JMIR Publications

ABSTRACT

Background:

Alzheimer's disease (AD) affects 55 million people globally, projected to reach 139 million by 2050. Early detection is essential for timely intervention, yet traditional diagnostic methods face accessibility barriers due to high costs and invasive procedures. Language impairment is an early AD symptom that can be assessed non-invasively through speech analysis. While large language models (LLMs) show promise for analyzing speech transcripts, existing approaches lack transparency in evaluation processes and standardized frameworks for systematic linguistic analysis, limiting their clinical applicability.

Objective:

This study aims to investigate the influence of linguistic features extracted from transcribed speech, as analyzed by large language models (LLMs), on the accuracy and interpretability of Alzheimer’s disease (AD) prediction.

Methods:

We propose a framework that leverages LLMs to analyze linguistic features extracted from transcribed speech for AD classification. Our approach focuses on four key aspects: readability, fluency, richness of detail, and keyword relevance. To enhance classification accuracy, the framework integrates transcript embeddings with feature explanation embeddings, forming a comprehensive linguistic representation. We conducted extensive ablation studies to evaluate the contributions of individual features and benchmarked our framework against existing LLM-driven methodologies through pairwise explainability evaluations.

Results:

Our framework achieved a precision of 92.07%, a sensitivity of 91.55%, and a specificity of 97.22% on the ADReSSo 2021 dataset, outperforming the performance of existing state-of-the-art LLM-based approaches. Ablation studies identified keyword relevance as the most influential feature, while the integration of transcript and feature embeddings significantly improved predictive performance compared to using either in isolation. Furthermore, explainability evaluations through both LLM-based and human expert assessments demonstrated that our method outperformed benchmark approaches.

Conclusions:

These findings highlight that a structured linguistic feature analysis utilizing LLMs provides a robust and interpretable framework for preliminary AD detection. Our approach offers a scalable and accessible solution that bridges AI-driven text analysis with clinical applications, supporting early detection of cognitive decline through non-invasive assessment methods.