Hannes Prosthesis Control Based on Regression Machine Learning Algorithms

The quality of life for upper limb amputees can be greatly improved by the adoption of poly-articulated myoelectric prostheses. Typically, in these applications, a pattern recognition algorithm is used to control the system by converting the recorded electromyographic activity (EMG) into complex multi-degrees of freedom (DoFs) movements. However, there is currently a trade-off between the intuitiveness of the control and the number of active DoFs. We here address this challenge by performing simultaneous multi-joint control of the Hannes system and testing several state-of-the-art classifiers to decode hand and wrist movements. The algorithms discriminated multi-DoF movements from forearm EMG signals of 10 healthy subjects reproducing hand opening-closing, wrist flexion-extension and wrist pronation-supination. We first explored the effect of the number of employed EMG electrodes on device performance through the classifiers optimization in terms of F1Score. We further improved classifiers by tuning their respective hyperparameters in terms of the Embedding Optimization Factor. Finally, three mono-lateral amputees tested the optimized algorithms to intuitively and simultaneously control the Hannes system. We found that the algorithms performances were similar to that of healthy subjects, particularly identifying the Non-Linear Regression classifier as the ideal candidate for prosthetic applications.