Electrocorticographic Representations of Semantic Information in the Human Cortex
Nicholas P. Szrama · Open Scholarship Institutional Repository (Washington University in St. Louis) · 2018
ABSTRACT OF THE DISSERTATION Electrocorticographic Representations of Semantic Information in the Human Cortex by Nicholas Paul Szrama Doctor of Philosophy in Biomedical Engineering Washington University in St. Louis, 2016 Professor Eric C. Leuthardt, Chair Semantics is a broad field of research that examines the literal meanings behind words, objects, concepts, and actions. Semantic memory is a critical component of our daily interactions with the world and it describes the cumulative knowledge that we acquire throughout our lifetime. Despite its importance, there is an incomplete understanding of how semantic information is organized, accessed, and retrieved in the brain. In this body of work, electrocorticographic (ECoG) potentials are recorded from five subjects while they perform a variant of the Deese, Roediger, and McDermott false memory experiment. By design, the auditory word stimuli used in this experiment were chosen to have strong semantic relationships with multiple non-presented semantic concepts. This design allowed the neural correlates to semantic information to be examined without an excessive repetition of identical word stimuli. In addition to these predefined semantic concepts, higher order semantic relationships across all sets of words were also studied using word2vec word embeddings. The DRM task was found to elicit strong high gamma ECoG power increases along the superior temporal gyrus during the auditory presentation of the stimuli. However, ECoG spectral power estimates did not distinguish individual semantic concepts or low-level semantic categories. Weak but significant correlations with semantic word embeddings particularly in the delta band were observed both in the within-electrode univariate ECoG spectral power patterns and the cross-electrode spatial ECoG power patterns. Additionally, repetition suppression patterns in ECoG power were unable to significantly discriminate different semantic concepts. Event-related potentials were also unable to distinguish individual semantic categories. In contrast to semantics, ECoG power at multiple frequencies were shown to reliably track the auditory envelope with a high correlation, statistically distinguish multiple vowels and consonants, and significantly (negatively) correlated with estimates of word imageability and word frequency. Taken together, these findings illustrate the degree of language-based information available within macroscale electrocorticographic recordings.