Reduced integrity in the frontal aslant tract's premotor connections is associated with dysfluency severity in stuttering

Abstract

Stuttering is a neurodevelopmental disorder marked by impaired speech production and substantial interindividual variability in symptom severity. Although neuroimaging studies have identified abnormalities in the basal ganglia-thalamocortical circuitry, the specific white matter pathways that underpin dysfluent speech production remain poorly understood. Here, we focus on the frontal aslant tract (FAT), a dorsoventral intralobar tract connecting the superior and inferior frontal cortices and proposed to support both cognitive and motor functions. Using diffusion MRI, we first compared the quantitative anisotropy of the FAT and relevant control tracts in individuals who stutter (n = 30, mean age = 35.67, SD = 13.03) and fluent speakers (n = 22, mean age = 35.64, SD = 10.87) and observed no significant group-level differences in either hemisphere. To delineate the role of FAT subbundles in speech dysfluency, we then segmented the tract at its dorsal and ventral cortical terminations and assessed their relationships with stuttering severity (Stuttering Severity Instrument-Fourth Edition). We identified left FAT subbundles connecting the posteroventral frontal structures, particularly the rostroventral precentral gyrus and, to a lesser extent, the pars opercularis to a dorsal prefrontal region whose integrity predicts stuttering severity. The left-lateralized nature of these associations suggests that stuttering results, in part, from abnormal connectivity between prefrontal systems governing volitional control and the motor circuits engaged in speech production. Together, these findings refine current neurobiological models of stuttering by implicating frontofrontal white matter pathways underlying speech dysfluency and highlight a potential anatomical target for individualized therapeutic interventions.

Significance statement.

Stuttering is a speech disorder that affects fluency and varies in severity among individuals. Neuroimaging research has shown abnormalities in speech and motor brain areas, but the nature of the association between white matter pathways and stuttering remains debated. This study focuses on the frontal aslant tract (FAT), a white matter pathway involved in motor and cognitive functions. We examined the integrity of different parts of the FAT in people who stutter and fluent speakers. Our findings indicate that only the left FAT fibers connecting the premotor cortex and the inferior frontal gyrus with the dorsal prefrontal control region are correlated with stuttering severity. This suggests that stuttering may result from issues in how the brain coordinates voluntary speech control and motor-speech program.

Introduction

Persistent developmental stuttering is a neurodevelopmental disorder affecting around 1% of the global adult population, roughly 70 million individuals (1, 2). It is characterized by disrupted speech fluency, with atypical rate, rhythm, and smoothness. Speech dysfluencies typically include repetitions or prolongations of words, syllables, and sounds, as well as speech blocks and difficulties initiating articulatory movements (3).

The biological origins of stuttering remain highly debated and appear to be multifactorial. In addition to genetic factors (4–9), atypical neurobiological mechanisms and structures have been proposed to be implicated in this fluency disorder. With respect to fluent speakers, people who stutter (PWS) exhibit decreased functional connectivity and gray matter volume, particularly affecting the basal ganglia-thalamocortical circuit (10–13) and frontal speech areas (14–18).

Diffusion-weighted MRI studies have explored the structural connectivity within the speech and language networks in PWS, identifying atypical microstructural properties in several major white matter tracts. Among those fiber tracts, differences have been reported in the cerebellar peduncles (19–21), in the arcuate and superior longitudinal fasciculi (22–27), in the corticospinal tract (CST) (28, 29), and in the frontal aslant tract (FAT) (20, 23, 29–31). However, the current evidence is often inconsistent, and no consensual interpretation has been reached yet. While some studies report changes in the left, language-dominant hemisphere, others report abnormalities in the right hemisphere, suggesting compensation mechanisms. Additionally, whereas some authors have reported decreased diffusivity patterns in PWS compared with fluent speakers, others have found increased connectivity strength in PWS.

Despite this mixed picture, there are reasons why certain tracts may be vulnerable in stuttering, especially the FAT. The FAT is an intralobar white matter fasciculus that connects the supplementary motor area (SMA) and pre-SMA to the inferior frontal gyrus (IFG) (32). SMA and pre-SMA are parts of the mesial frontal system, which contribute to voluntary motor initiation, control, and sequencing. The IFG, including the pars opercularis and triangularis, is well-known for its role in speech and language production. By connecting these regions, the FAT is considered part of the speech and language networks, as it is associated with speech initiation, verbal fluency, and lexical decision (33–36). Direct electrostimulation studies during awake neurosurgeries for glioma resection have provided compelling evidence for the role of the FAT in stuttering. Kemerdere et al. (37) reported that direct axonal stimulation of the left FAT causes reversive intraoperative stuttering in fluent speakers, strongly suggesting the causal involvement of this tract in speech fluency. In the same line, Kinoshita et al. (38) demonstrated that tumor resection surrounding the left FAT is associated with transient speech initiation disorders postsurgery. It is worth noting, however, that the neural mechanisms underlying dysfluency episodes triggered by axonal stimulation or tumor resection might be different from those inherent to the pathophysiology of developmental stuttering.

Recently, an expanded version of the FAT has been proposed, featuring dorsal cortical projections that extend more anteriorly into the superior frontal gyrus and ventral projections reaching more posterior regions, including the ventral premotor cortex (vPMC) (39–43). Beyond these anatomical refinements, recent work has focused on the functional relevance of distinct segments of the FAT. Collectively, they suggest a functional organization along an anteroposterior gradient, with the anterior segment of the FAT supporting lexico-semantic functions and the posterior segment contributing to motor-speech functions (39, 40, 44, 45). The current study aims to examine the possible contribution of different FAT segments to speech fluency. Building on the evidence mentioned above, we hypothesized that stuttering primarily stems from disruption of left FAT projections to the vPMC, a brain area critical for articulatory planning (46–48), rather than from diffuse dysfunction of the tract as a whole.

To test this hypothesis, we reconstructed white matter pathways using generalized q-sampling imaging (GQI) (49), which estimates the spin distribution function at each voxel, allowing precise characterization of complex fiber configurations. From these reconstructions, we derived quantitative anisotropy (QA) values along the FAT and other language- or motor-related white matter tracts to quantify tract-specific diffusion properties. QA reflects the amplitude of anisotropic diffusion along a given fiber orientation and serves as a macroscopic index of tract integrity sensitive to alterations in axonal organization. First, we compared the QA metric for all included tracts between 30 PWS and 22 fluent control speakers, with the hypothesis that no difference would be observed, irrespective of hemispheric laterality. To disentangle the specific contributions of the FAT subcomponents, we divided the tract according to its dorsal and ventral cortical projections, and we examined whether the integrity of these subbundles correlated with stuttering severity in PWS.

Here, we show that stuttering severity, particularly during spontaneous speech, is associated with left-hemisphere FAT segments connecting the rostral ventral precentral gyrus and, with more limited involvement, the adjacent pars opercularis to a dorsal prefrontal region. This left-lateralized pattern suggests that stuttering could result from disrupted integration between prefrontal mechanisms of volitional control and the motor execution systems governing speech production.

Results

Behavioral assessment

Participants’ behavioral assessment consisted of stuttering severity evaluation, additional measurements of verbal fluency, and a visuospatial attention test. Differences in behavioral scores between PWS (n = 30) and fluent speakers (n = 22) were assessed with a permutation-based analysis of covariance (ANCOVA; Freedman–Lane procedure, 10,000 permutations) controlling for age and sociocultural background. Full statistics for all behavioral comparisons are provided in Table S1.

Speech fluency was first analyzed using the Stuttering Severity Instrument-Fourth Edition (SSI-4) (50), which provides indices of stuttering severity during spontaneous speech, reading, and a global severity score combining measures of the frequency of dysfluencies, the duration of the longest stuttering events, and the associated physical concomitants. Participants’ speech fluency was evaluated under two different conditions: (i) during spontaneous speech, where they described a picture or engaged in an open-topic conversation for 2 min, and (ii) during reading, which required them to read aloud a standardized 150-word passage that was designed to ensure consistency in phonetic and syntactic complexity. PWS exhibited significantly higher severity scores than controls across all speech fluency measures, including total SSI-4 score (F(1,48) = 16.41, P_FDR = 0.001), spontaneous speech (F(1,48) = 26.02, P_FDR = 0.0008), and reading subscores (F(1,46) = 5.28, P_FDR = 0.028; Fig. 1A). It is worth noting that the group of PWS exhibits an important dispersion of stuttering severity scores.

Figure 1.

Comparison of PWS and controls’ behavioral scores on speech and verbal fluency tasks. A) Violin plots illustrating group comparisons on the SSI-4 total score and subtask scores. B) Violin plots illustrating group comparisons on the verbal fluency and visuospatial attention scores. Stars indicate significance levels, where ns, P > 0.05, *P ≤ 0.05, **P ≤ 0.01, and ***P ≤ 0.001.

Participants next completed a verbal fluency task that included phonological and semantic conditions. In the semantic fluency task, participants were given 2 min to say aloud as many words as they could that fell under a specific semantic category, which in this instance were animals. In the phonological fluency task, they were asked to speak aloud as many words as possible starting with a [p], within 2 min. The control group outperformed PWS in phonological fluency (F(1,45) = 9.15, P_FDR = 0.009), whereas no significant group difference was observed for semantic fluency (F(1,45) = 2.52, P_FDR = 0.12; Fig. 1B, left panel).

The neuropsychological measure of visuospatial exploration and attention, also known as the Bell test, required participants to circle all the instances of a specific object (a bell) among various distractors displayed on an A4-sized sheet. A group effect was observed on this visuospatial attention test (F(1,46) = 5.96, P_FDR = 0.03), with PWS making significantly more omissions than controls (Fig. 1B, right panel).

To systematically assess potential associations between SSI-4 scores, verbal fluency, and visuospatial attention performance, Spearman correlation analyses were used to examine between-task relationships. The resulting correlation matrix, including R values and significance levels, is provided in Fig. 2. In brief, the SSI-4 total score and its subscores systematically correlated with verbal fluency performance but not with the performance on the visuospatial attention test. The results of each correlation are provided in Table S2.

Figure 2.

Correlation matrix between SSI-4 scores, verbal fluency scores, and the visuospatial attention score. The values correspond to Spearman correlation coefficients. Stars indicate significance levels, where ns, P > 0.05, *P ≤ 0.05, **P ≤ 0.01, and ***P ≤ 0.001.

Together, these behavioral findings confirm that PWS are impaired in speech fluency, more markedly during the spontaneous speech condition, and that this deficit extends to verbal fluency in the phonological domain. Visuospatial attention was also affected, although to a more subtle degree, and does not appear to be associated with speech or verbal fluency performance.

Whole tract results

Diffusion-weighted MRI data were acquired for all participants using a 3 Tesla scanner, and tractography analysis was performed on four white matter fasciculi of interest. To test the hypothesis that axonal integrity of the FAT is not globally altered in PWS, we reconstructed it bilaterally using a QA-based deterministic fiber tracking algorithm in DSI studio (see Materials and methods), along with a language-control tract (i.e. the inferior longitudinal fasciculus; ILF) and two motor-control tracts (i.e. the corticospinal and frontal-striatal tracts; CST and FST, respectively). The ILF was included as part of the ventral language network, since in the left hemisphere, it supports lexico-semantic processing and language comprehension (51), rather than speech production. The CST served as a purely motor reference tract running posterior to the FAT, while the fronto-striatal tract connects the SMA and the pre-SMA with the striatum, providing a subcortical control pathway closely related to FAT in both anatomy (shares dorsal cortical projections) and function (speech control). We investigated tract-specific effects of group, hemispheric laterality, and their interaction on QA measurements using a nonparametric permutation framework. For each bilateral tract (i.e. FAT, ILF, FST, and CST), left- and right-hemisphere QA values were extracted to compute bilateral mean (L + R)/2 and a laterality index defined as the right–left difference (R−L). Analyses followed a mixed two-factor design with group (between-subjects) and laterality (within-subjects, R–L), while controlling for additional covariates (age and sociocultural level). No significant main effects or interactions were found in QA in any tract, indicating preserved macrostructural integrity and hemispheric organization across language and motor-related pathways in PWS (Fig. 3). It is noteworthy, however, that a near-significant interaction between group and hemisphere was found for the ILF and the FST (P_FDR = 0.09 and P_FDR = 0.08, respectively). For both tracts, PWS tend to have a lower mean QA in the left hemisphere than in the right. Detailed statistics for each tract are provided in Tables S3–S6.

Figure 3.

Whole tract reconstruction and violin plots illustrating left- and right-hemisphere QA comparisons between groups. FAT, frontal aslant tract; FST, fronto striatal tract; ILF, inferior longitudinal fasciculus; CST, corticospinal tract.

In a second step, we examined associations between tract-specific QA values and behavioral measurements using partial Spearman rank correlations, controlling for age and sociocultural background. Statistical significance was determined through permutation testing (10,000 iterations per test), with false discovery rate (FDR) correction applied across predictors within each behavioral domain. No partial correlation survived FDR correction for the FAT, ILF, FST, or CST (Tables S7–S10). The same results were observed when considering uniquely PWS (Tables S11–S14).

FAT subbundles results

To test the hypothesis that only the premotor projections of the FAT are associated with dysfluency severity, the tract was subdivided according to its cortical projection sites. Cortical regions were defined using the Human Connectome Project Multimodal Parcellation (HCP-MMP) atlas (52). Nine dorsal regions of interest (ROIs) were selected to cover the dorsolateral prefrontal cortex, the SMA/pre-SMA complex, and dorsal premotor areas (SFL, 6ma, 8BL, s6-8, 8BM, SCEF, 6a, i6-8, and 8A), and 10 ventral ROIs were chosen to include ventral premotor regions, the IFG, and the frontal operculum (area 44, area 45, 6v, 6r, IFJa, IFJp, IFSp, IFSa, FOP4, and FOP5). The selected ROIs are depicted in Fig. 4A.

Figure 4.

FAT subbundle reconstruction and correlations with SSI-4 scores and subscores. A) Dorsal and ventral ROIs from the HCP-MMP atlas for the FAT cortical projections. Superior panels show lateral and medial views of the dorsal cortical projections, while the inferior panels depict lateral and medial views of the ventral cortical projections. B) Circular plots showing the correlations between QA values and SSI-4 scores in the left FAT subbundles. Continuous lines represent significant P-corrected correlations. Bold dashed lines represent the subbundles existing in at least 30/52 participants, for which the P-corrected values were not significant. Light dashed lines represent subbundles present in fewer than 30/52 participants and were therefore not included in the analysis (Table S11). In any case, the width of the lines is proportional to the correlation coefficient. C) Reconstruction of FAT fibers connecting dorsal to ventral ROIs, averaged over all participants. D) Left FAT subbundles correlated with stuttering severity as measured by the total SSI-4 score. E) Correlation slopes between QA values and spontaneous speech score in all participants and PWS only for subbundles s6-8-6r and s6-8-44.

For each hemisphere and each participant, a 9 × 10 connectivity matrix was constructed representing QA values for all pairwise connections among these ROIs. The resulting matrices (9 × 10 × n) were analyzed using edge-wise partial Spearman correlations between QA and behavioral measures, controlling for age, sociocultural background, and the number of streamlines in pairwise ROI connections. Permutation inference (10,000) was used to assess statistical significance, and multiple comparisons were controlled across all edges within each behavioral outcome using the FDR correction (two-sided, q < 0.05). Only edges for which the connection exists for ≥30 participants (i.e. 57.7% of the subjects) were analyzed to maintain sufficient statistical power. According to this criterion, a total of eight subbundles were analyzed in the left hemisphere versus nine in the right hemisphere (Fig. S1).

We observed a significant negative correlation between total SSI-4 scores and the QA values of three subbundles of the left FAT: one connecting the dorsal ROI s6-8 and the ventral ROI 6r (r = −0.43, P_FDR = 0.022), one connecting the dorsal ROI s6-8 and the ventral ROI 44 (r = −0.40, P_FDR = 0.022) and another one connecting the dorsal ROI SFL and the ventral ROI 6r (r = −0.39, P_FDR = 0.032). A significant negative correlation was also found between SSI-4 spontaneous speech scores and two subbundles of the left FAT: one connecting the dorsal ROI s6-8 and the ventral ROI 6r (r = −0.43, P_FDR = 0.038), and the other connecting the dorsal ROI s6-8 and the ventral ROI 44 (r = −0.37, P_FDR = 0.038). Importantly, no correlation was significant for SSI-4 reading scores. The correlations with SSI-4 scores are illustrated in Fig. 4B. No correlation between FAT subbundles’ QA and verbal fluency or visuospatial attention scores remained significant after FDR correction. The circular plots illustrating correlations across all behavioral tasks are shown in Fig. S2, and the corresponding statistics are described in Tables S15–S21. Likewise, no significant correlations were observed for right-hemisphere subbundles after FDR corrections (Fig. S3 and Tables S22–S28).

To assess whether these associations merely reflected group differences between PWS and healthy speakers, we reran the correlation analyses within the PWS group alone. For the left FAT, the main correlation coefficients showed only modest attenuation and retained a broadly similar pattern, suggesting that the observed effects were not exclusively driven by the inclusion of healthy speakers (Tables S29–S35). However, these within-group correlations no longer reached statistical significance in the reduced PWS-only sample, likely reflecting the loss of statistical power associated with halving the sample size. This was particularly evident for the association between left bundle s6-8/6r QA and the spontaneous speech score, which remained moderate in magnitude but did not reach significance (r = −0.35, P = 0.078). Thus, although the PWS-only analysis does not provide definitive statistical confirmation, it argues against a purely group-driven effect and supports the possibility that these associations reflect clinically meaningful variability within developmental stuttering. Beyond, a positive correlation was observed between the semantic fluency score and a subbundle of the left FAT connecting dorsal area s6-8 to ventral area 6r (r = 0.54, P_FDR = 0.047). In the right hemisphere, only a tendency for positive correlations was observed between the semantic fluency score and subbundles connecting dorsal area SFL to ventral area 6v (r = 0.62, P_FDR = 0.062) and dorsal area 6ma to ventral area 6r (r = 0.51, P_FDR = 0.067). Full statistics are provided in Tables S36–S42 for the right FAT.

Lastly, for the three FAT subbundles showing a significant association with stuttering severity (SSI-4 scores), we conducted a posteriori group comparisons of QA values. Given the high heterogeneity of the PWS group in terms of stuttering severity, participants were subdivided into two subgroups: PWS−, corresponding to individuals with no evidence of abnormal disfluencies on the SSI-4 (scores ranging from 0 to 9), and PWS+, corresponding to individuals showing abnormal disfluencies, albeit with varying severity levels (scores ≥10, corresponding to the cutoff). A significant group effect was observed only for the 6r-s6-8 subbundle (F(2,36) = 3.15, P = 0.047). Post hoc analyses indicated that the difference between controls and PWS+ was significant (P = 0.02), while the comparisons between controls and PWS− (P = 0.54) and between PWS+ and PWS− (P = 0.34) did not reach significance (Tables S43). No significant effects were observed for the other two subbundles: 44-s6-8 (F(2,43) = 2.93, P = 0.061) and 6r-SFL (F(2,37) = 1.81, P = 0.18).

Discussion

This study investigated whether reduced white matter integrity of the FAT contributes to stuttering severity in PWS. Using QA-based tractography combined with detailed behavioral assessment, we found that speech dysfluency severity—especially during spontaneous speech—was selectively associated with subbundles of the left FAT interconnecting the rostro-vPMC and, to a lesser extent, the posterior IFG with an area of the dorsal prefrontal cortex. In contrast, no integrity differences emerged across the entire FAT or in control tracts related to language or motor function (ILF, CST, and FST), and no diffusion–behavior relationships were observed in these tracts. Collectively, these findings indicate that stuttering does not stem from a diffuse white matter abnormality but rather from a focal disruption within left-hemispheric FAT subconnections, delineating a distinct structural signature of the disorder.

From a behavioral standpoint, PWS unsurprisingly exhibited greater stuttering severity than fluent speakers during both spontaneous and reading speech, with the strongest effects observed in spontaneous speech. Beyond overt speech dysfluency, PWS showed poorer performance on phonological verbal fluency and visuospatial exploration tasks, whereas semantic fluency remained comparable between groups. SSI-4 scores were negatively correlated with phonological fluency, such that individuals with more severe stuttering produced fewer words under time-limited retrieval demands.

This pattern builds on prior evidence of phonological fluency impairments in adults who stutter (53) and demonstrates that individual differences in stuttering severity parallel variation in phonological lexical-access efficiency, consistent with shared constraints on executive control, lexical retrieval, and speech-motor coordination. Moreover, convergent neurodevelopmental and connectivity studies reveal aberrant coupling between prefrontal cortico-subcortical and sensorimotor speech networks in PWS (54, 55), in agreement with behavioral evidence that they exhibit reduced executive function capacities, such as working memory (56, 57) and action initiation (58). Together, these findings support the view that impaired integration between higher-order control systems and speech-motor planning mechanisms may underlie the cognitive–motor difficulties characteristic of PWS.

The identified FAT segments link the rostral and ventral part of the premotor cortex (area 6r), repeatedly involved in speech articulatory sequencing and coordination (46–48), and to a lesser extent the adjacent posterior IFG (area 44), with a dorsal frontal region (area s6-8), defined as a transitional area between the dorsal premotor cortex and the prefrontal cortex, lateral to the pre-SMA, and engaged in higher-order cognitive mechanisms, such as cognitive control, planning, and working memory (59–61). This connectivity profile suggests that these FAT subconnections serve as a structural interface between higher-order executive systems and motor-related systems supporting speech planning and production. Our observation that FAT integrity predicted dysfluency severity during spontaneous but not reading-based speech or verbal fluency performances further supports the view that stuttering primarily reflects impaired volitional control of speech initiation and production, rather than general articulatory or linguistic deficits. Spontaneous speech relies more heavily on endogenous selection and initiation mechanisms, whereas reading engages externally guided and more automatized speech-motor programs. The specificity of our findings to spontaneous speech thus highlights the FAT's role in the flexible and voluntary control of self-initiated speech action. PWS often report a feeling of “loss of control.” This feeling occurs when speakers know exactly what they want to say, but speech articulators seem to freeze, triggering involuntary disruptions in speech flow (62, 63). The reported reduced integrity of particular FAT segments, allowing communication between areas governing volitional cognitive control in the dorsal frontal cortex to speech initiation and articulatory coordination regions in the premotor and the inferior frontal cortex, represents an important contribution to better understanding the neuroanatomical mechanisms underlying the feeling of “loss of control” experienced by PWS.

The left-hemisphere lateralization of this effect aligns with previous findings of left-dominant activation in prefrontal and premotor cortices during fluent speech (64–66) and suggests that disruption within this hemisphere-specific circuit thus compromises efficient speech-motor control. Notably, the same left FAT subconnections were consistently reconstructed across participants, whereas their right-hemisphere counterparts were more difficult to identify (Tables S11 and S18), suggesting both the structural and functional specialization of this pathway for speech. In line with recent proposals that the FAT follows an anteroposterior functional gradient in the left hemisphere, with anterior segments mediating executive–linguistic functions and posterior segments supporting motor-speech control (42, 43), our results localize stuttering-related abnormalities to the posterior, speech-motor domain of this pathway, while emphasizing its integrative prefrontal–premotor nature.

Our findings align with contemporary neuroanatomical and computational models positing that fluent speech depends on the coordinated interaction between domain-general executive control networks and specialized motor systems (10, 67, 68). In this framework, the FAT acts as a structural bridge between volitional intention and articulatory execution, supporting hierarchical control over speech initiation and sequencing. When the integrity of this bridge is reduced, the dynamic coupling between prefrontal control signals and premotor output becomes unstable, possibly leading to the characteristic halting and effortful production observed in stuttering.

Several limitations should be acknowledged. First, although QA provides a robust measure of anisotropic diffusion amplitude and is sensitive to changes in white matter integrity, it remains a macroscopic diffusion metric and cannot distinguish among specific microstructural features, such as axonal density, myelination, or extracellular composition. In contrast, multicompartment diffusion models such as NODDI can estimate parameters associated with neurite density and orientation dispersion, enabling a more specific characterization of microstructural substrates. Second, while our sample size is generally greater compared with prior diffusion studies of stuttering, larger cohorts will be required to confirm the generalizability of the identified structure–function relationship. Third, the present study focused on structural connectivity; integrating these findings with functional and dynamic connectivity measures, particularly during speech production tasks, will be essential to establish causal mechanisms.

In conclusion, the findings reported here contribute to a refined neurobiological model of stuttering, in which dysfluency arises from atypical white matter connectivity—and, possibly, reduced functional coordination and integration—between cognitive control regions and speech-motor execution systems, within a left-lateralized oblique frontal circuit. Specific segments of the FAT, by bridging dorsal prefrontal and ventral premotor regions, appear to serve as a structural gateway linking volitional speech action control over speech-motor output. Disruption of this pathway could thus compromise the fluid initiation and sequencing of speech, leading to the characteristic halting and effortful production observed in stuttering. Future studies combining structural, functional, and neurostimulation approaches may clarify whether enhancing the efficiency of this pathway could alleviate speech dysfluency in PWS.

Materials and methods

Participants

Thirty PWS and 22 fluent speakers (control group) were recruited. Sociodemographic data and group comparison are presented in Table 1. All participants were native French speakers with no known neurological or psychiatric diseases. The Edinburgh Handedness Inventory (Oldfield, 1971) was used to confirm participants’ handedness. This study was approved by the French national ethical committee (CPP: 2-20-097 id9652) and followed the principles of the Declaration of Helsinki. All participants gave written informed consent for their voluntary involvement in this study.

Table 1.

Descriptive sociodemographic statistics by groups.

	n	Mean age ± SD (range)	Education ± SD (range)	Handedness	Gender: n (%)
Control	22	35.64 ± 10.87 (19–58)	16.26 ± 2.28 (13–21)	R = 18; L = 4	Males: 14 (63.64) Females: 8 (36.36)
PWS	30	35.67 ± 13.03 (18–59)	14.76 ± 2.26 (12–18)	R = 24; L = 4; A = 2	Males: 21 (70.0) Females: 9 (30.0)
Group difference		NS t = 0.009, P = 0.993	Control > PWS t = −2,246, P = 0.03	NS χ2(3, 52) = 2.35, P = 0.502	NS χ2(1, 52) = 0.03, P = 0.854

R, right-handed; L, left-handed; A, ambidextrous.

Behavioral assessment

The stuttering severity instrument (SSI-4; 51) was used to evaluate speech dysfluencies. The SSI-4 provides standardized measures of stuttering severity across three components: frequency of dysfluencies, duration of the longest stuttering events, and physical concomitants. Each participant was assessed during two speaking conditions: (i) spontaneous speech, which involved describing a picture or open-topic conversation for ∼2 min, and (ii) reading, which involved reading aloud a 150-word passage standardized for phonetic and syntactic complexity. Speech samples were video recorded, and stuttering events were identified and scored according to the SSI-4 manual by an experienced phonetician blinded to group identity. A general dysfluency score was established for both conditions and further combined to yield a total severity index. According to the SSI-4 severity levels, nine PWS were categorized as very mild, five as mild, two as moderate, and two as severe, and one as very severe. The remaining scores fell below the detectable stuttering severity threshold.

Additionally, participants completed semantic and phonological fluency tasks (69), as well as a visuospatial exploration task. In the semantic fluency task, participants were asked to recall as many words as possible belonging to a given semantic category (in this case, animals) within a 2-min time limit, while in the phonological fluency task, they were instructed to recall as many words as possible beginning with a specified sound (here, [p]), within 2 min. The visuospatial exploration task (the Bells test; (70)) required participants to identify all instances of a target object (a bell) among numerous distractors arranged on an A4-sized sheet.

Due to technical contingencies affecting the timing, the data collection of performance on behavioral tasks was incomplete for some participants: two participants did not complete the reading task and therefore lack an SSI-4 total score, three participants failed to complete the verbal fluency tasks, and two participants lack a score on the Bell test.

MRI data acquisition

MRI data were acquired for all participants using a 3 Tesla Siemens Prisma scanner (Siemens Healthineers, Erlangen, Germany) equipped with a 64-channel head/neck coil, at the I2FH platform of the Department of Neuroradiology, Montpellier University Medical Center, Gui de Chauliac Hospital (Montpellier, France).

T₁-weighted structural imaging was acquired using a standard 3D magnetization-prepared rapid acquisition gradient echo sequence with the following parameters: repetition time (TR) = 2,300 ms, echo time (TE) = 2.88 ms, inversion time (TI) = 900 ms, flip angle = 9°, voxel size = 1 × 1 × 1 mm.

Diffusion-weighted MRI was acquired using a spin-echo EPI sequence with the following parameters: TR = 4,500 ms, TE = 72 ms, multiband acquisition factor = 2, parallel acquisition factor = 2, voxel size 2 × 2 × 2 mm. A multishell diffusion scheme was used, incorporating three b-values (300, 1,000, and 2,000 s/mm²), with 6, 32, and 32 diffusion sampling directions, respectively, three acquisitions with b = 0 s/mm² with an anterior–posterior phase encoding direction. In addition, a diffusion-weighted acquisition with the same parameters was acquired with a b = 0 s/mm² with a posterior–anterior phase encoding direction.

Diffusion-weighted MRI preprocessing

Preprocessing of both anatomical and diffusion-weighted data was carried out using the QSIPrep pipeline (version 0.19.1, (71)), which is built on Nipype (version 1.8.6; (72)). The analysis pipeline was executed on a Windows Subsystem for Linux environment running on Ubuntu 22.04.4 LTS.

The 3D T₁-weighted (T1w) image was corrected for intensity nonuniformity using N4BiasFieldCorrection (73) and served as the anatomical reference throughout the preprocessing pipeline. The anatomical reference image was reoriented into AC-PC alignment through a 6-DOF (degree of freedom) transform obtained from an affine registration to the MNI152NLin2009cAsym template. Subsequently, a full nonlinear registration from AC-PC to the template space was estimated using symmetric nonlinear registration via antsRegistration. Brain extraction was performed on the T1w image using SynthStrip (74), and automated anatomical segmentation was conducted using SynthSeg (75) from FreeSurfer version 7.3.1.

Restricted diffusion was quantified using restricted diffusion imaging (76), and data reconstruction was performed using generalized GQI (49) with a diffusion sampling length ratio of 1.25. Tensor metrics were calculated using the subset of diffusion-weighted images with b-values <1,750 s/mm². Images with a b-value <100 s/mm² were treated as a b = 0 image. Preprocessing included MP-PCA denoising, implemented in MRtrix3's dwidenoise (77) with a 5-voxel window. Following denoising, the mean intensity of the DWI series was adjusted such that the mean intensity value of all b = 0 images matched across each separate DWI scanning sequence. B1 field inhomogeneity correction was applied using dwibiascorrect from MRtrix3 with the N4 algorithm (73) after resampling the bias-corrected images.

FSL (version 6.0.5.1:57b01774)'s eddy was used for head-motion correction and Eddy current correction (78). The Eddy configuration included a q-space smoothing factor of 10, 5 iterations, and 1,000 voxels used to estimate the hyperparameters. A linear first-level model and a linear second-level model were applied to model spatial distortions related to Eddy currents. Q-space coordinates were forcefully assigned to a discrete shell. Eddy's outlier replacement was enabled (79). Data were grouped by slice, including only values from slices determined to contain at least 250 intracerebral voxels. Within each group, voxels with signal exceeding 4 SDs from model predictions were identified as outliers and replaced with imputed values. Final data interpolation was performed using the Jacobian-based (jac) method.

Several confounding time series were calculated based on the preprocessed DWI. Framewise displacement was computed using the Nipype implementation following the definitions by Power et al. (80). In addition, the head-motion estimates obtained during motion correction were included in the corresponding confounds file. Slicewise cross-correlation was also calculated. The diffusion-weighted MRI time series were resampled to the anterior commissure-posterior commissure (AC-PC)-aligned space, generating a preprocessed diffusion-weighted MRI dataset in ACPC space with 1.3 mm isotropic voxels. Upsampling to a resolution of 1.3 × 1.3 × 1.3 mm was performed using B-spline interpolation.

Diffusion tractography processing

Diffusion orientation distribution functions were reconstructed using generalized GQI (62) with a ratio of mean diffusion distance of 1.25 in DSI Studio (version 94b9c79). A deterministic fiber tracking algorithm (81) was used with augmented tracking strategies (82) to improve reproducibility. The FAT was automatically delineated using the DSI's autotrack function, which matches trajectories to a tractography atlas in ICBM152 space with a distance tolerance of 24 mm (83). The tracking parameters were set as follows: anisotropy threshold = 0.06; angular threshold = 45°; step size = 1 mm. Fibers shorter than 30 mm or longer than 200 mm were discarded. A total of 1,000,000 seed points were generated for tracking. Topology-informed pruning iteration was set to 0. Tract shape analysis (82) was conducted to derive shape metrics for tractography. The same parameters were used to track the ILF, the FST, and the CST automatically. For each tract, QA values were extracted. QA reflects the amplitude of anisotropic diffusion along a given fiber orientation and serves as a macroscopic index of tract integrity sensitive to alterations in axonal organization.

Subdivision of the FAT subbundles

To test the hypothesis that only the posterior portion of the FAT is associated with dysfluency severity, the bundles were subdivided based on their cortical projection sites. To this end, we used the HCP-MMP dorsal and ventral ROIs depicted in Fig. 4A. To analyze the structural properties of the FAT subbundles, we extracted the QA values of each existing tract interconnecting the dorsal ROIs and the ventral ROIs. For each hemisphere and each participant, a 9 × 10 connectivity matrix was constructed, representing all pairwise connections (dorsal-ventral) among the defined ROIs for subsequent analyses.

Statistical analyses

Behavioral scores

Behavioral measures were compared across groups using a permutation-based one-way ANCOVA. For each behavioral variable, we fit a linear model including the group factor (controls versus PWS) and covariates (age and educational background). Group effects were assessed using Freedman–Lane permutation tests with 10,000 permutations and a fixed random seed.

The relationship between speech-related measures and other task performances (verbal fluency and visuospatial exploration) was analyzed using Spearman rank correlations and grouped in a 6 × 6 correlation matrix (Fig. 2).

Group comparisons of QA

For each tract, we evaluated: (i) a “group” main effect on the hemispheric mean $(L+R)/2$, (ii) a “laterality” main effect on asymmetry $(R-L)$, and (iii) a “group × laterality” interaction. All models included group as a between-subject factor and “age” and “education” as covariates. Group and interaction effects were tested using permutation ANCOVA with the Freedman–Lane procedure. The laterality main effect was tested using a sign-flip permutation test applied to the residualized asymmetry values (two-sided). All permutation P-values were computed using 10,000 permutations with a fixed random seed (41) and were FDR-corrected within each family of tests (group, laterality, interaction). All analyses were performed in Python using custom code.

Correlations between whole tract properties and behavioral scores

Associations between tract measures and behavioral scores were evaluated using partial Spearman correlations, controlling for “age” and “education.” For each predictor–behavior pair, variables were rank transformed and then residualized with respect to the covariates; the correlation was computed on the resulting residual ranks. Statistical significance was assessed using two-sided permutation tests (10,000 permutations, seed = 42) generated by permuting the residualized behavioral ranks across participants. For each behavioral measure, Benjamini–Hochberg FDR correction was applied across predictors.

FAT subbundles microstructural properties

For each behavioral measure, we evaluated associations with tract edges (QA) from subject-specific matrices restricted to predefined dorsal × ventral ROIs (see above). Edge values were taken from 9 × 10 matrices, with zeros treated as missing connections (no QA value to exploit). Associations were tested using partial Spearman correlations, controlling for “age,” “education,” and “number of streamlines connecting pairwise ROIs”: edge values and behaviors were rank transformed, both were residualized with respect to ranked covariates, and the correlation was computed on the residual ranks. Significance was assessed with two-sided permutation tests (10,000 permutations), permuting the behavioral residual ranks across participants for each edge. Only edges for which the connection exists for ≥30 participants were analyzed to maintain sufficient statistical power (Tables S15 and S22). Within each behavior, Benjamini–Hochberg FDR was applied across all tested edges, and we report FDR-adjusted q-values.

Supplementary material

Supplementary material is available at PNAS Nexus online.

Competing interests

The authors declare no competing interests.

Funding

F.H.: Agence Nationale de la Recherche ANR-18-CE36-0008. G.H.: Agence Nationale de la Recherche ANR-18-CE36-0008.

Author contributions

Maëva Michon (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Writing—original draft, Writing—review & editing), Fabrice Hirsch (Conceptualization, Data curation, Funding acquisition, Investigation, Methodology, Project administration, Supervision, Writing—review & editing), Ivana Didirková (Data curation, Formal analysis, Investigation, Methodology, Validation, Writing—review & editing), Davide Giampiccolo (Investigation, Methodology, Validation, Writing—review & editing), Marie-Claude Monfrais-Pfauwadel (Data curation, Investigation, Methodology), Anne-Laure Lemaitre (Data curation, Investigation, Methodology), Jérémy Deverdun (Data curation, Investigation, Methodology, Validation), Emmanuelle Lebars (Data curation, Investigation, Methodology, Validation), and Guillaume Herbet (Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Supervision, Writing—original draft, Writing—review & editing)

Data availability

The codes used for statistical analyses are available on the Open Science Framework repository at https://doi.org/10.17605/OSF.IO/JMB6F. Due to the sensitive nature of the clinical and neuroimaging data involved, and because participants did not provide consent for unrestricted public access or general external sharing of raw data, the raw dataset cannot be made publicly available. However, access to pseudonymized individual-level data may be considered upon request from qualified researchers. Such access is contingent upon approval from the study sponsor/data controller. Additionally, it must comply with relevant French ethical and data protection requirements, including the General Data Protection Regulation and the provisions of the National Commission on Informatics and Liberty (CNIL). Where necessary, approval from the competent French Comité de Protection des Personnes will also be required (contact details: cppsoom2@ars.sante.fr; https://www.chu-toulouse.fr/comite-de-protection-des-personnes-cpp). Any granted access will necessitate the completion of an appropriate data-use agreement.