Adversarial Text Generation using Large Language Models for Dementia Detection.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Medical Informatics

Background:

• Large language models (LLMs) show promise in text classification but face challenges in specialized tasks like dementia detection from image descriptions.
• Standard prompting methods are insufficient due to the subtle linguistic markers of dementia and LLMs' difficulty in connecting internal knowledge to this specific diagnostic task.

Purpose of the Study:

• To develop an accurate and interpretable classification approach for dementia detection using image descriptions.
• To introduce a novel decoding strategy, Adversarial Text Generation (ATG), to enhance LLM performance in dementia detection.

Main Methods:

• Adversarial Text Generation (ATG) was employed as a novel decoding strategy to link dementia detection with other related tasks.
• A comprehensive set of task-specific instructions was developed and utilized to guide the ATG process.
• Feature context was introduced to provide human-understandable explanations for the LLM's classification decisions.

Main Results:

• The proposed ATG approach achieved a top accuracy of 85% in dementia detection.
• This represents a significant improvement of over 10% compared to conventional prompting strategies.
• Feature context analysis revealed that dementia detection correlates with assessing attention to detail, language, and clarity, linked to environmental and character-specific features.

Conclusions:

• Adversarial Text Generation (ATG) offers a more accurate and interpretable method for dementia detection using LLMs compared to standard prompting.
• The feature context provides valuable insights into the decision-making process of LLMs, highlighting key discriminative features for dementia.
• Future research will explore multi-modal LLMs to integrate both speech and visual information for enhanced dementia assessment.