Scholarlink

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder, emphasizing the critical need for early detection and intervention. Traditional clinical methods are often limited by their inefficiency and high costs. In contrast, automated AD screening systems utilizing speech analysis offer a scalable, non-invasive alternative. These systems typically depend on large annotated datasets to fine-tune pretrained language models for classification. However, the heterogeneity and complexity of the AD population, coupled with insufficient data from diverse groups, frequently lead to less effective detection across demographic lines. This study investigates the application of large language models in zero-shot and few-shot learning contexts for AD detection. Tailored prompt engineering techniques have been designed to address these challenges, with models such as GPT-3.5 and GPT-4 tested on the ADReSSo dataset. These models achieved an 83.3% accuracy rate, surpassing conventional, data-intensive methods. Additionally, a ‘thought chain’ framework was employed to enable stepwise analysis of AD symptoms, producing not only accurate but also interpretable results. The findings highlight that, with strategic prompt engineering, large language models can enhance AD detection using minimal data, offering promising prospects for future diagnostic advancements.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Interpretable Alzheimer’s Disease Detection with Minimal Data: Zero-Shot and Few-Shot Approaches Using Large Language Models

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Long Guo,
  • Hong Liu,
  • Hongyu Yang,
  • Hu Chen,
  • Wenchao Du,
  • Yu Liu

摘要

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder, emphasizing the critical need for early detection and intervention. Traditional clinical methods are often limited by their inefficiency and high costs. In contrast, automated AD screening systems utilizing speech analysis offer a scalable, non-invasive alternative. These systems typically depend on large annotated datasets to fine-tune pretrained language models for classification. However, the heterogeneity and complexity of the AD population, coupled with insufficient data from diverse groups, frequently lead to less effective detection across demographic lines. This study investigates the application of large language models in zero-shot and few-shot learning contexts for AD detection. Tailored prompt engineering techniques have been designed to address these challenges, with models such as GPT-3.5 and GPT-4 tested on the ADReSSo dataset. These models achieved an 83.3% accuracy rate, surpassing conventional, data-intensive methods. Additionally, a ‘thought chain’ framework was employed to enable stepwise analysis of AD symptoms, producing not only accurate but also interpretable results. The findings highlight that, with strategic prompt engineering, large language models can enhance AD detection using minimal data, offering promising prospects for future diagnostic advancements.