An Avatar-Based Framework Using Large Language Models For Assessing Patient Cognitive State

The goal of this project was to create an avatar-based framework that uses a large language model to communicate with patients in a natural conversational format for the purpose of doing a cognitive assessment. The framework includes a webpage featuring an avatar with computer-generated voice and a large language model to navigate a conversation through a series of standardized cognitive assessment questions while having the conversation feel natural to the patient. The two cognitive assessments include a mini-cog and mini-mental state exam. While generalized chat bots exist, determining if currently available large language models are adequate to perform these assessments and can redirect questions to complete an exam for patients that may have dementia would be useful for doing computer automated assessments on a more frequent basis without human interaction. Six large language models were compared and scored to determine which did best at performing this type of cognitive assessment. It was found that Chat GPT4 was best at performing cognitive assessments.