Toward Minimal-Sensing Locomotion Mode Recognition for a Powered Knee-Ankle Prosthesis

Area of Science:

• Biomedical Engineering
• Robotics
• Machine Learning

Background:

• Locomotion mode recognition (LMR) is crucial for seamless transitions in powered prostheses.
• Current LMR systems can be complex, requiring extensive features and sensors.

Purpose of the Study:

• To develop an optimization framework for LMR that minimizes feature set size without compromising performance.
• To create more efficient and less complex LMR systems for transfemoral prostheses.

Main Methods:

• Utilized multi-objective biogeography-based optimization to balance performance and feature set size.
• Collected experimental data from four transfemoral users with powered knee-ankle prostheses.
• Compared LMR performance using optimal feature subsets versus the full feature set with a deep neural network classifier across six locomotion modes.

Main Results:

• Classifier performance with optimal feature subsets was statistically equivalent to using the full feature set.
• LMR with optimal subsets achieved low error rates (1.98% steady-state, 4.09% transitional).
• The optimized LMR utilized only ~41% of features and ~53% of sensors.

Conclusions:

• The proposed framework enables accurate and low-complexity LMR for transfemoral powered prostheses.
• This optimization can lead to reduced clinical visits, healthcare costs, and user burden.