Decoding the Neural Basis of Postictal Aphasia: A Case Study in Contextualizing Human Language Processing with Artificial Neural Networks

Abstract

Postictal aphasia is a language disorder that commonly results after seizures localized in the left-hemisphere of the human brain, near language processing regions, thus disrupting the neural circuits at play during language cognition. Various types of postictal aphasia may occur, yet the neural basis of language dysfunction after seizures is vaguely understood. The current work presents a computational analysis that compares the encoding performance of two large language models (LLMs): GloVe and GPT-2, while simultaneously examining the fluctuations in the spatiotemporal nature of human language cognition surrounding word onset during naturalistic speech as well as before, during, and after an aphasic attack. By evaluating encoding model performance utilizing next-word prediction LLMs trained on spontaneous speech data, this study compares aphasic and non-aphasic speech while modulating context integration window size in hopes to advance the understanding of how competent deep neural networks (DNNs) are at encoding language processing in the human brain. Through utilizing high spatiotemporal resolution ECoG (intracranial electroencephalogram) data from the left- hemisphere of a single postoperative epilepsy patient experiencing language deficits consistent with conduction aphasia, the employment of an encoding model which utilizes GPT-2 and GloVe to generate word embeddings is able to roughly clarify where, when, and how the brain produces and comprehends language during aphasia as opposed to when the patient is engaged in non-aphasic speech. The results collected from this case study are limited in their extrapolative value due to small sampling sizes, but the theoretical concepts at play with respect to decoding language processing using next-word predictor LLMs motivates future research in this field of inquiry using more complex computational methodologies.