Machine learning for activity recognition: hip versus wrist data.

Area of Science:

• Physical Activity Recognition
• Biomedical Engineering
• Wearable Technology

Background:

• Wrist-worn accelerometers offer high compliance for activity monitoring.
• Validated algorithms for wrist-based activity recognition are limited.
• Comparison of hip vs. wrist accelerometer data for activity classification is needed.

Purpose of the Study:

• To compare the activity recognition accuracy of algorithms trained on wrist-worn versus hip-worn accelerometers.
• To evaluate the performance of activity classifiers in children and adolescents.
• To determine the feasibility of using wrist accelerometers for activity type prediction.

Main Methods:

• 52 children and adolescents completed 12 activity trials across 7 classes (lying, sitting, standing, walking, running, basketball, dancing).
• Tri-axial accelerometers (ActiGraph GT3X+) were worn simultaneously on the hip and wrist.
• A regularized logistic regression model (Glmnet + L1) was trained on features extracted from 10-second windows.

Main Results:

• Hip-worn accelerometers achieved 91.0% ± 3.1% classification accuracy, while wrist-worn accelerometers achieved 88.4% ± 3.0%.
• Both placements demonstrated excellent accuracy for sitting, standing, and walking.
• The hip model performed better for running and lying down, while the wrist model showed higher accuracy for sitting.

Conclusions:

• Both hip and wrist accelerometer placements can achieve acceptable classification accuracy for activity recognition.
• Researchers can consider either placement for activity monitoring, with wrist placement potentially offering better compliance.
• Further refinement of algorithms may improve accuracy for specific activities like dancing and running on wrist-worn devices.