Why machine learning (ML) has failed physical activity research and how we can improve.

Area of Science:

• Biomechanics
• Computer Science
• Health Sciences

Background:

• Accurate measurement of physical activity is essential for understanding human movement and health benefits.
• Machine learning (ML), particularly using accelerometer data, is increasingly utilized for physical activity assessment.

Purpose of the Study:

• To identify critical limitations hindering the effective application and adoption of machine learning in physical activity measurement research.
• To propose areas for improvement to advance the field of physical activity measurement using ML.

Main Methods:

• The study critically reviews the current state of ML in physical activity research, highlighting four key areas of deficiency.
• Analysis focuses on the adoption of computer science principles, prioritization of methods, software integration, and training.

Main Results:

• Physical activity research has not fully adopted computer science principles like benchmark datasets, impeding direct comparison of ML approaches.
• Overemphasis on ML methods has created blind spots, potentially overlooking superior alternative measurement techniques.
• ML methods are rarely integrated into usable software, limiting their practical application by researchers.
• Insufficient training and accessible software hinder the widespread adoption and application of ML in physical activity research.

Conclusions:

• Advancing ML for physical activity measurement requires adopting computer science standards, such as benchmark datasets.
• Prioritizing diverse measurement methods beyond ML and integrating ML into accessible software are essential.
• Enhanced training and software development are critical for the broader uptake of ML in applied physical activity research.