Improvement of EMG Pattern Recognition Model Performance in Repeated Uses by Combining Feature Selection and Incremental Transfer Learning

Electromyography (EMG) pattern recognition is one of the widely used methods to control the rehabilitation robots and prostheses. However, the changes in the distribution of EMG data due to electrodes shifting results in classification decline, which hinders its clinical application in repeated uses. Adaptive learning can solve this problem but takes additional time. To address this, an efficient scheme is developed by comparing the performance of 12 combinations of three feature selection methods [no feature selection (NFS), sequential forward search (SFS), and particle swarm optimization (PSO)] and four classification methods [non-adaptive support vector machine (N-SVM), incremental SVM (I-SVM), SVM based on TrAdaBoost (T-SVM), and I-SVM based on TrAdaBoost (TI-SVM)] in the classification of EMG data of 12 subjects for 5 consecutive days. Our results showed that TI-SVM achieved the highest classification accuracy among the classification methods (p < 0.05). The SFS method achieved the same classification accuracy as that of the scheme trained with the feature vectors selected by the NFS method (p = 0.999) while achieving a lower training time than that of TI-SVM combined with the NFS method (p = 0.043). Although the PSO method outperformed the NFS and SFS methods by achieving reduced training and response times (p < 0.05), the PSO method achieved a considerably lower classification accuracy than that of the scheme trained with the feature vectors selected by the NFS (p = 0.001) or SFS (p = 0.001) method. Furthermore, TI-SVM combined with the SFS method outperformed the CNN method with fine-tuning in classification accuracy on a small data set (p = 0.001). The results indicate that TI-SVM combined with the SFS method is suitable for improving the performance of EMG pattern recognition in repeated uses.

Performance Evaluation of EMG Pattern Recognition Techniques While Increasing The Number of Movement Classes

:In the past few years of research done in the field of myoelectric control, many researchers have proposed several models imploying a combination of different features and classifiers to increase the movement classes, but all that work fails to explain if there is any correlation between multi-class classification and its accuracy. This paper focuses on finding the factors that decide the limit of movement classes that machine learning algorithms can accurately differentiate and to evaluate the performance of pattern classification techniques using the sEMG signal when the number of movement classes is increased while keeping the simplicity of the system. The results were obtained for eight channels sEMG signal using 7 independent time-domain features and four feature set combinations over 4 classifiers (Support Vector Machine(SVM), K-Nearest Neighbour(K-NN), Decision Tree(DT), and Naïve Bayes(NB)). Then the number of classes was increased in the manner of 5, 7, 10, 12, and 15 classes to determine the highest number of movement classes that the sEMG system with above-described features can classify efficiently. And the effect of increasing the number of movement classes on system accuracy was observed. The highest accuracies for all five class progression were obtained for SVM with the MFL feature, and for DT using MAV, it was successfully observed that the NB classifier had minimum performance depletion for the features used in this work

Recent Advances in EMG Pattern Recognition for Prosthetic Control

Limb loss results in significant debilitation and reduces the quality of life of the affected individuals [ 1 ]. To restore the lost limb’s function, myoelectric systems have been widely used in powered prostheses [ 2 ]. With this approach, the motor intent is estimated from the electromyogram (EMG) signals recorded by electrodes which are placed on the skin surface above the residual muscles [ 1 ]. The principle of commercial myoelectric schemes has not changed in several decades, and is referred to as conventional control [ 2 ]. This technique uses a measure of amplitude (such as mean absolute value over a time window) of the EMG signals recorded by electrodes placed at two control sites, preferably over a pair of antagonist muscles of the residual limb, to control a single motion i.e. degree of freedom (DoF), for example hand opening closing [ 2 ]. To change the DoF, a mode switch is conducted by muscle co-contraction or a hardware switch [ 2 ]. The mode switch, however, results in an unnatural control of multiple DoFs [ 2 ]. To overcome this challenge, a significant body of research has been conducted on pattern recognition techniques [ 3 ]. With this approach, a classifier is trained to discriminate between different DoFs, using patterns from multi-channel EMG input data. Promising results have been achieved in the literature for classification of several DoFs [ 2 ]. Since activities of daily living include simultaneous movements of multiple DoFs, combined motions must be also included as separate classes, and they have to be conducted in the training set [ 4 ]. The limitation of this approach, however, is that it does not allow the DoFs in combined motions to have different magnitudes. As a solution to this problem, regression-based systems have been proposed [ 5 , 6 ], where a regressor is trained to estimate each DoF, using data from single and combined motions. This strategy provides independent simultaneous control, because it does not limit the DoFs to have the same amplitude. Classification and regression based systems are the two categories of pattern recognition methods. Due to the high dimensionality of EMG signals, the EMG instantaneous values are not directly used as the inputs to classifiers/regressors [ 1 ]. Instead, a set of features is extracted from a time window (100-200 ms) of EMG signals [ 7 ]. Feature engineering is the process of design and extraction of features with the highest amount of useful information to maximize the classification/regression accuracy [ 8 ] Among various EMG features proposed in the literature, the Time Domain (TD) set [ 9 ] is the most popular set and includes mean absolute value, waveform length, zero-crossings, and slope sign changes. The past few years have seen the advent of deep learning-based myoelectric control [ 4 , 10 ]. Deep learning can perform classification/regression tasks directly from high-dimensional raw data, without feature engineering [ 8 ]. Convolutional neural network (CNN) [ 11 ] is one of the most widely used deep learning frameworks. The successive convolution layers of CNNs can learn useful features from the EMG data to estimate the motor intent [ 4 ]. As the outcomes of the previous studies [ 4 , 10 ] confirm, CNNs outperform classical models such as support vector machines (SVMs) with engineered feature sets. EMG pattern recognition schemes have yet to be deployed in commercial prostheses. The major challenge is performance degradation due to disturbances such as electrode shift, skin impedance change, muscle size variations, and learning effect [ 2 ]. Recent studies (e.g. [ 12 , 13 ]) have proposed methods to improve the robustness of EMG pattern recognition to such disturbances. These methods as well as new deep learning schemes that eliminate feature engineering, may pave the way for commercial implementation of myoelectric pattern recognition prostheses. Moreover, independent simultaneous control can be achieved by using regression deep learning models. These promising methods have the potential to significantly outperform existing commercial systems. Consequently, the missing functions in people with limb loss can be restored more efficiently by delivering a more natural and intuitive control.