Conceptual

LLM Prompt-Tuning with LoRA for Speech-Based Alzheimer's Detection

A non-invasive Alzheimer's-disease detection method that classifies speech-derived text with a pretrained large language model (LLaMA2) adapted via a prompt fine-tuning / prompt-based learning strategy and conditional learning. To work under limited compute, it uses parameter-efficient LoRA (low-rank adaptation) instead of full fine-tuning. Under 10-fold cross-validation it attains 81.31% accuracy, a 4.46% absolute gain over a BERT baseline, and is proposed as a route to accessible at-home smartphone screening for dementia.