Neural correlates of verbal fluency revealed by longitudinal T1, T2 and FLAIR imaging in stroke

Highlights

• •Anatomical images can serve as good proxy sources to derive neural correlates of verbal fluency.
• •Damage risk to left white matter tracts negatively related to semantic fluency scores.
• •Subcortical grey matter volume positively related to phonemic fluency scores.
• •Subcortical grey matter volume is a cross-validated variable related to phonemic fluency.

Abstract

Diffusion-weighted imaging has been widely used in the research on post-stroke verbal fluency but acquiring diffusion data is not always clinically feasible. Achieving comparable reliability for detecting brain variables associated with verbal fluency impairments, based on more readily available anatomical, non-diffusion images (T1, T2 and FLAIR), enables clinical practitioners to have complementary neurophysiological information at hand to facilitate diagnosis and treatment of language impairment. Meanwhile, although the predominant focus in the stroke recovery literature has been on cortical contributions to verbal fluency, it remains unclear how subcortical regions and white matter disconnection are related to verbal fluency. Our study thus utilized anatomical scans of ischaemic stroke survivors (n = 121) to identify longitudinal relationships between subcortical volume, white matter tract disconnection, and verbal fluency performance at 3- and 12-months post-stroke. Subcortical grey matter volume was derived from FreeSurfer. We used an indirect probabilistic approach to quantify white matter disconnection in terms of disconnection severity, the proportion of lesioned voxel volume to the total volume of a tract, and disconnection probability, the probability of the overlap between the stroke lesion and a tract. These disconnection variables of each subject were identified based on the disconnectome map of the BCBToolkit. Using a linear mixed multiple regression method with 5-fold cross-validations, we correlated the semantic and phonemic fluency scores with longitudinal measurements of subcortical grey matter volume and 22 bilateral white matter tracts, while controlling for demographic variables (age, sex, handedness and education), total brain volume, lesion volume, and cortical thickness. The results showed that the right subcortical grey matter volume was positively correlated with phonemic fluency averaged over 3 months and 12 months. The finding generalized well on the test data. The disconnection probability of left superior longitudinal fasciculus II and left posterior arcuate fasciculus was negatively associated with semantic fluency only on the training data, but the result aligned with our previous study using diffusion scans in the same clinical population. In sum, our results presented evidence that routinely acquired anatomical scans can serve as a reliable source for deriving neural variables of post-stroke verbal fluency performance. The use of this method might provide an ecologically valid and more readily implementable analysis tool.

1. Introduction

Acute ischaemic stroke is associated with high mortality and significant long-term morbidity. Stroke survivors have been reported to have cognitive impairments in multiple domains, including cognitive control, visuospatial function, and memory (Ramsey et al., 2017). Around 20–25 % of stroke patients may experience acute language deficits, of which a further 33 % sustain long-term language impairments (Simmons-Mackie & Cherney, 2018). Stroke patients may develop various language abnormalities as a function of lesions in cortical, subcortical and white matter (WM) regions (Hillis et al., 2018; Kuljic‐Obradovic, 2003).

In clinical settings, the language and related cognitive abilities of stroke patients are often evaluated in terms of oral verbal fluency (Abrahams et al., 2000, Baldo et al., 2006). To measure verbal fluency, a short test typically consisting of category fluency and letter fluency tasks is administered (Lezak et al., 2012, Shao et al., 2014). In the standard versions of the tasks, category fluency (also called semantic fluency) is assessed by prompting the participants to produce as many unique words in a semantic category as possible within one minute (Benton, 1968). Letter fluency (also called lexical or phonemic fluency) is assessed by the number of unique words starting with a given letter (e.g., “the letter A”) produced in one minute (Newcombe, 1969). The two tasks are believed to draw on different cognitive processes, with semantic fluency on the semantic system, consisting of taxonomic and thematic relationships, and phonemic fluency on both the lexical and phonological systems mediated by education and literacy, as well as processes of higher cognitive control (Ardila et al., 2010, Ratclif et al., 1998).

The relationship between verbal fluency and cortical functions has been investigated in many neuroimaging studies. Semantic fluency is reported to be closely related to the left temporal cortex and right inferior frontal regions, while phonemic fluency is associated with the left prefrontal cortex (Li et al., 2017). However, several large-scale studies have shown that greater than 85 % of stroke lesions are subcortical. These subcortical lesions contribute to neuropsychological dysfunctions across multiple cognitive domains (Corbetta et al., 2015), including executive deficits (Jokinen et al., 2006) and impaired verbal fluency (Jacova et al., 2012; Kuljic-Obradovic, 2003) in both healthy bilingual (Burgaleta et al., 2016, Grogan et al., 2009) and clinical populations with a variety of cognitive disorders (Parkinson’s disease: Ellfolk et al., 2014, Foley et al., 2021; social adversity: Thames et al., 2018; apraxia of speech: Peach & Tonkovich, 2004). Compared to the more in-depth research on cortically based language impairment, our understanding of how subcortical regions contribute to different aspects of verbal fluency of stroke patients remains rather limited, although general theories of subcortical language function that cortico-subcortical circuits support visual/auditory language processing, monitoring and speech production have been proposed early in the 1980 s (Crosson, 1985, Wallesch and Papagno, 1988). Researchers have reported inconsistent language impairment syndromes even if patients had lesions in the same subcortical structure (Nadeau and Crosson, 1997, Kreisler et al., 2000). Structural disruption of different subcortical regions is also reported to induce deficits in both executive and responsive language functions (Copland, 2003, Crosson, 2021, Mega and Alexander, 1994, Radanovic and Mansur, 2017). For example, lesions in structures like the thalamus, basal ganglia, putamen and caudate can lead to reduced phonemic fluency (Andrade et al., 2012) and compromised ability to switch between semantic and phonemic fluency tasks (Tröster et al., 1998). The lack of convergent empirical evidence on the role of subcortical regions implies the necessity of adding structural measurements, such as subcortical volume, as a complementary dimension to probe the involvement of these regions in post-stroke verbal fluency.

In addition, more investigations are needed to explore how the longitudinal change of subcortical volume influences verbal fluency performances as stroke patients transition into different stroke stages. Although a stroke frequently occurs as a focal event, damage to subcortical, cortical, and WM can appear greater over time due to processes like Wallerian degeneration (i.e., retrograde degeneration of the distal end of an axon). Recent evaluations have shown that stroke lesions likely become larger with time as a result of the slow degeneration of axon bodies, dendrites, and axons (Seghier et al., 2014), which, in turn, is likely to further impact verbal fluency performances of the individuals with stroke.

The association between post-stroke verbal fluency performance and WM tract disconnections has been more intensively investigated in recent years due to the advancement of the diffusion-weighted imaging (DWI) technique (Kümmerer et al., 2013, Li et al., 2017). However, multidirectional, high beta-value DWIs are not routinely acquired in clinical scenarios, which may constrain clinical practitioners from accurately understanding the neurophysiological bases of stroke patients’ language and cognitive dysfunction. Identification of WM tract variables derived from more readily available non-diffusion-weighted anatomical images, such as T1, T2 and fluid attenuated inversion recovery (FLAIR) sequences, would allow greater application clinically for correlating post-stroke verbal fluency performance with the underlying neural substrates. The emerging field of WM disconnectome mapping offers a promising technique to characterize WM disconnections without collecting tractography data, by basing tract disconnection caused by stroke lesions on a tract disconnectome probability map constructed from a group of normative healthy controls (Foulon et al., 2018, Thiebaut de Schotten et al., 2011, Thiebaut De Schotten et al., 2014, Thiebaut de Schotten et al., 2020). The method has demonstrated high reliability in matching populations with different types of brain damage. Compared to the state-of-the-art approach of using direct DWI that tends to capture the spared connectome, the disconnectome method estimates how probable a WM tract is disconnected by a lesion placed in healthy connectomes.

In the current study, we sought to address two outstanding questions using anatomical T1, T2 and FLAIR scans. First, we specifically examined whether subcortical grey matter (GM) volume was associated with verbal fluency impairment when cortical and lesion factors, such as cortical thickness, total brain volume and lesion volume, were controlled. We evaluated the extent to which damage to subcortical GM volume impaired verbal fluency in a large sample of stroke survivors participating in the Cognitive And Neocortical Volume After Stroke (CANVAS) project (Brodtmann et al., 2014). We expected that verbal fluency performance would be associated with both the right and left subcortical GM volume due to bilateral cortico-subcortical connections. Additionally, we explored how the change in subcortical GM volume from the early chronic to late recovery periods would longitudinally impact verbal fluency.

Second, we investigated the relationships between WM tract disconnection derived from the disconnectome mapping method and verbal fluency, and compared them with those identified in our recent study (Egorova-Brumley et al., 2022) using a direct DWI approach. A rich repertoire of structural connectivity literature reported nine WM tracts frequently involved in language processing, including bilateral arcuate fasciculus (long, anterior, posterior) (AF), inferior fronto-occipital fasciculus (IFOF), inferior longitudinal fasciculus (ILF), superior longitudinal fasciculus (SLF I, II, III) and uncinate fasciculus (UF) (see reviews of Catani and Mesulam, 2008, Egorova-Brumley et al., 2022, Jang, 2013, Kümmerer et al., 2013, Parker et al., 2005, Saura et al., 2008, Smits et al., 2014, Yagmurlu et al., 2016). However, the exact functional contributions of these anatomical pathways to verbal fluency remain elusive. The ventral tracts (i.e. left IFOF, ILF and UF) have been reported as potential candidates to support lexical-semantic processing, and more dorsal streams including the left AF and SLF have been identified to be related to semantic fluency (Acosta-Cabronero et al., 2011, Agosta et al., 2010, Almairac et al., 2015, Costentin et al., 2019, Han et al., 2013, Hope et al., 2018, Li et al., 2017, Schwindt et al., 2013). Verbal dysfluency was also reported to be more associated with the disruptions to the left side of the above WM tracts, regardless of the lesion sites, than the tracts in the right hemisphere (Biesbroek et al., 2021). In addition, recent research on the frontal aslant tract (FAT) and frontal striatal tract (FST) revealed that the bilateral FAT, which connects the lateral inferior frontal gyrus with the (pre-)supplementary motor areas (SMA) of the medial frontal gyrus, was highly related to spontaneous speech production and executive control for action (Zyryanov et al., 2020; also see the reviews of Dick et al., 2019, La Corte et al., 2021). FST connecting the SMA to the caudate nucleus was found to be an important tract in the motor control network supporting speech movement, semantic and phonemic verbal fluency (Kinoshita et al., 2015, Costentin et al., 2019). Using the indirect probabilistic disconnectome mapping approach, we expected to find that disconnection to these 11 tracts, especially in the left hemisphere, would also be associated with verbal dysfluency in our study. The increased disconnection of the left WM tracts could be correlated with worse verbal fluency performances of stroke survivors transitioning from the early chronic to the chronic stage. The results were compared with our previous DWI study on the same clinical population (Egorova-Brumley et al., 2022) to assess the comparable reliability of detecting WM tract variables based on anatomical scans.

2. Materials and methods

2.1. Participants and dataset information

Data from 128 individuals with ischemic stroke 3 months post-stroke were extracted from the CANVAS project, which is a multi-center longitudinal cohort study of ischaemic stroke survivors and healthy controls in Melbourne, Australia. Images and behavioral data of 113 subjects were available at 12 months post-stroke due to experimental attrition. The details of recruiting and screening procedures, study design, and methodology were discussed in Brodtmann et al. (2014). The study was approved by the ethics committees of three Stroke Units (Austin Hospital, Box Hill Hospital, and the Royal Melbourne Hospital) in line with the Declaration of Helsinki. Only participants without psychiatric history were included. Informed written consent was obtained for each participant before the study.

Our data included the behavioral assessment of oral verbal fluency (Category Fluency Animals Test and Controlled Oral Word Association Tests with letters F, A, and S) (Loonstra et al., 2001) and anatomical MRI scans (isotropic T1, T2, and FLAIR scans) at 3 months and 12 months post-stroke. The cohort at the first time point included 16 participants with recurrent stroke and 105 with a first-ever clinical event. The detailed descriptive statistics of the dataset are listed in Table 1.

Table 1.

Descriptive statistics of the dataset.

Type of dataset	Longitudinal (3 months/12 months)
Demographics
Age, mean $\pm$ SD (range)	3 monthsa: 68.3 $\pm$ 11.9 (30–89)
	12 monthsb: 68.9 $\pm$ 11.8 (30–89)
Male, n (%)	82 (67.8)
Education in years, median (IQR)	12 (5)
Hand preference
Right, n (%)	112 (92.6)
Stroke severity
NIHSS, median (IQR)	3 monthsc: 0 (1)
	12 monthsd: 0 (1)
Sample size, n	3 months: 121
	12 months: 113
Brain MRI
Prior stroke lesions, n (%)	16 (13.2)
Lesion volume, mean$\pm$ SD (range) (cm3)	3 months: 18.19 $\pm$ 42.45 (0.13 – 405.84)
	12 months: 23.17 $\pm$ 53.94 (0.51 – 499.55)
Verbal fluency (Z-score)
Semantic fluency (3 months), mean $\pm$ SD (range)	0.08e$\pm$ 1.27 (-2.65 – 3.40) (7)
Semantic fluency (12 months), mean $\pm$ SD (range)	0.16f$\pm$ 1.22 (-2.19 – 4.56) (6)
Phonemic fluency (3 months), mean $\pm$ SD (range)	−0.44 g$\pm$ 1.08 (-2.98 – 2.66) (6)
Phonemic fluency (12 months), mean $\pm$ SD (range)	−0.30 h$\pm$ 0.98 (-2.56 – 2.57) (5)

IQR = interquartile range.

^a: Data missing in 1 case.

^b: Data missing in 12 cases.

^c: Data missing in 10 cases.

^d: Data missing in 13 cases.

^e: Data missing in 7 cases.

^f: Data missing in 6 cases.

^g: Data missing in 6 cases.

^h: Data missing in 5 cases.

2.1.1. Imaging data acquisition

A Siemens 3 T Tim Trio scanner (Erlangen, Germany) with a 12-channel head coil was used to collect MR images at the three Stroke Units. T1-weighted MPRAGE images were acquired with the following parameters: number of slices = 160 sagittal slices, TR = 1900 ms, TE = 2.55 ms, flip angle = 9°, slice thickness = 1 mm, matrix size = 256 × 256, and voxel size = 1 × 1 × 1 mm³. T2 images were acquired with the following parameters: number of slices = 176 sagittal slices, repetition time (TR) = 3390 ms, echo time (TE) = 390 ms, flip angle = 120°, matrix size = 256 × 204, and voxel size = 1 × 1 × 1 mm³. FLAIR images were collected with the following parameters: number of slices = sagittal 160, TR = 6000 ms, TE = 388 ms, flip angle = 120°, matrix size = 256 × 254, voxel size 0.5 × 0.5 × 1 mm³. Stroke infarcts were manually traced on FLAIR images (M.S.K) and verified by a stroke neurologist (A.B.).

2.2. Data preprocessing

2.2.1. Data exclusion and imputation

The raw dataset included 128 individuals with stroke. Data exclusion was performed on verbal fluency and MRI data based on three criteria: 1) participants with missing verbal fluency assessment scores at both time points (n = 4); 2) participants with missing MRI scans at both time points (n = 1); 3) participants with poor tissue segmentation results (i.e., a high number of pial surface errors, topological defects and WM segmentation errors) and failed normalization (n = 2), which were decided based on the consensus of two raters (Y. X. and B.C.S.) (see supplementary 1B for an example). The total number of excluded participants was seven, approximately 5.5% of the raw dataset. The dataset used in further analyses contained data from 121 individuals.

There was one missing case of the age attribute at the 3-month time point and 12 missing cases at the 12-month time point. As the data were collected at a fixed interval of 9 months between the two time points, the missing age data of the participants were imputed as the same age number of the other time point. For the missing cases of verbal fluency scores and other MRI variables, no explicit imputation and data exclusion was performed as the linear mixed-effect model of the nlme R package (used in the modeling procedure) could handle the missing values.

2.2.2. Preprocessing MRI data

The T1 image of each participant was first transformed into an “unlesioned” T1 image using the cerebral spinal fluid (CSF) probability maps from the Clinic Toolbox (https://www.nitrc.org/projects/clinicaltbx), which specifically aims at normalizing CT scans or MRI scans for elderly people or people with brain juries (Rorden et al., 2012). We first resized and binarized the lesion image of each subject (512 × 512 × 160) with reference to its correspondent T1 image (256 × 256 × 160) using the FSL flirt function with the degree of freedom 12 (version 6.0.1). To minimize the effect of the brain lesions and achieve more accurate spatial normalization, we created enantiomorphic T1 images (Nachev et al., 2008) by first reorienting the lesion and T1 images on the MNI coordinates and deriving the transformation matrix, which was then applied to the lesion and T1 images respectively. The lesion image was flipped in the left–right direction and used to mask the T1 image in order to carve out a healthy enantiomorphic filling. The healthy tissue was flipped back and applied with the inverse of the transformation matrix to change back to the individual native space. Meanwhile, the flipped healthy tissue image was binarized with 0 s on the lesion site and 1 s in other regions and multiplied with the original T1 image to create 0 values on the lesion site of the T1 image. Lastly, the unbinarized healthy tissue image in the native space was added to the T1 image to obtain the unlesioned T1 image.

Skull-stripping was performed based on the GM, WM and CSF probability maps from the Clinic Toolbox (Rorden et al., 2012). These maps were derived from 50 healthy individuals with a mean age of 73 and with both larger ventricles and a certain degree of atrophy. These features matched the anatomical characteristics of the participants in our dataset. To improve the downstream results of brain extraction and image registration, we resampled the probability masks (2 × 2 × 2 mm, 91 × 109 × 91) to the voxel size and dimension of the MNI152_wskull template (1 × 1 × 1 mm, 182 × 218 × 182) using the ResampleImage function of ANTS (version 2.3.1) with B-spline interpolation. The three maps were then added together using the FSL fslmaths function to create the whole brain-cerebellum probability mask. The sct1_unsmooth.nii template of the Clinic Toolbox and the probability mask were used as the input arguments of the ANTS antsBrainExtraction function to extract the brain.

Spatial registration was performed by taking the skull-stripped “unlesioned” T1 image of each subject at each time point as a fixed image and aligning it with the MNI152_wskull.nii.gz template of BCBToolKit (https://toolkit.bcblab.com/) using the ANTS antsRegistration function (Foulon et al., 2018). The registration included three steps: rigid, affine and non-linear (Syn) transformation. At each step, registration was performed at four levels of resolution with decrease order of iterations for each level (1000 × 500 × 250 × 100). The specified quantiles of the minimum and maximum intensities of the image histograms were between 0.01 and 0.99. The intensities of the two registered images were matched based on their histograms of the affined and non-linear registration. We applied the default gradient steps of the rigid and affine transformation (0.1) and the diffeomorphic deformations warping (0.05) for the initial registration. The similarities between the two images were measured in terms of 32-bin mutual information (MI) for the rigid and affine transformation, and local cross-correlation (CC) for the non-linear registration. The latter was computed as the correlations between every voxel’s neighborhood with a radius of 4. The optimization procedure of the multi-resolution algorithms stopped if the improvement of MI and CC in the last 10 iterations was no more than 1e-6 (rigid), 1e-8 (affine) and 1e-9 (non-linear) respectively. The inverse of the three transformation matrices was then applied to the lesioned T1 image and lesion mask to normalize them in the standard MNI space.

Each of the preprocessing steps was followed by a manual quality check. There were six skull-stripping failures of the T1 images at the 3-month post-stroke time point and 13 at the 12-month post-stroke time point. The failures were corrected using the alternative NKI template of the BCBToolKit. We separated the three registration steps of problematic coregistered images to examine which transformation led to the normalization failures. We found that six 3-month images and one 12-month image failed at the affine transformation step. Eleven 3-month images and five 12-month images failed at the non-linear step. Better registration results were achieved for all these images by adjusting the gradient descent optimization steps of either affine transformation (0.05 or 0.01) and/or diffeomorphic deformation warping (0.01 or 0.005). Currently, there is no empirical research to systematically compare the choice of the different gradient steps on the comparability of the downstream results. However, Avants et al. (2008) pointed out that the default parameter range may need to be adjusted based on the nature of the segmentation problem for the optimal performance of the regularization algorithm. We observed that smaller gradient steps helped achieve more accurate results for the stroke population of our study although the convergence process took longer. Meanwhile, our manual inspections also found that these problematic images often showed abnormal curvatures in gyri/sulci and deviated image intensities due to brain atrophy and tissue death. Smaller search steps were found to solve these issues better (see supplementary Fig. 2A and 2B for comparing the coregistration results of two subjects using the default and adjusted gradient descent optimization parameters).

Fig. 2.

Phonemic fluency in relation to time and the right subcortical GM volume. A) The phonemic fluency Z scores at 3 months and 12 months after stroke (blue dots). Individual trajectories of the score changes were represented by the connected green lines between the two time points (data with missing values were not connected). The red line was the averaged trajectory and the grey ribbon represented the standard error. B) The positive linear relationship between the right subcortical GM volume and the phonemic fluency scores. The light red and green lines were the fitted linear regression lines with the correspondent color ribbons representing the 95 % confidence interval. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

2.3. Calculating white matter disconnection probability and severity

WM tract disconnections were quantified using two metrics: disconnection severity, which was calculated as the proportion of the lesioned voxel volume to the total volume of a tract, and disconnection probability, which was calculated by multiplying the normalized lesion mask of each patient by each tract and taking the maximum probability of the overlapping cluster of voxels as the disconnection probability of that tract passing the lesion (Thiebaut De Schotten et al., 2014). To calculate the WM disconnections, we used the BCBToolKit (Thiebaut de Schotten et al., 2011) which contains 68 WM tract atlases in the MNI space derived from the high-resolution tractography of a normative group of 47 healthy subjects (Rojkova et al., 2016). These atlases provided a probability of the tract assignment in each voxel. Different thresholds (30%, 50%, 70% and 90%) were adopted to examine the certainty of the tracts overlapping with the subject’s lesion. If the probability was estimated to be above a certain threshold, the tract was considered disconnected for that subject. As the study focused on those WM tracts reported to be most frequently involved in language processing based on the prior literature discussed in the Introduction, only 11 key tracts of interest including the bilateral AF (long, anterior, posterior), IFOF, ILF, SLF (I, II, III), UF, FAT and FST were used in the statistical analyses.

2.4. Calculating volumetric measurements and cortical thickness

As previous research indicates that the global volumetric (lesion volume and subcortical volume) and cortical thickness measurements are associated with post-stroke cognitive symptoms, we aimed to control for them in our analyses (Schellhorn et al., 2021, Vonk et al., 2019). We obtained the information for these variables using the longitudinal pipeline of the automated FreeSurfer algorithm (version 7.1.1, https://surfer.nmr.mgh.harvard.edu/). The recon-all function segments the T1-weighted image of each participant at each time point using the corresponding T2 image to improve the pial surface reconstruction. While FLAIR images show more sensitivity to differentiate CSF and lesions, the comparison of segmentation results of the two types of MRI scans across the participants showed that T2 images helped achieve better segmentation of the pial surface. The -bigventricles flag was used to correct the enlarged ventricles present in most of the participants. The recon-all function outputs the surface thickness and volume information of 44 cortical and subcortical structures for the whole brain based on the default DKT40 atlas (Desikan et al., 2006). The seven subcortical structures included the thalamus, caudate, putamen, pallidum, hippocampus, amygdala and accumbens area in each hemisphere. The average structural volume of these regions was calculated for each hemisphere.

Manual intervention was performed to resolve topological surface errors in an iterative manner. We implemented four types of corrections for less ideal segmentation results: 1) if the skull stripping was suboptimal with skull and/or dura assigned as WM in more than 20 slices in the sagittal, coronal or axial orientation, the skull-stripping subprocess was run one more time with a lowered watershed threshold between 25 and 1; 2) manually edited the brain mask to remove the leftover skull or dura from the previous correction and then rerun recon-all -autorecon-pial step; 3) checked whether the lesions were assigned correct tissue types and manually edit the intensity values of the lesion voxels on the WM mask if the lesion site was on the WM (this was the most common type of error compared to the opposite type); 4) added other WM mask points to improve the segmentation effect of the pial surface by making the curvature map more closely to the T1 image and then rerun recon-all -autorecon2-wm and -autorecon3 subprocesses. Manual corrections were performed on five 3-month images (4.1%) and the eleven 12-month images (9.2%). Lastly, we used the asegstats2table and aparcstats2table functions of FreeSurfer to convert the volumetric and cortical thickness stats files of all subjects into tables. Each line of the table represents a subject with columns listing the volumetric (mm³) or thickness (mm²) data of the 44 cortical and subcortical brain structures.

For better visualization of the locus of lesion sites across the participants, we presented 3-month and 12-month lesion prevalence heatmaps. Only the 3-month map is illustrated in Fig. 1 due to its high resemblance to the 12-month map. The highest overlap was found in the right subcortical regions.

Fig. 1.

Lesion prevalence maps. A rendered lesion heat map to the MNI 152 template in the radiological orientation across 121-stroke patients. The maximum value (red) indicates the highest lesion overlap among the participants (N = 18). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

2.5. Statistical modeling

The linear mixed-effects regression model (LMM) in R (version 4.1.2) was used to characterize the relationship between the dependent variables (phonemic or semantic fluency scores) and independent variables (left/right subcortical GM volume and WM disconnection measurements). The models characterized the effects of subcortical GM volume and/or WM tract disconnections on the phonemic (or semantic) fluency scores of the stroke patients over time (3 months post-stroke to 12 months post-stroke). The demographic features (age, sex, handedness, and education), total brain volume, lesion volume, and cortical thickness were used as covariates. Following the method used by Pustina et al. (2017), we performed the model selection using the whole sample and cross-validated the generalizability of the models by dividing 80% of the total sample into five folds. The hold-out 20 % of the data served as the test set. The model selection procedure included model comparisons and evaluation of model performance. All models were constructed with a random intercept to better account for the individual differences between the stroke participants.

We used the 11 WM tracts of interest and their counterparts in the right hemisphere to model the predictive effects of tract disconnections on verbal fluency. The theory-based approach helped circumvent typical problems inherent in the supervised and unsupervised feature selection approaches, such as overfitting and increased difficulties in interpretability (Hope et al., 2018). These tracts included the bilateral AF (long, anterior, posterior), IFOF, ILF, SLF (I, II, III), UF, FAT and FST.

Each full model was constructed including one time predictor, two subcortical GM volume predictors (left/right), 22 WM tracts, and the seven covariates described earlier (e.g., age). The “full” model evaluated the extent to which all predictors all-together influenced verbal fluency. The full model is presented mathematically in the following equation. $f_{\mathit{ij}}$ is the phonemic or semantic fluency score of the ith subject at the jth time point. $\gamma$ is the regression coefficient of the corresponding predictor.

$$f_{\mathit{ij}}={\gamma }_{00}+{\gamma }_{10}{\mathit{Time}}_{\mathit{ij}}+{\gamma }_{01}{\mathit{Age}}_i+{\gamma }_{02}{\mathit{Gender}}_i+{\gamma }_{03}{\mathit{Handedness}}_i+{\gamma }_{04}{\mathit{Education}}_i+{\gamma }_{20}{\mathit{BrainVolume}}_{\mathit{ij}}+{\gamma }_{30}{\mathit{LesionVolume}}_{\mathit{ij}}+{\gamma }_{40}{\mathit{CorticalThickness}}_{\mathit{ij}}+{\gamma }_{50}{SubCoritcalVolume\_Left}_{\mathit{ij}}+{\gamma }_{60}{SubCoritcalVolume\_Right}_{\mathit{ij}}+{\gamma }_{70}{Tract\_1\_Left}_{\mathit{ij}}+{\gamma }_{80}{Tract\_1\_Right}_{\mathit{ij}}\cdots +{\gamma }_{270}{Tract\_11\_Left}_{\mathit{ij}}+{\gamma }_{280}{Tract\_11\_Right}_{\mathit{ij}}$$

Three reduced models — the base model, the left WM model, and the right WM model, were derived from the above full model. The base model only included time, the left and right subcortical GM volume and the seven covariates. It tested the effects of the bilateral subcortical volume without accounting for the effect of WM tract disconnections. The left and right WM models aimed to identify the distinct contributions of the anatomical pathways in the left and right hemispheres to verbal fluency. They were set up by adding the predictors of interest for either the left or right WM tracts to the base model. Model comparisons were performed between these four models to examine whether including information about tract disconnections in one or both hemispheres would significantly enhance the goodness of fit of the models (i.e., whether the left and/or right WM models outperformed the base model). Model performances were evaluated based on both loglikelihood ratio tests (LRT) and AIC (Akaike information criterion) values in order to achieve the optimal trade-off between model complexity and goodness-of-fit (Baayen et al., 2008). The statistical significance of LRT was set at p < 0.05.

To further examine the effects of the individual WM tracts of interest, the progressive model reduction was performed using LRT. For the significant predictors of the best model, the subject-mean centering method was used to separate the effect of the individual change across time (within-subject effect) from the effect of the differences across participants (between-subject effect) (Rabe-Hesketh & Skrondal, 2008).

To evaluate the generalizability of the models, we performed 100 iterations of 5-fold cross-validations using the nlme (version 3.1–153), groupdata2  (version 2.0.0), lmerTest (version 3.1.3), and hydroGOF (version 0.4.0) packages. The total dataset was divided into a training set (80% of subjects) and a test set (20% of subjects). For each iteration of the 5-fold CV on the 80 % training set, the model parameters estimated from the training fold were applied to the four test folds to calculate the root mean square error (RMSE) of each model. The values of RMSE across the iterations were averaged on the training set. Non-parametric Kruskal-Wallis rank sum tests were conducted to examine the significant differences in the RMSE between the base model, the left WM model, the right WM model and the full models. The estimated models were then applied to the 20% test set to assess the overall performance of each model. Note that the iterations could only be performed at the cross-validation level, not at the partition level to circumvent the issue of data leakage between the training and test data.

To guard against the violations of normality and assure homoscedasticity assumptions of model residuals, we scaled and centered the seven continuous predictors (age, education, total brain volume, lesion volume, left/right subcortical GM volume, mean cortical thickness) to the total sample means, as their raw data are at 10¹- to 10⁵-fold magnitude of the 22 WM tract predictors (bounded between 0 and 1.0). Both assumptions were met for the best models after the data transformation. The covariates were checked for collinearity. The technical details of model comparisons and evaluations are discussed in the Supplementary Materials.

3. Results

We applied 30%, 50%, 70% and 90% thresholds of disconnection probability. Phonemic fluency did not show sensitivity to different thresholds and the results were similar across the levels. Thresholding had some effects on the relationships between WM disconnection and semantic fluency. The 30% and 50% thresholds identified the left model as the best fit model. The higher thresholds (70% and 90%) tend to eliminate the effects of WM tract disconnections and found the base model as the best fit. One possible reason for the discrepancies at different thresholds is that higher cut-offs (e.g., 70%, 90%) lead to more zero values in the models, therefore, less WM tract information was used to calculate the correlations between the WM tract predictors and semantic fluency. Additionally, more stringent thresholds imply that fewer individual differences in WM tract anatomy are accounted for. To capture a better trade-off between the accuracy of model estimation and flexibility, we presented the results at the 50% threshold in the following sections. The results of other thresholds were described in the supplementary materials.

3.1. Phonemic fluency

The model comparisons showed that the random-intercept base model was the best fit model of disconnection probability for phonemic fluency with the lowest AIC (AIC = 533.3). Therefore, no WM tract was significantly related to the change in phonemic fluency scores. The model suggested that salient inter-individual variabilities in phonemic fluency already existed among the stroke patients at the early chronic stage. As Fig. 2A shows, participants did not show much difference in their recovery trajectories between 3 months and 12 months after the stroke (t = 1.94, p > 0.05). The phonemic fluency scores at 3 months dispersed widely within 3 standard deviations of the z-scores but showed a generally increasing trend across time. We calculated both marginal and conditional pseudo-R-squared of the base model using the R MuMIn package (Nakagawa et al., 2013). The variance explained by the fixed effects was 10 % and 82 % of the variance was explained by the entire model, including both fixed and random effects, indicating that the linear mixed effect model was effective in capturing a great portion of inter-subject variabilities in our data sample.

To further explore the different effects of the left and right subcortical volume on phonemic fluency, we compared the two reduced versions of the base model. One only included the left subcortical GM volume and the other only included the right subcortical GM volume. The results showed that the base model with the right subcortical GM volume was the best fit (p < 0.05). There was a significant intercept effect (estimate = -0.88, p < 0.05) and a right subcortical GM volume effect (estimate = 0.36, p < 0.05) after multiple comparison corrections based on the False Discovery Rate method (Table 2), suggesting that an increase in the right subcortical GM volume was related to increased phonemic fluency scores. Fig. 2B illustrates that the right subcortical GM volume had a significant positive correlation with phonemic fluency at 12 months post-stroke (r = 0.18, p < 0.05), but not at 3 months after stroke (r = 0.12, p < 0.05).

Table 2.

The linear mixed effects of the predictors for phonemic fluency in the best fit base model.

Predictor	Estimate	t	p.adjusted
Intercept	−0.88, 95 % CI [-1.54, −0.22]	−2.58	0.01*
Time	0.13 95 % CI [0.001, 0.25]	1.95	0.05
Age	0.13, 95 % CI [-0.06, 0.33]	1.30	0.20
Sex	0.06, 95 % CI [-0.36, 0.48]	0.27	0.79
Handedness	0.44, 95 % CI [-0.23, 1.14]	1.28	0.21
Education	−0.13, 95 % CI [-0.31, 0.04]	−1.47	0.14
Total brain volume	−0.08, 95 % CI [-0.35, 0.20]	−0.54	0.59
Lesion volume	0.04, 95 % CI [-0.16, 0.23]	0.35	0.72
Mean cortical thickness	−0.07, 95 % CI [-0.26, 0.12]	−0.69	0.49
Subcortical GM volume (right)	0.36, 95 % CI [0.08, 0.65]	2.51	0.01*

*<0.05 (FDR corrected).

The follow-up subject-mean centering analysis showed that neither the between-effect (t = 1.66, p > 0.05) nor the within-effect (t = 1.97, p > 0.05) of the right subcortical GM volume was significant. This finding implies that its overall significant effect on phonemic fluency was driven by the combined effects of the chronic volume change within each subject and the volume variabilities between the participants.

The disconnection severity analysis also found the base model was the best fit model. The detailed results were the same as disconnection probability and not repeated here.

We performed 5-fold cross-validations on the 80% of the dataset to further examine whether the base model, identified as the best fit model in the model selection procedure, yielded the lowest RMSE on the test folds compared with the other three models and if it led to a decreased RMSE on the 20% test set. We first performed the Levene’s test to check the homogeneity of variance assumption of the RMSE values of the four models (the base, left WM, right WM and full models) generated by the 100 iterations using the R car package. The variances between the models were found to be non-homogeneous for disconnection probability and severity (p < 0.001). Non-parametric Kruskal-Wallis rank sum tests were therefore used as an alternative to the parametric one-way ANOVA. Results showed that there was a significant difference in the RMSE of the models (disconnection probability: ${\chi }^2$ = 297.23, df = 3, p < 0.001; disconnection severity: ${\chi }^2$ = 297.57, df = 3, p < 0.001). The post-hoc pairwise Wilcoxon tests with Bonferroni corrections for multiple testing found that the base model RMSE was significantly lower than the other three models (base vs left WM: p < 0.001; base vs right WM: p < 0.001; base vs full: p < 0.001), indicating that the base model generalized well on the test folds for both disconnection metrics. The application of the model parameters on the 20 % hold-out test set found the biggest decrease of RMSE (1.01) of the base model for disconnection probability and severity (Table 3). The right subcortical volume maintained a significant positive correlation with phonemic fluency on the 20% test set (t = 2.16, p < 0.05) (supplementary Table 7). The marginal pseudo-R-squared of the based model was 0.09 and the conditional pseudo-R-squared was 0.92, suggesting that the entire model explained most of the variance in the 20% test data. The above cross-validated results supported the predictive role of the right subcortical GM volume on phonemic fluency.

Table 3.

Phonemic fluency - RMSE of the four models on the training/test set.

	80 % Training Set				20 % Test Set
Phonemic Fluency	Base Model	Left WM Model	Right WM Model	Full Model	Best Model
Probability	1.36	1.38	1.37	1.41	1.01 (Base model)
Severity	1.36	2.38	1.82	2.62	1.01 (Base model)

3.2. Semantic fluency

The best fit model of semantic fluency for disconnection probability was the random-intercept left WM model (p < 0.01) (Supplementary Table 5), demonstrating greater inter-subject variabilities at 3 months post-stroke, but an overall increase of the semantic fluency scores across the participants during the nine-month recovery period. Fig. 3A presents the individual and averaged trajectories of change in semantic fluency scores between 3 months and 12 months. The marginal pseudo-R-squared of the left WM model was 0.15. The conditional pseudo-R-squared representing the variance explained by the entire model reached 0.59.

Fig. 3.

Semantic fluency in relation to time and the disconnection probability of the left SLF II. A) The semantic fluency Z scores at 3 months and 12 months after stroke (black dots). Individual trajectories of the score changes were represented by the connected blue lines between the two time points (data with missing values were not connected). The red line indicated the averaged trajectory and the grey ribbon represented the standard error. B) The negative linear relationship between the disconnection probability (at the 50 % threshold) of the left SLF II and the semantic fluency scores. The violet and blue lines were the fitted linear regression lines with the correspondent color ribbon representing the 95 % confidence interval. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

To explore the effects of the 11 left WM tracts on semantic fluency, we conducted model comparisons by progressively reducing the 11 tracts from the complete left WM model using LRT. The reduced left WM model only including the left posterior AF (estimate = -1.14, p < 0.05), long AF (estimate = 1.87, p < 0.05) and SLF II (estimate = -1.67, p < 0.05) showed the best model performance (Table 4). The negative association between the left posterior AF/SLF II and semantic fluency indicated that a greater probability of damage to this WM tract may result in decreased semantic fluency. Fig. 3B illustrates that the disconnection probability of the left SLF II was more negatively associated with semantic fluency at 3 months (violet line) (r = -0.23, p < 0.05) than at 12 months (blue line) (r = 0.11, p > 0.05). Note that we applied a 50% threshold and set the values to 0 if the overlapping probability of the lesion sites of each stroke patient with a WM tract was less than 0.5.

Table 4.

The linear mixed-effects of the predictors on semantic fluency in the best fit left WM model.

Predictor	Estimate	t	p.adjusted
Intercept	0.20, 95 % CI [-0.56, 0.97]	0.52	0.84
Time	0.10, 95 % CI [-0.13, 0.34]	0.83	0.43
Age	0.03, 95 % CI [-0.20, 0.25]	0.22	0.83
Sex	0.15, 95 % CI [-0.34, 0.67]	0.56	0.58
Handedness	−0.09, 95 % CI [-0.85, 0.70]	−0.23	0.82
Education	−0.02, 95 % CI [-0.22, 0.18]	−0.17	0.87
Total volume	0.02, 95 % CI [-0.37, 0.40]	0.08	0.94
Lesion volume	−0.15, 95 % CI [-0.39, 0.10]	−1.17	0.25
Mean cortical thickness	−0.05, 95 % CI [-0.30, 0.20]	−0.39	0.70
Subcortical GM volume (left)	0.13, 95 % CI [-0.29, 0.55]	0.57	0.57
Subcortical GM volume (right)	0.01, 95 % CI [-0.41, 0.44]	0.06	0.95
Long AF (left)	1.87, 95 % CI [0.13, 3.61]	2.07	0.04*
Posterior AF (left)	−1.14, 95 % CI [-2.04, −0.24]	−2.44	0.03*
SLF II (left)	−1.67, 95 % CI [-3.17, −0.16]	−2.13	0.04*

*<0.05 (FDR corrected).

The follow-up subject-mean centering analysis found that only the between-subject effect of the left SLF II was significant (estimate = -2.26, p < 0.05), indicating that, on average, the stroke patients with a higher risk of damage to the left SLF II tended to perform worse on the semantic fluency task. The within-subject effects were not significant (SLF: estimate = 1.59, p > 0.05; long AF: estimate = -0,84, p > 0.05; posterior AF: estimate = -1.51, p > 0.05), suggesting that, on average, the effects of these tract disconnections on semantic fluency were similar at 3 months and 12 months poststroke. Fig. 4A presents the anatomical overlap between the left SLF II tract template and the lesions of the subjects above the 50% disconnection probability at 3 months post-stroke.

Fig. 4.

A) The sagittal, coronal and axial views of the overlap between the left SLF II and the lesions of the subjects above the 50 % disconnection probability at the 3 months post-stroke. The left SLF tract was based on the MNI space template reconstructed from 1065 subjects proposed by Yeh and Tseng (2011) using a multishell diffusion scheme. Both the lesion mask and the SLF tract were superimposed upon the MNI152_T1_0.5 mm.nii anatomical image in Dsi_studio ((https://dsi-studio.labsolver.org). There were 1883 fiber tracts of the left SLF II overlapping with the lesion mask. B) The left SLF related to semantic fluency as reported by Egorova-Brumley et al. (2022).

For disconnection severity, no significant WM tract predictors were identified for the best fit model (the base model), suggesting that disconnection severity was not as related to the behavioral scores as the disconnection probability of the WM tracts.

The 5-fold cross-validations identified the left WM model as the best model (1.30) for semantic fluency on the training set for disconnection probability and the base model as the best model (1.32) for disconnection severity (Table 5). The non-parametric Kruskal-Wallis rank sum tests showed that there was a significant difference in the RMSE of the four models (disconnection probability: ${\chi }^2$ = 112.04, df = 3, p < 0.001; disconnection severity: ${\chi }^2$ = 297.54, df = 3, p < 0.001). The post-hoc pairwise Wilcoxon tests with Bonferroni corrections revealed that the RMSE of the left WM model was significantly lower than the other three models for disconnection probability (left WM vs base: p < 0.001; left WM vs right WM: p < 0.001; left WM vs full: p < 0.001). The base model RMSE was significantly lower for disconnection severity (base vs left WM: p < 0.001; base vs right WM: p < 0.001; base vs full: p < 0.001). However, both trained models produced increased RMSE (1.54 and 1.39) on the 20% hold-out test data, suggesting that out-of-sample generalization was not strong. The left SLF II was still significantly correlated with semantic fluency in the positive direction (t = -2.41, p < 0.05) for the left WM model of disconnection probability, however, the effects of the left long and posterior AF were not significant on the smaller sample of the test data (p > 0.05) (Supplementary Table 8). The variance explained by the left model was 18.6% with the fixed effects and 88% with the entire model. No predictors were significant for the base model of disconnection severity for the test set (Supplementary Table 9).

Table 5.

Semantic fluency - RMSE of the four models on the training/test set.

	80 % Training Set				20 % Test Set
Semantic Fluency	Base Model	Left WM Model	Right WM Model	Full Model	Best Model
Probability	1.32	1.30	1.36	1.35	1.54 (Left WM model)
Severity	1.32	2.04	2.81	3.07	1.39 (Base model)

In general, because the model performances on the test set did not deteriorate dramatically from the training set, this suggested some extension of our findings of the importance of left WM to semantic fluency outside of data, but also suggested that our findings were at least partially constrained by the specific features of the project dataset. Nevertheless, the effects of the left SLF and AF in the current study were corroborated by our previous study using the direct DWI approach (Egorova-Brumley et al., 2022), which also identified that the two tracts were associated with semantic fluency. Fig. 4A illustrates the disconnected left SLF II of the current study as the overlap between the tract template and the lesion mask of the subjects. Fig. 4B presents the portion of the SLF identified to be associated with semantic fluency in Egorova-Brumley et al. (2022).

4. Discussion

We investigated the effects of subcortical GM volume and damage to WM fiber bundles on verbal fluency of individuals at 3 months and 12 months post-stroke utilizing non-diffusion-weighted anatomical scans (T1, T2 and FLAIR) routinely acquired in clinical settings. We found that disruptions to the left SLF and posterior AF were negatively related to semantic fluency on the training set. Although no improvements on the test data suggested partial constraint of this finding to the specific data within this study, the results were aligned with the findings of the high-resolution DWI study by Egorova-Brumley et al. (2022) on the same clinical population (SLF: pFDR = 0.035; AF: pFDR = 0.037), which reported the similar significant associations of the two tracts with semantic fluency. In contrast, we did not identify any WM tracts as significantly correlated to phonemic fluency, which also corroborated findings from our team’s DWI study (Egorova-Brumley et al. 2022). Interestingly, we found that the volume of right subcortical GM was positively related to phonemic fluency but not semantic fluency, when total brain volume, lesion volume and cortical thickness were controlled, and this relationship generalized well on the test data. This suggests that non-DWI anatomical scans can serve as good proxy sources to derive neuroimaging variables related to verbal fluency of stroke survivors, potentially generalizing beyond the specific characteristics of the stroke sample in this study.

4.1. White matter tract disconnection and semantic fluency

Our study echoed the findings of prior diffusion-weighted literature that certain left WM tracts contribute significantly to semantic fluency (Li et al., 2017). We identified the left SLF and the posterior segment of AF as significant predictors, similarly reported in other DWI studies. The present study provides complementary results, without relying on the microstructural integrity information only accessible from the DWI data (e.g., fractional anisotropy, mean diffusivity, fiber density and cross-section, etc.), offering the potential application of non-diffusion-weighted clinical scans (e.g., T1, T2, FLAIR) in detecting the neural substrates related to semantic fluency post-stroke.

The SLF and AF connect the temporo-parieto-occipital lobe to the ipsilateral frontal lobe and have long been treated as a uniform tract due to their close anatomical locations and structure (Petrides & Pandya, 1984). Contemporary dorsal–ventral models describing the organization of the brain’s language network consider the left SLF and AF as crucial dorsal tracts involved in a wide range of language functions in aging populations and individuals diagnosed with aphasia (Bajada et al., 2015, Bornkessel-Schlesewsky et al., 2015, Dick et al., 2014, Hickok and Poeppel, 2004, Ivanova et al., 2016, Madhavan et al., 2014), such as compromised function in speech fluency (Bates et al., 2003, Fridriksson et al., 2013, Marchina et al., 2011, Wang et al., 2013), repetition (Breier et al., 2008, Forkel et al., 2020, Geva et al., 2015, Kümmerer et al., 2013), naming (Geva et al., 2015, Marchina et al., 2011, Wang et al., 2013) and comprehension at the word and sentence levels (Friederici and Gierhan, 2013, Grossman et al., 2013, Ivanova et al., 2016, Wilson et al., 2011). With the development of more advanced DWI techniques, many studies have revealed that the SLF and AF are separate entities, and each is composed of multiple segments with distinct origins, courses, and terminations (Catani et al., 2005, Catani and Mesulam, 2008, Dick and Tremblay, 2012, Makris et al., 2005). However, the functional roles of these segment components in language processing have been poorly delineated in the literature. Recent work by Ivanova et al., 2021, Forkel et al., 2020 reported that the long segment of left AF was related to naming, the anterior segment to speech rate/naming, and the left posterior segment to repetition/ comprehension.

The present study provides evidence that the left SLF and the posterior segment of the left AF contribute to semantic fluency. These findings support the segment model of Catani et al. (2005) which suggests that the left posterior temporo-parietal periventricular WM is a converging projection locus of both tracts and is closely related to lexical-semantic processing. Compared to the phonemic fluency task, the semantic fluency task demands taxonomic and thematic lexical retrieval, engaging working memory and lexical-semantic mapping. It is possible that damage to the left SLF disrupts the working memory function in the prefrontal region from providing information to regulate the lexical retrieval process in the inferior parietal region. Damage to the posterior segment of AF around Wernicke’s area may disrupt the lexical-semantic mapping process, per se. Interestingly, we found that only left WM tracts to be associated with semantic fluency, despite that the fact that the most frequently lesioned regions in our dataset were in the right hemisphere. This suggests that left WM tracts play a dominant role in lexical-semantic mapping, which corroborates most modern models of lexical retrieval (Hickok, 2012, Hickok and Poeppel, 2004).

The significant between-subject effect of the left SLF confirms that stroke patients with a higher risk of damage to the tract tend to perform worse on the semantic fluency task, but individuals with a greater disconnection probability of the left posterior AF did not necessarily present with worsening semantic fluency performance. This result suggests that the left SLF could be a more robust imaging variable related to semantic fluency performance at the population level. Additionally, the lack of significant within-subject effects of the two tracts indicates that the effects of tract disconnections on semantic fluency may have already plateaued at the early chronic stage of stroke (3 months). This possibility needs to be validated by other data of stroke patients at the acute stage or already well into the chronic stage.

Cautions need to be exercised when interpreting the left SLF and the left posterior AF as predictive of semantic fluency in general because we found that the generalizability of the left was not strong. Although the present study and our previous DWI study (Egorova-Brumley et al., 2022) demonstrate converging evidence underlining the importance of these two tracts on post-stroke semantic fluency, our cross-validation procedure showed the risk of overfitting the training data due to high variance. The reliability of the tracts in predicting semantic fluency should therefore be further validated on other independent datasets.

4.2. Right subcortical grey matter volume and phonemic fluency

Our results confirmed the second hypothesis that beyond the cortical neural underpinnings, subcortical GM volume would be independently predictive of verbal fluency. We found that the right subcortical GM volume was positively associated with phonemic fluency of stroke survivors. The result can be generalized to the test data in our study, as our cross-validation procedure found consistent improvement of accuracy of model prediction. This finding also agrees with Ellfolk et al., 2014, Grogan et al., 2009 who reported that right subcortical GM volume was linked to phonemic fluency for patients with early Parkinson’s disease and a bilingual neurotypical group of adults, respectively.

Several functional imaging studies have found that both the basal ganglia and the caudate nucleus were more involved in tasks relying on phonological processing (Tettamanti et al., 2005, Watkins et al., 2002). Disruptions of these regions lead to repetition and naming difficulties. Our results suggest that reduced subcortical GM volume of stroke patients may result in greater impairment in phonemic fluency. However, we could not exclude another possibility that the significant effect of right subcortical GM volume on phonemic fluency was mediated by other factors, such as executive functioning. Abundant evidence from subcortical aphasia and bilingual studies has shown that damage to subcortical structures may lead to dysfunctions in concept generation, semantic monitoring, lexical selection, and speech planning (Abutalebi et al., 2008, Crosson, 1985, Nadeau and Crosson, 1997). As the phonemic fluency task requires participants to produce words with a given initial letter, it needs the above control processes to select a word from the competing phonetic and lexical alternatives and formulate the appropriate articulatory utterance. Those stroke patients with more severe loss of GM volume in these subcortical regions may be more likely to be impaired in their performances.

One more point worth noting is that both phonemic and semantic fluency tasks involve multiple cognitive and motor processes. Although the CANVAS project did not collect data on the behavioral control of the motor/articulatory aspects of speech production, we did perform follow-up analyses with the bilateral cortico-spinal tracts as neural controls for motor effects in the models. In other words, one could argue that the left WM tracts and the right subcortical GM identified in the study were predicting unique motor components related to verbal fluency, rather than phonological or semantic components of verbal. However, when adding in the ‘neutral’ bilateral cortico-spinal WM tract, which would not be anticipated to be correlated with verbal fluency but instead with motor function, we found that the results were similar. The base model as it was in the main analysis was again the best model for predicting phonemic fluency, with the right subcortical volume (t = 2.62, p < 0.05) as a significant predictor after the FDR multiple comparison corrections. The model including the left long and posterior AF and SLF II still performed best for semantic fluency, such that the left long (t = 2.07, p < 0.05) and posterior (t = -2.47, p < 0.05) AF and SLF II (t = -2.04, p < 0.05) remained significant predictors. These findings suggest that the motor/articulatory aspects of speech production were less likely to be confounding factors.

4.3. Limitations

The current study had some limitations that should be addressed for the reference of future studies. One is that our dataset comprised relatively mild language impairment (and stroke) severity, which may yield very limited change in both neural substrates and behavioral correlates over time. The data were analyzed only at 3 months and 12 months poststroke. We chose to evaluate a change between these two time periods because stroke recovery mechanisms (e.g., inflammation) have typically plateaued, revealing more stable brain anatomy and an accurate reflection of the extent of the lesion damage. Future research on datasets with greater stroke severity and evaluated across multiple recovery stages may help to further validate the results. We also encourage further validation in clinical datasets (i.e., datasets with lower resolution anatomical scans) to evaluate whether our findings hold true for data collected as part of typical clinical imaging protocols (e.g., on 1.5 T MRI scanners compared to the data analyzed here, which was collected on a 3 T system with 1 mm resolution). In terms of cross-validations, the best fit models of semantic and phonemic fluency were tested on the validation sets separated from the same data pool. It is possible that the general data features will artificially reduce RMSE. More rigorous testing needs to be performed on other independent datasets in future studies to evaluate the performance and generalizability of the models. In addition, cerebellum was not examined in the present study, despite its contributions to oral fluency being documented (Jossinger et al., 2023). We had a very small number of participants with cerebellar strokes (N = 7), thus preventing us from reliably evaluating its effect on fluency.

5. Conclusions

Our study utilized anatomical, non-diffusion images to identify that semantic fluency was supported by the left SLF and the posterior segment of the left AF for chronic stroke patients. Greater disconnection probabilities of the two tracts were correlated to reduced semantic fluency. In contrast, phonemic fluency pointed to the right subcortical GM volume as the crucial neural substrate. The results lend further support to the claim that semantic and phonemic fluency may engage different neural correlates.

Funding

This work was supported by NHMRC grant [APP1020526], Brain Foundation, Wicking Trust, Collie Trust, and Sidney and Fiona Myer Family Foundation. N.E.B. was supported by ARC grant [DE180100893]. A.B. was supported by National Heart Foundation Future Leader Fellowships [100784 and 104748].

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

Data will be made available on request.