Tracking priming-induced language recovery in aphasia with pre-trained language models

Area of Science:

• Computational linguistics
• Neurorehabilitation
• Artificial Intelligence in Medicine

Background:

• Aphasia, a language disorder post-stroke, presents challenges in tracking recovery.
• Current methods for assessing language recovery in aphasia can be subjective and time-consuming.
• Pre-trained language models (PLMs) offer novel computational approaches for linguistic analysis.

Purpose of the Study:

• To evaluate the utility of PLM-derived surprisals as a measure of language change in aphasia following priming treatment.
• To develop and assess a prompting-based pipeline using PLMs for automated clinical classification tasks in aphasia.
• To explore the potential of Generative AI (GenAI) in clinical rehabilitation settings for language recovery.

Main Methods:

• Utilized PLM-derived surprisal scores (negative log-probabilities) to quantify language complexity and change.
• Analyzed surprisal score changes in participants receiving structural priming treatment for aphasia.
• Developed and validated a prompting-based pipeline for classifying sentence correctness and error types in aphasia.

Main Results:

• Surprisal scores significantly decreased after structural priming treatment, indicating language improvement.
• The decrease in surprisal was more pronounced in participants with severe sentence production impairments.
• The prompting-based pipeline achieved high performance in classifying aphasia sentence correctness (F1=0.967) and error detection (accuracy=0.846).

Conclusions:

• PLM-derived surprisals provide a continuous and interpretable measure for tracking treatment-induced language recovery in aphasia.
• PLMs demonstrate significant potential for automated clinical classification and modeling of language rehabilitation.
• This study highlights a practical application of GenAI in clinical settings for personalized aphasia therapy and progress monitoring.