Predicting dementia from spontaneous speech using large language models

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Linguistics

Background:

• Language impairment is a key biomarker for neurodegenerative diseases like Alzheimer's disease (AD).
• Artificial intelligence (AI) and natural language processing (NLP) are increasingly used for early AD prediction via speech.
• Few studies have explored large language models (LLMs), such as GPT-3, for early dementia diagnosis.

Purpose of the Study:

• To investigate the efficacy of GPT-3 in predicting dementia from spontaneous speech.
• To utilize GPT-3's semantic knowledge for generating text embeddings from speech data.
• To assess the potential of GPT-3 text embeddings for early Alzheimer's disease diagnosis.

Main Methods:

• Transcribed spontaneous speech data was used to generate text embeddings via the GPT-3 model.
• Text embeddings were analyzed to distinguish individuals with AD from healthy controls.
• The ability of text embeddings to infer cognitive testing scores from speech data was evaluated.

Main Results:

• GPT-3 text embeddings reliably distinguished individuals with Alzheimer's disease from healthy controls based solely on speech.
• Text embeddings accurately inferred subjects' cognitive testing scores from speech data.
• This approach outperformed conventional acoustic feature-based methods and competed with fine-tuned models.

Conclusions:

• GPT-3 based text embedding is a viable method for assessing Alzheimer's disease directly from speech.
• This AI-driven technique holds significant potential for improving the early diagnosis of dementia.
• Leveraging LLMs for speech analysis offers a novel pathway for cognitive health assessment.