Wearable-based estimation of continuous 3D knee moments during running using a convolutional neural network

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning

Background:

• Estimating 3D knee moments during running is crucial for understanding injury mechanisms.
• Previous methods often require extensive laboratory setups and complex data processing.

Purpose of the Study:

• To develop and validate a machine learning (ML) method for estimating continuous 3D knee moments during running using wearable sensor data.
• To assess the accuracy and reliability of this ML method across different running conditions and body segments.

Main Methods:

• A convolutional neural network (CNN) was trained using data from 7 inertial measuring units (IMUs) and 2 pressure insoles.
• Data were collected from 19 recreational runners on a treadmill under varying slopes, speeds, and footwear conditions.
• Performance was evaluated using intraclass correlation (ICC) and normalized root mean squared error (nRMSE) over continuous time windows and stance phases.

Main Results:

• The ML model demonstrated good to excellent agreement (ICC: 0.84-0.98) and low error (nRMSE: 0.05-0.11) for sagittal plane knee moments.
• Accuracy was lower for non-sagittal planes (frontal: ICC 0.19-0.90; transverse: ICC 0.72-0.94).
• Accuracy decreased during stance phases (PHSS) compared to continuous analysis (CONT).

Conclusions:

• The developed ML approach provides a viable, near real-time method for wearable-based running kinetics assessment.
• This method shows comparable or superior accuracy to existing techniques with less preprocessing.
• Further improvements in addressing inter-individual variability could enhance precision for assessing frontal plane injury risk factors.