Estimation of ground reaction force waveforms during fixed pace running outside the laboratory

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning

Background:

• Wearable sensors and machine learning show promise for collecting biomechanical data.
• Current machine learning models have not reached their full potential in analyzing gait and kinetic waveforms.
• Accurate ground reaction force (GRF) estimation is crucial for understanding running mechanics.

Purpose of the Study:

• To propose and evaluate a Long Short Term Memory (LSTM) network for estimating GRF data from inertial measurement unit (IMU) data in a semi-uncontrolled running environment.
• To assess the accuracy of the LSTM model in identifying gait events and kinetic waveforms compared to force sensing insoles.

Main Methods:

• Fifteen healthy runners of varying experience and age participated.
• Three IMUs were placed on each participant's feet and sacrum.
• Force sensing insoles measured ground reaction forces as the ground truth.
• An LSTM network was trained to map IMU data to GRF data.

Main Results:

• The LSTM network estimated 4-second temporal windows of GRF data with Root Mean Square Error (RMSE) ranging from 0.189-0.288 BW.
• Estimation of foot contact achieved an R-squared value of 0.795.
• Estimation of peak force yielded the best results with an R-squared value of 0.614.

Conclusions:

• A Long Short Term Memory network can effectively estimate ground reaction force data during running using wearable inertial sensors.
• This approach is viable in semi-uncontrolled environments at controlled paces over level ground.
• The findings suggest a significant advancement in leveraging machine learning for non-invasive biomechanical analysis in runners.