Comparing Machine Learning Methods and Feature Extraction Techniques for the EMG Based Decoding of Human Intention.

Area of Science:

• Biomedical Engineering
• Robotics
• Machine Learning

Background:

• Intuitive Muscle-Machine Interfaces (MuMIs) are crucial for controlling robotic and prosthetic devices.
• Machine learning methods analyzing myoelectric activations decode user intention from Electromyography (EMG) signals.
• Optimal feature extraction and machine learning model selection are key challenges.

Purpose of the Study:

• To compare the performance of five machine learning methods and eight time-domain feature extraction techniques for gesture discrimination in EMG-based MuMIs.
• To identify the most effective combination of feature extraction and machine learning for intention decoding.
• To validate the best performing methods in a real-time robotic hand control application.

Main Methods:

• Five distinct machine learning algorithms were evaluated.
• Eight time-domain feature extraction techniques were applied to raw EMG signals.
• Gesture classification accuracy and inter-subject variance were used as performance metrics.

Main Results:

• The Willison Amplitude feature extraction method consistently outperformed others across all tested machine learning algorithms.
• Random Forests achieved the highest classification accuracies with the lowest inter-subject variance.
• Zero Crossings and Variance were identified as less effective feature extraction techniques for certain algorithms.

Conclusions:

• The combination of Random Forests and Willison Amplitude offers a robust and accurate approach for EMG-based MuMI development.
• Real-time decoding of operator gestures using this method successfully controlled a robotic hand.
• This research provides valuable insights for designing more intuitive and effective human-machine interactions.