Feedback affects how users improve when training machine learning control (Original title: Different feedback during user training for pattern recognition control lead to similar performance, but in different ways)

Conference object OPEN
Kristoffersen, Morten Bak ; Franzke, Andreas ; Murgia, Alessio ; van der Sluis, Corry ; Bongers, Raoul (2018)

Abstract Different feedback during user training for pattern recognition control lead to similar performance, but in different ways M. B. Kristoffersen, A. Franzke, A. Murgia, R. Bongers, C. van der Sluis; University Medical Center Groningen, University of Groningen, Groningen, Netherlands. Background: Pattern recognition (PR) control has been proposed as a clinical alternative to the direct control scheme used in upper limb prosthetics. The intent of PR control is that distinct electromyography (EMG) patterns over several muscles generated by phantom movements will provide intuitive control. However in many patients the EMG patterns from a set of movements are indistinguishable and inconsistent leading to non-intuitive control. Arguably, training would improve the EMG patterns, but it is unknown which training strategy achieves the most consistent and distinct patterns. In this study users trained with three methods, proposed in the literature, that differed in the levels of feedback. This approach might reveal which training leads to the best EMG patterns for PR control. We study (1) the effect of feedback during training on online performance and (2) which training leads to the most consistent and distinct EMG patterns in feature space. Materials and methods: Able bodied volunteers (N=37; mean age 21.6, 18 females) trained using a PR system with 8 electrodes based on the Linear Discriminant Analysis classifier using the Hudgins feature set. Participants were divided in groups with No Feedback (NF), Visual Feedback (VF) and Visual + Coaching Feedback (VCF). NF trained following oral prompts and never received feedback on their performance. VF trained following targets on a screen. VCF trained as VF, but in addition received coaching on how to improve. A pre/post-test design with five training sessions on five consecutive days was used. Outcome measures were online accuracy, number of completed movements measured with the Motion Test and distinguishability/consistency of EMG patterns in feature space. Results: Both online accuracy as well as completed movements showed a significant improvement from pre- to post-test with no significant group effect. Analyses of feature space metrics revealed that NF and VCF achieved more distinct patterns. Unlike VCF, NF achieved distinct patterns by using more force. VF improved performance by producing more consistent patterns. Conclusions: Following training all groups had a similar performance. Surprisingly NF achieved as good online performance as VCF. Interestingly groups behaved differently in feature space after training; meaning that different training methods lead to different outcomes in the feature space. It remains to be seen if these results are applicable in individuals with an upper limb defect.
Share - Bookmark

  • Download from
    ZENODO via ZENODO (Conference object, 2018)
  • Funded by
  • Cite this publication