Provision of real-time bioelectric feedback to enhance training of upper limb prostheses

Artificial limbs are hard to control, and learning to use a pattern recognition based myoelectric prosthesis implies learning to produce high-quality electromyogram (EMG) patterns with distinct, not too variable and highly reliable features. Training is required to become proficient and generate consistent and separable patterns for different movements. During the experiment carried out in this work, real-time visual feedback were provided to the users through a spider plot that was implemented on MATLAB and integrated on the open source software BioPatRec. The spider plot was developed in such a way that based on the most conflicting neighbour, a new target pattern was calculated for every movement performed. The participants could see in real-time how the patterns in the spider plot changed in accordance to their muscle contractions, and they could modify their movements to match the new target shape with the aim of generating more separate patterns. Ten able-bodied participants were recruited, and the experiment was conducted by dividing them in two groups that trained with two different kinds of feedback for three consecutive days. The spider plot method was compared to a traditional training method, i.e. training with raw EMG signals. A significant difference was not found between the two feedback approaches. For some of the investigated variables, both groups increased their performance. Although the increase was not significant, separability index improved with training, thus meaning that training had a positive effect on it. On the first day, the group training with spider plot feedback reported an average separability index of 2.64 ± 0.27, and on the last training the separability improved to 2.74 ± 0.30. An improvement in the average offline accuracy was also detected for both groups. Despite the hypothesis that more distinct patterns lead to better control performance, a weak correlation between online accuracy and separability index was found (0.22). This evidence shows that the correlation between these two metrics is not straightforward. A strong correlation was found between online accuracy and completion time (-0.94). The overall results reported that training had a positive effect on performance accuracy and separation of patterns, notwithstanding the type of feedback, and that training with spider plot feedback have the potential to improve pattern recognition control performance.