Exploring the Impact of the NULL Class on In-the-Wild Human Activity Recognition.

Area of Science:

• Machine Learning
• Human-Computer Interaction
• Gerontology

Background:

• Monitoring activities of daily living (ADLs) is crucial for assessing and responding to an individual's basic physical needs.
• Existing ADL recognition systems struggle with naturalistic, infrequently occurring activities and imbalanced datasets common in real-world environments.
• Current methods often focus on controlled activities or balanced datasets, limiting applicability to authentic settings.

Purpose of the Study:

• Investigate the challenges of applying machine learning to imbalanced, in-the-wild datasets for ADL monitoring.
• Develop and evaluate techniques to improve the performance of ADL recognition systems in realistic scenarios.
• Enhance the reliability and deployment readiness of machine learning models for ADL monitoring.

Main Methods:

• Utilized a fully in-the-wild dataset containing naturalistic activities with infrequent occurrences.
• Applied a combination of preprocessing techniques to enhance recall and postprocessing techniques to improve precision.
• Conducted a user-independent evaluation to assess model performance on diverse, real-world data.

Main Results:

• The developed approach significantly improved ADL recognition accuracy on imbalanced, in-the-wild data.
• Achieved an event-based F1-score exceeding 0.9 for activities like brushing teeth, combing hair, walking, and washing hands.
• Demonstrated the effectiveness of combined preprocessing and postprocessing for robust ADL monitoring.

Conclusions:

• Preprocessing and postprocessing techniques are vital for creating effective machine learning models for ADL monitoring in real-world settings.
• Addressing data imbalance and activity infrequency is key to deploying reliable ADL recognition systems.
• This research tackles fundamental machine learning challenges for practical application in healthcare and assisted living.