Related Experiment Video
Updated: Dec 20, 2025
08:24
Sit-to-stand-and-walk from 120% Knee Height: A Novel Approach to Assess Dynamic Postural Control Independent of Lead-limb
Published on: August 30, 2016
10.6K
Lower Limb Kinematics Trajectory Prediction Using Long Short-Term Memory Neural Networks
Abdelrahman Zaroug1, Daniel T H Lai1,2, Kurt Mudie3
1Institute for Health and Sport, Victoria University, Melbourne, VIC, Australia.
Frontiers in Bioengineering and Biotechnology
|May 28, 2020
Summary
Long short-term memory (LSTM) neural networks accurately predict lower limb kinematics during walking. This motion prediction model shows potential for fall prevention and improved human-machine interfaces.
Area of Science:
- Biomechanics
- Machine Learning
- Wearable Technology
Background:
- Accurate prediction of lower limb kinematics is crucial for understanding human movement and developing assistive technologies.
- Traditional methods for kinematic analysis can be complex and time-consuming.
- The application of artificial intelligence, specifically neural networks, offers a novel approach to kinematic prediction.
Purpose of the Study:
- To determine if long short-term memory (LSTM) neural networks can reliably extrapolate lower limb kinematic trajectories during walking.
- To investigate the capability of LSTM autoencoders in forecasting multiple time-step trajectories of linear acceleration (LA) and angular velocity (AV) of the lower limb.
- To evaluate the generalizability of the LSTM model across different participants.
Main Methods:
- Utilized 3D motion capture to record lower limb position-time coordinates (100 Hz) from six male participants walking on a treadmill.
- Employed an LSTM model with a sliding window of 25 samples, analyzing four kinematic variables: thigh and shank LA and AV.
- Trained the LSTM model on 2,665 strides from five participants and tested its predictive accuracy on one stride from a sixth participant, forecasting five time steps ahead.
Main Results:
- The LSTM model successfully learned and predicted lower limb kinematic trajectories, demonstrating generalization across participants.
- Forecasting accuracy was higher for earlier time steps within the prediction horizon.
- The model achieved a mean absolute error (MAE) of 0.047 m/s² for thigh LA, 0.047 m/s² for shank LA, 0.028 deg/s for thigh AV, and 0.024 deg/s for shank AV.
- Predicted trajectories showed high correlation with measured trajectories (correlation coefficients > 0.98).
Conclusions:
- LSTM neural networks provide a reliable method for extrapolating lower limb kinematics during walking.
- The developed motion prediction model has significant potential applications in fall risk mitigation and enhancing human-machine interfaces for wearable devices.
- This study highlights the efficacy of deep learning in biomechanical motion analysis and prediction.
