A Multi-Agent LLM Network for Suggesting and Correcting Human Activity and Posture Annotations.

Area of Science:

• Computer Science
• Artificial Intelligence
• Health Informatics

Background:

• Accurate human activity recognition (HAR) is vital for health monitoring and behavior-aware systems.
• Collecting large, labeled datasets for HAR in free-living settings is challenging due to reliance on self-reports prone to errors.
• Context-assisted recall aids memory but requires burdensome data triangulation.

Purpose of the Study:

• To adapt the GLOSS multi-agent LLM system for assisting human activity recall and annotation.
• To evaluate the effectiveness of LLM-generated activity suggestions in improving data quality for HAR studies.

Main Methods:

• Adapted GLOSS, a multi-agent LLM system, to triangulate self-reports and passive sensing data.
• Used GLOSS to suggest the most likely activities, assisting participants in recall and annotation.
• Evaluated the agreement between GLOSS suggestions and human recall.

Main Results:

• GLOSS provided activity suggestions with 63-75% agreement with human recall.
• The system effectively identified and corrected common human annotation errors.
• Demonstrated potential for LLM-powered, human-in-the-loop approaches in HAR.

Conclusions:

• LLM-powered human-in-the-loop systems can enhance the quality and scalability of activity annotation.
• GLOSS shows promise for improving HAR dataset creation in real-world studies.
• This approach offers a practical solution to overcome limitations of traditional self-reporting methods.