Statement of Research Significance
Research Question(s) or Topic(s): This study explores the white matter circuitry associated with semantic and phonemic verbal fluency in post-stroke aphasia. Main Findings: The findings suggest a partially segregated yet interconnected neural architecture for semantic and phonemic fluency within left hemisphere language networks and domain-general pathways. Study Contributions: Building on prior lesion-based evidence, the study contributes to a more comprehensive understanding of the structural neural networks supporting verbal executive functions. In particular, the study implies both the contribution of language networks and the role of domain-general control in verbal executive functions with visual inspection of the connectomes and supplementary cross-control analyses suggesting differences in the relative emphasis of specific pathways across fluency tasks. These observations may help refine theoretical accounts of language processing and executive functioning and could inform future approaches to clinical assessment and rehabilitation strategies.
Introduction
Verbal fluency tasks are extensively used in both clinical and research settings across diverse clinical and nonclinical populations, as they evaluate a range of cognitive functions, including executive functioning, language abilities, working memory, attention, and processing speed within a short time frame. In stroke patients with and without aphasia, verbal fluency is routinely assessed to advance understanding of brain–behavior relationships and incorporated into broader clinical evaluations of cognitive–linguistic functioning. Verbal fluency has been indicated as a sensitive clinical marker for the prognosis of executive functioning that closely parallels general functional change after stroke (e.g., Levine et al., Reference Levine, Galecki, Langa, Unverzagt, Kabeto, Giordani and Wadley2015).
The primary task types of verbal fluency are semantic fluency, which involves listing words from a specific category, and phonemic fluency, which requires generating words beginning with a specific letter. Studying the neural basis of fluency – especially in post-stroke aphasia (PSA) – is critical for advancing our understanding of how distributed brain networks support verbal executive abilities. Evidence from lesion-symptom mapping (LSM) suggests partially overlapping but differentially weighted substrates, with semantic fluency more consistently linked to left posterior and middle temporal regions, and phonemic fluency showing stronger associations with left inferior frontal and premotor regions (Baldo et al., Reference Baldo, Schwartz, Wilkins and Dronkers2006; Biesbroek et al., Reference Biesbroek, van Zandvoort, Kappelle, Velthuis, Biessels and Postma2016, Reference Biesbroek, Lim, Weaver, Arikan, Kang, Kim, Kuijf, Postma, Lee, Lee, Yu, Bae and Biessels2021; Chouiter et al., Reference Chouiter, Holmberg, Manuel, Colombo, Clarke, Annoni and Spierer2016; Thye et al., Reference Thye, Szaflarski and Mirman2021). Proposals regarding the precise anatomical correlates have been mixed, however, with converging evidence suggesting that subcortical contributions may be particularly important for both task types (Egorova-Brumley et al., Reference Egorova-Brumley, Liang, Khlif and Brodtmann2022; Li et al., Reference Li, Zhang, Song, Huang, Ding, Fang, Xu and Han2017; Thye et al., Reference Thye, Szaflarski and Mirman2021). Similarly, connectivity studies (Biesbroek et al., Reference Biesbroek, Lim, Weaver, Arikan, Kang, Kim, Kuijf, Postma, Lee, Lee, Yu, Bae and Biessels2021; Egorova-Brumley et al., Reference Egorova-Brumley, Liang, Khlif and Brodtmann2022; Li et al., Reference Li, Zhang, Song, Huang, Ding, Fang, Xu and Han2017) have yielded inconsistent results on which white matter pathways are specifically associated with phonological and semantic fluency, especially within the dual-stream (dorsal-ventral) network of speech processing (Hickok & Poeppel, Reference Hickok and Poeppel2007; Saur et al., Reference Saur, Kreher, Schnell, Kümmerer, Kellmeyer, Vry, Umarova, Musso, Glauche, Abel, Huber, Rijntjes, Hennig and Weiller2008). Notably, these studies have predominantly relied on lesion-based measures, such as lesion load, or tractography approaches limited to predefined pathways, or traditional metrics such as fractional anisotropy (FA). The specificity of different tracts for phonemic versus semantic fluency remains ambiguous, and discrepancies in previous findings further complicate interpretations.
Here, we aimed to define the microstructure of white matter tracts associated with semantic and phonemic fluency in a sample of 44 patients with PSA. To this end, we obtained multi-shell diffusion-weighted MRI (DW-MRI) and employed whole-brain quantitative anisotropy (QA)-based connectometry derived from correlational tractography. As opposed to fiber tracking methods based on FA, this approach has been shown to be less susceptible to biases introduced by, e.g., crossing fibers and partial volume effects (Yeh et al., Reference Yeh, Badre and Verstynen2016) and therefore, provides higher reliability and specificity in assessing white matter connectivity. Rather than capturing lesion effects per se, this whole-brain approach identifies behavioral associations with inter-individual variation in white matter architecture.
Materials and methods
Participants
Forty-four participants with chronic PSA (>6 months since onset) due to left-hemisphere ischemic or hemorrhagic stroke participated in the study between 2018 and 2020. All participants were over 18, spoke Finnish as a native language, and had no hearing deficits or psychiatric or neurological comorbidities. The study adhered to the Declaration of Helsinki and was approved by the Ethics Committee of Helsinki-Uusimaa Hospital District. All participants provided informed consent. As part of a broader neuropsychological assessment, verbal fluency was evaluated using one-minute tasks that involved naming as many animals (semantic fluency) and words beginning with the letter “s” (phonemic fluency) as possible, excluding proper nouns. Aphasia severity was assessed with Boston Diagnostic Aphasia Examination (BDAE) severity rating. There were no missing data. Demographic and clinical information of the patients is reported in Table 1.
Demographic and clinical characteristics of the participants (N = 44)
Table 1. Long description
A table summarizing demographic and clinical characteristics of 44 participants. The table has 11 rows and 2 columns. The first column lists the characteristics, and the second column lists the corresponding values. The characteristics include age, sex, handedness, education level, time from stroke, lesion volume, BDAE severity rating, semantic fluency, phonemic fluency, and BDAE verbal agility. The values are presented as means with standard deviations and ranges or as counts. For example, the age of participants is 64.4 years with a standard deviation of 10.3 and a range of 49. The table provides a detailed overview of the participants’ demographic and clinical information.
Data are mean (SD: range) unless otherwise reported. BDAE = Boston Diagnostic Aphasia Examination.
a Education level according to the Unesco International Standard Classification of Education: range 1 (primary education) to 6 (doctoral or equivalent level).
b Correct unique words, excluding rule breaks.
MRI data acquisition, preprocessing, and reconstruction
Patients underwent MRI scanning on a 3-Tesla Siemens Skyra at facilities in the Helsinki and Uusimaa Hospital District or the Southwest Finland Medical Imaging Centers. Each patient received high-resolution MPRAGE T1-weighted anatomical imaging (TR = 1800 ms; TE = 2.27 ms; TI = 900 ms; FOV = 250 × 250 mm; voxel size = 1 × 0.98 × 0.98 mm3) and multi-shell HARDI DW-MRI with 13 non-diffusion and 130 diffusion-weighted volumes (30 unique directions at b = 1,000 s/mm2 and 100 at b = 2,500 s/mm2; TR = 5,000 ms, TE = 104 ms, FOV = 240 × 240 mm; voxel size = 2.0 × 2.0 × 2.0 mm3).
Preprocessing of individual T1 images and lesion reconstruction are described in Supplementary Methods and lesion overlap map shown in Supplementary Figure 1.
DW-MRI data were denoised using the Marchenko-Pastur PCA method (https://cai2r.net/resources/denoising-using-marchenko-pastur-principal-component-analysis/) in the MRTrix3 (https://www.mrtrix.org/). Gibbs ringing artifact correction was applied using local subvoxel-shifts (Kellner et al., Reference Kellner, Dhital, Kiselev and Reisert2016). Data reconstruction into MNI space was performed with DSI Studio (http://dsi-studio.labsolver.org, version April 7, 2021) using q-space diffeomorphic reconstruction (QSDR), employing a diffusion sampling length ratio of 1.25 to construct spin distribution functions (SDFs; Yeh et al., Reference Yeh, Badre and Verstynen2016). QSDR reconstructs diffusion information from local diffusion orientation properties and applies a white matter mask, restricting QA estimation to white matter voxels and improving robustness to cerebrospinal fluid and non-white-matter tissue while reducing sensitivity to signal abnormalities associated with focal lesion.
Fiber coherence index was used for automatic quality control of b-table orientation (Schilling et al., Reference Schilling, Yeh, Nath, Hansen, Williams, Resnick, Anderson and Landman2019), and R2 values assessed the template-to-anisotropy map fit for each participant. SDF maps had a resolution of 2.0 × 2.0 × 2.0 mm3. Forceps major and minor were examined for normalization quality (Hula et al., Reference Hula, Panesar, Gravier, Yeh, Dresang, Dickey and Fernandez-Miranda2020), and resulting anisotropy maps visually inspected in MNI space to confirm adequate whole-brain alignment in each participant. Restricted diffusion imaging quantified restricted diffusion. QA was extracted from SDF peak orientations for connectometry analysis as the local connectome fingerprint (Yeh et al., Reference Yeh, Badre and Verstynen2016).
Statistical and connectometry analysis
Word counts in semantic and phonemic fluency were compared using the Wilcoxon signed-rank test. The behavioral correlations between verbal fluency measures and demographic and clinical factors were explored using two-tailed Spearman’s rank correlation coefficient.
DW-MRI correlational tractography employed connectometry (Yeh et al., Reference Yeh, Badre and Verstynen2016) analyses in DSI Studio, which statistically maps connectivity between adjacent voxels within a white matter pathway, defined by SDFs. Connectometry identifies significant associations in white matter segments, rather than calculating mean FA of the entire tract. Using standard space reconstruction and automatized tracking further minimizes manual tracking bias. Two multiple regression models determined the local connectomes associated with semantic and phonemic verbal fluency. The analysis was restricted to positive correlations, as negative correlations are of limited interpretability in this context. Four iterations of topology-informed pruning were applied to remove false fiber trajectories. A T statistic threshold of 3 was used, selecting only tracks longer than 30 voxels, with age and lesion volume controlled as nuisance variables. With a sample size of 44, a T-value threshold of 3 in DSI Studio corresponds to an assumed effect size of 0.44, which at an alpha level of 0.05 yields an estimated power of 0.86. This is consistent with prior white matter connectometry studies in aphasia that have reported power estimates of at least 0.81 as sufficient for detecting meaningful structure-behavior associations (Hula et al., Reference Hula, Panesar, Gravier, Yeh, Dresang, Dickey and Fernandez-Miranda2020). Accordingly, T-value of 3 was selected to ensure adequate sensitivity while maintaining consistency with previous work and supporting robust statistical inference. QA values were normalized to standardize data across the two scanning sites. A false discovery rate (FDR) threshold of 0.05 was employed, applying 2000 randomized permutations to estimate the FDR from the null distribution of track length. Significant tract segments were manually inspected and verified, consulting Human Connectome Project template (HCP 1,065) based tracking as needed to ensure accurate fiber bundle classification. For reporting, tracts were classified as association (intrahemispheric cortical), projection (cortico-subcortical), and commissural (interhemispheric) tracts (Catani & Thiebaut de Schotten, Reference Catani and Thiebaut de Schotten2012).
Data availability
Anonymized data can be requested from the corresponding author pending regulatory approval. No custom codes were used. Software details are outlined in preceding sections.
Results
Demographic and clinical correlates of verbal fluency
Patients produced significantly more words in the semantic task than in the phonemic task (W = 61.5, P < 0.001). Semantic and phonemic fluency showed strong mutual correlation (rs = 0.779, P = 0.001) and both were linked to lesion volume (semantic: rs = −0.590, P < 0.001; phonemic: rs = −0.642, P < 0.001) and BDAE severity rating (semantic: rs = 0.804, P < 0.001; phonemic: rs = 0.761, P = 0.001). Neither semantic nor phonemic task showed statistically significant correlations to age (rs = −0.097, P = 0.533; rs = −0.044, P = 0.777, respectively), education (rs = −0.161, P = 0.297; rs = −0.052, P = 0.736), or time from stroke (rs = 0.254, P = 0.096; rs = 0.146, P = 0.345).
Analysis of semantic fluency tracts
Semantic fluency was positively associated (FDR < 0.05) with greater QA in the left association tracts (arcuate fasciculus [AF], inferior fronto-occipital fasciculus [IFOF]) and left projection pathways (corticostriatal, corticopontine and corticobulbar tracts and thalamic radiation; Figure 1A).
White matter connectomes positively associated with (A) semantic fluency and (B) phonemic fluency, and (C) visual comparison of tracts associated with each verbal fluency type and their overlap. L = left; QA = quantitative anisotropy; R = right. 597 × 689mm (118 × 118 DPI).
Analysis of phonemic fluency tracts
Tracts positively associated with phonemic fluency (FDR < 0.05) included the left association tracts (AF, IFOF, superior longitudinal fasciculus III [SLF-III]) and bilateral projection pathways (left corticopontine and corticostriatal tracts, left thalamic radiation and right middle cerebellar peduncle) (Figure 1B).
Comparison of tracts
A visual illustration of shared and distinct tracts positively associated with semantic and phonemic fluency are shown in Figure 1C. To complement this descriptive comparison, we conducted cross-control analyses to partial out shared variance between the fluency measures, with the main tract associations largely retained while some changes emerged (see Supplementary Results).
Controlling for speech-motor difficulties via BDAE verbal agility yielded a similar overall pattern of results, except for loss or attenuation of some phonemic fluency associations (see Supplementary Results).
Discussion
Using whole-brain QA-based white matter connectometry analysis, the present study aimed to determine the hodological foundation of verbal fluency in PSA. We found that both semantic and phonemic fluency were associated with a white matter connectome comprising association tracts linked to the left-hemisphere language network of speech processing, but also domain-general projection pathways. However, the emphasis on specific white matter tracts and their distribution diverged between the tasks.
Prevailing views on speech processing suggest a dual organization around the perisylvian areas, especially in the left hemisphere (Hickok & Poeppel, Reference Hickok and Poeppel2007; Saur et al., Reference Saur, Kreher, Schnell, Kümmerer, Kellmeyer, Vry, Umarova, Musso, Glauche, Abel, Huber, Rijntjes, Hennig and Weiller2008). Although the cortical layout of this dual-stream network is relatively well established (Hickok & Poeppel, Reference Hickok and Poeppel2007), consensus on the exact white matter tracts remains elusive. However, the association tracts implicated in the current study are typically considered components of the dual-stream system, with AF and SLF-III contributing to the dorsal and the IFOF to the ventral stream of speech processing. Both fluency types were associated with, but not restricted to, segments of dorsal and ventral streams. This corroborates previous evidence for both verbal fluency (Egorova-Brumley et al., Reference Egorova-Brumley, Liang, Khlif and Brodtmann2022; Li et al., Reference Li, Zhang, Song, Huang, Ding, Fang, Xu and Han2017) and speech production more generally (e.g., Hula et al., Reference Hula, Panesar, Gravier, Yeh, Dresang, Dickey and Fernandez-Miranda2020) that both dorsal and ventral accounts, but also other pathways beyond the dual-stream model, are essential for successful verbal production in PSA. The overlap of cognitive processes in semantic and phonemic fluency have been previously debated (Schmidt et al., Reference Schmidt, Schumacher, Römer, Leonhart, Beume, Martin, Dressing, Weiller and Kaller2017; Thye et al., Reference Thye, Szaflarski and Mirman2021); although only pointing to a degree of neuroanatomical overlap and not necessarily shared processes, the observed common pattern could reflect a shared demand set by both verbal fluency tasks for the interplay of efficiently coordinated motor-articulatory control and access to lexical-semantic representations.
The current methodology allowed for the exploration of the fiber bundle segments that showed statistically significant correlation to fluency measures, further characterizing potential differences and similarities between the tasks. Overall, the white matter connectome for both fluency types included anterior frontal connections of associative pathways terminating in the inferior frontal regions, which have previously been associated with strongest lesion weights (Thye et al., Reference Thye, Szaflarski and Mirman2021), and are thought to be an integral part of the neural networks governing language and executive functions (Schumacher et al., Reference Schumacher, Halai and Lambon Ralph2022). Phonemic fluency appeared to show relatively stronger involvement of dorsal stream pathways, whereas semantic fluency showed a more apparent ventral emphasis. The present findings complement previous lesion-based reports, with methodological differences warranting caution in drawing direct equivalencies. The present pattern of results may nevertheless help contextualize discrepancies in the previous literature, with some studies reporting largely overlapping correlates (e.g., Thye et al., Reference Thye, Szaflarski and Mirman2021) and some more selective associations (e.g. Baldo et al., Reference Baldo, Schwartz, Wilkins and Dronkers2006; Biesbroek et al., Reference Biesbroek, van Zandvoort, Kappelle, Velthuis, Biessels and Postma2016; Biesbroek et al., Reference Biesbroek, Lim, Weaver, Arikan, Kang, Kim, Kuijf, Postma, Lee, Lee, Yu, Bae and Biessels2021; Chouiter et al., Reference Chouiter, Holmberg, Manuel, Colombo, Clarke, Annoni and Spierer2016). In these latter studies, phonemic fluency has been linked predominantly to inferior frontal and precentral regions, consistent with AF and SLF-III terminations, whereas semantic fluency has been more strongly associated with anterior and middle temporal lesions, consistent with the role of the IFOF in linking temporal semantic regions with frontal control areas. The deviating emphasis may indicate the stronger reliance of semantic fluency on word retrieval based on meaning, engaging lexical-semantic processing more deeply, compared to phonemic fluency, which is more reliant on retrieval based on sounds, entailing especially phonological processing.
The involvement of corticostriatal and thalamic radiation align with focal findings indicating that verbal fluency can be affected by basal ganglia and thalamic lesions (Biesbroek et al., Reference Biesbroek, Lim, Weaver, Arikan, Kang, Kim, Kuijf, Postma, Lee, Lee, Yu, Bae and Biessels2021; Chouiter et al., Reference Chouiter, Holmberg, Manuel, Colombo, Clarke, Annoni and Spierer2016). Corticostriatal systems have been suggested to play a domain-general role in language regulation, particularly in tasks requiring increased control through selection, sequencing, prediction, and planning (Copland et al., Reference Copland, Brownsett, Iyer and Angwin2021) that are well represented by verbal fluency tasks in, e.g., requirement of rule application, set-shifting and self-monitoring. While the subcortico-cortical pathways associated with both fluency types were anterior projections, they were not restricted to thalamic radiation (Li et al., Reference Li, Zhang, Song, Huang, Ding, Fang, Xu and Han2017) but also comprised corticopontine and corticostriatal tracts. Posterior projections of these tracts were also traced for semantic fluency, aligning with prior results of posterior thalamic contributions (Egorova-Brumley et al., Reference Egorova-Brumley, Liang, Khlif and Brodtmann2022; Thye et al., Reference Thye, Szaflarski and Mirman2021) and, in a broader picture, with propositions that semantic fluency engages a more extensively distributed subcortico-cortical connectivity (Biesbroek et al., Reference Biesbroek, Lim, Weaver, Arikan, Kang, Kim, Kuijf, Postma, Lee, Lee, Yu, Bae and Biessels2021; Thye et al., Reference Thye, Szaflarski and Mirman2021). Semantic fluency was additionally associated with a segment of the corticobulbar tract and phonemic fluency with middle cerebellar peduncle. For phonemic fluency, the cerebellar association no longer emerged after controlling for verbal agility, and projection pathway associations also more generally attenuated after controlling for verbal agility or semantic fluency.
Patients produced more words for the semantic task than the phonemic task. Whilst diverging engagement of articulatory-motor processing via AF may have impacted this difference in patients with primarily nonfluent PSA, disparity in this direction appears to be typical also in healthy populations (Vaughan et al., Reference Vaughan, Coen, Kenny and Lawlor2016). In fact, semantic and phonemic fluency have been suggested to have stronger association in PSA than in healthy populations (Schmidt et al., Reference Schmidt, Schumacher, Römer, Leonhart, Beume, Martin, Dressing, Weiller and Kaller2017), potentially reflecting that despite task demand differences, shared lexical, memory, and cognitive control processes may play a larger role in aphasia (Thye et al., Reference Thye, Szaflarski and Mirman2021). Coherently, and corroborating previous findings (Biesbroek et al., Reference Biesbroek, van Zandvoort, Kappelle, Velthuis, Biessels and Postma2016; Chouiter et al., Reference Chouiter, Holmberg, Manuel, Colombo, Clarke, Annoni and Spierer2016; Schmidt et al., Reference Schmidt, Schumacher, Römer, Leonhart, Beume, Martin, Dressing, Weiller and Kaller2017; Thye et al., Reference Thye, Szaflarski and Mirman2021), semantic and phonemic fluency showed strong mutual correlation in the present study.
There are important limitations to consider in this study. Each fluency type was assessed only with a single task, aligning with standard clinical practice yet potentially limiting reliability of task performance. Additionally, the sample of chronic PSA patients, all with left-hemisphere strokes, might overlook post-stroke neuroplasticity and right-hemisphere effects, though time from stroke did not significantly affect task performance. Moreover, prior studies including bilateral lesions detected only left-hemisphere contributions to verbal fluency (Egorova-Brumley et al., Reference Egorova-Brumley, Liang, Khlif and Brodtmann2022; Li et al., Reference Li, Zhang, Song, Huang, Ding, Fang, Xu and Han2017), whereas a study with only left-lateralized lesions found additional right frontotemporal associations (Schumacher et al., Reference Schumacher, Halai and Lambon Ralph2022). As fluency tasks engage a wide range of cognitive domains, this study does not allow precise identification of the neural correlates of each individual process involved, nor of their specific deficits. Aphasia severity was strongly linked to both fluency types, making its influence important to acknowledge. However, because severity was also highly correlated with lesion volume and the fluency measures themselves, adjusting for it would risk removing variance intrinsic to the constructs of interest. Although verbal executive tests have been shown to capture executive difficulties beyond language (Schumacher et al., Reference Schumacher, Halai and Lambon Ralph2022), it must be borne in mind that aphasia severity most likely also contributes to performance. Even with an assumed effect size of at least 0.44, the modest sample size may have limited sensitivity to detect smaller or task-specific effects, thereby increasing the potential risk of Type II error. Further research should confirm these findings in larger samples and determine whether the observed neural networks apply to other clinical and non-clinical populations.
In conclusion, this study contributes to understanding of white matter networks supporting verbal fluency in PSA, suggesting a partially segregated yet interconnected foundation within left hemisphere language regions and domain-general pathways. Mapping fluency-related tracts bridges neuroscientific information and clinical practice, providing insights into the neural organization of verbal executive functioning. The findings could be used in prospective longitudinal clinical studies to examine whether these structural brain patterns have value for prognostic modeling, guiding intervention stratification, and characterizing recovery trajectories in patients with stroke and aphasia.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S1355617726102240.
Acknowledgments
We express our gratitude to the subjects of the study for their contribution. We gratefully thank Sini-Tuuli Siponkoski, Linda Kuusela, Katja Paukkunen, Johanna Pekkola, and Viljami Sairanen for their contributions to the planning and conduct of this study.
Funding statement
This work was financially supported by the Academy of Finland (grant nos 299044, 306625, and 327996); European Research Council (grant no. 803466); Maire Taponen Foundation; and Signe and Ane Gyllenberg Foundation.
Competing interests
The authors report no competing interests.
Open access
