RESEARCH ARTICLE Functional neural circuits that underlie developmental stuttering Jianping Qiao1,2☯ , Zhishun Wang2☯ *, Guihu Zhao3 *, Yuankai Huo2 , Carl L. Herder4 , Chamonix O. Sikora4 , Bradley S. Peterson5,6 1 School of Physics and Electronics, Shandong Province Key Laboratory of Medical Physics and Image Processing Technology, Institute of Data Science and Technology, Shandong Normal University, Jinan, China, 2 Department of Psychiatry, Columbia University, New York, NY, United States of America, 3 School of Information Science and Engineering, Central South University, Changsha, China, 4 American Institute for Stuttering, New York, NY, United States of America, 5 Institute for the Developing Mind, Children’s Hospital Los Angeles, CA, United States of America, 6 Department of Psychiatry, Keck School of Medicine at the University of Southern California, Los Angeles, CA, United States of America ☯ These authors contributed equally to this work. * wangz@nyspi.columbia.edu (ZW); zhaoguihu@foxmail.com (GZ) Abstract The aim of this study was to identify differences in functional and effective brain connectivity between persons who stutter (PWS) and typically developing (TD) fluent speakers, and to assess whether those differences can serve as biomarkers to distinguish PWS from TD controls. We acquired resting-state functional magnetic resonance imaging data in 44 PWS and 50 TD controls. We then used Independent Component Analysis (ICA) together with Hierarchical Partner Matching (HPM) to identify networks of robust, functionally connected brain regions that were highly reproducible across participants, and we assessed whether connectivity differed significantly across diagnostic groups. We then used Granger Causality (GC) to study the causal interactions (effective connectivity) between the regions that ICA and HPM identified. Finally, we used a kernel support vector machine to assess how well these measures of functional connectivity and granger causality discriminate PWS from TD controls. Functional connectivity was stronger in PWS compared with TD controls in the supplementary motor area (SMA) and primary motor cortices, but weaker in inferior frontal cortex (IFG, Broca’s area), caudate, putamen, and thalamus. Additionally, causal influences were significantly weaker in PWS from the IFG to SMA, and from the basal ganglia to IFG through the thalamus, compared to TD controls. ICA and GC indices together yielded an accuracy of 92.7% in classifying PWS from TD controls. Our findings suggest the presence of dysfunctional circuits that support speech planning and timing cues for the initiation and execution of motor sequences in PWS. Our high accuracy of classification further suggests that these aberrant brain features may serve as robust biomarkers for PWS. PLOS ONE | https://doi.org/10.1371/journal.pone.0179255 July 31, 2017 1 / 17 a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Qiao J, Wang Z, Zhao G, Huo Y, Herder CL, Sikora CO, et al. (2017) Functional neural circuits that underlie developmental stuttering. PLoS ONE 12(7): e0179255. https://doi.org/ 10.1371/journal.pone.0179255 Editor: J. Bruce Morton, Western University, CANADA Received: September 26, 2016 Accepted: May 28, 2017 Published: July 31, 2017 Copyright: © 2017 Qiao et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: All relevant data are within the paper and its Supporting Information files. Funding: This work was supported by the McGue Millhiser Family Trust, National Natural Science Foundation of China (61603225)(http://www.nsfc. gov.cn/), Natural Science Foundation of Shandong Province (ZR2016FQ04), and China Postdoctoral Science Foundation (2016M602182). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Introduction Developmental stuttering is a speech disorder in which sounds, syllables, or words are repeated or prolonged, disrupting the normal flow of speech [1–2]. These speech disruptions may be accompanied by behaviors representing effortful motor control, such as rapid eye blinking or lip tremor. Stuttering affects approximately 5 percent of all children at some time in their life, most often between 2 and 5 years of age, when speech and language skills are developing rapidly. Symptoms typically last from several weeks to several years, with adult persistence noted in less than 25% of those affected [3]. Stuttering therefore affects people of all ages, and it often impairs overall quality of life [4]. Prior neuroimaging studies have revealed structural and functional differences in people who stutter (PWS) compared with fluent speakers. The preponderance of evidence currently suggests that the pathogenesis of stuttering involves neural systems involved in the production and control of speech, those within Broca’s region of the inferior frontal gyrus and associated portions of the arcuate fasciculus, as well as their interconnections with cortico-striato-thalamo-cortical (CSTC) loops [5–11]. Voxel-based morphometry studies using anatomical Magnetic Resonance Imaging (MRI) have reported atypical leftward asymmetry [8, 12–13] and increased white matter volumes in the superior temporal gyrus (STG), middle temporal gyrus (MTG), inferior frontal gyrus (IFG), middle frontal gyrus (MFG) and corpus callosum [14–16] of adults who stutter compared with controls. Moreover, children who stutter are reported to have less grey matter volume (GMV) in the IFG, caudate, and putamen [8, 17–19] but more GMV in the right rolandic operculum and STG relative to fluent children [17]. Diffusion Tensor Imaging (DTI) has shown that children who stutter have reduced fractional anisotropy in the region of the left superior longitudinal fasciculus and bilateral corticospinal tracts [18, 20]. Other DTI studies have also shown that adults who stutter have reduced fractional anisotropy (FA) in the left sensorimotor cortex [21–22]. Our recent spectroscopy study in the same participants of the present study reported reduced levels of N-acetyl aspartate (NAA), an index of neural density, in inferior and superior frontal cortices (including Broca’s region) and caudate nucleus, and increased NAA in the posterior cingulate, lateral parietal, and parahippocampal cortices, as well as the hippocampus and amygdala [5]. Functional imaging studies have suggested the presence of a dysfunctional feedforward CSTC network in language learning [23], with these functions extending to linguistic sequencing in a domain-specific manner [24]; because sequencing difficulties during spoken language are the defining hallmark of stuttering, these studies suggest the importance of CSTC circuits in the pathogenesis of developmental stuttering. Positron Emission Tomography [25–27] and functional Magnetic Resonance Imaging (fMRI) studies of stuttering have reported similar patterns of deactivation in the left IFG (Broca’s region), as well as hyperactivation in right frontal operculum, motor areas, basal ganglia, and thalamus in PWS, implicating disturbances in speech planning [28] and speech execution [29] during various speech [26, 30–32] and nonspeech tasks [33–34], even for PWS who are non-English speakers [35–36]. In our prior perfusion study of the same participants included in the present study, we detected lower perfusion at rest in PWS compared to fluent controls in Broca’s region bilaterally and in the superior frontal gyrus [10]. Perfusion in Broca’s area correlated inversely with the severity of stuttering and extended posteriorly into other portions of the arcuate fasciculus. Resting state fMRI studies in stuttering children and adults have reported reduced intrinsic functional connectivity in CTSC and auditory-motor cortical loops of the left hemisphere [37– 39], suggesting that the neural systems supporting complex and dynamic speech production is dysfunctional in PWS [6, 33]. These studies support the central role that Broca’s region within the IFG, the basal ganglia, and CSTC loops play in the pathogenesis of dysfluent speech. Functional circuits in developmental stuttering PLOS ONE | https://doi.org/10.1371/journal.pone.0179255 July 31, 2017 2 / 17 Competing interests: The authors have declared that no competing interests exist. Despite this progress from neuroimaging in elaborating the brain basis for stuttering, the neural systems that underlie developmental stuttering have not been fully identified. Analyses of functional connectivity using resting-state functional MRI data can be a powerful tool for identifying the intrinsic neural circuits that underlie pathogenic disturbances in neuropsychiatric illnesses. Therefore, we applied independent component analysis (ICA) with hierarchical partner matching (HPM) [40–42] to resting-state fMRI data, aiming to identify altered intrinsic connectivity within cortical-subcortical circuits in a large sample of PWS compared with age- and sex-matched, typically developing (TD) fluent speakers. ICA combined with HPM identifies robust networks of functionally connected brain regions that are reproducible across participants, and it can identify significant connectivity differences across diagnostic groups. We also used Granger Causality (GC) to study the causal interactions (effective connectivity) among the regions that ICA-HPM identified. One of the main goals of brain imaging and neuroscience is to uncover neural circuits that govern human behaviors and psychiatric disorders, and thereby to identify biomarkers that can robustly predict and identify these behaviors and disorders. Many studies have demonstrated that machine learning techniques are useful for finding potential biomarkers of psychological disorders [43–45]. However, to our knowledge, no studies have been reported that classify PWS and TD controls based on intrinsic brain connectivity. Therefore, we used machine learning to assess how well the ICA measures of functional connectivity and Granger Causality differentiate PWS from TD controls. Based on prior theoretical models and empirical studies of stuttering, we hypothesized that we would detect brain connectivity differences in PWS in Broca’s region and the associated language loop, and within the cortical-subcortical neural systems that support the components for phonetic encoding, sensory feedback, motor planning, and motor execution of human speech. We further hypothesized that measures of functional and effective connectivity in these systems would discriminate among PWS and TD controls with high accuracy, thereby providing a set of putative biomarkers for stuttering that can be useful for future studies of the pathophysiology and treatment of PWS. Materials and methods Participants We recruited 44 participants with developmental stuttering (PWS: 27 males, 17 females, mean age 24.6±11.2 years, range 9–49 years) via advertisements posted on the internet or at local clinics, hospitals, and stuttering support groups. We recruited participants across this wide age range intentionally, to permit assessment of the stability of group differences in imaging measures across the lifespan. All PWS participants, including the adults, were currently symptomatic. A licensed speech-language pathologist diagnosed all stuttering participants before enrollment. The Kiddie-Schedule for Affective Disorders and Schizophrenia, Present and Lifetime Version was administered to a parent for participants under age 18 [46], whereas the Structured Clinical Interview for DSM-IV-TR Axis I Disorders was administered to participants over 18 [47]. Stuttering participants with comorbid disorders (chronic motor tic disorder, N = 2; attention deficit hyperactivity disorder, N = 2; social anxiety disorder, N = 1) were allowed into the study because these are highly prevalent in stuttering speakers; excluding comorbidities would have impaired the generalizability of our findings. The Assessment of the Child’s Experience of Stuttering (ACES) for children who stuttered and the corresponding scale for adults, the Overall Assessment of the Speaker's Experience of Stuttering (OASES), were used to evaluate stuttering severity [48], which ranged from mild to severe, with an average severity of 50.8. Approximately 70% of PWS participants endorsed Functional circuits in developmental stuttering PLOS ONE | https://doi.org/10.1371/journal.pone.0179255 July 31, 2017 3 / 17 symptoms of moderate severity. We recruited 50 TD controls (31 males, 19 females, mean age 23.1±11.3 years, ranging from 8 to 49 years) randomly from a telemarketing list of 10,000 names in the local community, excluding those with lifetime Axis I or language disorder, group-matching to the PWS group on age and sex. The Institutional Review Board of the New York State Psychiatric Institute approved the study procedures. We obtained written informed consent from all participants and parental consent for participants under age 18. Image acquisition Participants were positioned in the scanner and head coil using canthomeatal landmarks. We instructed the participants to keep their eyes closed and not to think about anything in particular for the duration of the scan. Images were acquired using a 3 Telsa GE Signa EXCITE scanner (General Electric, Milwaukee, USA) equipped with an 8-channel head coil. Resting-state fMRI data were acquired using a gradient-echo echo-planar imaging (EPI) sequence (matrix 64×64, field of view 24cm×24cm, repetition time 2200msec, echo time 30msec, flip angle 90˚, slice thickness 3.5mm without gap). Thirty-four interleaved slices were acquired along the AC-PC plane to provide whole brain coverage, with 146 images acquired per participant, for a scan duration of 5 minutes 21 seconds. Data analysis Preprocessing. Functional MRI data were preprocessed using SPM8 (Welcome Department of Imaging Neuroscience, London, United Kingdom; see http://www.fil.ion.ucl.ac.uk/ spm/) that was run under MATLAB. The first six volumes of scans were excluded from analysis to allow for signal stabilization following onset transients. The remaining functional scans (140 volumes) were slice timing-corrected, realigned to the first scan using rigid-body transformations to correct for head movements. Realigned functional volumes were motionadjusted. The outlier volumes were identified based on the first-order derivative of the motion parameters between successive EPI images and global mean BOLD signal. We used linear interpolation to replace the identified outlier volumes with the closest non-outlier volumes in ArtRepair toolbox version 4 (http://spnl.stanford.edu/tools/ArtRepair) [49]. Finally, functional images were normalized to the Montreal Neurological Institute (MNI) coordinate system [50] and spatially smoothed using an isotropic 8 mm full-width at half-maximum Gaussian kernel. Independent component analysis with hierarchical partner matching. We elected to employ single-subject ICA in this study because, compared with group-based ICA [51], it better preserves the specificity and independence of independent components (ICs) identified within individual participants. Furthermore, ICA combined with HPM is able to identify networks of functionally connected brain regions that are reproducible across participants and that are highly robust with respect to the number of ICs generated for each person [40–41], a considerable advantage that group-based ICA by definition cannot provide. The combination of ICA with HPM, and the validation of the ICs they identify, are provided in detail elsewhere [40–41], but we briefly summarize them here. Determining rationally and accurately the number of components that should be extracted in any given dataset is notoriously difficult. We therefore combined individual-level ICA with HPM to ensure that we identified ICs that were robust with respect to the number of ICs generated in each participant. HPM first identifies ICs that are robust and reproducible across participants for a given number of extracted ICs, and then it identifies those ICs that are independent of the number of ICs extracted in each participant. Thus in the first HPM step we combined Minimum Description Length and Akaike’s Information Criterion to establish the Functional circuits in developmental stuttering PLOS ONE | https://doi.org/10.1371/journal.pone.0179255 July 31, 2017 4 / 17 lower and upper bounds (determined to be 20 and 130, respectively) for the number of ICs likely present in each participant. Next, starting from the lower bound of 20, we increased by increments of 10 the number of IC’s extracted in each person (i.e., 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, and 130 IC’s per set), yielding 12 sets of ICs for each participant. We used Principal Component Analysis (PCA) to do the data reduction prior to the ICA. We reduced the data X with a dimension of T×M to a dimension of N×M, where T represented the number of time points and N represented the number of independent components to be generated (N<