Poster A58, Tuesday, August 20, 2019, 10:15 am – 12:00 pm, Restaurant Hall
Complete Cortical Dynamics of Single Word Articulation
Nitin Tandon1,2, Kiefer J Forseth1, Xaq Pitkow3,4;1McGovern Medical School, 2Memorial Hermann Hospital, 3Baylor College of Medicine, 4Rice University
Speech production involves an integrated multistage process that seamlessly translates conceptual representations in the brain to an articulatory plan. Evaluating this progression of cognitive operations requires high-resolution recordings combined with an analytic approach to model discrete neural states. We used a large-scale electrocorticographic dataset with complete coverage of cortical structures in the language dominant hemisphere to identify the timing of activation of regions relative to articulation and subsequently applied an Autoregressive Hidden Markov Model (ARHMMs) to resolve trial-by-trial state transition sequences in distributed networks. Intracranial electrodes (n=22311,129 patients), including either surface grid electrodes and penetrating stereotactic depth electrodes, were implanted for the evaluation of drug-resistant epilepsy. Patients performed picture naming of common objects from the Snodgrass and Vanderwart image set. A surface-based mixed-effects multilevel analysis of broadband gamma activity in the language-dominant hemisphere was used to identify loci with significant activity. This revealed 12 regions, listed here in order of activation: early visual cortex, fusiform gyrus, intraparietal sulcus, parahippocampal gyrus, supplementary motor area, pars triangularis, superior frontal sulcus, pars opercularis, subcentral gyrus, posterior insula, superior temporal gyrus, and posterior middle temporal gyrus. In a second analysis, we included additional psycholinguistic parameters in a mixed-effects model to capture broadband gamma power driven by complexity specific to visual, semantic (familiarity), lexical (frequency), and phonologic (weighted phonemic positional probability) domains. This was used to ground the observed network nodes in extant theories of speech production. To investigate the dynamic trial-by-trial evolution of neural states from stimulus presentation to articulation, we implemented an ARHMM. This framework combines the interpretability of multivariate autoregressive analysis with the sophistication of nonlinear analysis embedded in the switching Markov characteristic. Each latent network state was defined as a 3rd order autoregressive tensor and associated covariance matrix, encoding dynamics conserved across the entire patient population. Importantly, we present a major novel advance in stitching datasets across patients to solve a pernicious problem in electrocorticography: sparse cortical sampling in each individual patient. With this framework, we identify 5 distinct cortical states in picture naming (as well as a baseline state). These states broadly correspond visual recognition, conceptualization, formulation, articulation, and monitoring. A consistent, though not strictly serial, sequence of states was observed across trials; furthermore, the duration of the formulation and articulation states predicted reaction time (p<0.01). The latent dynamics of each state were quantified using pairwise partial directed coherence, revealing distinct networks driven by (Granger) causal outflow from a sparse subset of network nodes. Our work pairs large-scale electrocorticography with sophisticated analyses to answer long-standing questions about models of language production.
Themes: Language Production, Computational Approaches
Method: Electrophysiology (MEG/EEG/ECOG)