149 Dynamical Network State Sequences for Human Language Production

Abstract

INTRODUCTION: Speech requires the selection of a conceptual representation, the construction of a word form, and the execution of a complex articulatory plan. Investigating this global system requires high-resolution recordings with an analytic approach to resolve discrete cognitive states. We integrated autoregressive hidden Markov models to resolve trial-by-trial state transition sequences in distributed networks derived from a large-scale electrocorticographic dataset with complete coverage of language-dominant cortex. METHODS: Intracranial electrodes (n = 25810, 134 patients), including both surface grids and stereotactic depths, were implanted for the evaluation of epilepsy. Patients performed picture naming of common objects and underwent systematic cortical stimulation mapping of language function. We used cross-validated autoregressive hidden Markov models – combining the interpretability of multivariate autoregressive analysis with the nonlinear embedding of the switching Markov characteristic – to distinguish cognitive states defined by causal interactional motifs of distributed cortical substrates and to simulate in silico lesions of dynamical network structures. RESULTS: We created a detailed spatiotemporal atlas of word production spanning the entire language-dominant cortex. From this map, we resolved network dynamics comprising sequences of state space transitions for each word articulated. This identified five essential states for speech production, each defined by a unique pattern of directed interactions within the language network. We then derived a computational lesion model for state dynamics and compared its predictions with causal perturbation by direct cortical stimulation. We assembled a global dynamical state model for language production and resolved the fine-scale interregional dynamics of conceptualization and lexical access. CONCLUSIONS: The distributed network dynamics detailed within this comprehensive interactional map of speech production advance our understanding of how both local and disconnection injuries yield complex neurological deficits, facilitating the development of novel therapeutic approaches for aphasia.