A comparison of neural networks algorithms for EEG and sEMG features based gait phases recognition

https://doi.org/10.1016/j.bspc.2021.102587Get rights and content

Highlights

  • Various dimensions of features and different classifiers applied for gait phases recognition.

  • The wider value distribution of features, the better accuracy of gait recognition.

  • Two-dimensional feature sets with KNN are suitable for online gait recognition.

  • Thirty-seven-dimensional feature sets achieved the highest classification accuracy.

Abstract

Surface electromyography (sEMG) and electroencephalogram (EEG) can be utilized to discriminate gait phases. However, the classification performance of various combination methods of the features extracted from sEMG and EEG channels for seven gait phase recognition has yet to be discussed. This study investigates the effectiveness of various dimensions of feature sets with different neural network algorithms in multiclass discrimination of gait phases. There are thirty-seven feature sets (slope sign change (SSC) of eight sEMG and twenty-one EEG channels, mean absolute value (MAV) of eight sEMG channels) and three classifiers (Linear Discriminant Analysis (LDA), K-nearest neighbor (KNN), Kernel Support Vector Machine (KSVM)) were utilized. The thirty-seven one-dimensional and six two-dimensional feature sets were applied to LDA and KNN, twenty-one-dimensional and thirty-seven-dimensional feature sets were applied to three optimized KSVM for gait phase recognition. We found that thirty-seven-dimensional feature sets with grid search KSVM achieved the highest classification accuracy (98.56 ± 1.34 %) and the time consumption was 26.37 s. The average time consumption of two-dimensional feature sets with KNN was the shortest (0.33 s). The SSC of sEMG with wider values distributions than others obtained a high performance. This indicates the wider the value distribution of features, the better accuracy of gait recognition. The findings suggest that a multi-dimensional feature set composed of EEG and sEMG features with KSVM achieved good performance. Considering execution time and recognition rate, two-dimensional feature sets with KNN are suitable for online gait recognition, thirty-seven-dimensional feature sets with KSVM are more likely to be used for off-line gait analysis.

Introduction

A suitable lower limb exoskeleton will improve the life quality of patients with lower limb disorder greatly [1]. The recognition of the gait phases correctly is critical for robots to assist timely [2]. Human gait usually is divided into stance and swing phases. The stance phase can be sub-split into loading response, mid-stance, terminal stance, and pre-swing. The swing phase can be sub-split into initial swing, mid-swing, and terminal swing [3]. More information about gait partitioning can be found in [3]. Surface electromyography (sEMG) and electroencephalogram (EEG) signals generate before motion and are a trigger of muscle contraction. Therefore, sEMG and EEG have the potential to predict motion intention of the lower limb [2,[4], [5], [6], [7]]. sEMG has been applied to lower limb rehabilitation robots such as HAL [5,8]. Nurhazimah Nazmi et al. utilized sEMG to recognize gait phases and a recognition rate of 87.4 % was achieved [9]. Surface electromyography signals were used to identify gait phases of children with spastic cerebral palsy by PENG-NA WEI et al. with an average classification rate of 89.40 % [10]. However, sEMG are susceptible to muscle fatigue. EEG signals, produce in the cerebral cortex, will make up for the disadvantage of sEMG [11]. Junhyuk Choi et al. decoded gait intention from EEG signals to control exoskeletons [12].

However, sEMG and EEG-based leg motion recognition methods have not been widely used. Especially in gait phase recognition. Most gait phase recognition research is based on sEMG or kinematics data [10,13,4,14]. Therefore, how to combine the sEMG and EEG channels for gait phase recognition has not been systematically discussed before. For example, SUSANNA YU. GORDLEEVA et al. just combined four sEMG and seven EEG channel features for classifying the two motions of the leg with an average recognition rate of 80 % [15]. We introduced four feature combination methods with a different number of channels (one-dimensional, two-dimensional, twenty-one and thirty-seven-dimensional) in this research to evaluate their effectiveness on gait phase recognition.

Linear discriminant analysis (LDA) [16] and K-nearest neighbor (KNN) [17] are two simple and classical classifiers in sEMG and EEG signals classification [18]. LDA classifiers are more suitable for linear separable signals [16]. KNN and LDA mostly are applied to low dimensional features [19,20]. A better classification accuracy usually is obtained using LDA with dimension reduction [21]. The Kernel support vector machine (KSVM) utilizes kernel function to map linear indivisible features to high dimensional space. Then the hyperplane is constructed to separate the high-dimensional features [22]. According to the characteristic of three classifiers, one-dimensional and two-dimensional feature sets were applied to LDA and KNN, twenty-one-dimensional and thirty-seven-dimensional feature sets were applied to KSVM. Grid Search (GS) [23], Genetic Algorithm (GA) [24], and Particle Swarm Optimization (PSO) [25] were used to optimize the parameters of the kernel function of KSVM.

In this study, the combination methods of the features extracted from sEMG and EEG channels for gait phase recognition were investigated. Firstly, the feature was extracted from all sEMG and EEG channels. And then four types of feature sets were constructed and applied to classifiers respectively. Finally, the effectiveness of feature-classifier combinations was analyzed.

Section snippets

Materials and methods

To acquire the sEMG and EEG signals of subjects during walking, a custom-designed and synchronous acquisition system for sEMG and EEG signals was utilized. The biomedical signals were acquired from seven healthy participants who walked on a treadmill. The objective of this study was to research the performance of different combinations of feature sets and neural network algorithms in multiclass discrimination of seven gait phases (loading response, mid-stance, terminal stance, pre-swing,

Results

The classification results of feature types of four different dimensions (one-dimensional, two-dimensional, twenty-one-dimensional and thirty-seven-dimensional)) were displayed to test the practicability of each feature type in gait phases. Furthermore, three experiments were conducted to verify the effectiveness of four feature types and three classifiers.

Discussion

We investigated the gait phase recognition performance of four combination methods of features (one-dimensional, two-dimensional, twenty-one-dimensional and thirty-seven-dimensional) extracted from sEMG and EEG channels in this study. According to the results of experiments, thirty-seven-dimensional feature sets with KSVM outperformed those from one-dimensional, two-dimensional and twenty-one-dimensional feature sets. Furthermore, thirty-seven-dimensional feature sets with KSVM obtained much

Conclusions

We investigated the different effects of feature sets with four different dimensions for gait phase recognition. The one-dimensional feature sets were applied to LDA, KNN and KSVM, and two-dimensional feature sets were applied to LDA and KNN, and thirty-seven-dimensional feature sets were applied to three optimized KSVM for gait phases recognition. We found that thirty-seven-dimensional feature sets with GS-KSVM were superior to other feature-classifier combinations and achieved the highest

CRediT authorship contribution statement

Pengna Wei: designed the study, conducted and supervised the experimental process, analyzed the experimental data and wrote this paper. Jinhua Zhang: Conceptualization, guided the experimental process. Feifei Tian: co-designed the study. Jun Hong: supervised data analysis. All authors revised and approved the manuscript.

Funding

This work was supported by the Fundamental Research Funds for the Central Universities [grant numbers sxzy022019050] and the Equipment Advance Research Foundation of China [grant numbers 61400030701].

Acknowledgements

The authors thank the participants who enrolled in this experiment. We thank the Xi’an Jiaotong University provide the lab space.

Declaration of Competing Interest

None.

References (40)

  • T. Lenzi et al.

    Intention-based EMG control for powered exoskeletons

    IEEE Trans. Biomed. Eng.

    (2012)
  • J. Choi et al.

    Detecting voluntary gait initiation/termination intention using EEG

    2018 6th Int. Conf. Brain-Computer Interface, BCI 2018. 2018-Janua

    (2018)
  • S. Tortora et al.

    Deep learning-based BCI for gait decoding from EEG with LSTM recurrent neural network

    J. Neural Eng.

    (2020)
  • D. Shi et al.

    A review on lower limb rehabilitation exoskeleton robots

    Chinese J. Mech. Eng.

    (2019)
  • P. na Wei et al.

    sEMG based gait phase recognition for children with spastic cerebral palsy

    Ann. Biomed. Eng.

    (2018)
  • J. Tryon et al.

    Performance evaluation of EEG/EMG fusion methods for motion classification

    2019 IEEE 16th Int. Conf. Rehabil. Robot.

    (2019)
  • J. Choi et al.

    Real-time decoding of EEG gait intention for controlling a lower-limb exoskeleton system

    7th Int. Winter Conf. Brain-Computer Interface, BCI 2019

    (2019)
  • Y. Li et al.

    Gait recognition based on EMG with different individuals and sample sizes

    Chinese Control Conf. CCC. 2016-Augus

    (2016)
  • A. Mannini et al.

    A hidden Markov model-based technique for gait segmentation using a foot-mounted gyroscope

    Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. EMBS.

    (2011)
  • S.Y. Gordleeva et al.

    Real-Time EEG-EMG human-machine interface-based control system for a lower-limb exoskeleton

    IEEE Access

    (2020)
  • Cited by (24)

    • Gait phases recognition based on lower limb sEMG signals using LDA-PSO-LSTM algorithm

      2023, Biomedical Signal Processing and Control
      Citation Excerpt :

      In the experiment, we chose the 2.5 km/h walking speed data as the training data because this speed has the highest accuracy in the previous section. In each experiment, we separately tested the effect of LDA [34], principal component analysis (PCA), independent component correlation algorithm (ICA) [35], and random forest (RF) [36,37] as data dimensionality reduction method. Then we tested and compared the effect of k-nearest neighbor [15](KNN), kernel support vector machine (KSVM) [38], RF, eXtreme gradient boosting (XGBoost)[39], MLP [40], back propagation neural network (BPNN) [6], RNN, LSTM [12], gate recurrent unit (GRU) [41] as classification models.

    View all citing articles on Scopus
    View full text