Gradient-Based Multi-Objective Feature Selection for Gait Mode Recognition of Transfemoral Amputees

Area of Science:

• Biomedical Engineering
• Robotics
• Machine Learning

Background:

• Seamless gait mode transitions are crucial for prosthetic leg control.
• User intent recognition (UIR) systems identify user activity for prosthetic control.
• Optimizing UIR for performance and complexity is a significant challenge.

Purpose of the Study:

• To propose a novel framework for designing optimal UIR systems.
• To achieve a balance between maximum performance and minimum complexity in gait mode recognition.
• To introduce a gradient-based multi-objective feature selection (GMOFS) method and apply advanced evolutionary multi-objective optimization (MOO) methods for UIR.

Main Methods:

• Developed a gradient-based multi-objective feature selection (GMOFS) method.
• Incorporated elastic net within multilayer perceptron neural network training for simultaneous feature selection and classification.
• Applied advanced evolutionary MOO methods, including multi-objective biogeography-based optimization (MOBBO) variants.
• Collected experimental data from six subjects (three able-bodied, three transfemoral amputees).

Main Results:

• GMOFS demonstrated competitive performance against four MOBBO methods.
• Achieved high classification accuracy: 97.14% ± 1.51% for able-bodied and 98.45% ± 1.22% for amputee subjects.
• Utilized only 23% of available features with the optimal selected subset, indicating significant complexity reduction.

Conclusions:

• Advanced optimization methods, like GMOFS, can effectively design accurate and compact UIR systems.
• The proposed framework offers a promising approach for improving locomotion mode detection in lower-limb amputees.
• Results highlight the potential for enhanced prosthetic leg functionality and user experience.