riccardi et al., 2020). Thus, future studies should examine whether semantic category or object type (e.g., whether the object is a manipulable artifact or natural object) modulates type of thematic semantic errors and determine whether these patterns are associated with specific lesions to cortical grey matter or white matter pathways. These data could further enhance our understanding of the neural networks underlying thematic relationships.

It is noteworthy that patients with semantic dementia and patients with post-stroke aphasia both demonstrate a similar pattern of more taxonomic than thematic errors despite the difference in lesion pattern. As mentioned previously, those with semantic dementia typically have focal deterioration of the ATL whereas those with post-stroke aphasia typically have lesions in MCA territory, which can include portions of the ATL, but often occur in the insula, along the temporal lobe, and temporoparietal cortex. It is possible that the similar behavioral pattern in error types reflects that in both populations the associative semantic network, as a whole, is disrupted, but at different locations. In other words, in both populations, the anterior-posterior connections are being disrupted, albeit in different cortical regions of the network, and resulting in similar deficits. Indeed, error patterns as they relate to semantic deficits have led researchers to theorize that the ATL (more damaged in semantic dementia) is involved in storage of semantic representations and other regions (inferior frontal gyrus and posterior MTG, for example) are involved in semantic control (Jefferies et al., 2020; Thompson et al., 2017). Our results suggest that people with aphasia may have impaired access to thematic associations due to disconnection of important pathways within the thematic subnetwork due to lesions in the MCA territory and not the temporal pole (as the temporal pole was not very affected in our participants, see Figure 1). People with semantic dementia may suffer a similar deficit because their lesions in the temporal pole/ATL disconnect these same important pathways and/or the semantic representations themselves, which would align with our findings that the temporal pole (MTGpole and STGpole, specifically) are involved in thematic associations.

Our study must be interpreted in the context of certain caveats. First, there is a level of subjectivity in coding semantic errors. Although we had high reliability between coders, the occasional ambiguous dissociations between thematic and other types of relationships (e.g., taxonomic) may have affected results. As has been done previously, we chose to err on the side of classifying relationships as non-thematic if there was debate as to whether it was a truly thematic association to arrive at a more homogenous collection of thematic errors (Kalénine & Buxbaum, 2016; Schwartz et al., 2011) given thematic relationships were the focus of this paper. However, this may have influenced outcomes. Indeed, distinguishing between the type of semantic relationships can be difficult as relationships between words are often complex and multifarious (Jackson & Bolger, 2014). This strategy may also explain why we did not find any significant white matter associations with taxonomic error proportions—since these two error types are not doubly dissociated, errors that could be interpreted as both taxonomic and thematic and were, in the case of this study, coded as taxonomic may have prevented the ability to find distinct networks associated with taxonomic associations. A related possibility is that superordinate, category coordinate, and subordinate errors have different neuroanatomical associates and grouping them together limited our ability to observe effects. Another limitation that we acknowledge is that we restricted our analysis to the left hemisphere and a specific subset of regions. We did so in order to retain enough power to reveal any existing associations between white matter connectivity and behavior and chose language-specific and domain-general regions that, according to literature, support language processing (Fedorenko, 2014; Fedorenko & Thompson, 2014; Fedorenko et al., 2010, 2011, 2012). A downside to this approach is that we could miss other important intra- or interhemispheric connections involved in the production of thematic errors. Future studies should employ methods, such as structural disconnectome mapping (Salvalaggio et al., 2020; Theibaut et al., 2020) that could reveal associated white matter connections to the right hemisphere and regions outside the traditional language network, which will likely require a larger sample size.

5. Conclusions

The results from the present study suggest that thematic associations in the context of language are supported by a distributed network linking temporal pole and inferior temporal regions to more posterior and peri-Sylvian structures. These findings add to the growing body of evidence that semantic thematic processing during language processing is dependent on a component of the ventral stream network, which involves distributed anterior-posterior cortical grey matter regions and their white matter connections.

Abbreviations:

ATL

anterior temporal lobe

CLSM

connectome-based lesion-symptom mapping

WAB-R

Western Aphasia Battery Revised

PNT

Philadelphia Naming Test

SLP

speech-language pathologist

PPTT

Pyramid and Palm Trees Tees

SVR

support vector regression

Appendix

Regions of Interest Related to Language Function

JHU Atlas Region Number	JHU Region Name (all left hemisphere)	Abbreviation
1	superior frontal gyrus (posterior segment)	postSFG
7	middle frontal gyrus (posterior segment)	postMFG
9	middle frontal gyrus (dorsal prefrontal cortex)	MFG-DPFC
11	inferior frontal gyrus pars opercularis	IFGoperc
13	inferior frontal gyrus pars orbitalis	IFGorbit
15	inferior frontal gyrus pars triangularis	IFGtriang
25	precentral gyrus	PreCG
27	superior parietal gyrus	SPG
29	supramarginal gyrus	SMG
31	angular gyrus	AG
35	superior temporal gyrus	STG
37	pole of superior temporal gyrus	STGpole
39	middle temporal gyrus	MTG
41	pole of the middle temporal gyrus	MTGpole
43	inferior temporal gyrus	ITG
49	fusiform gyrus	FG
69	posterior cingulate gyrus	postCing
71	insula	Insula
184	posterior superior temporal gyrus	postSTG
186	posterior middle temporal gyrus	postMTG

Data availability:

In accordance with the National Institute of Health policy for data sharing, upon completion of the POLAR trial and dissemination of primary study results, the analysis data files will be made available to the public, along with the final version of the study protocol, the data dictionary, and brief instructions.