Using Intelligent Personal Annotations to Improve Human Activity Recognition for Movements in Natural Environments

Area of Science:

• Human-computer interaction
• Machine learning for health monitoring
• Wearable computing

Background:

• Accurate health tracking algorithms are crucial for understanding lifestyle choices in natural environments (NE).
• Training these algorithms requires annotated data from NE, which is burdensome to collect.
• Algorithms trained in labs often perform poorly in real-world NE settings.

Purpose of the Study:

• To develop an intelligent scoring algorithm that minimizes user annotation requests for health tracking.
• To enhance tracking algorithm performance using context data from nearable sensors.
• To introduce a hierarchical annotation approach for fine-grained activity differentiation.

Main Methods:

• An intelligent scoring algorithm was developed to automatically annotate high-confidence predictions, reducing user input.
• Nearable sensor context data was integrated to improve the tracking algorithm's accuracy and confidence.
• A hierarchical annotation strategy was implemented for detailed activity differentiation.

Main Results:

• Accuracy in natural environments (NE) improved by 8% when limiting annotation requests to 20 per day.
• The F1-score for activity recognition improved by 11% using hierarchical annotations.
• The approach was validated in a diet monitoring case study, demonstrating practical application.

Conclusions:

• The proposed intelligent scoring and hierarchical annotation methods significantly improve health tracking accuracy in natural environments.
• Integrating context data from nearable sensors enhances algorithm performance and confidence.
• This approach offers a more efficient and effective way to collect annotated data for training health monitoring algorithms.