Unsupervised learning for real-time and continuous gait phase detection.

Area of Science:

• Biomedical Engineering
• Rehabilitation Robotics
• Machine Learning in Healthcare

Background:

• Lower limb impairment following stroke or spinal cord injury necessitates effective rehabilitation strategies.
• Current robotic rehabilitation systems lack real-time continuous gait phase detection, limiting therapeutic efficacy.
• Existing gait phase detection methods often rely on complex algorithms or are not suitable for real-time application.

Purpose of the Study:

• To develop an unsupervised learning method for real-time and continuous gait phase detection.
• To enhance the effectiveness of robotic-assisted lower limb rehabilitation.
• To provide a reliable gait phase detection system applicable to overground locomotion.

Main Methods:

• An unsupervised learning approach utilizing windows of real-time kinematic trajectories.
• A pre-trained neural network model leveraging treadmill walking data.
• Application of the model to detect continuous gait phases during overground locomotion.

Main Results:

• The developed neural network model achieved an average time error of less than 11.51 ms across various walking conditions.
• Overground walking exhibited a lower average time error (11.20 ms) compared to treadmill walking (12.42 ms).
• The method successfully predicts real-time gait phases using a pre-trained model, eliminating the need for complex overground data acquisition.

Conclusions:

• The proposed unsupervised learning method offers accurate and real-time continuous gait phase detection.
• This technology can significantly improve the precision and adaptability of robotic rehabilitation devices.
• The system's ability to use pre-trained models from laboratory data simplifies its application in clinical settings.