Unsupervised Gait Phase Estimation With Domain-Adversarial Neural Network and Adaptive Window

Area of Science:

• Biomechanics
• Machine Learning
• Wearable Technology

Background:

• Previous machine learning models for gait phase estimation require subject-specific ground truth data, limiting their practical application.
• Obtaining ground truth gait phase data necessitates expensive equipment and is not feasible for individual user optimization.
• Existing models suffer from performance degradation due to variations in stride time and lack optimal subject selection strategies.

Purpose of the Study:

• To develop an unsupervised domain adaptation technique for accurate gait phase estimation without requiring target subject ground truth.
• To introduce an adaptive window method to mitigate accuracy loss caused by stride time variations.
• To propose a novel method for selecting the optimal source subject based on sequential embedding feature similarity.

Main Methods:

• Modified a domain-adversarial neural network for regression on continuous gait phases, enabling unsupervised domain adaptation.
• Developed an adaptive window method to dynamically adjust to changes in stride time during motion analysis.
• Implemented a subject selection strategy based on the similarity of sequential embedding features to identify the most suitable training subjects.

Main Results:

• The proposed unsupervised domain adaptation technique significantly reduces estimation errors without needing target subject-specific ground truth.
• The adaptive window method demonstrated considerable reduction in gait phase estimation errors for both walking and running motions.
• The optimal source subject selection method improved model performance by identifying the most relevant training data.

Conclusions:

• This study presents a robust and practical approach for gait phase estimation, overcoming limitations of previous methods.
• The developed techniques enhance the generalizability and accuracy of machine learning models for gait analysis in real-world scenarios.
• The findings pave the way for more accessible and personalized gait monitoring systems using wearable sensors.