Gait Phase Detection in Walking and Stairs Using Machine Learning.

Area of Science:

• Biomechanics
• Robotics
• Machine Learning

Background:

• Machine learning algorithms are crucial for powered motion assistive devices, informing control of motorized components.
• Accurate activity and gait phase recognition is essential for effective control of lower limb assistive technologies.

Purpose of the Study:

• Develop a supervised multiclass classifier for simultaneous activity and gait phase detection (stance, swing).
• Utilize inertial measurement data from the thigh and shank for real-world walking, stair ascent, and stair descent.
• Inform the control of knee-local motion assistive devices.

Main Methods:

• Evaluated decision tree and long short-term memory (LSTM) models using diverse feature sets.
• Introduced a novel performance metric: proportion of perfectly classified strides (PPCS).
• Developed both simultaneous six-state classifiers and activity-specific binary classifiers for stance/swing phases.

Main Results:

• The superior activity-specific model achieved 98.0% accuracy and 55.7% PPCS.
• The best six-phase model (using filtered data and median filtering) reached 92.1% accuracy and 22.9% PPCS.
• A binary classifier for stance/swing phases across all activities showed 97.1% accuracy.

Conclusions:

• Activity-specific models demonstrate high performance for precise gait phase recognition.
• Simultaneous activity and gait phase classification shows potential, especially when simplified to binary stance/swing detection.
• These findings support the use of machine learning with inertial sensors for advanced control of lower limb exoskeletons.