Poster: Dementia Detection with a Lightweight Language Model Using DistilBERT and Linguistic Feature Fusion from Participant Speech

Abstract

The increasing number of affected people with neurodegenerative disorders, like Alzheimer’s disease, has made an urgent demand for cost-effective and accessible screening methods [1] . The Current gold standard diagnostic methods are considered impractical for widespread population screening due to their high costs, invasive nature, and requirement for specialized facilities [1] , [2] . Recent advancements in speech and language analysis provide a new opportunity for an alternative diagnostic method, which identifies cognitive decline through natural language processing methods and language models [3] . Linguistic digital biomarkers that measure lexical diversity, hesitation, and difficulty in retrieving words have demonstrated a strong connection with cognitive decline [1] , [4] . On the other hand, transformer-based language models have been proven effective in capturing linguistic features [2] , [5] . However, large language models have limitations in clinical settings, such as their high demand for computational power, lack of a substantial high-quality dataset, and interpretability challenges. To overcome these limitations, the lightweight transformer architectures were developed to reduce computational cost and time while keeping comparable performance [2] . This study aims to evaluate and compare the diagnostic performance of a lightweight transformer model with machine learning approaches for detection of mild cognitive impairment (MCI) from transcribed language samples. The model could be used for real-world clinical applications.