Cog-TiPRO: Iterative Prompt Refinement with LLMs to Detect Cognitive Decline via Longitudinal Voice Assistant...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Speech Technology

Background:

• Early detection of cognitive decline is vital for managing neurodegenerative diseases.
• Current diagnostic methods are often labor-intensive and unsuitable for frequent monitoring.
• Voice assistant systems (VAS) present a potential non-invasive tool for continuous health assessment.

Purpose of the Study:

• To investigate the efficacy of VAS in detecting cognitive decline through longitudinal speech analysis.
• To develop and validate a novel framework (Cog-TiPRO) for analyzing short, unstructured voice commands.
• To identify specific linguistic and acoustic speech features indicative of cognitive decline.

Main Methods:

• Collected 18-month voice command data from 35 older adults, including daily at-home VAS interactions.
• Developed Cog-TiPRO: a framework integrating LLM-driven linguistic feature extraction, HuBERT-based acoustic feature extraction, and transformer-based temporal modeling (iTransformer).
• Utilized iterative prompt refinement for enhanced feature extraction from noisy, short voice commands.

Main Results:

• The Cog-TiPRO framework achieved 73.80% accuracy and 72.67% F1-score in detecting mild cognitive impairment (MCI).
• Our approach significantly outperformed baseline methods by 27.13%.
• Identified unique linguistic patterns in everyday voice commands associated with cognitive decline.

Conclusions:

• VAS, analyzed with advanced AI frameworks like Cog-TiPRO, show promise for non-invasive, frequent cognitive decline monitoring.
• The Cog-TiPRO framework effectively extracts relevant features from challenging voice data for MCI detection.
• LLM-driven analysis of speech patterns can reveal subtle linguistic markers of cognitive impairment.