GRM-038 Optimizing Prompts for Alzheimer's Speech Classification Using LLM

Large Language Models (LLMs) are widely used in Alzheimer's disease research to classify speech patterns. However, there is no standardized framework to ensure the reliability of prompts used in these classifications. This study investigates the sensitivity of Alzheimer’s disease classification prompts to small variations and finds that these prompts are indeed sensitive, leading to inconsistencies in model performance. To address this, we implement an automatic prompt optimization framework to refine the base prompt. Experimental results demonstrate that the optimized prompt improves classification accuracy by 12.83% compared to the baseline, underscoring the significance of systematic prompt engineering in enhancing the reliability of LLM-based Alzheimer’s disease detection. Although the optimized prompt remained sensitive to variations, it consistently showed improved overall accuracy.