Machine Learning Approach to Multi-Subject EMG Classification of Hand Gestures in Varied Forearm Orientation

Abstract

The electromyogram (EMG) is a bio signal which manifests in conjunction with muscle contraction. It can be used for neuro-muscular diagnosis, ergonomic analysis and machine control including prosthetics and teleoperation. Although in recent years there has been advances in detection and gesture prediction methods including artificial intelligence, machine learning still plays a fundamental role as the EMG is a waveform-based signal. As a biological signal, the EMG signal can vary throughout the day depending on the placement of the electrodes, muscle contraction level and fatigue of the individual. Moreover, due to the variability between individuals, a prediction model trained on an individual may not provide accurate prediction for another individual. Therefore, there is a strong motivation to study the variation of the EMG waveform and identify the most universally classifiable gestures. The main objective of this work is to measure variability of the forearm EMG signals between individuals and to classify the gestures based on training data from various forearm positions. In this work, the EMG signals of nine gestures of the hand was performed with the forearm in neutral, pronation and supination. The variability analysis was performed with normalized cross-correlation (NCC), then training models were developed with various time and frequency domain features and classifiers. Five-fold cross-validation was used to validate the classification accuracy. The main classification results show that the best classification accuracy can be achieved with the use of 10-Hz linear envelope and linear dependent analysis (LDA) classifier which yielded 89-92% accuracy for the forearm flex and extend motion. The NCC of these opposing gestures, also yielded a coefficient of 0.78 which shows a significant difference in the gestures.