Developing Predictive Models with Large Language Models for Enhancing Language Learning and Recovery in Stroke-induced Aphasia Patients

Language learning and recovery involve dynamic interactions among multiple brain regions, with previous studies supporting potential similarities in their neural activity. This study proposes a novel approach to understanding the neural mechanisms underlyinglanguage learning and recovery by linking brain neural representations with embeddings from large language models (LLMs) such as GPT-2 and BERT. It will involve healthy participants engaging in a structured artificial language learning task over 7 days andstroke patients undergoing similar learning with real native language materials appropriate to their impairment levels. Task-based fMRI data will be collected at multiple time points to observe neural dynamics over the course of learning and recovery. By analyzing neuralactivation patterns in multiple brain regions defined by standardized brain atlases and comparing them with semantic representations generated by large language models (LLMs), we aim to identify regions where neural activity closely correlates with LLMscomputational outputs. Subsequently, we will train predictive regression models that map word embeddings to neural activity based on these regions and apply them to compare the neural activity of stroke patients at different time points. We hypothesize that strokepatients whose early neural activity resembles the predicted neural activity of healthy controls at later learning point will demonstrate better language recovery. Findings from this study could inform clinical interventions by identifying neural biomarkers predictive ofrehabilitation outcomes, enabling tailored therapeutic strategies to improve language learning and recovery in stroke patients.