Related Experiment Video
Updated: Jan 13, 2026
07:24
Home-Based Monitor for Gait and Activity Analysis
Published on: August 8, 2019
7.2K
Machine and Deep Learning for Detection of Moderate-to-Vigorous Physical Activity From Accelerometer Data: Systematic
Yahua Zi1, Sjors Rb van de Ven2, Eco Jc de Geus2
1School of Exercise and Health, Shanghai University of Sport, Shanghai, China.
Interactive Journal of Medical Research
|January 8, 2026
Summary
Machine learning (ML) and deep learning (DL) show promise for accurately estimating moderate-to-vigorous physical activity (MVPA) using accelerometers. While effective in labs, real-world performance varies, highlighting needs for better generalizability and open science practices in physical activity research.
Area of Science:
- Wearable technology and sensor data analysis
- Biomedical engineering and public health research
- Artificial intelligence in health and fitness
Background:
- Accurate monitoring of moderate-to-vigorous physical activity (MVPA) is crucial for public health and personalized interventions.
- Traditional accelerometry methods struggle with accuracy and generalizability in free-living conditions.
- Machine learning (ML) and deep learning (DL) offer advanced automated MVPA detection capabilities.
Purpose of the Study:
- To conduct a scoping review of ML and DL techniques for MVPA estimation using accelerometer data.
- To analyze the performance, bias, sensor configurations, and translational potential of these advanced methods.
- To synthesize current evidence on AI-driven physical activity assessment.
Main Methods:
- Systematic literature search following PRISMA-ScR guidelines across major scientific databases (PubMed, IEEE Xplore, Web of Science).
- Screening of titles, abstracts, and full texts by two independent reviewers.
- Data extraction and narrative synthesis guided by predefined research questions, with rigorous author review.
Main Results:
- 40 studies met inclusion criteria; traditional ML models showed high lab performance but declined in real-world settings.
- Deep learning (DL) architectures demonstrated robust free-living performance, with hybrid models achieving state-of-the-art results.
- Wrist-worn sensors were common, but multi-sensor configurations (e.g., wrist + hip) showed higher accuracy; algorithmic bias and lack of data sharing were key challenges.
Conclusions:
- ML and DL significantly improve MVPA monitoring by automating feature extraction and adapting to real-world variability.
- Gaps in generalizability, validation consistency, and transparency impede the translation of these technologies.
- Future research should focus on inclusive training, standardized reporting, and open science to ensure equitable AI in physical activity assessment.
