Leveraging Large Language Models for Identifying Interpretable Linguistic Markers and Enhancing Alzheimer’s Disease Diagnostics

Alzheimer’s Disease (AD) is a progressive irreversible neurodegenerative disorder. Early AD detection is crucial for timely intervention. This study proposes a novel LLM framework to identify interpretable linguistic markers from LLMs and incorporate them to supervised AD detection transformers, while evaluating corresponding model performance and interpretability. Our work carries three major novelties: First, we design in-context Few-Shot and Zero-Shot prompting strategies to facilitate LLMs in summarising the high-level linguistic markers discriminative of AD from Normal Control (NC), providing interpretation and assessment of their strength, reliability and relevance for AD classification. Second, we incorporate LLM-summarised linguistic markers into a smaller transformer model to enhance the performance of AD detection. Third, we investigate whether the LLM-summarised linguistic markers can enhance accuracy and interpretability of our downstream supervised transformer model when used with the original speech transcripts. Our findings have shown that using LLM-summarised linguistic markers solely may yield lower accuracy using supervised learning, while combining these markers with the features extracted from the original speech transcripts can enhance the AD detection model’s diagnostic capabilities. Our hybrid speech-based AD detection framework capitalizes both the interpretability of linguistic markers and the rich information embedded within the original transcripts. When used as complementary features, the distinguishable AD linguistic markers have enhanced the AD detection performance, while providing clinically meaningful insights into speech-based AD detection. These insights have enabled healthcare professionals to gain deeper understandings about what linguistic variations have dictated individuals suffering from AD, enabling the clinical professionals to make more informed interventions and healthcare decision-makings.