Related Experiment Video
Updated: Jun 28, 2025
06:37
Author Spotlight: Addressing Technical and Subjective Challenges in Measuring Classroom Attention
Published on: December 15, 2023
3.7K
Self-supervised learning for human activity recognition using 700,000 person-days of wearable data
Hang Yuan1,2,3, Shing Chan1,2, Andrew P Creagh2,4
1Nuffield Department of Population Health, University of Oxford, Oxford, UK.
NPJ Digital Medicine
|April 12, 2024
Summary
This study uses self-supervised learning on a large UK Biobank dataset to improve human activity recognition models. The new models show better accuracy and generalizability across diverse conditions, aiding health research.
Area of Science:
- Biomedical Engineering
- Computer Science
- Public Health
Background:
- Accurate physical activity monitoring is crucial for understanding health impacts.
- Current human activity recognition algorithms are limited by small labeled datasets.
Purpose of the Study:
- To develop highly generalizable and accurate human activity recognition models.
- To leverage self-supervised learning on the UK Biobank accelerometer dataset.
Main Methods:
- Utilized self-supervised learning techniques.
- Trained models on the 700,000 person-days UK Biobank unlabelled accelerometer data.
Main Results:
- Models demonstrated significant performance improvements over baselines (2.5-130.9% F1 relative improvement).
- Achieved superior generalizability across external datasets, cohorts, environments, and sensor devices.
Conclusions:
- Self-supervised learning effectively addresses data limitations in human activity recognition.
- Open-sourced models offer valuable tools for research with scarce labeled data or diverse sampling needs.
