Generating Alzheimer's Narratives Using Large Language Models

Large Language Models (LLMs) have demonstrated strong potential in Alzheimer's Disease (AD) research, particularly through their ability to analyze semi-spontaneous speech. However, their clinical utility is limited by the scarcity of annotated data. This study investigates whether LLMs, including T5, GPT, LLaMA, and Mistral, can generate narrative descriptions that mimic speech from AD patients and healthy controls during the Cookie Theft Picture task. We introduce a framework with two configurations:(1) Human-to-Bot, where an LLM responds to real interviewer prompts, and (2) Bot-to-Bot, where LLMs simulate both interviewer and participant. Generated narratives are used to augment training data for AD classification. Results show that models like Mistral and LLaMA produce high-quality, semantically coherent responses, leading to improved F1 scores when real and synthetic data are combined. Human evaluations confirm their fluency and clinical realism. This work offers a scalable approach for data augmentation in cognitive assessments and highlights future directions for improving synthetic dialogue quality and diagnostic reliability.