Performance potential of classical machine learning and deep learning classifiers for isometric upper-body myoelectric control of direction in virtual reality with reduced muscle inputs

Abstract

Electromyography (EMG) signals can be classified by machine learning (ML) algorithms to command prosthetic devices that functionally assist persons after neuromuscular traumas, including amputation and spinal cord injury. This pilot study evaluated several ML algorithms in mapping isometric EMG signals from the upper body (dominant-side arm, chest, back) of able-bodied participants to directional commands across multiple muscle recording sets. Each set (up to 14 muscles) was based on muscles presumed under volitional control following various levels of nerve lesion or amputation. Among the evaluated ML algorithms were those that did and did not rely on feature extraction. The ML algorithms included: support vector machine, adaptive boosting, bootstrap aggregating, Naïve Bayes, linear discriminant analysis, and variations of neural networks (NN). Specifically, we examined a shallow (single-layer feedforward) NN and two ‘deep’ NN structures (ten-layer feedforward network, convolutional NN). The ML algorithms were evaluated according to classification accuracy and performance in a maze navigation task in virtual reality. Adaptive boosting and bootstrap aggregating demonstrated significantly greater (p < 0.05) classification accuracy across most muscle sets. Maze task performance depended on the combination of classifier and muscle set utilized. Advantages in classification accuracy from adaptive boosting and bootstrap aggregating should be balanced against the cost of increased time to train EMG control for persons with motor impairment.