Deep Domain Adaptation, Pseudo-Labeling, and Shallow Network for Accurate and Fast Gait Prediction of Unlabeled...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Robotics

Background:

• Personalized gait phase prediction is crucial for real-time control systems like wearable robots.
• Acquiring accurate gait data typically requires expensive experiments, hindering model development.
• Existing domain adaptation (DA) models present a trade-off between prediction accuracy and inference speed.

Purpose of the Study:

• To develop a novel dual-stage domain adaptation (DA) framework for personalized gait phase prediction.
• To achieve both high accuracy and fast inference speeds, overcoming limitations of current DA models.
• To enable efficient and accurate gait phase prediction for real-time applications.

Main Methods:

• A semi-supervised DA approach is employed to minimize feature discrepancies between source and target subjects.
• The proposed framework utilizes a deep network in the first stage for precise DA and pseudo-label generation.
• A shallow, fast network is trained in the second stage using the pseudo-labels, avoiding DA computation for speed.

Main Results:

• The dual-stage DA framework significantly reduces prediction error compared to traditional shallow DA models.
• The proposed method achieves a 1.04% reduction in prediction error while maintaining fast inference speeds.
• The framework demonstrates the potential for accurate gait phase prediction with computational efficiency.

Conclusions:

• The dual-stage DA framework successfully balances accuracy and inference speed for personalized gait prediction.
• This approach offers a viable solution for developing fast, personalized gait prediction models for real-time systems.
• The proposed method can enhance the performance and applicability of wearable robots and other real-time control systems.