Unsupervised learning for real-time and continuous gait phase detection

This study addresses the challenge of real-time gait phase detection for rehabilitation robots used by individuals with lower limb impairments. Current robotic systems struggle to accurately detect continuous gait phases in real time, limiting their effectiveness. The researchers propose an unsupervised learning method using a pre-trained model from treadmill walking data to detect the continuous gait phase of humans during overground locomotion. This method aims to eliminate the need for challenging-to-obtain overground walking data. The neural network model developed exhibits an average time error of less than 11.51 ms across all walking conditions, indicating its suitability for real-time applications. This allows for precise and timely control of walking patterns, potentially improving rehabilitation outcomes.

• 1
  The developed neural network model exhibits an average time error of less than 11.51 ms across all walking conditions, indicating its suitability for real-time applications.
• 2
  The average time error during overground walking (11.20 ms) is lower than that during treadmill walking (12.42 ms), suggesting the model's effectiveness in real-world scenarios.
• 3
  Models trained on specific treadmill speeds demonstrate significantly lower time errors in detecting gait events compared to models trained on combined treadmill data across all speeds.