Brain-behavior correlates of speech sequencing and stuttering: leveraging functional MRI and machine learning
Embargo Date
2027-11-25
OA Version
Citation
Abstract
Speech production is a complex motor behavior that typically appears effortless, yet in people who stutter (PWS), this finely tuned system is prone to involuntary disfluencies. Persistent developmental stuttering is a neurological disorder with considerable heterogeneity. To advance understanding of the neural and behavioral mechanisms underlying stuttering, this dissertation investigated motor speech sequencing in PWS and stuttering subtypes through three complementary studies. The first study employed functional magnetic resonance imaging (fMRI) to examine multi-syllabic sequence learning in 30 adults who stutter (AWS) and 30 age-matched neurotypical controls (CON) performing a nonword repetition task. Although AWS were less accurate than CON, learning trajectories were comparable. Neuroimaging results revealed increased activation for novel sequences in regions implicated in phonological working memory, speech planning, auditory processing, and articulation, with decreased connectivity between interhemispheric inferior frontal sulcus (IFs). Interestingly, IFs exhibited minimal change in activation with learning, contrasting with prior studies involving learning of non-native syllables, suggesting it processes all syllables in a multi-syllabic word regardless of familiarity. AWS demonstrated reduced basal ganglia-thalamo-cortical loop activity compared to CON, with higher stuttering severity associated with decreased connectivity between basal ganglia and cortical speech-motor areas. The second study adapted the same behavioral multi-syllabic sequence learning paradigm with shorter stimuli for 12 children who stutter (CWS) and 8 age-matched children with no stutter (CNS), ages 6-13, with no neuroimaging. Results mirrored the first study: CWS were less accurate than CNS but demonstrated similar learning trajectories with longer durations and reaction times, providing further evidence that stuttering is an impairment in execution, rather than learning. The final study applied supervised machine learning to classify stuttering subtypes using speech samples and resting-state functional connectivity data from 112 AWS and 108 CON. Models achieved ~65.5% accuracy in predicting predominant disfluency types (50% chance level), while severity classification was only slightly above chance. Based on these models, disfluency subtypes may be associated with potential neural connectivity differences, whereas stuttering severity showed less consistent patterns. Together, these studies provide insights into speech motor sequence learning and variability in stuttering. By integrating neuroimaging and machine learning approaches, this work elucidates the intricate brain-behavior dynamics in PWS and highlights future directions for personalized interventions.
Description
2026
License
Attribution-NonCommercial-NoDerivatives 4.0 International