Abstract
This study explores the use of pre-trained language models (PLMs) in tracking priming treatment induced language recovery in aphasia. We evaluate PLM-derived surprisals, the negative log-probabilities of a word or a sequence of words calculated by a PLM given its preceding context, as a continuous and interpretable measure of treatment-induced language change. We found that surprisal scores decreased following structural priming treatment, especially in participants with more severe sentence production impairments. We also introduce a prompting-based pipeline for clinical classification tasks. It achieved promising results in classifying aphasia sentence correctness (F1 = 0.967) and detecting error categories in aphasia (accuracy = 0.846). Such use of PLMs for modeling, tracking, and automatically classifying language recovery in aphasia represents a promising deployment of GenAI in a clinical rehabilitation setting. Together, our PLM-based analyses offer a practical approach for modeling language rehabilitation, tracking not only language structure but also individual change over time in clinical contexts. CLINICAL TRIAL REGISTRATION: Identifier NTC05415501.