Advancing Alzheimer Disease Prediction With Large Language Model-Based Linguistic Feature Analysis: Development and...

Area of Science:

• Artificial Intelligence in Medicine
• Computational Linguistics
• Neurodegenerative Disease Diagnostics

Background:

• Alzheimer disease (AD) is a growing global health concern, necessitating early detection for effective intervention.
• Current diagnostic methods for AD are often invasive and costly.
• Speech analysis offers a noninvasive alternative, as AD impacts linguistic abilities.

Purpose of the Study:

• To investigate the efficacy of linguistic features extracted by large language models (LLMs) for Alzheimer disease (AD) prediction.
• To enhance the accuracy and clinical interpretability of automated AD detection using speech data.
• To develop a transparent and clinically applicable AI framework for early AD identification.

Main Methods:

• Proposed a framework leveraging LLMs to analyze linguistic features (readability, fluency, detail richness, keyword relevance) from transcribed speech for AD classification.
• Integrated transcript and feature explanation embeddings to improve classification accuracy.
• Conducted ablation studies, benchmarked against existing LLM methods, and assessed output stability and privacy-preserving deployment feasibility (Llama 3 8B + nomic-embed-text).

Main Results:

• Achieved high performance on the ADReSSo 2021 dataset (91.52% precision, 91.08% sensitivity, 96.29% specificity, 91.05% F1-score) across three runs.
• Demonstrated framework transferability to privacy-preserving environments with a fully local configuration achieving an 81.58% F1-score.
• Keyword relevance emerged as the most influential feature; LLM-based explainability favored the proposed method over a benchmark (49/54 cases).

Conclusions:

• Structured linguistic feature analysis using LLMs provides a robust and interpretable framework for preliminary Alzheimer disease (AD) detection.
• The developed approach bridges AI-driven text analysis with clinical applications, supporting early cognitive decline detection.
• Offers a scalable and accessible noninvasive method for early AD identification through speech assessment.