Exploring LLM Embedding Potential for Dementia Detection Using Audio Transcripts

Abstract

Dementia is a neurodegenerative disorder characterized by progressive cognitive impairment that significantly affects daily living. Early detection of Alzheimer’s disease—the most common form of dementia—remains essential for prompt intervention and treatment, yet clinical diagnosis often requires extensive and resource-intensive procedures. This article explores the effectiveness of automated Natural Language Processing (NLP) methods for identifying Alzheimer’s indicators from audio transcriptions of the Cookie Theft picture description task in the PittCorpus dementia database. Five NLP approaches were compared: a classical Tf–Idf statistical representation and embeddings derived from large language models (GloVe, BERT, Gemma-2B, and Linq-Embed-Mistral), each integrated with a logistic regression classifier. Transcriptions were carefully preprocessed to preserve linguistically relevant features such as repetitions, self-corrections, and pauses. To compare the performance of the five approaches, a stratified 5-fold cross-validation was conducted; the best results were obtained with BERT embeddings (84.73% accuracy) closely followed by the simpler Tf–Idf approach (83.73% accuracy) and the state-of-the-art model Linq-Embed-Mistral (83.54% accuracy), while Gemma-2B and GloVe embeddings yielded slightly lower performances (80.91% and 78.11% accuracy, respectively). Contrary to initial expectations—that richer semantic and contextual embeddings would substantially outperform simpler frequency-based methods—the competitive accuracy of Tf–Idf suggests that the choice and frequency of the words used might be more important than semantic or contextual information in Alzheimer’s detection. This work represents an effort toward implementing user-friendly software capable of offering an initial indicator of Alzheimer’s risk, potentially reducing the need for an in-person clinical visit.