1437 - Large Language Model Generated Prompting as Speech Therapy Adjuvant for a Patient with Expressive Aphasia

Case Diagnosis: 65-year-old male with left frontal cerebrovascular accident

Case Description: A 65-year-old male with a history of gastric ulcer presented with right-sided weakness and difficulty speaking. The patient was non-verbal, with strength 0/5 in the right upper and lower extremities, 5/5 on the left. Brain MRI demonstrated acute left anterior cerebral artery and anterior cerebral artery-middle cerebral artery watershed territory infarcts. The patient received tenecteplase therapy without thrombectomy. The patient was admitted to acute inpatient rehab (AIR) seven days later. In speech therapy, he was noted to have moderate-severe expressive aphasia and echolalia with severely impaired speech initiation, while the language comprehension, repetition, and articulation remained intact.

Discussions: Word retrieval was tested using standard sets of sentence descriptions: five sets of ten target word starting with a different letter, five excluded for apparent difficulty. The patient correctly spoke the target word in 8/45 cases (18%). Of the remaining 37 words after being supplied with the cue from the Large Language Models (LLMs), the patient said ten words correctly. Subsequently, utility of LLM-prompts in simulated functional tasks was assessed by testing food, clothing, and furniture ordering tasks. The patient excelled in identifying food items and clothing even without LLM-prompts 30/35 correct; 4/5 remaining targets correct with LLM-prompt assist. Furniture: only 4/8 from picture, additional three with LLM-prompt. The sentence completion prompt was strongest when the target word was at the end of a phrase or sentence. Opposites (e.g., “Not a smile but a frown”) and antonyms were also strong cues. In addition, LLM-assist was helpful in simulated functional tasks.

Conclusions: LLMs may be helpful as adjuvants in speech therapy with patients who have expressive aphasia, and apps based on these models may have functional and social utility and be particularly helpful when speech therapists do not speak the patient’s language.