Generating Alzheimer's narratives using large language models.

Area of Science:

• Computational linguistics
• Artificial intelligence in healthcare
• Neuroscience

Background:

• Semi-spontaneous speech analysis shows promise for Alzheimer's disease (AD) assessment.
• Clinical data scarcity currently limits progress in AD assessment using speech.
• Large Language Models (LLMs) offer a novel approach to generate synthetic speech data.

Purpose of the Study:

• To evaluate the ability of various LLMs (GPT, T5/Flan-T5, LLaMA, Mistral, Qwen) to generate clinically plausible narratives for AD assessment.
• To assess synthetic narrative quality using automated metrics and human expert ratings.
• To determine the impact of LLM-generated data on AD classification model performance.

Main Methods:

• LLMs were fine-tuned on the DementiaBank Pitt Corpus.
• Two configurations were tested: Human-to-Bot and Bot-to-Bot interactions.
• Generated narratives were evaluated using lexical/semantic metrics, human ratings, and for augmenting a BERT-based AD classifier.

Main Results:

• Mistral, LLaMA, and Qwen demonstrated superior performance in generating fluent, plausible, and diagnostically informative narratives.
• LLM-generated data augmented classifier training achieved a higher F1-score (0.84) compared to real data alone (0.74).
• Human-to-Bot synthetic data yielded the most significant diagnostic improvements, while Bot-to-Bot data showed less clinical realism.

Conclusions:

• LLMs can generate high-quality synthetic narratives for cognitive assessment, enhancing AD classification.
• LLM-generated data offers a scalable solution to data scarcity in dementia research.
• Future research should focus on improving synthetic dialogue quality and evaluation frameworks for clinical relevance.