A Feature-Augmented Transformer Model to Recognize Functional Activities From in-the-Wild Smartwatch Data.

Area of Science:

• Human-Computer Interaction
• Wearable Computing
• Digital Health

Background:

• Traditional human activity recognition (HAR) focuses on atomic movements.
• Recognizing functional activities is crucial for healthcare but complex in real-world settings.
• Existing methods struggle with the variability and data sparsity of in-the-wild functional HAR.

Purpose of the Study:

• To investigate methods for functional HAR.
• To introduce a novel approach augmenting feature representations with feature token-transformer embeddings.
• To improve classification performance for complex, goal-directed behaviors.

Main Methods:

• Compared various machine learning and deep learning methods for functional HAR.
• Developed a novel approach using feature token-transformer embeddings.
• Utilized the large-scale ArWISE dataset (n=503, 32M+ labeled points) for longitudinal collection.

Main Results:

• Feature embeddings significantly improved functional HAR model performance.
• The proposed method demonstrated effectiveness in handling real-world variability and data sparsity.
• Experiments showed generalization capabilities across a diverse population.

Conclusions:

• Incorporating feature embeddings is advantageous for functional HAR.
• This work bridges the gap between atomic movement and functional behavior recognition.
• The findings support advanced, behavior-aware applications in digital health and human-centered AI.