Genetic Algorithm for Feature Selection in Lower Limb Pattern Recognition.

Area of Science:

• Biomedical Engineering
• Rehabilitation Engineering
• Signal Processing

Background:

• Electromyography (EMG) signals are crucial for lower limb prosthetic control but are inherently noisy, complicating feature extraction.
• Numerous features are employed in myoelectric control, necessitating research into the most effective ones for lower limb applications.
• Identifying optimal feature combinations is vital for enhancing the performance of lower limb pattern recognition systems.

Purpose of the Study:

• To demonstrate the capability of genetic algorithms in optimizing feature selection for lower limb myoelectric control.
• To identify a superior feature set compared to the current state-of-the-art for prosthetic applications.
• To evaluate the performance of algorithm-selected features against established methods.

Main Methods:

• Collected EMG and kinematic data from ten able-bodied subjects during gait-related activities.
• Employed a genetic algorithm to search feature spaces and select optimal features based on training data performance.
• Evaluated the selected feature sets on a test set and the online ENABL3S benchmark dataset.

Main Results:

• The genetic algorithm-selected feature set demonstrated superior performance compared to the state-of-the-art feature set.
• Observed up to an 11.6% relative decrease in overall errors and a 14.1% relative decrease in transitional errors.
• While not statistically significant, the results indicated promising performance improvements.

Conclusions:

• Genetic algorithms are effective tools for navigating large feature spaces in myoelectric control research.
• Systematic feature selection using genetic algorithms shows significant potential for advancing lower limb myoelectric control.
• The study highlights a promising direction for improving prosthetic functionality through optimized feature extraction.