Predicting continuous ground reaction forces from accelerometers during uphill and downhill running: a recurrent...

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning

Background:

• Ground reaction forces (GRFs) are crucial for analyzing human movement but are typically measured in labs.
• Previous neural network models predicted GRFs during level-ground running only.
• Accurate GRF prediction outside the lab could enable real-time biomechanical analysis.

Purpose of the Study:

• Develop a recurrent neural network (RNN) to predict continuous normal GRFs.
• Utilize accelerometer data from wearable devices.
• Enable predictions across diverse running speeds and slopes.

Main Methods:

• Recruited 19 subjects to run on a force-measuring treadmill.
• Collected accelerometer data from sacral and shoe-mounted sensors.
• Trained an RNN to predict normal GRF waveforms frame-by-frame across various speeds and slopes.

Main Results:

• The RNN accurately predicted normal GRF waveforms.
• Achieved an average RMSE of 0.16 ± 0.04 BW and relative RMSE of 6.4 ± 1.5%.
• Demonstrated higher accuracy than previous neural network models.

Conclusions:

• The developed RNN model effectively predicts normal GRFs during running.
• This approach allows for out-of-lab biomechanical predictions in near real-time.
• Enhances the quantification of external forces during running activities.