Development and Validation of Accelerometer-Based Machine Learning Models for Classifying Walking, Running, and...

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning

Background:

• Accurate quantification of mechanical loading is crucial for bone health.
• Wearable accelerometers offer potential for free-living load estimation.
• Current methods require prior activity identification, limiting real-world application.

Purpose of the Study:

• To develop and validate machine learning models for automatic activity classification using accelerometer data.
• To differentiate between walking, running, and jumping without prior knowledge of the activity.
• To establish a foundation for activity-specific mechanical loading predictions in free-living settings.

Main Methods:

• Forty-eight healthy adults performed walking, running, and jumping tasks.
• ActiGraph GT9X Link accelerometers were worn at the ankle, lower back, and hip.
• Random Forest, Support Vector Machine, and K-Nearest Neighbors algorithms were trained and evaluated.

Main Results:

• All machine learning models demonstrated excellent classification accuracy.
• Percent agreement ranged from 93.8% to 97.7%.
• Receiver operating characteristic area under the curve values exceeded 0.97, and Kappa coefficients surpassed 0.89.

Conclusions:

• Accelerometer-based activity classification reliably distinguishes walking, running, and jumping.
• This provides a practical framework for applying mechanical loading prediction equations.
• The developed models support objective assessment of physical activity for bone health research.