Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Coherence-based connectivity analysis of EEG and EMG signals during reach-to-grasp movement involving two weights

Coherence-based connectivity analysis of EEG and EMG signals during reach-to-grasp movement... BRAIN-COMPUTER INTERFACES 2022, VOL. 9, NO. 3, 140–154 https://doi.org/10.1080/2326263X.2022.2029308 ORIGINAL RESEARCH Coherence-based connectivity analysis of EEG and EMG signals during reach-to- grasp movement involving two weights Cristian D. Guerrero-Mendez and Andres F. Ruiz-Olaya Bioengineering Research Group, Faculty of Mechanical, Electronic and Biomedical Engineering, Antonio Nariño University, Bogotá, Colombia ABSTRACT ARTICLE HISTORY Received 17 August 2021 Corticomuscular coherence allows studying the mechanism of the cerebral cortex’s control of Accepted 7 January 2022 muscle activity, which reveals the communication in corticospinal pathways between the primary motor cortex and muscles. The present study aims to quantify the connectivity between the motor KEYWORDS cortex (EEG signals) and five muscles of the right upper limb (EMG signals) during the manipulation Corticomuscular coherence; of an object. A public dataset (WAY-EEG-GAL) was used which recorded EEG and EMG of twelve object manipulation; healthy subjects who performed movements of reaching, grasping, holding, and replacing objects functional connectivity; of two different weights. Corticomuscular connectivity was established using the coherence hybrid brain-computer interface; motor execution; algorithm for 3 EEG channels and 5 upper-limb muscles varying two objects' weights (0.165 kg electroencephalography and 0.660 kg). Results show that the 0.165 kg weight shows greater coherence between the signals (EEG); electromyography for all analyses than the 0.660 kg weight. Furthermore, the results show that there is a contralateral (EMG); reach-to-grasp and ipsilateral behavior in the EEG-EMG coherence. 1. Introduction Normally, cortical events propagate to the periphery, and the motor cortex also receives input from the Previous reports have shown that synchronization between periphery [6]. neurons in the motor cortex and motor units occurs dur- Corticomuscular coherence is task-dependent, that is ing the performance of a motor task [1]. That mechanism reflects attention and precision, compliance of the was shown using one Magnetoencephalography (MEG) gripped objects, displacement, magnitude of the force, channel recording the cortical motor activity and the sur- and learning processes [2]. In the literature, it has been face electromyogram (EMG) of a contralateral active mus- reported that cortico-muscular coherence is higher in cle during the execution of a muscular voluntary beta (15–30 Hz) and gamma (30–80 Hz) frequency contraction; Corticomuscular connectivity between corti- bands of EEG signals [7,8]. Furthermore, coherence of cal rhythms and rectified EMG confined to the beta (15– the beta band increases during postural tasks by main- 30 Hz) frequency range was evidenced, applying coherence taining sustained motor contractions [6]. Cortico- analysis [2]. Furthermore, movements have long been muscular coherence has also been observed at higher known to induce frequency-specific changes in gamma band frequencies during dynamic movements Electroencephalography (EEG) [3]. Those works evidence or during increasing muscle contraction strength [3]. that there is corticomuscular connectivity defined as Other studies have demonstrated that grasp is a relationship, association, or statistical dependence encoded by neural activity [9]. Kim et al. describe the between EEG and EMG signals resulting from the func- association between EEG and EMG signals in healthy tional integration of the neural and muscular systems [4,5]. individuals when subjects performed active exercise EEG–EMG coherence could be used to examine (finger motion) with movement intention and passive a functional connection between a human brain and mus- exercise without movement intention. These findings cles by calculating the linear relationship of frequency show how movement intention in the patients enhances domain components of EEG and EMG signals. association EEG–EMG, to be implemented in Corticomuscular coherence allows studying of the a rehabilitative training system [10,11]. Various works mechanism of the cerebral cortex’s control of muscle have quantified connectivity in tasks related to holding activity, which reveals the communication in corticospinal positions and manipulating objects by applying differ - pathways between the primary motor cortex and muscles. ent forces [4,12,13]. Nevertheless, brain connectivity CONTACT Cristian D. Guerrero-Mendez crguerrero69@uan.edu.co Bioengineering Research Group, Faculty of Mechanical, Electronic and Biomedical Engineering, Antonio Nariño University, Bogotá, Colombia. © 2022 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. BRAIN-COMPUTER INTERFACES 141 studies have been reported to assess selective attention to evaluate the effect of manipulating an object that [14], and others have used connectivity to identify requires the generation of a low and a high muscle movements, where improvements for classification contraction. rates have been found with connectivity-based methods First of all, in this study, in order to characterize the over standard methods [15,16]. cortical response, Event-Related Desynchronization Neuroprosthetics using neural information from (ERD) and Event-Related Synchronization (ERS) were the patient using EMG and EEG, could be envisaged used. ERD: Increased cellular excitability in thalamocor- to enhance control of rehabilitation devices and tical systems results in a low-amplitude desynchronized improve usability. Thus, the EMG signal from EEG [21,22]. In normal participants, ERD and ERS are a voluntary muscle contraction allows the user to considered to indicate the activation and subsequent transmit the intention of a command, which is recovery of the motor cortex during planning, executing, related to EEG neuronal activity to perform and completing a movement. Therefore, the amount of a motor action. Likewise, with the use of multimo- the cortical activation during a sensory, motor or cogni- dal information, conventional BCI systems could tive task could be indexed by the ERD and ERS concern- improve their performance in terms of classification ing a baseline [23]. To characterize the muscular rate or accuracy, using what is currently called electrical activity, the EMG signals were processed using hybrid Brain–Computer Interfaces (hBCI) [17,18]. segmentation and amplitude estimation techniques by The restoration of functional manipulative activity is calculating the Root Mean Square (RMS) metric. After of special importance in the rehabilitation of people characterizing the cortical responses and muscle electrical with amputation or motor disability at the level of the activation, the estimation of coherence was performed by upper limb, taking into account their involvement in the evaluating different studies mentioned above, where the performance of activities of daily living. The manipula- frequency bands with greater coherence were identified tion of an object requires executing a series of motor and the muscles that present greater coherence with the tasks that include reaching, grasping, holding, and EEG channels. replacing [19]. Such tasks require the coordinated and This article is organized as follows: Section II pre- timed activation of the muscles that intervene in them. sents the experimental methods, including the experi- Identifying and predicting motor tasks during reach and mental protocol, the description of the data processing, grasp for manipulating an object increases the usability, and the algorithm proposal; Section III presents the comfort, and controllability of the prosthetic device results obtained; and finally, Section IV presents the [20]. This translates into greater user acceptance in discussion that includes the conclusions, impact of the rehabilitation systems based on Brain–Computer results, and future works. Interface (BCI) systems, so it is of great importance to identify and predict these events. An adequate identifi - 2. Materials and methods cation of the different stages during the manipulation of an object would allow a lower latency in the response of 2.1. EEG–EMG dataset BCI systems, compared to the identification that is per- This work was implemented using the WAY-EEG-GAL formed when the upper limb is already in the final dataset, which is an open and free available EEG–EMG posture. dataset [24]. The dataset consists of EEG and EMG To our knowledge, currently, it is not clear how the recordings, as well as 3D hand and object position EEG–EMG corticomuscular coherence is related to measurements. Twelve healthy right-handed subjects reach-grasp-lift-hold and replace tasks during the (8 females and 4 males, aged 19–35 years) were recorded manipulation of an object. This work focuses on quan- using 32 EEG channels located at Fp1, Fp2, F7, F3, Fz, tifying the coherence EEG–EMG, under several condi- F4, F8, FC5, FC1, FC2, FC6, T7, C , C , C , T8, TP9, tions that include: 1) Evaluating EEG-bands of extended 3 z 4 CP5, CP1, CP2, CP6, TP10, P7, P3, Pz, P4, P8, PO9, O1, alpha ðαÞ ð6 13Þ Hz, beta ðβÞ ð14 30Þ Hz and Oz, O2, PO10 according to the 10–20 international EEG gamma ðγÞ ð35 50Þ Hz that contributes to coher- placement system. Reference and ground electrodes ence; 2) Evaluating five upper-limb muscles during were connected to FCz and AFz locations, respectively. manipulation of an object involving two weights. In addition, five EMG channels from the following A public dataset was used that records EEG and EMG muscles: 1. Anterior Deltoid (AD), 2. Brachioradialis information during the performance of tasks of reach- (B), 3. Flexor Figitorum (FD), 4. Common Extensor ing and grasping objects of different weights. The ana- Digitorum (CED), and 5. First Dorsal Interosseous lysis was defined at two weights: 0:165 kg and 0:660 kg, 142 C. D. GUERRERO-MENDEZ AND A. F. RUIZ-OLAYA (FDI). EEG signals were recorded with the ActiCap was kept constant with sandpaper. Five series of weights device at a sampling rate of 500 Hz. On the other were used where each series included 22 trials, for a total hand, EMG signals were recorded using five sensors at of 110 trials per participant. Finally, data were taken for a sampling frequency of 4 kHz. Figure 1 shows the each trial until the subject performed the task of repla- experimental setup for the dataset acquisition. cing the object. In the protocol, initially, there is a rest period of 2 seconds before starting the movement where subjects 2.2. Dataset integrity validation maintain the right upper limb leaning on a table, next, the subject receives a visual indication from a LED to Data integrity validation of EEG signals was performed start performing a reaching movement of the right hand to verify contralateral behavior and characterize cortical toward an object. Then, the user grasps it with the index responses when performing the movements with the and thumb fingers; afterward lifts it and holds the object right upper limb. For this purpose, the EEG signals steadily within a circle that is about 5 cm from the table were filtered using an 8th order bandpass Butterworth for 2 seconds until the LED turned off, and subsequently filter between 8 30 Hz, and a Common Average replacing the object and returns the upper limb to the Reference Filter (CAR) to remove related noise on the position indicator, as shown in Figure 1. The object electrodes located in the cortico-motor area. varied in weight and contact surface randomly in 3 Subsequently, the EEG signals’ Power Spectral Density different conditions, the variation of weights was kept (PSD) is calculated for all 32 channels (Figure 2) at (0:165, 0:330, 0:660 kg) and contact surface at (sand- recorded in the two weights to evidence the frequency paper, chamois, silk). To module the objects’ weight, an distribution across the cortico-motor area. For this, the electromagnet was used to present the same weight PSD was calculated using the Fast Fourier Transform between 1 and 4 times and then change it. These weight (FFT) with a Hanning window of 1 second in the fre- variations were performed using two electromagnets of quency range 8 30 Hz for the execution data (2– 0:165 kg and 0:330 kg that to vary the weight in some 8 seconds). conditions only one electromagnet is activated and in In addition, Event-Related Synchronization and others both. For the variation of the contact surface, the Desynchronization (ERS/ERD) was calculated for 3 same logic was maintained, but in this condition, an EEG channels (C , C , and C ) to identify the short- 3 z 4 external person had to intervene on the object. Finally, lasting and localized amplitude attenuation of EEG ten series of approximately 32 trials were recorded, for rhythms within the α and β bands [23]. EEG signals a total of 328 trials per participant in which the weight of were processed for each weight using the open-source the object (0:165, 0:330, 0:660 kg), the contact surface FieldTrip Toolbox (https://www.fieldtriptoolbox.org/ ). (sandpaper, chamois, silk), or both was changed. The ERS/ERD was implemented with a time-sliding This study used 3 EEG channels (C , C , and C ) and window in time steps of 50 ms with frequency intervals 3 z 4 five EMG channels for calculating corticomuscular con- of 1 Hz from 8 to 30 Hz using the Morlet Wavelet nectivity, following the international 10–20 system of method for a time-frequency representation. The ERD EEG electrode placement and muscle location for EMG, is related to a decrease in amplitude and power and the as presented in Figure 2. Data from two different ERS is related to an increase in amplitude and power, weights were used when the subject manipulated the which was calculated using equation (1) with a baseline object (0:165 and 0:660 kg) where the contact surface period of 1 second before the stimulus [23]. ERS/ERD Figure 1. Experimental setup for data acquisition. Modified image and adapted from [24] available under the terms of Attribution 4.0 International Creative Commons License (https://creativecommons.org/licenses/by/4.0/). BRAIN-COMPUTER INTERFACES 143 data is presented as a percentage of the baseline period. when the subject starts to perform the movement, the Nevertheless, only the significant data of the increase or EMG amplitude should be greater for the 0:660 kg decrease power (ERS/ERD) with respect to a reference weight than for the 0:165 kg weight, where the ampli- time of 1 second were taken for the analysis. For this, the tude of the signal estimates the muscle activation when Bootstrap algorithm was used with a significance level of a task is performed [27]. For this, a statistical analysis 0:05 (p< 0:05) according to the description made by was applied using the Mann–Whitney U-test with p ¼ [25]. Additionally, to evaluate contralateral behavior in 0:05 using the mean of all subjects, due to the muscle the execution of the movement, a statistical analysis was activity data of the five muscles and the two weights do performed using the ERS/ERD data to identify signifi - not present a normal distribution according to the cant differences between channels and weights using the Kolmogorov–Smirnov test. In equation (2), N is the Mann–Whitney U-test with p value of 0:05, after length of the window, and X is the EMG signal for checking for abnormal distribution of the data. each sample n. vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi R A u X ERD% ¼ � 100 (1) t RMS ¼ X (2) A n n¼1 In equation (1), A is the power related to the step time after the event between the interest frequency range and R is the reference baseline period. 2.3. Methodology for EEG–EMG coherence analysis Data integrity validation of EMG signals was per- formed to verify the amplitude and characterize the After dataset integrity validation the methodology for muscle electrical activity when the movement was per- EEG–EMG coherence analysis presented in Figure (3) formed. For this, EMG signals for the five muscles was implemented. First, the study is delimited using two (Figure 2) were filtered using an 8th order bandpass weights and selecting subjects. After, artifacts are Butterworth filter between 10 500 Hz due to in that detected in the EEG signal, and channels are selected. frequency band there is higher energy and power of the Next, the signals are pre-processed using filters. After, EMG signal [26]. Subsequently, EMG signals were ana- the EEG and EMG signals are segmented by using lyzed through the signal amplitude estimation using the a Hanning window of 1 second overlapped to 25% for RMS value following the equation (2) with a sliding calculating the coherence, which is represented using window of 300 ms overlapped 50%. In the analysis, significant coherence values. And finally, statistical Figure 2. EEG electrodes layout following the international 10–20 system used in the original WAY-EEG-GAL dataset and EMG electrode placement following the locations of five upper limb muscles. 144 C. D. GUERRERO-MENDEZ AND A. F. RUIZ-OLAYA Figure 3. Block diagram of the implemented methodology in this study to estimate corticomuscular connectivity between EEG and EMG signals during the movement of reach-grasp-lift-hold and replace an object. analysis is performed to find significant differences signal that is 4 kHz, such as reported in the literature between the coherence results for the two weights [30]. For this purpose, only the execution movement is using the paired t-test. analyzed, because at rest there is no movement intention and therefore muscle activation is negligible [4]. Some authors describe that coherence in the resting is insig- 2.3.1. Subjects and channels selection nificant [31]. In this study, a trial rejection process was performed by identifying artifacts in EEG signals, using threshold cri- teria. For this, the EEG signals should be within the range 2.3.3. Connectivity analysis between EEG and EMG of � 350μV [28]; otherwise, trials were rejected contain- signals ing high outliers that were outside the range. If outliers Connectivity analysis between EEG and EMG signals are detected in an EEG channel, the trial is rejected for all was calculated using the coherence algorithm, a measure EEG and EMG channels. EEG channels C , C , and C 3 z 4 of connection or correlation between two signals in the and all EMG channels were analyzed. The subject selec- frequency domain that determines the strength of cor- tion criterion was that amount rejects trials no should be relation in the range of 0–1 [6]. Coherence was imple- greater than 10% of the total trials [29]. In this study, no mented to evaluate the connectivity between the subject was rejected, selecting all participants. electrical activity of the brain and muscles when per- forming an object manipulation movement. To calcu- late coherence, the frequency components of EEG and 2.3.2. Pre-Processing EMG signals in the range 6 50 Hz were extracted by To estimate corticomuscular connectivity, the EEG sig- calculating the auto-spectrum and cross-spectrum using nals for each trial and each subject were filtered using MATLAB (Version R2020b, MathWorks Inc). EEG and a bandpass Butterworth filter of 8th order between 6 EMG signals were segmented by a 1 second Hanning 50 Hz and a CAR filter. The EMG signals were filtered window with 25% overlap at a frequency resolution of 2 using a bandpass Butterworth filter of 8th order the 6 Hz between 6 and 50 Hz for each trial and subject [32]. 50 Hz. Contaminated trials in both EEG and EMG Combinations between channels were obtained by relat- signals were rejected for each subject. A resampling of ing the 3 EEG channels (C , C , C ) with each EMG EEG signal was performed at the sample rate of EMG 3 4 z BRAIN-COMPUTER INTERFACES 145 channel (5 Channels), which generate a total of 15 2.3.4. Coherence analysis combinations between channels, as shown in Figure 3. Three different studies were performed to evaluate the The coherence was calculated according to equation coherence between EEG and EMG signals during the (3) [33]. movement. In the first study, segmentation was per- formed in the frequency bands (α, β, and γ) where all � � � � the EMG channels were averaged for each EEG channel P ðfÞ xy Coh ðfÞ ¼ (3) xy and each frequency band analyzed to determine in P ðfÞ� P ðfÞ xx yy which frequency band significant differences exist between weights. For the second study, the coherence where Coh ðfÞ is the correlation specified in the range xy was evaluated for the combination of each EEG and 0–1 between two signals in the frequencies f. P ðfÞ is xy EMG channel in the frequency band from 6 to 50 Hz the cross-spectrum of x and y, while P ðfÞ and P ðfÞ is xx yy to determine in which muscles there are significant the auto-spectrum of x and y, respectively, which uses differences between weights. Finally, coherence is eval- Welch’s averaged and modified periodogram method of uated for the 0:165 kg weight in each frequency band spectral estimation. In addition, a significant threshold and in each muscle to determine under similar weight was calculated using the coherence values to use in the conditions which muscles are mostly in communication study the significant results of the proposed coherence with each side of the motor cortex during these actions method, equation (4) [33]. using information from EEG channels C and C . For 3 4 each study, the distribution of the data was assessed by L 1 Confidence Limit ¼ 1 ð1 εÞ (4) applying the Shapiro-Wilk test with a p value of 0:05, where all data were found to have a high probability of where L is the number of segments and ε is the con- following a normal distribution. fidence level. However, in this study, L is the number of windows for each weight, and the confidence level used 2.3.5. Statistical analysis is 95% for both weights. Statistical analysis was performed to identify signifi - To extract significant coherence, a conditional- cant differences between the connectivity of the two based classification was implemented where the weights analyzed when the subject performs the coherence spectrum above the confidence limit was manipulation of an object. For this, first, the analyzed. Then, equation (5) was used to obtain the Shapiro–Wilk test with a p value of 0:05 was applied average coherence per subject, channel combination, to determine the normal distribution of the data and and frequency band (α, β, γ). However, before apply- the Levene’s test was applied to determine the homo- ing equation (5) the distribution of the significant geneity of variances between the data. Based on these coherence features was evaluated using the Shapiro- results, the two-sample paired t-test was subsequently Wilk test with a p value of 0:05 because the vector applied using Matlab because the data follow a normal size was kept less than 50 samples. In the results distribution and equal variances. Significance levels for obtained, it is concluded that the significant coher- this test were set at 5% (**) and 10% (*). The alter- ence data in the 3 frequency bands (α, β, γ), the 2 native hypothesis is that one of the two weights present weight conditions (0:165 kg, 0:660 kg), the channels greater coherence than the other weight in the first two combination and the subjects follow a normal distri- studies performed according to section 2.3.4, and the bution with p value< 0:05. null hypothesis is the opposite. For the third study, the alternative hypothesis is that one of the two EEG channels (C and C ) presents greater coherence in 3 4 CS ¼ jXðfÞj (5) each of the 3 frequency bands and the 5 muscles, and f¼1 the null hypothesis is the opposite. For these purposes, where the CS is the average coherence spectrum sig- the results are presented in bar figures with confidence nificant, f is the specific frequency significant, N is the intervals of 95% to evaluate the significant variation number of significant coherences by the bands (α, β between subjects. and γ) and X is the value of significant coherence. In this study, equation (5) was used for each indepen- 3. Results dent frequency band, where to estimate coherence over the entire (6 50) Hz spectrum, connectivity Subjects perform an upper-limb movement that information was extracted from each frequency band involves reaching, grasping, lifting, holding, and repla- and averaged. cing an object of two different weights. Taking into 146 C. D. GUERRERO-MENDEZ AND A. F. RUIZ-OLAYA account that each subject performs the motor tasks at other hand, the duration of the movement for the a different time, quantification of the time variation weight of 0:165 kg among the mean of the subjects between subjects and between the tasks was obtained corresponds to 6:88� 0:40 and for the weight of 0:660 during the manipulation of the object. kg corresponds to 7:55� 0:57. Figure 4 shows the time variation between subjects In this article, data were taken for each trial until the for each task performed in the manipulation of the subject performed the task of replacing the object. First, object (Reach, Grasp, Lift, Hold and Replace), the time obtained results of the data integrity validation were variation of the whole task performed until the replace- presented. Next, coherence between EEG and EMG ment of the object between subjects (blue for the weight signals was shown in frequency bands and upper-limb of 0:165 kg and red for the weight of 0:660 kg), and the muscles involved and finally, the study is presented to distribution of each task along with the movement of demonstrate which muscles are mostly in communica- manipulating an object for the weight of 0:165 kg, where tion with the hemispheres of the motor cortex. the red color represents the reaching task, blue repre- sents the grasp, green represents the lift, yellow repre- 3.1. ERD/ERS quantification sents holding and magenta represents replacing the object. In Figure 4, the rest period corresponds to the Figure 5 presents the head maps for each analyzed first two seconds, and the time after the replace task weight using all trials and subjects spanning the fre- corresponds to the release task. In the figure, the task quency range 8 30 Hz from 2 8 seconds. with the longest duration within the manipulation of the According to the figure, the movement was distributed object was the holding that lasts approximately 2 sec- in all cortical cortex by the power decrease presented in onds, and the task with the shortest duration is the the head map, where this effect could be presented by lifting that lasts approximately 0:30 seconds. On the Figure 4. Duration of segments and trials. (A) Average of all subjects in each task involved in the movement. (B) Average of all subjects for the total duration of trials taken until the replacing the object. (C) Distribution of duration of the 0:165 kg weight for each task involved in the manipulation of the object. BRAIN-COMPUTER INTERFACES 147 Figure 5. Power Spectral Density (PSD) for all EEG channels (32 channels) for both weights using all trials and subjects analyzed. PSD magnitude extends in the range ð0 20Þ of magnitude expressed in the head map for the execution movement (2 8 seconds) in the 8 30 Hz frequency band. the movement complexity. Finally, the EEG channels before the vertical dotted line, as shown in Figure 6. chosen for the coherence analysis were C ; C and C After 2 seconds, a LED turns on and subjects start 3 4 z due to the main location in the cortical-motor cortex. performing the movement; the ERD/ERS data reveals The ERD/ERS (relative power) was calculated to a significant ERD after the 2 seconds in all 3 channels observe the change in power when the subject performs according to the results of the bootstrap algorithm [25]. the movement with the two weights. The EEG time- Figure 7 shows the ERD/ERS as a representation of frequency representations were calculated for C ; C the relative amplitude (%) versus the time (s) taken in 3 4 and C channels in the 8 30Hz frequency band the execution of the movement. The results are pre- throughout the movement. Figure 6 shows time- sented for the 3 EEG channels (C , C y C ) segmenting 3 z 4 frequency maps of ERD/ERS percentage values across the information α (8 13 Hz) and β (14 30 Hz) all trials (55 trials per weight) and subjects (12 subjects) frequency bands, and presenting for the same channel recorded for each weight. The figure is divided by α and the variation of the two weights, where the weight 0:165 β bands with a bar color between ð 50%; 50%Þ. During kg is presented in blue color and the weight of 0:660 kg the rest segment (0 2 seconds) there is no ERD/ERS in red. However, for each frequency band, the mean and Figure 6. Time-frequency maps ERS/ERD for C ; C and C channels for two weights using all trials and subjects analyzed. The black 3 z 4 vertical dotted line represents the time turning on the LED and subjects start to perform the movement. The black horizontal dotted line presents α and β bands separation. 148 C. D. GUERRERO-MENDEZ AND A. F. RUIZ-OLAYA Figure 7. ERD/ERS value and its statistics. (A) Relative amplitude (%) for the 3 channels and two weights by segmenting the information into (α) and (β) frequency bands. The black line of vertical dots represents the moment when the LED is turned on and the subjects start to perform the movement. (B) Average and standard deviation of the relative amplitude of the 3 channels and the two weights for the two frequency bands. (**) Represents that there are significant differences (p< 0:05) between the channels. standard deviation as shaded regions for the 3 channels was performed. In the literature, it is reported that when in the two weights are presented, as shown in the bars of a user performs a right-hand motor task, an ERD is Figure 7. presented in the contralateral side (C channel) [23]. To evaluate the contralateral behavior in the execu- tion of the movement, a statistical analysis was per- 3.2. Estimation and analysis of EMG amplitude formed to identify significant differences between the channels and between the weights. First, an analysis was Changes in the electrical activity of the upper limb performed using the Kolmogorov–Smirnov test muscles (5 muscles) when the subject performs the (Samples > 50 data) to confirm that the behavior of movement was quantified. This muscle activity is the ERD/ERS for each channel and weight have a high expected to provide greater activation in the 0:660 kg probability of having a normal distribution. weight than in the 0:165 kg weight. The RMS value of Subsequently, Mann-Whitney U-test was applied with the EMG signal in the five muscles was calculated, as a value of p ¼ 0:05 since the data did not present shown in Figure 8, where the 0:165 kg weight is pre- a normal distribution. According to the results, there sented by a blue line and the 0:660 kg weight is pre- are significant differences (p< 0:05) in the relative sented by a red line. The average value of muscle amplitude of the two weights in the 3 channels for the electrical activity for the 5 muscles of all recorded sub- 2 frequency bands. Additionally, in α there are signifi - jects is presented along with the standard deviation in cant differences (p< 0:05) between channels C and C , shaded regions for the two weights, as shown in 3 z and C and C for the two weights, and in β there are Figure 8. This allows us to observe how the change in 4 z significant differences (p< 0:05) in the relative ampli- muscle activation is for all subjects. However, according tude for the 3 channels (C and C , C and C , C to the figure and the statistical analysis performed using 3 z 3 4 4 and C ). the Mann-Whitney U-test with a significance value of Considering the results presented in Figures 6 and 7, 5%, the 0:660 kg weight presents greater muscle activa- the 0:660 kg weight presents a greater decrease in power tion than the 0:165 kg weight evaluated for each muscle than the 0:165 kg weight and that in the C channel individually with a p value< 0:05. As expected, there there is a greater decrease in power compared to the C is no muscle activation before initiating the movement and C channels, which could confirm the contralateral presented as a vertical dotted black line at second 2. All effect that occurs when movements are performed with EMG channels (5 channels or muscles) were selected to the right hand. That is, the C channel presents greater evaluate the coherence between EEG and EMG signals prominence than the C channel when the movement in the analyzed movement. 4 BRAIN-COMPUTER INTERFACES 149 Figure 8. Root Mean Square (RMS) value for all subjects with the standard deviation presented in shaded regions for the two weights in the 5 muscles recorded. The black line of vertical dots represents the time at which the LED is turned on and at which the movement is initiated. The blue line presents the RMS value for the 0:165 kg weight and the red line for the 0:660 kg weight. 3.3. Coherence for frequency bands and muscles shown in Figure 9(A). As an error bar, confidence intervals of 95% were calculated with respect to all Connectivity was quantified over the time interval from subjects for each weight and EEG channel. In addition, the start of the reach task (LED on) to the end of the the results of the statistical analysis using the two- replace task (see Figure 4). For this, the coherence sample paired t-test are presented, where (*) corre- spectrum for each EEG channel averaged over all sponds to the 10% significance level and (**) corre- EMG channels in the frequency bands (α, β, and γ) is sponds to the 5% significance level. Figure 9. Coherence and statistical analysis for two weights quantified over the time interval from the start of the reach task (LED on) to the end of the replace task. (A) Coherence between each EEG channel (C , C , C ) and all EMG channels in the 3 frequency bands for 3 z 4 all subjects in the two weights. (B) Coherence between each EEG and EMG channel in the averaged of (α, β and γ) frequency bands for all subjects in the two weights. The confidence intervals of 95% regarding all subjects are presented. (*) corresponds to significant differences between weights at a significant level of 10% and (**) corresponds to a significant level of 5%. 150 C. D. GUERRERO-MENDEZ AND A. F. RUIZ-OLAYA The α band for the 3 EEG channels has a low coher- Finally, for C channel, there were significant differences ence concerning other bands. On the other hand, the with a 5% of significance considering the B and FD greatest coherence was found in the β band for the 3 muscles with p value ðp< 0:05; p ¼ 0:00Þ and channels, as shown in Figure 9(A). The γ band presents ðp< 0:05; p ¼ 0:04Þ, respectively. Brachioradial muscle a coherence spectrum between the values of α and β presents significant differences between weights for the band coherence. It is important to highlight that the 3 EEG channels indicating that B muscle has an impor- coherence with the 0:165 kg weight is greater than the tant change regarding the connectivity when the weight is 0:660 kg weight for all bands at the 3 channels (C , C , changed. 3 z C ). Therefore, statistically significant differences were found for some cases as in the C channel where there 3.4. Analysis of coherence according to are significant differences ðp< 0:1; p ¼ 0:08Þ in the hemispheric laterality coherence between weights for α band. Moreover, for the C channel there are significant differences of Figure 10 presents the muscles that are mostly in com- ðp< 0:05; p ¼ 0:04Þ in the α band between weights and munication with each side of the motor cortex during in the C channel are significant differences in the α and the manipulation of an object using EEG–EMG con- β band with p values of ðp< 0:1; p ¼ 0:09Þ and nectivity quantified over the time interval from the start ðp< 0:1; p ¼ 0:08Þ, respectively. of the reach task (LED on) to the end of the replace task However, the connectivity where muscle connections (see Figure 4). Initially, only the coherence calculated were varied regard each EEG channel (C , C , C ) aver- for the weight of 0:165 kg was used, where the variation 3 z 4 aged over the 3 frequency bands (α, β, γ) is presented in of coherence for the 5 recorded muscles and the 3 Figure 9(B). This figure allows identifying how connec- frequency bands in channels C and C are presented. 3 4 tivity changes for each muscle, taking into account that In the figure, C channel is represented using yellow each muscle presents a different electrical activation color and C channel using green color. A statistical depending on the task performed in the manipulation analysis using the two-sample paired t-test is presented of the object which is controlled by the Central Nervous in order to show whether there are significant differ - System (CNS). The muscles are presented and labeled as ences in the coherence of the 2 EEG channels with Anterior Deltoid (AD), Brachioradial (B), Flexor respect to each frequency band and recorded muscles, Digitorum (FD), Common Extensor Digitorum where (*) corresponds to the 10% significance level and (CED), and First Dorsal Interosseous (FDI). (**) corresponds to the 5% significance level. Connectivity was quantified over the time interval The EEG channels (C and C ) present significant 3 4 from the start of the reach task (LED on) to the end of differences for the AD muscle in the α band at the replace task (see Figure 4). a significance level of 5% with p value The coherence for the weight of 0:165 kg in the (p< 0:05; p ¼ 0:03). For the B muscle these differences Brachioradialis muscle was higher than other muscles are found in the β band with 5% significance and in the for all EEG channels. Coherence of the weight of 0:660 γ band with 10% significance with p value of kg for Brachioradialis muscle is lower than others in the 3 (p< 0:05; p ¼ 0:02) and (p< 0:1; p ¼ 0:09), respec- EEG channels. Nevertheless, the coherence for the 0:660 tively. However, for the FDI muscle the difference is in kg weight was lower than the 0:165 kg weight for all the α band at the 10% significance level with p value muscles and EEG channels, except for the connectivity (p< 0:05; p ¼ 0:07). The muscles that do not present of C with the FDI muscle where the coherence for both significant differences between the channels are the FD weights is approximately equal, and the connectivity and CED in the 3 frequency bands. According to the between C and the AD muscle where the coherence at results, there is a contralateral behavior in the α band of the 0:660 kg weight is higher than the 0:165 kg weight. the AD muscle and in the γ band of the B muscle, and However, these results are not significant according to the there is an ipsilateral behavior in the β band of the statistical analysis. Significant differences were found as B muscle and in the α band of the FDI muscle. follows: For the C channel, there were differences with a significant level of 5% for the AD and B muscle with a 4. Discussion p value for both of ðp< 0:05; p ¼ 0:04Þ; for the C channel, there were significant differences with a 5% of Obtained results of EEG–EMG coherence allow deter- significance considering the B muscle with a p value of mine how functional corticomuscular connectivity ðp< 0:05; p ¼ 0:03Þ and differences with a significant occurs between brain cortical area and specific muscles level of 10% for the CED muscle ðp< 0:1; p ¼ 0:09Þ. when a reach-grasp-lift-hold and replace an object BRAIN-COMPUTER INTERFACES 151 Figure 10. Coherence for the 0:165 kg weight in the EEG channels (C and C ), the 5 muscles and the 3 frequency bands (α, β and γ). 3 4 The confidence intervals of 95% regarding all subjects are presented. Connectivity was quantified over the time interval from the start of the reach task (LED on) to the end of the replace task. (*) corresponds to significant differences between weights at a significant level of 10% and (**) corresponds to a significant level of 5%. motor task was performed. The study aims to investi- signals. Specifically, significant differences were found gate how EEG and EMG connectivity changes during of; a) coherence of C with Anterior Deltoid and the manipulation of an object having a low and high Brachioradial muscles; b) coherence of C with weight (0:165 kg and 0:660 kg). Brachioradial muscle; c) coherence of C with Flexor Cortical responses characterization of EEG data Digitorum and Brachioradial muscles. Furthermore, the showed that when subjects perform the activity with the greatest coherence was found in the β band for the 3 0:660 kg weight presents a greater decrease in power than EEG channels, which is in agreement with the results of the 0:165 kg weight in the α and β frequency bands previous studies [10,12,13,35]. Those studies reported presenting a contralateral behavior. As mentioned that subjects showed a peak in the coherence spectra above, the manipulation of an object involves different between 15 – 30 Hz bandwidth during a hold task motor tasks such as reaching, grasping, lifting, holding, involving stable force production where the coherence and replacing, which generate changes in brain signals between signals may be affected by the performance in according to the cellular excitability that can form each the development of motor tasks. On the other hand, in motor task. Being a complex task because multiple joints future studies, we propose to evaluate how the connec- are involved in performing the movement, larger brain tivity is for each task involved in the manipulation of the neural networks are needed in the processing of the object, in order to determine the delay times that occur information presented in the movement, so an ERD is in sending the information to execute each action present around the entire execution of the movement involved in the movement [4]. [23]. On the other hand, this type of task involves Studies have been published in the literature that has a number of different muscles, where muscle activation determined contralaterality with respect to corticomus- differs throughout the movement because each muscle is cular connectivity in a single muscle for upper limb in charge of specific functions in the movement. As it is movements. Effects have been found that relate to a combination of different tasks, the muscle synergies there being greater connectivity quantified using the (muscle groups) are coactivated in a coordinated way coherence algorithm, in the contralateral hemisphere through the CNS to perform a complex task. In other recorded by the channels (C and C ) establishing con- 3 4 words, the CNS activates the muscle groups to generate nectivity with only the Flexor-Digitorum-Superficialis a different motor command [34]. (FDS) muscle when performing right and left-hand Significant differences in the movement of manipula- movements in healthy subjects [36]. However, that tion of an object with two different weights were found study quantifies corticomuscular connectivity in an by quantifying the connectivity of EEG and EMG upper limb movement when it reaches its final posture. 152 C. D. GUERRERO-MENDEZ AND A. F. RUIZ-OLAYA This study involves a multi-joint movement since the [17,41]. Therefore, these provide a benefit and a clear subjects perform a movement of control and precision advantage over any single brain signal acquisition with a synergy of synchronized movements to perform modality. the action of manipulation of an object, which accord- In this work, EMG-EEG coherence is higher in the ing to cortical topographies shows an ERD in the entire beta frequency band were demonstrated, individualized motor cortex of the brain. differences in coherence have been found according to Corticomuscular coherence has been shown to be each muscle involved during the reaching and grasping task-dependent, that it reflects attention and precision, movement, showing that the brachioradialis muscle is compliance of the gripped objects, displacement, mag- the most involved in the connectivity due to the signifi - nitude of the force, and learning processes [2]. Thus, cant differences found in the EEG channels. effects such as motor adaptation or neuronal noise Additionally, it has been demonstrated how the EMG- sources may be taking place in the modulation of con- EEG coherence could change depending on the force nectivity during the manipulation of the object invol- exerted to grasp an object of different weight, and it has ving two weights [21,37]. Furthermore, corticomuscular been determined which muscles are mostly in commu- coherence has widely been considered a motor mechan- nication with each side of the cerebral hemispheres. ism reflecting efferent oscillatory activity descending Finally, the results presented in this study allow us to from the primary motor cortex via corticospinal path- conclude that coherence is significantly higher at ways, but it has been reported works that suggest that a weight of 0:165 kg than at a weight of 0:660 kg for ascending sensory activity also contributes to the cou- the α band in the C , C and in α and β for the C 3 z 4 pling between sensorimotor brain regions and muscle channel. In addition, these differences are found in the [38]. Such phenomenon could be taking part in C -AD, C -B, C -B, C -CED, C -B and C -FD channel 3 3 z z 4 4 obtained significant differences of coherence during combinations. These results are of great importance in reach-to-grasp movement for two different weights. rehabilitation engineering applications because cortico- The results from the coherence analysis show that muscular connectivity can be used as a descriptor to there are significant differences in EEG–EMG connectiv- improve the classification rates, usability, and control of ity in the manipulation of an object of two different prostheses based on BCI systems. In this case, the appli- weights (low and high) mainly in the α and β band, and cation of prostheses for weightlifting identification. in the AD, B, FD and CED muscles in different EEG However, as future studies, we propose to evaluate channels. These differences are centered on the 0:165 kg other methods such as Granger Causality to establish weight presents higher coherence than the other weight. connectivity and delay times between EEG and EMG Therefore, by applying these data in a multimodal acqui- information when performing other types of move- sition system, a high classification could be obtained in ments involved in activities of daily living. As well as the identification of motor tasks, not only when the upper using corticomuscular connectivity as a rehabilitation limb has already reached its final position but also in the method for people with disabilities, for implementing identification of different tasks that are involved in computational methods based on connectivity to a given action such as the manipulation of an object. improve identification rates using techniques such as The findings of this study could be of great impor- filter banks. tance in the development of Hybrid Brain-Computer Interface (hBCI) systems because currently, the imple- Acknowledgment mentation of a BCI system based exclusively on EEG signals generates low accuracy in pattern decoding, due The authors would like to thank the Antonio Nariño to the fact that it only generates one control command University, for the support of the development in this work. [4,18]. However, hBCIs have 3 different ways to be implemented characterized in multiple biosignals Disclosure statement such as EEG–EMG, paradigms and multisensory sti- muli [39] which allows increasing the reliability of the No potential conflict of interest was reported by the author(s). system in terms of the percentage of command classi- fication in specific tasks and the generalization of applications in fields such as Neuro-Rehabilitation Funding [17, 18, 39, 40]. Mainly, the advantages found encom- This work was supported by Antonio Nariño University pass (1) higher BCI classification accuracy, (2) higher (UAN) under the grant N.2021020 ‘Model based on multi- number of brain commands for control application, modal EEG–EMG information to improve motion intention and (3) shorter brain command detection time decoding for the control of a BCI system’. BRAIN-COMPUTER INTERFACES 153 [14] Borhani S, Abiri R, Jiang Y, et al. Brain connectivity ORCID evaluation during selective attention using EEG-based Cristian D. Guerrero-Mendez http://orcid.org/0000-0003- brain-computer interface. Brain-Comput Interfaces. 1019-9423 2019;6(1–2):25–35. Andres F. Ruiz-Olaya http://orcid.org/0000-0002-5883- [15] Guerrero-Mendez CD, Blanco-Diaz CF, and Ruiz- Olaya AF. 2021. Identification of motor imagery tasks using power-based connectivity descriptors from EEG signals. In 2021 XXIII symposium on image, signal pro- cessing and artificial vision (STSIVA) Popayán, References Colombia (pp. 1–6). [1] Conway B, Halliday D, Farmer S, et al. Synchronization [16] Li X, Samuel OW, Zhang X, et al. A motion- classifica - between motor cortex and spinal motoneuronal pool tion strategy based on sEMG-EEG signal combination during the performance of a maintained motor task in for upper-limb amputees. J Neuroeng Rehabil. 2017;14 man. J Physiol. 1995;489(3):917–924. (1):1–13. [2] Andrykiewicz A, Patino L, Naranjo JR, et al. [17] Hong K-S, and Khan MJ. Hybrid brain–computer Corticomuscular synchronization with small and large interface techniques for improved classification accu- dynamic force output. BMC Neurosci. 2007;8(1):1–12. racy and increased number of commands: a review. [3] Rizzo R, Zhang X, Wang JW, et al. Network physiology Front Neurorobot. 2017 July;11 35 . of cortico–muscular interactions. Front Physiol. [18] Ma T, Li H, Deng L, et al. The hybrid BCI system for 2020;11. DOI:10.3389/fphys.2020.558070 movement control by combining motor imagery and [4] Kim B, Kim L, Kim Y-H, et al. Cross-association ana- moving onset visual evoked potential. J Neural Eng. lysis of EEG and EMG signals according to movement 2017 February;14(2):026015. intention state. Cognit Syst Res. 2017;44:1–9. [19] Betti S, Zani G, Guerra S, et al. Reach-to-grasp move- [5] Zhuang KZ, Lebedev MA, Nicolelis MA. Joint ments: a multimodal techniques study. Front Psychol. cross-correlation analysis reveals complex, 2018;9:990. time-dependent functional relationship between corti- [20] Batzianoulis I, El-Khoury S, Pirondini E, et al. cal neurons and arm electromyograms. J Neurophysiol. EMG-based decoding of grasp gestures in 2014;112(11):2865–2887. reaching-to-grasping motions. Rob Auton Syst. [6] Liu J, Sheng Y, Liu H. Corticomuscular coherence and 2017;91:59–70. its applications: a review. Front Hum Neurosci. [21] Mendez-Balbuena I, Huethe F, Schulte-Mönting J, et al. 2019;13:100. Corticomuscular coherence reflects interindividual dif- [7] Mima T, Hallett M. Electroencephalographic analysis of ferences in the state of the corticomuscular network cortico-muscular coherence: reference effect, volume during low-level static and dynamic forces. Cereb conduction and generator mechanism. Clin Cortex. 2012;22(3):628–638. Neurophysiol. 1999;110(11):1892–1899. [22] Steriade M, Llinás RR. The functional states of the [8] Rossiter HE, Eaves C, Davis E, et al. Changes in the thalamus and the associated neuronal interplay. location of cortico-muscular coherence following Physiol Rev. 1988;68(3):649–742. stroke. NeuroImage Clin. 2013;2:50–55. [23] Pfurtscheller G, Da Silva FL. Event-related EEG/MEG [9] Sburlea AI, Müller-Putz GR. Exploring representations synchronization and desynchronization: basic principles. of human grasping in neural, muscle and kinematic Clin Neurophysiol. 1999;110(11):1842–1857. signals. Sci Rep. 2018;8(1):1–14. [24] Luciw MD, Jarocka E, Edin BB. Multi-channel EEG [10] Boonstra TW, van Wijk BC, Praamstra P, et al. recordings during 3,936 grasp and lift trials with vary- Corticomuscular and bilateral EMG coherence reflect ing weight and friction. Sci Data. 2014;1(1):1–11. distinct aspects of neural synchronization. Neurosci [25] Graimann B, Huggins JE, Levine SP, et al. Lett. 2009;463(1):17–21. Visualization of significant ERD/ERS patterns in mul- [11] Lóopez-Larraz E, Birbaumer N, and Ramos- tichannel EEG and ECoG data. Clin Neurophysiol. Murguialday A. 2018. A hybrid EEG- EMG BMI 2002;113(1):43–47. improves the detection of movement intention in cor- [26] Merletti R, Botter A, Troiano A, et al. Technology and tical stroke patients with complete hand paralysis. In instrumentation for detection and conditioning of the 2018 40th annual international conference of the IEEE surface electromyographic signal: state of the art. Clin engineering in medicine and biology society (EMBC) Biomech. 2009;24(2):122–134. Honolulu, HI, USA (pp. 2000–2003). [27] Renshaw D, Bice MR, Cassidy C, et al. A comparison of [12] Gao L, Wu H, Cheng W, et al. Enhanced descending three computer-based methods used to determine EMG corticomuscular coupling during hand grip with static signal amplitude. Int J Exercise Sci. 2010;3(1):43. force compared with enhancing force. Clin EEG [28] Dimigen O. Optimizing the ICA-based removal of ocu- Neurosci. 2020;52(6):436–443. lar EEG artifacts from free viewing experiments. [13] Kilner J, Baker S, Salenius S, et al. Task- dependent NeuroImage. 2020;207:116117. modulation of 15-30 Hz coherence between rectified [29] Krauledat M, Dornhege G, Blankertz B, et al. EMGs from human hand and forearm muscles. Robustifying EEG data analysis by removing outliers. J Physiol. 1999;516(2):559–570. Chaos Complexity Letters. 2007;2(3):259–274. 154 C. D. GUERRERO-MENDEZ AND A. F. RUIZ-OLAYA [30] Xi X, Ma C, Yuan C, et al. Enhanced EEG-EMG coher- [36] Chowdhury A, Raza H, Meena YK, et al. An EEG- EMG ence analysis based on hand movements. Biomed Signal correlation-based brain-computer interface for hand Process Control. 2020;56:101727. orthosis supported neuro-rehabilitation. J Neurosci [31] Hashimoto Y, Ushiba J, Kimura A, et al. Correlation Methods. 2019;312:1–11. between EEG-EMG coherence during isometric con- [37] Trenado C, Mendez-Balbuena I, Manjarrez E, et al. traction and its imaginary execution. Acta Neuro biol Enhanced corticomuscular coherence by external sto- Exp (Wars). 2010;70(1):76–85. chastic noise. Front Hum Neurosci. 2014;8:325. [32] Bortel R, Sovka P. EEG-EMG coherence enhancement. [38] Beck MM, Spedden ME, Lundbye-Jensen J. Signal Process. 2006;86(7):1737–1751. Reorganization of functional and directed cortico- [33] Rosenberg J, Amjad A, Breeze P, et al. The Fourier muscular connectivity during precision grip from approach to the identification of functional coupling childhood to adulthood. Sci Rep. 2021;11(1):1. between neuronal spike trains. Prog Biophys Mol Biol. November. 1989;53(1):1–31. [39] Li Z, Zhang S, Pan J. Advances in hybrid brain-computer [34] Torricelli D, Barroso F, and Coscia M, et al. 2015 interfaces: principles, design, and applications. Comput October. Muscle synergies in clinical practice: theore- Intell Neurosci. 2019 October;2019:1–9. tical and practical implications Pons, J, Raya, R, and [40] Amiri S, Fazel-Rezai R, Asadpour V. A review of hybrid González, J. In: Biosystems & biorobotics. 251–272. Switzerland: Springer International Publishing. brain-computer interface systems. Adv Hum Comput [35] Kristeva R, Patino L, Omlor W. Beta-range cortical motor Interact. 2013;2013:1–8. spectral power and corticomuscular coherence as [41] Banville H, Falk T. Recent advances and open chal- a mechanism for effective corticospinal interaction during lenges in hybrid brain- computer interfacing: steady-state motor output. Neuroimage. 2007;36 a technological review of non-invasive human (3):785–792. research. Brain-Comput Interfaces. 2016;3(1):9–46. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Brain-Computer Interfaces Taylor & Francis

Coherence-based connectivity analysis of EEG and EMG signals during reach-to-grasp movement involving two weights

Coherence-based connectivity analysis of EEG and EMG signals during reach-to-grasp movement involving two weights

Abstract

Corticomuscular coherence allows studying the mechanism of the cerebral cortex’s control of muscle activity, which reveals the communication in corticospinal pathways between the primary motor cortex and muscles. The present study aims to quantify the connectivity between the motor cortex (EEG signals) and five muscles of the right upper limb (EMG signals) during the manipulation of an object. A public dataset (WAY-EEG-GAL) was used which recorded EEG and EMG of twelve healthy subjects...
Loading next page...
 
/lp/taylor-francis/coherence-based-connectivity-analysis-of-eeg-and-emg-signals-during-XPVSW3vpeL
Publisher
Taylor & Francis
Copyright
© 2022 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group.
ISSN
2326-2621
eISSN
2326-263x
DOI
10.1080/2326263X.2022.2029308
Publisher site
See Article on Publisher Site

Abstract

BRAIN-COMPUTER INTERFACES 2022, VOL. 9, NO. 3, 140–154 https://doi.org/10.1080/2326263X.2022.2029308 ORIGINAL RESEARCH Coherence-based connectivity analysis of EEG and EMG signals during reach-to- grasp movement involving two weights Cristian D. Guerrero-Mendez and Andres F. Ruiz-Olaya Bioengineering Research Group, Faculty of Mechanical, Electronic and Biomedical Engineering, Antonio Nariño University, Bogotá, Colombia ABSTRACT ARTICLE HISTORY Received 17 August 2021 Corticomuscular coherence allows studying the mechanism of the cerebral cortex’s control of Accepted 7 January 2022 muscle activity, which reveals the communication in corticospinal pathways between the primary motor cortex and muscles. The present study aims to quantify the connectivity between the motor KEYWORDS cortex (EEG signals) and five muscles of the right upper limb (EMG signals) during the manipulation Corticomuscular coherence; of an object. A public dataset (WAY-EEG-GAL) was used which recorded EEG and EMG of twelve object manipulation; healthy subjects who performed movements of reaching, grasping, holding, and replacing objects functional connectivity; of two different weights. Corticomuscular connectivity was established using the coherence hybrid brain-computer interface; motor execution; algorithm for 3 EEG channels and 5 upper-limb muscles varying two objects' weights (0.165 kg electroencephalography and 0.660 kg). Results show that the 0.165 kg weight shows greater coherence between the signals (EEG); electromyography for all analyses than the 0.660 kg weight. Furthermore, the results show that there is a contralateral (EMG); reach-to-grasp and ipsilateral behavior in the EEG-EMG coherence. 1. Introduction Normally, cortical events propagate to the periphery, and the motor cortex also receives input from the Previous reports have shown that synchronization between periphery [6]. neurons in the motor cortex and motor units occurs dur- Corticomuscular coherence is task-dependent, that is ing the performance of a motor task [1]. That mechanism reflects attention and precision, compliance of the was shown using one Magnetoencephalography (MEG) gripped objects, displacement, magnitude of the force, channel recording the cortical motor activity and the sur- and learning processes [2]. In the literature, it has been face electromyogram (EMG) of a contralateral active mus- reported that cortico-muscular coherence is higher in cle during the execution of a muscular voluntary beta (15–30 Hz) and gamma (30–80 Hz) frequency contraction; Corticomuscular connectivity between corti- bands of EEG signals [7,8]. Furthermore, coherence of cal rhythms and rectified EMG confined to the beta (15– the beta band increases during postural tasks by main- 30 Hz) frequency range was evidenced, applying coherence taining sustained motor contractions [6]. Cortico- analysis [2]. Furthermore, movements have long been muscular coherence has also been observed at higher known to induce frequency-specific changes in gamma band frequencies during dynamic movements Electroencephalography (EEG) [3]. Those works evidence or during increasing muscle contraction strength [3]. that there is corticomuscular connectivity defined as Other studies have demonstrated that grasp is a relationship, association, or statistical dependence encoded by neural activity [9]. Kim et al. describe the between EEG and EMG signals resulting from the func- association between EEG and EMG signals in healthy tional integration of the neural and muscular systems [4,5]. individuals when subjects performed active exercise EEG–EMG coherence could be used to examine (finger motion) with movement intention and passive a functional connection between a human brain and mus- exercise without movement intention. These findings cles by calculating the linear relationship of frequency show how movement intention in the patients enhances domain components of EEG and EMG signals. association EEG–EMG, to be implemented in Corticomuscular coherence allows studying of the a rehabilitative training system [10,11]. Various works mechanism of the cerebral cortex’s control of muscle have quantified connectivity in tasks related to holding activity, which reveals the communication in corticospinal positions and manipulating objects by applying differ - pathways between the primary motor cortex and muscles. ent forces [4,12,13]. Nevertheless, brain connectivity CONTACT Cristian D. Guerrero-Mendez crguerrero69@uan.edu.co Bioengineering Research Group, Faculty of Mechanical, Electronic and Biomedical Engineering, Antonio Nariño University, Bogotá, Colombia. © 2022 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. BRAIN-COMPUTER INTERFACES 141 studies have been reported to assess selective attention to evaluate the effect of manipulating an object that [14], and others have used connectivity to identify requires the generation of a low and a high muscle movements, where improvements for classification contraction. rates have been found with connectivity-based methods First of all, in this study, in order to characterize the over standard methods [15,16]. cortical response, Event-Related Desynchronization Neuroprosthetics using neural information from (ERD) and Event-Related Synchronization (ERS) were the patient using EMG and EEG, could be envisaged used. ERD: Increased cellular excitability in thalamocor- to enhance control of rehabilitation devices and tical systems results in a low-amplitude desynchronized improve usability. Thus, the EMG signal from EEG [21,22]. In normal participants, ERD and ERS are a voluntary muscle contraction allows the user to considered to indicate the activation and subsequent transmit the intention of a command, which is recovery of the motor cortex during planning, executing, related to EEG neuronal activity to perform and completing a movement. Therefore, the amount of a motor action. Likewise, with the use of multimo- the cortical activation during a sensory, motor or cogni- dal information, conventional BCI systems could tive task could be indexed by the ERD and ERS concern- improve their performance in terms of classification ing a baseline [23]. To characterize the muscular rate or accuracy, using what is currently called electrical activity, the EMG signals were processed using hybrid Brain–Computer Interfaces (hBCI) [17,18]. segmentation and amplitude estimation techniques by The restoration of functional manipulative activity is calculating the Root Mean Square (RMS) metric. After of special importance in the rehabilitation of people characterizing the cortical responses and muscle electrical with amputation or motor disability at the level of the activation, the estimation of coherence was performed by upper limb, taking into account their involvement in the evaluating different studies mentioned above, where the performance of activities of daily living. The manipula- frequency bands with greater coherence were identified tion of an object requires executing a series of motor and the muscles that present greater coherence with the tasks that include reaching, grasping, holding, and EEG channels. replacing [19]. Such tasks require the coordinated and This article is organized as follows: Section II pre- timed activation of the muscles that intervene in them. sents the experimental methods, including the experi- Identifying and predicting motor tasks during reach and mental protocol, the description of the data processing, grasp for manipulating an object increases the usability, and the algorithm proposal; Section III presents the comfort, and controllability of the prosthetic device results obtained; and finally, Section IV presents the [20]. This translates into greater user acceptance in discussion that includes the conclusions, impact of the rehabilitation systems based on Brain–Computer results, and future works. Interface (BCI) systems, so it is of great importance to identify and predict these events. An adequate identifi - 2. Materials and methods cation of the different stages during the manipulation of an object would allow a lower latency in the response of 2.1. EEG–EMG dataset BCI systems, compared to the identification that is per- This work was implemented using the WAY-EEG-GAL formed when the upper limb is already in the final dataset, which is an open and free available EEG–EMG posture. dataset [24]. The dataset consists of EEG and EMG To our knowledge, currently, it is not clear how the recordings, as well as 3D hand and object position EEG–EMG corticomuscular coherence is related to measurements. Twelve healthy right-handed subjects reach-grasp-lift-hold and replace tasks during the (8 females and 4 males, aged 19–35 years) were recorded manipulation of an object. This work focuses on quan- using 32 EEG channels located at Fp1, Fp2, F7, F3, Fz, tifying the coherence EEG–EMG, under several condi- F4, F8, FC5, FC1, FC2, FC6, T7, C , C , C , T8, TP9, tions that include: 1) Evaluating EEG-bands of extended 3 z 4 CP5, CP1, CP2, CP6, TP10, P7, P3, Pz, P4, P8, PO9, O1, alpha ðαÞ ð6 13Þ Hz, beta ðβÞ ð14 30Þ Hz and Oz, O2, PO10 according to the 10–20 international EEG gamma ðγÞ ð35 50Þ Hz that contributes to coher- placement system. Reference and ground electrodes ence; 2) Evaluating five upper-limb muscles during were connected to FCz and AFz locations, respectively. manipulation of an object involving two weights. In addition, five EMG channels from the following A public dataset was used that records EEG and EMG muscles: 1. Anterior Deltoid (AD), 2. Brachioradialis information during the performance of tasks of reach- (B), 3. Flexor Figitorum (FD), 4. Common Extensor ing and grasping objects of different weights. The ana- Digitorum (CED), and 5. First Dorsal Interosseous lysis was defined at two weights: 0:165 kg and 0:660 kg, 142 C. D. GUERRERO-MENDEZ AND A. F. RUIZ-OLAYA (FDI). EEG signals were recorded with the ActiCap was kept constant with sandpaper. Five series of weights device at a sampling rate of 500 Hz. On the other were used where each series included 22 trials, for a total hand, EMG signals were recorded using five sensors at of 110 trials per participant. Finally, data were taken for a sampling frequency of 4 kHz. Figure 1 shows the each trial until the subject performed the task of repla- experimental setup for the dataset acquisition. cing the object. In the protocol, initially, there is a rest period of 2 seconds before starting the movement where subjects 2.2. Dataset integrity validation maintain the right upper limb leaning on a table, next, the subject receives a visual indication from a LED to Data integrity validation of EEG signals was performed start performing a reaching movement of the right hand to verify contralateral behavior and characterize cortical toward an object. Then, the user grasps it with the index responses when performing the movements with the and thumb fingers; afterward lifts it and holds the object right upper limb. For this purpose, the EEG signals steadily within a circle that is about 5 cm from the table were filtered using an 8th order bandpass Butterworth for 2 seconds until the LED turned off, and subsequently filter between 8 30 Hz, and a Common Average replacing the object and returns the upper limb to the Reference Filter (CAR) to remove related noise on the position indicator, as shown in Figure 1. The object electrodes located in the cortico-motor area. varied in weight and contact surface randomly in 3 Subsequently, the EEG signals’ Power Spectral Density different conditions, the variation of weights was kept (PSD) is calculated for all 32 channels (Figure 2) at (0:165, 0:330, 0:660 kg) and contact surface at (sand- recorded in the two weights to evidence the frequency paper, chamois, silk). To module the objects’ weight, an distribution across the cortico-motor area. For this, the electromagnet was used to present the same weight PSD was calculated using the Fast Fourier Transform between 1 and 4 times and then change it. These weight (FFT) with a Hanning window of 1 second in the fre- variations were performed using two electromagnets of quency range 8 30 Hz for the execution data (2– 0:165 kg and 0:330 kg that to vary the weight in some 8 seconds). conditions only one electromagnet is activated and in In addition, Event-Related Synchronization and others both. For the variation of the contact surface, the Desynchronization (ERS/ERD) was calculated for 3 same logic was maintained, but in this condition, an EEG channels (C , C , and C ) to identify the short- 3 z 4 external person had to intervene on the object. Finally, lasting and localized amplitude attenuation of EEG ten series of approximately 32 trials were recorded, for rhythms within the α and β bands [23]. EEG signals a total of 328 trials per participant in which the weight of were processed for each weight using the open-source the object (0:165, 0:330, 0:660 kg), the contact surface FieldTrip Toolbox (https://www.fieldtriptoolbox.org/ ). (sandpaper, chamois, silk), or both was changed. The ERS/ERD was implemented with a time-sliding This study used 3 EEG channels (C , C , and C ) and window in time steps of 50 ms with frequency intervals 3 z 4 five EMG channels for calculating corticomuscular con- of 1 Hz from 8 to 30 Hz using the Morlet Wavelet nectivity, following the international 10–20 system of method for a time-frequency representation. The ERD EEG electrode placement and muscle location for EMG, is related to a decrease in amplitude and power and the as presented in Figure 2. Data from two different ERS is related to an increase in amplitude and power, weights were used when the subject manipulated the which was calculated using equation (1) with a baseline object (0:165 and 0:660 kg) where the contact surface period of 1 second before the stimulus [23]. ERS/ERD Figure 1. Experimental setup for data acquisition. Modified image and adapted from [24] available under the terms of Attribution 4.0 International Creative Commons License (https://creativecommons.org/licenses/by/4.0/). BRAIN-COMPUTER INTERFACES 143 data is presented as a percentage of the baseline period. when the subject starts to perform the movement, the Nevertheless, only the significant data of the increase or EMG amplitude should be greater for the 0:660 kg decrease power (ERS/ERD) with respect to a reference weight than for the 0:165 kg weight, where the ampli- time of 1 second were taken for the analysis. For this, the tude of the signal estimates the muscle activation when Bootstrap algorithm was used with a significance level of a task is performed [27]. For this, a statistical analysis 0:05 (p< 0:05) according to the description made by was applied using the Mann–Whitney U-test with p ¼ [25]. Additionally, to evaluate contralateral behavior in 0:05 using the mean of all subjects, due to the muscle the execution of the movement, a statistical analysis was activity data of the five muscles and the two weights do performed using the ERS/ERD data to identify signifi - not present a normal distribution according to the cant differences between channels and weights using the Kolmogorov–Smirnov test. In equation (2), N is the Mann–Whitney U-test with p value of 0:05, after length of the window, and X is the EMG signal for checking for abnormal distribution of the data. each sample n. vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi R A u X ERD% ¼ � 100 (1) t RMS ¼ X (2) A n n¼1 In equation (1), A is the power related to the step time after the event between the interest frequency range and R is the reference baseline period. 2.3. Methodology for EEG–EMG coherence analysis Data integrity validation of EMG signals was per- formed to verify the amplitude and characterize the After dataset integrity validation the methodology for muscle electrical activity when the movement was per- EEG–EMG coherence analysis presented in Figure (3) formed. For this, EMG signals for the five muscles was implemented. First, the study is delimited using two (Figure 2) were filtered using an 8th order bandpass weights and selecting subjects. After, artifacts are Butterworth filter between 10 500 Hz due to in that detected in the EEG signal, and channels are selected. frequency band there is higher energy and power of the Next, the signals are pre-processed using filters. After, EMG signal [26]. Subsequently, EMG signals were ana- the EEG and EMG signals are segmented by using lyzed through the signal amplitude estimation using the a Hanning window of 1 second overlapped to 25% for RMS value following the equation (2) with a sliding calculating the coherence, which is represented using window of 300 ms overlapped 50%. In the analysis, significant coherence values. And finally, statistical Figure 2. EEG electrodes layout following the international 10–20 system used in the original WAY-EEG-GAL dataset and EMG electrode placement following the locations of five upper limb muscles. 144 C. D. GUERRERO-MENDEZ AND A. F. RUIZ-OLAYA Figure 3. Block diagram of the implemented methodology in this study to estimate corticomuscular connectivity between EEG and EMG signals during the movement of reach-grasp-lift-hold and replace an object. analysis is performed to find significant differences signal that is 4 kHz, such as reported in the literature between the coherence results for the two weights [30]. For this purpose, only the execution movement is using the paired t-test. analyzed, because at rest there is no movement intention and therefore muscle activation is negligible [4]. Some authors describe that coherence in the resting is insig- 2.3.1. Subjects and channels selection nificant [31]. In this study, a trial rejection process was performed by identifying artifacts in EEG signals, using threshold cri- teria. For this, the EEG signals should be within the range 2.3.3. Connectivity analysis between EEG and EMG of � 350μV [28]; otherwise, trials were rejected contain- signals ing high outliers that were outside the range. If outliers Connectivity analysis between EEG and EMG signals are detected in an EEG channel, the trial is rejected for all was calculated using the coherence algorithm, a measure EEG and EMG channels. EEG channels C , C , and C 3 z 4 of connection or correlation between two signals in the and all EMG channels were analyzed. The subject selec- frequency domain that determines the strength of cor- tion criterion was that amount rejects trials no should be relation in the range of 0–1 [6]. Coherence was imple- greater than 10% of the total trials [29]. In this study, no mented to evaluate the connectivity between the subject was rejected, selecting all participants. electrical activity of the brain and muscles when per- forming an object manipulation movement. To calcu- late coherence, the frequency components of EEG and 2.3.2. Pre-Processing EMG signals in the range 6 50 Hz were extracted by To estimate corticomuscular connectivity, the EEG sig- calculating the auto-spectrum and cross-spectrum using nals for each trial and each subject were filtered using MATLAB (Version R2020b, MathWorks Inc). EEG and a bandpass Butterworth filter of 8th order between 6 EMG signals were segmented by a 1 second Hanning 50 Hz and a CAR filter. The EMG signals were filtered window with 25% overlap at a frequency resolution of 2 using a bandpass Butterworth filter of 8th order the 6 Hz between 6 and 50 Hz for each trial and subject [32]. 50 Hz. Contaminated trials in both EEG and EMG Combinations between channels were obtained by relat- signals were rejected for each subject. A resampling of ing the 3 EEG channels (C , C , C ) with each EMG EEG signal was performed at the sample rate of EMG 3 4 z BRAIN-COMPUTER INTERFACES 145 channel (5 Channels), which generate a total of 15 2.3.4. Coherence analysis combinations between channels, as shown in Figure 3. Three different studies were performed to evaluate the The coherence was calculated according to equation coherence between EEG and EMG signals during the (3) [33]. movement. In the first study, segmentation was per- formed in the frequency bands (α, β, and γ) where all � � � � the EMG channels were averaged for each EEG channel P ðfÞ xy Coh ðfÞ ¼ (3) xy and each frequency band analyzed to determine in P ðfÞ� P ðfÞ xx yy which frequency band significant differences exist between weights. For the second study, the coherence where Coh ðfÞ is the correlation specified in the range xy was evaluated for the combination of each EEG and 0–1 between two signals in the frequencies f. P ðfÞ is xy EMG channel in the frequency band from 6 to 50 Hz the cross-spectrum of x and y, while P ðfÞ and P ðfÞ is xx yy to determine in which muscles there are significant the auto-spectrum of x and y, respectively, which uses differences between weights. Finally, coherence is eval- Welch’s averaged and modified periodogram method of uated for the 0:165 kg weight in each frequency band spectral estimation. In addition, a significant threshold and in each muscle to determine under similar weight was calculated using the coherence values to use in the conditions which muscles are mostly in communication study the significant results of the proposed coherence with each side of the motor cortex during these actions method, equation (4) [33]. using information from EEG channels C and C . For 3 4 each study, the distribution of the data was assessed by L 1 Confidence Limit ¼ 1 ð1 εÞ (4) applying the Shapiro-Wilk test with a p value of 0:05, where all data were found to have a high probability of where L is the number of segments and ε is the con- following a normal distribution. fidence level. However, in this study, L is the number of windows for each weight, and the confidence level used 2.3.5. Statistical analysis is 95% for both weights. Statistical analysis was performed to identify signifi - To extract significant coherence, a conditional- cant differences between the connectivity of the two based classification was implemented where the weights analyzed when the subject performs the coherence spectrum above the confidence limit was manipulation of an object. For this, first, the analyzed. Then, equation (5) was used to obtain the Shapiro–Wilk test with a p value of 0:05 was applied average coherence per subject, channel combination, to determine the normal distribution of the data and and frequency band (α, β, γ). However, before apply- the Levene’s test was applied to determine the homo- ing equation (5) the distribution of the significant geneity of variances between the data. Based on these coherence features was evaluated using the Shapiro- results, the two-sample paired t-test was subsequently Wilk test with a p value of 0:05 because the vector applied using Matlab because the data follow a normal size was kept less than 50 samples. In the results distribution and equal variances. Significance levels for obtained, it is concluded that the significant coher- this test were set at 5% (**) and 10% (*). The alter- ence data in the 3 frequency bands (α, β, γ), the 2 native hypothesis is that one of the two weights present weight conditions (0:165 kg, 0:660 kg), the channels greater coherence than the other weight in the first two combination and the subjects follow a normal distri- studies performed according to section 2.3.4, and the bution with p value< 0:05. null hypothesis is the opposite. For the third study, the alternative hypothesis is that one of the two EEG channels (C and C ) presents greater coherence in 3 4 CS ¼ jXðfÞj (5) each of the 3 frequency bands and the 5 muscles, and f¼1 the null hypothesis is the opposite. For these purposes, where the CS is the average coherence spectrum sig- the results are presented in bar figures with confidence nificant, f is the specific frequency significant, N is the intervals of 95% to evaluate the significant variation number of significant coherences by the bands (α, β between subjects. and γ) and X is the value of significant coherence. In this study, equation (5) was used for each indepen- 3. Results dent frequency band, where to estimate coherence over the entire (6 50) Hz spectrum, connectivity Subjects perform an upper-limb movement that information was extracted from each frequency band involves reaching, grasping, lifting, holding, and repla- and averaged. cing an object of two different weights. Taking into 146 C. D. GUERRERO-MENDEZ AND A. F. RUIZ-OLAYA account that each subject performs the motor tasks at other hand, the duration of the movement for the a different time, quantification of the time variation weight of 0:165 kg among the mean of the subjects between subjects and between the tasks was obtained corresponds to 6:88� 0:40 and for the weight of 0:660 during the manipulation of the object. kg corresponds to 7:55� 0:57. Figure 4 shows the time variation between subjects In this article, data were taken for each trial until the for each task performed in the manipulation of the subject performed the task of replacing the object. First, object (Reach, Grasp, Lift, Hold and Replace), the time obtained results of the data integrity validation were variation of the whole task performed until the replace- presented. Next, coherence between EEG and EMG ment of the object between subjects (blue for the weight signals was shown in frequency bands and upper-limb of 0:165 kg and red for the weight of 0:660 kg), and the muscles involved and finally, the study is presented to distribution of each task along with the movement of demonstrate which muscles are mostly in communica- manipulating an object for the weight of 0:165 kg, where tion with the hemispheres of the motor cortex. the red color represents the reaching task, blue repre- sents the grasp, green represents the lift, yellow repre- 3.1. ERD/ERS quantification sents holding and magenta represents replacing the object. In Figure 4, the rest period corresponds to the Figure 5 presents the head maps for each analyzed first two seconds, and the time after the replace task weight using all trials and subjects spanning the fre- corresponds to the release task. In the figure, the task quency range 8 30 Hz from 2 8 seconds. with the longest duration within the manipulation of the According to the figure, the movement was distributed object was the holding that lasts approximately 2 sec- in all cortical cortex by the power decrease presented in onds, and the task with the shortest duration is the the head map, where this effect could be presented by lifting that lasts approximately 0:30 seconds. On the Figure 4. Duration of segments and trials. (A) Average of all subjects in each task involved in the movement. (B) Average of all subjects for the total duration of trials taken until the replacing the object. (C) Distribution of duration of the 0:165 kg weight for each task involved in the manipulation of the object. BRAIN-COMPUTER INTERFACES 147 Figure 5. Power Spectral Density (PSD) for all EEG channels (32 channels) for both weights using all trials and subjects analyzed. PSD magnitude extends in the range ð0 20Þ of magnitude expressed in the head map for the execution movement (2 8 seconds) in the 8 30 Hz frequency band. the movement complexity. Finally, the EEG channels before the vertical dotted line, as shown in Figure 6. chosen for the coherence analysis were C ; C and C After 2 seconds, a LED turns on and subjects start 3 4 z due to the main location in the cortical-motor cortex. performing the movement; the ERD/ERS data reveals The ERD/ERS (relative power) was calculated to a significant ERD after the 2 seconds in all 3 channels observe the change in power when the subject performs according to the results of the bootstrap algorithm [25]. the movement with the two weights. The EEG time- Figure 7 shows the ERD/ERS as a representation of frequency representations were calculated for C ; C the relative amplitude (%) versus the time (s) taken in 3 4 and C channels in the 8 30Hz frequency band the execution of the movement. The results are pre- throughout the movement. Figure 6 shows time- sented for the 3 EEG channels (C , C y C ) segmenting 3 z 4 frequency maps of ERD/ERS percentage values across the information α (8 13 Hz) and β (14 30 Hz) all trials (55 trials per weight) and subjects (12 subjects) frequency bands, and presenting for the same channel recorded for each weight. The figure is divided by α and the variation of the two weights, where the weight 0:165 β bands with a bar color between ð 50%; 50%Þ. During kg is presented in blue color and the weight of 0:660 kg the rest segment (0 2 seconds) there is no ERD/ERS in red. However, for each frequency band, the mean and Figure 6. Time-frequency maps ERS/ERD for C ; C and C channels for two weights using all trials and subjects analyzed. The black 3 z 4 vertical dotted line represents the time turning on the LED and subjects start to perform the movement. The black horizontal dotted line presents α and β bands separation. 148 C. D. GUERRERO-MENDEZ AND A. F. RUIZ-OLAYA Figure 7. ERD/ERS value and its statistics. (A) Relative amplitude (%) for the 3 channels and two weights by segmenting the information into (α) and (β) frequency bands. The black line of vertical dots represents the moment when the LED is turned on and the subjects start to perform the movement. (B) Average and standard deviation of the relative amplitude of the 3 channels and the two weights for the two frequency bands. (**) Represents that there are significant differences (p< 0:05) between the channels. standard deviation as shaded regions for the 3 channels was performed. In the literature, it is reported that when in the two weights are presented, as shown in the bars of a user performs a right-hand motor task, an ERD is Figure 7. presented in the contralateral side (C channel) [23]. To evaluate the contralateral behavior in the execu- tion of the movement, a statistical analysis was per- 3.2. Estimation and analysis of EMG amplitude formed to identify significant differences between the channels and between the weights. First, an analysis was Changes in the electrical activity of the upper limb performed using the Kolmogorov–Smirnov test muscles (5 muscles) when the subject performs the (Samples > 50 data) to confirm that the behavior of movement was quantified. This muscle activity is the ERD/ERS for each channel and weight have a high expected to provide greater activation in the 0:660 kg probability of having a normal distribution. weight than in the 0:165 kg weight. The RMS value of Subsequently, Mann-Whitney U-test was applied with the EMG signal in the five muscles was calculated, as a value of p ¼ 0:05 since the data did not present shown in Figure 8, where the 0:165 kg weight is pre- a normal distribution. According to the results, there sented by a blue line and the 0:660 kg weight is pre- are significant differences (p< 0:05) in the relative sented by a red line. The average value of muscle amplitude of the two weights in the 3 channels for the electrical activity for the 5 muscles of all recorded sub- 2 frequency bands. Additionally, in α there are signifi - jects is presented along with the standard deviation in cant differences (p< 0:05) between channels C and C , shaded regions for the two weights, as shown in 3 z and C and C for the two weights, and in β there are Figure 8. This allows us to observe how the change in 4 z significant differences (p< 0:05) in the relative ampli- muscle activation is for all subjects. However, according tude for the 3 channels (C and C , C and C , C to the figure and the statistical analysis performed using 3 z 3 4 4 and C ). the Mann-Whitney U-test with a significance value of Considering the results presented in Figures 6 and 7, 5%, the 0:660 kg weight presents greater muscle activa- the 0:660 kg weight presents a greater decrease in power tion than the 0:165 kg weight evaluated for each muscle than the 0:165 kg weight and that in the C channel individually with a p value< 0:05. As expected, there there is a greater decrease in power compared to the C is no muscle activation before initiating the movement and C channels, which could confirm the contralateral presented as a vertical dotted black line at second 2. All effect that occurs when movements are performed with EMG channels (5 channels or muscles) were selected to the right hand. That is, the C channel presents greater evaluate the coherence between EEG and EMG signals prominence than the C channel when the movement in the analyzed movement. 4 BRAIN-COMPUTER INTERFACES 149 Figure 8. Root Mean Square (RMS) value for all subjects with the standard deviation presented in shaded regions for the two weights in the 5 muscles recorded. The black line of vertical dots represents the time at which the LED is turned on and at which the movement is initiated. The blue line presents the RMS value for the 0:165 kg weight and the red line for the 0:660 kg weight. 3.3. Coherence for frequency bands and muscles shown in Figure 9(A). As an error bar, confidence intervals of 95% were calculated with respect to all Connectivity was quantified over the time interval from subjects for each weight and EEG channel. In addition, the start of the reach task (LED on) to the end of the the results of the statistical analysis using the two- replace task (see Figure 4). For this, the coherence sample paired t-test are presented, where (*) corre- spectrum for each EEG channel averaged over all sponds to the 10% significance level and (**) corre- EMG channels in the frequency bands (α, β, and γ) is sponds to the 5% significance level. Figure 9. Coherence and statistical analysis for two weights quantified over the time interval from the start of the reach task (LED on) to the end of the replace task. (A) Coherence between each EEG channel (C , C , C ) and all EMG channels in the 3 frequency bands for 3 z 4 all subjects in the two weights. (B) Coherence between each EEG and EMG channel in the averaged of (α, β and γ) frequency bands for all subjects in the two weights. The confidence intervals of 95% regarding all subjects are presented. (*) corresponds to significant differences between weights at a significant level of 10% and (**) corresponds to a significant level of 5%. 150 C. D. GUERRERO-MENDEZ AND A. F. RUIZ-OLAYA The α band for the 3 EEG channels has a low coher- Finally, for C channel, there were significant differences ence concerning other bands. On the other hand, the with a 5% of significance considering the B and FD greatest coherence was found in the β band for the 3 muscles with p value ðp< 0:05; p ¼ 0:00Þ and channels, as shown in Figure 9(A). The γ band presents ðp< 0:05; p ¼ 0:04Þ, respectively. Brachioradial muscle a coherence spectrum between the values of α and β presents significant differences between weights for the band coherence. It is important to highlight that the 3 EEG channels indicating that B muscle has an impor- coherence with the 0:165 kg weight is greater than the tant change regarding the connectivity when the weight is 0:660 kg weight for all bands at the 3 channels (C , C , changed. 3 z C ). Therefore, statistically significant differences were found for some cases as in the C channel where there 3.4. Analysis of coherence according to are significant differences ðp< 0:1; p ¼ 0:08Þ in the hemispheric laterality coherence between weights for α band. Moreover, for the C channel there are significant differences of Figure 10 presents the muscles that are mostly in com- ðp< 0:05; p ¼ 0:04Þ in the α band between weights and munication with each side of the motor cortex during in the C channel are significant differences in the α and the manipulation of an object using EEG–EMG con- β band with p values of ðp< 0:1; p ¼ 0:09Þ and nectivity quantified over the time interval from the start ðp< 0:1; p ¼ 0:08Þ, respectively. of the reach task (LED on) to the end of the replace task However, the connectivity where muscle connections (see Figure 4). Initially, only the coherence calculated were varied regard each EEG channel (C , C , C ) aver- for the weight of 0:165 kg was used, where the variation 3 z 4 aged over the 3 frequency bands (α, β, γ) is presented in of coherence for the 5 recorded muscles and the 3 Figure 9(B). This figure allows identifying how connec- frequency bands in channels C and C are presented. 3 4 tivity changes for each muscle, taking into account that In the figure, C channel is represented using yellow each muscle presents a different electrical activation color and C channel using green color. A statistical depending on the task performed in the manipulation analysis using the two-sample paired t-test is presented of the object which is controlled by the Central Nervous in order to show whether there are significant differ - System (CNS). The muscles are presented and labeled as ences in the coherence of the 2 EEG channels with Anterior Deltoid (AD), Brachioradial (B), Flexor respect to each frequency band and recorded muscles, Digitorum (FD), Common Extensor Digitorum where (*) corresponds to the 10% significance level and (CED), and First Dorsal Interosseous (FDI). (**) corresponds to the 5% significance level. Connectivity was quantified over the time interval The EEG channels (C and C ) present significant 3 4 from the start of the reach task (LED on) to the end of differences for the AD muscle in the α band at the replace task (see Figure 4). a significance level of 5% with p value The coherence for the weight of 0:165 kg in the (p< 0:05; p ¼ 0:03). For the B muscle these differences Brachioradialis muscle was higher than other muscles are found in the β band with 5% significance and in the for all EEG channels. Coherence of the weight of 0:660 γ band with 10% significance with p value of kg for Brachioradialis muscle is lower than others in the 3 (p< 0:05; p ¼ 0:02) and (p< 0:1; p ¼ 0:09), respec- EEG channels. Nevertheless, the coherence for the 0:660 tively. However, for the FDI muscle the difference is in kg weight was lower than the 0:165 kg weight for all the α band at the 10% significance level with p value muscles and EEG channels, except for the connectivity (p< 0:05; p ¼ 0:07). The muscles that do not present of C with the FDI muscle where the coherence for both significant differences between the channels are the FD weights is approximately equal, and the connectivity and CED in the 3 frequency bands. According to the between C and the AD muscle where the coherence at results, there is a contralateral behavior in the α band of the 0:660 kg weight is higher than the 0:165 kg weight. the AD muscle and in the γ band of the B muscle, and However, these results are not significant according to the there is an ipsilateral behavior in the β band of the statistical analysis. Significant differences were found as B muscle and in the α band of the FDI muscle. follows: For the C channel, there were differences with a significant level of 5% for the AD and B muscle with a 4. Discussion p value for both of ðp< 0:05; p ¼ 0:04Þ; for the C channel, there were significant differences with a 5% of Obtained results of EEG–EMG coherence allow deter- significance considering the B muscle with a p value of mine how functional corticomuscular connectivity ðp< 0:05; p ¼ 0:03Þ and differences with a significant occurs between brain cortical area and specific muscles level of 10% for the CED muscle ðp< 0:1; p ¼ 0:09Þ. when a reach-grasp-lift-hold and replace an object BRAIN-COMPUTER INTERFACES 151 Figure 10. Coherence for the 0:165 kg weight in the EEG channels (C and C ), the 5 muscles and the 3 frequency bands (α, β and γ). 3 4 The confidence intervals of 95% regarding all subjects are presented. Connectivity was quantified over the time interval from the start of the reach task (LED on) to the end of the replace task. (*) corresponds to significant differences between weights at a significant level of 10% and (**) corresponds to a significant level of 5%. motor task was performed. The study aims to investi- signals. Specifically, significant differences were found gate how EEG and EMG connectivity changes during of; a) coherence of C with Anterior Deltoid and the manipulation of an object having a low and high Brachioradial muscles; b) coherence of C with weight (0:165 kg and 0:660 kg). Brachioradial muscle; c) coherence of C with Flexor Cortical responses characterization of EEG data Digitorum and Brachioradial muscles. Furthermore, the showed that when subjects perform the activity with the greatest coherence was found in the β band for the 3 0:660 kg weight presents a greater decrease in power than EEG channels, which is in agreement with the results of the 0:165 kg weight in the α and β frequency bands previous studies [10,12,13,35]. Those studies reported presenting a contralateral behavior. As mentioned that subjects showed a peak in the coherence spectra above, the manipulation of an object involves different between 15 – 30 Hz bandwidth during a hold task motor tasks such as reaching, grasping, lifting, holding, involving stable force production where the coherence and replacing, which generate changes in brain signals between signals may be affected by the performance in according to the cellular excitability that can form each the development of motor tasks. On the other hand, in motor task. Being a complex task because multiple joints future studies, we propose to evaluate how the connec- are involved in performing the movement, larger brain tivity is for each task involved in the manipulation of the neural networks are needed in the processing of the object, in order to determine the delay times that occur information presented in the movement, so an ERD is in sending the information to execute each action present around the entire execution of the movement involved in the movement [4]. [23]. On the other hand, this type of task involves Studies have been published in the literature that has a number of different muscles, where muscle activation determined contralaterality with respect to corticomus- differs throughout the movement because each muscle is cular connectivity in a single muscle for upper limb in charge of specific functions in the movement. As it is movements. Effects have been found that relate to a combination of different tasks, the muscle synergies there being greater connectivity quantified using the (muscle groups) are coactivated in a coordinated way coherence algorithm, in the contralateral hemisphere through the CNS to perform a complex task. In other recorded by the channels (C and C ) establishing con- 3 4 words, the CNS activates the muscle groups to generate nectivity with only the Flexor-Digitorum-Superficialis a different motor command [34]. (FDS) muscle when performing right and left-hand Significant differences in the movement of manipula- movements in healthy subjects [36]. However, that tion of an object with two different weights were found study quantifies corticomuscular connectivity in an by quantifying the connectivity of EEG and EMG upper limb movement when it reaches its final posture. 152 C. D. GUERRERO-MENDEZ AND A. F. RUIZ-OLAYA This study involves a multi-joint movement since the [17,41]. Therefore, these provide a benefit and a clear subjects perform a movement of control and precision advantage over any single brain signal acquisition with a synergy of synchronized movements to perform modality. the action of manipulation of an object, which accord- In this work, EMG-EEG coherence is higher in the ing to cortical topographies shows an ERD in the entire beta frequency band were demonstrated, individualized motor cortex of the brain. differences in coherence have been found according to Corticomuscular coherence has been shown to be each muscle involved during the reaching and grasping task-dependent, that it reflects attention and precision, movement, showing that the brachioradialis muscle is compliance of the gripped objects, displacement, mag- the most involved in the connectivity due to the signifi - nitude of the force, and learning processes [2]. Thus, cant differences found in the EEG channels. effects such as motor adaptation or neuronal noise Additionally, it has been demonstrated how the EMG- sources may be taking place in the modulation of con- EEG coherence could change depending on the force nectivity during the manipulation of the object invol- exerted to grasp an object of different weight, and it has ving two weights [21,37]. Furthermore, corticomuscular been determined which muscles are mostly in commu- coherence has widely been considered a motor mechan- nication with each side of the cerebral hemispheres. ism reflecting efferent oscillatory activity descending Finally, the results presented in this study allow us to from the primary motor cortex via corticospinal path- conclude that coherence is significantly higher at ways, but it has been reported works that suggest that a weight of 0:165 kg than at a weight of 0:660 kg for ascending sensory activity also contributes to the cou- the α band in the C , C and in α and β for the C 3 z 4 pling between sensorimotor brain regions and muscle channel. In addition, these differences are found in the [38]. Such phenomenon could be taking part in C -AD, C -B, C -B, C -CED, C -B and C -FD channel 3 3 z z 4 4 obtained significant differences of coherence during combinations. These results are of great importance in reach-to-grasp movement for two different weights. rehabilitation engineering applications because cortico- The results from the coherence analysis show that muscular connectivity can be used as a descriptor to there are significant differences in EEG–EMG connectiv- improve the classification rates, usability, and control of ity in the manipulation of an object of two different prostheses based on BCI systems. In this case, the appli- weights (low and high) mainly in the α and β band, and cation of prostheses for weightlifting identification. in the AD, B, FD and CED muscles in different EEG However, as future studies, we propose to evaluate channels. These differences are centered on the 0:165 kg other methods such as Granger Causality to establish weight presents higher coherence than the other weight. connectivity and delay times between EEG and EMG Therefore, by applying these data in a multimodal acqui- information when performing other types of move- sition system, a high classification could be obtained in ments involved in activities of daily living. As well as the identification of motor tasks, not only when the upper using corticomuscular connectivity as a rehabilitation limb has already reached its final position but also in the method for people with disabilities, for implementing identification of different tasks that are involved in computational methods based on connectivity to a given action such as the manipulation of an object. improve identification rates using techniques such as The findings of this study could be of great impor- filter banks. tance in the development of Hybrid Brain-Computer Interface (hBCI) systems because currently, the imple- Acknowledgment mentation of a BCI system based exclusively on EEG signals generates low accuracy in pattern decoding, due The authors would like to thank the Antonio Nariño to the fact that it only generates one control command University, for the support of the development in this work. [4,18]. However, hBCIs have 3 different ways to be implemented characterized in multiple biosignals Disclosure statement such as EEG–EMG, paradigms and multisensory sti- muli [39] which allows increasing the reliability of the No potential conflict of interest was reported by the author(s). system in terms of the percentage of command classi- fication in specific tasks and the generalization of applications in fields such as Neuro-Rehabilitation Funding [17, 18, 39, 40]. Mainly, the advantages found encom- This work was supported by Antonio Nariño University pass (1) higher BCI classification accuracy, (2) higher (UAN) under the grant N.2021020 ‘Model based on multi- number of brain commands for control application, modal EEG–EMG information to improve motion intention and (3) shorter brain command detection time decoding for the control of a BCI system’. BRAIN-COMPUTER INTERFACES 153 [14] Borhani S, Abiri R, Jiang Y, et al. Brain connectivity ORCID evaluation during selective attention using EEG-based Cristian D. Guerrero-Mendez http://orcid.org/0000-0003- brain-computer interface. Brain-Comput Interfaces. 1019-9423 2019;6(1–2):25–35. Andres F. Ruiz-Olaya http://orcid.org/0000-0002-5883- [15] Guerrero-Mendez CD, Blanco-Diaz CF, and Ruiz- Olaya AF. 2021. Identification of motor imagery tasks using power-based connectivity descriptors from EEG signals. In 2021 XXIII symposium on image, signal pro- cessing and artificial vision (STSIVA) Popayán, References Colombia (pp. 1–6). [1] Conway B, Halliday D, Farmer S, et al. Synchronization [16] Li X, Samuel OW, Zhang X, et al. A motion- classifica - between motor cortex and spinal motoneuronal pool tion strategy based on sEMG-EEG signal combination during the performance of a maintained motor task in for upper-limb amputees. J Neuroeng Rehabil. 2017;14 man. J Physiol. 1995;489(3):917–924. (1):1–13. [2] Andrykiewicz A, Patino L, Naranjo JR, et al. [17] Hong K-S, and Khan MJ. Hybrid brain–computer Corticomuscular synchronization with small and large interface techniques for improved classification accu- dynamic force output. BMC Neurosci. 2007;8(1):1–12. racy and increased number of commands: a review. [3] Rizzo R, Zhang X, Wang JW, et al. Network physiology Front Neurorobot. 2017 July;11 35 . of cortico–muscular interactions. Front Physiol. [18] Ma T, Li H, Deng L, et al. The hybrid BCI system for 2020;11. DOI:10.3389/fphys.2020.558070 movement control by combining motor imagery and [4] Kim B, Kim L, Kim Y-H, et al. Cross-association ana- moving onset visual evoked potential. J Neural Eng. lysis of EEG and EMG signals according to movement 2017 February;14(2):026015. intention state. Cognit Syst Res. 2017;44:1–9. [19] Betti S, Zani G, Guerra S, et al. Reach-to-grasp move- [5] Zhuang KZ, Lebedev MA, Nicolelis MA. Joint ments: a multimodal techniques study. Front Psychol. cross-correlation analysis reveals complex, 2018;9:990. time-dependent functional relationship between corti- [20] Batzianoulis I, El-Khoury S, Pirondini E, et al. cal neurons and arm electromyograms. J Neurophysiol. EMG-based decoding of grasp gestures in 2014;112(11):2865–2887. reaching-to-grasping motions. Rob Auton Syst. [6] Liu J, Sheng Y, Liu H. Corticomuscular coherence and 2017;91:59–70. its applications: a review. Front Hum Neurosci. [21] Mendez-Balbuena I, Huethe F, Schulte-Mönting J, et al. 2019;13:100. Corticomuscular coherence reflects interindividual dif- [7] Mima T, Hallett M. Electroencephalographic analysis of ferences in the state of the corticomuscular network cortico-muscular coherence: reference effect, volume during low-level static and dynamic forces. Cereb conduction and generator mechanism. Clin Cortex. 2012;22(3):628–638. Neurophysiol. 1999;110(11):1892–1899. [22] Steriade M, Llinás RR. The functional states of the [8] Rossiter HE, Eaves C, Davis E, et al. Changes in the thalamus and the associated neuronal interplay. location of cortico-muscular coherence following Physiol Rev. 1988;68(3):649–742. stroke. NeuroImage Clin. 2013;2:50–55. [23] Pfurtscheller G, Da Silva FL. Event-related EEG/MEG [9] Sburlea AI, Müller-Putz GR. Exploring representations synchronization and desynchronization: basic principles. of human grasping in neural, muscle and kinematic Clin Neurophysiol. 1999;110(11):1842–1857. signals. Sci Rep. 2018;8(1):1–14. [24] Luciw MD, Jarocka E, Edin BB. Multi-channel EEG [10] Boonstra TW, van Wijk BC, Praamstra P, et al. recordings during 3,936 grasp and lift trials with vary- Corticomuscular and bilateral EMG coherence reflect ing weight and friction. Sci Data. 2014;1(1):1–11. distinct aspects of neural synchronization. Neurosci [25] Graimann B, Huggins JE, Levine SP, et al. Lett. 2009;463(1):17–21. Visualization of significant ERD/ERS patterns in mul- [11] Lóopez-Larraz E, Birbaumer N, and Ramos- tichannel EEG and ECoG data. Clin Neurophysiol. Murguialday A. 2018. A hybrid EEG- EMG BMI 2002;113(1):43–47. improves the detection of movement intention in cor- [26] Merletti R, Botter A, Troiano A, et al. Technology and tical stroke patients with complete hand paralysis. In instrumentation for detection and conditioning of the 2018 40th annual international conference of the IEEE surface electromyographic signal: state of the art. Clin engineering in medicine and biology society (EMBC) Biomech. 2009;24(2):122–134. Honolulu, HI, USA (pp. 2000–2003). [27] Renshaw D, Bice MR, Cassidy C, et al. A comparison of [12] Gao L, Wu H, Cheng W, et al. Enhanced descending three computer-based methods used to determine EMG corticomuscular coupling during hand grip with static signal amplitude. Int J Exercise Sci. 2010;3(1):43. force compared with enhancing force. Clin EEG [28] Dimigen O. Optimizing the ICA-based removal of ocu- Neurosci. 2020;52(6):436–443. lar EEG artifacts from free viewing experiments. [13] Kilner J, Baker S, Salenius S, et al. Task- dependent NeuroImage. 2020;207:116117. modulation of 15-30 Hz coherence between rectified [29] Krauledat M, Dornhege G, Blankertz B, et al. EMGs from human hand and forearm muscles. Robustifying EEG data analysis by removing outliers. J Physiol. 1999;516(2):559–570. Chaos Complexity Letters. 2007;2(3):259–274. 154 C. D. GUERRERO-MENDEZ AND A. F. RUIZ-OLAYA [30] Xi X, Ma C, Yuan C, et al. Enhanced EEG-EMG coher- [36] Chowdhury A, Raza H, Meena YK, et al. An EEG- EMG ence analysis based on hand movements. Biomed Signal correlation-based brain-computer interface for hand Process Control. 2020;56:101727. orthosis supported neuro-rehabilitation. J Neurosci [31] Hashimoto Y, Ushiba J, Kimura A, et al. Correlation Methods. 2019;312:1–11. between EEG-EMG coherence during isometric con- [37] Trenado C, Mendez-Balbuena I, Manjarrez E, et al. traction and its imaginary execution. Acta Neuro biol Enhanced corticomuscular coherence by external sto- Exp (Wars). 2010;70(1):76–85. chastic noise. Front Hum Neurosci. 2014;8:325. [32] Bortel R, Sovka P. EEG-EMG coherence enhancement. [38] Beck MM, Spedden ME, Lundbye-Jensen J. Signal Process. 2006;86(7):1737–1751. Reorganization of functional and directed cortico- [33] Rosenberg J, Amjad A, Breeze P, et al. The Fourier muscular connectivity during precision grip from approach to the identification of functional coupling childhood to adulthood. Sci Rep. 2021;11(1):1. between neuronal spike trains. Prog Biophys Mol Biol. November. 1989;53(1):1–31. [39] Li Z, Zhang S, Pan J. Advances in hybrid brain-computer [34] Torricelli D, Barroso F, and Coscia M, et al. 2015 interfaces: principles, design, and applications. Comput October. Muscle synergies in clinical practice: theore- Intell Neurosci. 2019 October;2019:1–9. tical and practical implications Pons, J, Raya, R, and [40] Amiri S, Fazel-Rezai R, Asadpour V. A review of hybrid González, J. In: Biosystems & biorobotics. 251–272. Switzerland: Springer International Publishing. brain-computer interface systems. Adv Hum Comput [35] Kristeva R, Patino L, Omlor W. Beta-range cortical motor Interact. 2013;2013:1–8. spectral power and corticomuscular coherence as [41] Banville H, Falk T. Recent advances and open chal- a mechanism for effective corticospinal interaction during lenges in hybrid brain- computer interfacing: steady-state motor output. Neuroimage. 2007;36 a technological review of non-invasive human (3):785–792. research. Brain-Comput Interfaces. 2016;3(1):9–46.

Journal

Brain-Computer InterfacesTaylor & Francis

Published: Jul 3, 2022

Keywords: Corticomuscular coherence; object manipulation; functional connectivity; hybrid brain-computer interface; motor execution; electroencephalography (EEG); electromyography (EMG); reach-to-grasp

References