Dynamic training protocol improves the robustness of PR-based myoelectric control

Abstract

In pattern recognition (PR)-based myoelectric control schemes, the classifier is generally trained in ideal laboratory conditions, due to which the classification accuracy might be affected by confounding factors such as force variations, limb positions, and inadvertent electromyography (EMG) activation. Many endeavors have been put forward to mitigate this effect by adopting new training protocols that consider only quite a few independent factors. In this note, we propose a dynamic protocol, which embraces multiple EMG variations in data collection, to train a classifier with improved generalization ability. A total of four training protocols are examined, wherein affecting factors like upper-limb movements, contraction levels and inadvertent EMG activations are differently considered. Based on receiver operating characteristic (ROC) analysis, we came up with a new performance metric, ROC area rate (RAR), to directly inspect the accuracy and robustness of the classifiers obtained through different training protocols. Our results show that, compared with the other three protocols, the protocol with dynamic limb postures and dynamic muscle contractions (termed as DPDE) obtains the highest RAR (73.3%, on-way analysis of variance, $p<0.005$). Our results suggest that there is no need to integrate every EMG variation in the training protocol for receiving a robust EMG pattern recognition. Online control experiments with three amputees manipulating a multiple-DOF prosthetic hand also verify our findings.

Introduction

Using engineering approaches to restoring human hand’s function has long been an ambitious goal in robotics society. In literature, a variety of robotic/prosthetic hand prototypes with different degrees of dexterity have been developed [1], [2]. Meanwhile, on the market, advanced hand prostheses with multiple degrees of freedom (DOF) and neural control/feedback interfaces have also started to become available [3], [4], [5], [6], [7]. However, the artificial hand still cannot be naturally and intuitively used as a biological hand with the same degree of skill and delicacy [8]. A fact is that, as the number of the active joints increases, the control of these dexterous prosthetic hands tends to be exponentially complicated [9], [10], [11].

For operating a prosthetic hand, a bidirectional interface (peripheral machine interface, PMI) [12] that links the mechanical hand to its user is required. Among various biomedical signals, the myoelectric signals (electromyography, EMG) collected from the residual arm of the amputee are widely used. The signal contains rich neural information and requires no surgery and hospitalization. In ideal laboratory conditions, several grip patterns and finder/wrist motions can be faithfully decoded based on the EMG activation patterns [13], [14], [15]. However, the pattern recognition (PR)-based control schemes still have many problems when being put into clinical practice [12], despite the considerable academic achievements [9], [16], [17].

One big problem in the PR-based myoelectric control is that one cannot ensure the learned classifier’s generalization ability in various application environments. Confounding factors, such as electrode shifting [18], [19], different contraction levels [20], [21] and changing body positions [22], would inevitably introduce some signal variations that dramatically affect the performance of the classifier trained on ideal laboratory conditions [11], [12]. For example, without considering the transient EMG signals (generally produced at the initiation stage of a contraction), there is only a weak correlation between the control system’s offline classification accuracy and its real usability [23]. Thus, if the EMG signals collected from dynamic contractions are included during training, the usability of the PR-based control scheme can be greatly improved [20], [24], [25]. Body posture is another annoying factor affecting the accuracy of the EMG pattern recognition [22], [26]. This factor can be alleviated by collecting training samples on several preselected body postures (especially, the arm postures) [27], [28], [29]. Another driving source for low online control accuracy is the inadvertent EMG activation, which are caused by involuntary muscular contractions and seen as one of the main reason of frustration during clinical testing [30]. This problem can be somehow resolved through extensive subject training, for obtaining more disciplined, repeatable EMG signals resisting the confounding factors.

To increase the PR-based EMG control’s usability, another solution is to improve the classifier’s generalization ability by directly enveloping some EMG variations during training [24]. However, it is impractical to consider all confounding factors at the same time in one training session. An effective, uniform and easy-to-perform training protocol is greatly desired when pushing PR-based EMG control into clinical practice. From this angle, this note initially proposes a new training protocol that synthesizes dynamic muscle contractions, dynamic limb postures, and inadvertent EMG activations together. This protocol can improve the generalization ability of the PR-based EMG control used in general living conditions.