Dementia Detection by In-Text Pause Encoding.

Area of Science:

• Computational linguistics
• Neuroscience
• Artificial intelligence

Background:

• Dementia, especially Alzheimer's Disease (AD), impairs communication, particularly vocabulary and pragmatics.
• Individuals with AD often use vague language, lacking specific nouns and verbs.
• While sentence structure may remain intact initially, this is debated; LLMs offer new analytical tools.

Purpose of the Study:

• To investigate novel in-text encoding strategies for enhancing Large Language Model (LLM) performance in analyzing dementia-related communication changes.
• To address the challenge of limited datasets in dementia detection for LLMs.
• To quantify the impact of special character embedding and frequency analysis on model accuracy.

Main Methods:

• Developed and tested a novel approach using LLMs to analyze textual data from individuals with dementia.
• Implemented in-text encoding strategies, embedding special characters and incorporating sequence/frequency analysis.
• Compared the performance of the novel approach against a baseline model.

Main Results:

• The best-performing model achieved an f1-score of 0.88 and an accuracy of 0.86.
• The baseline model achieved significantly lower scores, with an f1-score of 0.42 and an accuracy of 0.56.
• The proposed encoding strategies demonstrated a substantial improvement in model performance.

Conclusions:

• Novel in-text encoding strategies, including special character embedding and frequency analysis, significantly enhance LLM performance for dementia communication analysis.
• This approach offers a promising method to overcome dataset limitations in training LLMs for AD research.
• The findings highlight the potential of advanced NLP techniques in understanding and potentially detecting communication changes in dementia.