Linguistic changes in spontaneous speech for detecting Parkinson's disease using large language models

Area of Science:

• Neuroscience
• Computational Linguistics
• Artificial Intelligence

Background:

• Parkinson's disease (PD) is a prevalent neurodegenerative disorder with complex symptom presentation.
• Early diagnosis is challenging due to heterogeneous and often delayed motor symptoms.
• Linguistic impairments can precede motor symptoms, offering a potential diagnostic window.

Purpose of the Study:

• To evaluate state-of-the-art large language models (LLMs) for automatic Parkinson's disease detection from spontaneous speech.
• To assess the capability of LLMs in predicting Parkinson's disease severity.
• To understand the feature extraction mechanisms driving LLM performance in PD detection.

Main Methods:

• Application of advanced large language models to analyze spontaneous speech data.
• Utilizing LLMs for both binary classification (PD detection) and regression (severity prediction).
• Comparative analysis of LLM feature extraction against traditional methods.

Main Results:

• Achieved up to 78% accuracy in automatically detecting Parkinson's disease from speech.
• Demonstrated LLMs' effectiveness in predicting PD severity through regression analysis.
• Confirmed that enhanced performance stems from superior linguistic feature extraction, not just increased dimensionality.

Conclusions:

• Large language models show significant promise as a tool for early Parkinson's disease detection and monitoring.
• LLMs can identify subtle linguistic markers indicative of Parkinson's disease, even in early stages.
• The ability of LLMs to extract nuanced linguistic features is key to their diagnostic and prognostic potential in Parkinson's disease.