ECG Artifact Removal from EMG Recordingsusing Independent ComponentAnalysis and Adapted Filter

Abstract ² Surface electromyography (sEMG) recordings from trunk or limb muscles are often easily corrupted by electrocardiography (ECG) signals. In order to remove or reduce ECG in sEMG so as to improve the practicability, a novel signal filtering method with joint independent component analysis (ICA) and adaptive filtering (AF) is proposed in this paper. The method is validated with synthetic noisy EMG signals derived from 8-channel real sEMG added with 8-channel ECG recordings. Two groups of sEMG signals and two groups of ECG signals were used to examine the performance of the proposed method in our validation study. Experimental results demonstrate that the ICA+AF signal filtering method achieves better performance on reduction ECG artifact than the conventional Butterworth High-pass filter with 30 Hz cutoff frequency. The proposed method also performed well with 8-channel real ECG contaminated sEMG signals.I. I NTRODUCTIONElectromyography (EMG) measures muscular activities represented by the summation of action potentials occurring in motor units. EMG analysis of a muscle serves as a practical means of collecting important physiological and pathophysiological behavior information. Compared to the routine intramuscular EMG examination using needle electrodes, the surface EMG (sEMG), due to its noninvasive fashion, has much broader application prospect in a wide range of fields from human movement and kinesiology through clinical diagnostics and pathology to rehabilitative treatment [1-4].Because of the complexity of sEMG measurement and its weak nature, sEMG is easily corrupted by noises such as the signals from other physiological origins or electromagnetic interference from ambient environment [5-6]. For example, the electrocardiography (ECG) signal is likely to contaminate the EMG signal recorded from truk or limb muscles, especially close to the heart [7]. This makes ECG interference a major obstacle encountered in EMG analysis.*Research supported by the National Nature Science Foundation of China under Grant 61271138.X. Chen is with the Department of Electronic Science and Technology, University of Science and Technology of China (USTC), Hefei 230027 China (corresponding author to provide phone: +86-0551-6360-1175; e-mail: xch@ ).Y. Li and X. Zhang are with the Department of Electronic Science and Technology, University of Science and Technology of China, Hefei 230027 China (e-mail: liyun5@).P. Zhou is with the Sensory Motor Performance Program, Rehabilitation Institute of Chicago, Chicago, IL 60611 USA, with the Department of Physical Medicine and Rehabilitation, Northwestern University, Chicago, IL 60611 USA, and also with the Institute of Biomedical Engineering, USTC, Hefei 230027 China (e-mail: p-zhou@).Many previous efforts have been made to retrieve the EMG signal from the ECG contaminated recordings. Removal of ECG interference from EMG signals cannot be simply achieved by applying conventional digital filters, since both of the true EMG signal and ECG interference overlapped each other in amplitude and frequency. Redfern et al. also suggested that a high-pass filter with a cutoff frequency of approximate 30Hz could yield relatively optimal performance in balancing ECG cancellation and excessive EMG degradation [10]. Another simplistic, yet potentially effective, procedure is based on the gating method which is designed to simply cut out the signal segments overlapping with QRS complexes in amplitude, thus suffering from the loss of portions of useful EMG signals.Recently, several advanced signal processing algorithms such as adaptive filtering (AF) and independent component analysis (ICA) have been developed and studied for removing ECG from EMG recordings[8][9][11]. The AF technique is able to adjust their parameters based on the statistical properties of inputs, which allow adaption to track the dynamic changes in the signal and noise components. However, the adaptive filters therein require a synchronic series of clean ECG noise as the reference signal for adaptive calculation, which can be further obtained in the following two ways. One is to derive the reference signal directly from the noisy EMG recordings by applying a band-pass filter. But sometimes unsatisfactory reference signal could result in residual ECG artifacts apparent in the denoised signal. The other way is to measure the ECG signal during EMG data collection. Likewise, such a solution may raise additional concerns and costs in experiments. On the contrary, ICA is the most popular blind source separation algorithm, which has shown its advantage for separation of an component from multi-channel EMG recordings by the different statistical properties of the ECG and EMG sources. ICA is always applied on the multi-channel recordings since it requires at least as many channels (sensors) as sources.With the multi-channel sEMG recordings, the ECG contamination in each channel comes from the same source (heart), hence should be synchronous, while EMG in different channels are supposed to be different in muscles. The application of an ICA algorithm to multi-channel sEMG recordings with ECG contamination is prone to extract the ECG component, which is exactly synchronized with ECG contamination in each channel, offering a qualified reference signal for performing adaptive filters on each channel. Taking advantage of such property, in this paper, we investigate the combined approach of ICA and AF technique to remove ECG contamination from multi-channel EMG signals.ECG Artifact Removal from EMG Recordings using IndependentComponent Analysis and Adapted Filter*Yun Li, Student Member, IEEE , Xiang Chen, Member, IEEE , Xu Zhang, Member, IEEE ,and Ping Zhou, Member, IEEE6th Annual International IEEE EMBS Conference on Neural Engineering San Diego, California, 6 - 8 November, 2013II.M ETHODA. Independent Component AnalysisICA is a multivariate statistical approach to blind source separation, which aims to recover a set of mutually independent components (ICs) when only mixtures of these sources with unknown coefficients are observed[12][13]. Generally, the raw sEMG recordings, especially from trunk muscle, are mixtures of useful signals (true EMG) and artifacts (ECG). Since EMG generated from skeletal muscle and ECG generated from cardiac muscle can be regarded as physiologically independent processes, the ICA approach is performed with the attempt to separate the ECG-artifact IC from raw multi-channel SEMG signals. Thereafter, it is easy to obtain the ECG-concentrated signal. The detailed concept and principle of ICA algorithm could be found in [12-14]. In this study, the ICA method is implemented using the well-known FastICA software package as a MATLAB toolbox.B. Adaptive FilterThe principle of the adaptive filter used for ECG interference removal is presented as follows. The original noisy EMG signal x(t) of an individual channel could be considered as a combination of clean EMG signal s(t) and ECG artifact n(t):x(t)= s(t)+ n(t) (1) The ECG artifact, n(t), is then estimated by modeling the reference signal, derived from ICA, nr(t) via a finite impulse response (FIR) digital filter, denoted as its impulse response h(t):Q (t)= h(t)* nr (t) (2) where * denotes the convolution operator.7KH GHQRLVHG (0* VLJQDO GHQRWHG DV V W FRXOG EH REWDLQHG E\ VXEWUDFWLQJ WKH HVWLPDWH RI (&* QRLVH Q (t) from the original contaminated EMGV W [ W - Q (t) (3) The FIR filter used for estimating ECG contamination is implemented in an adaptive way that the filter parameters (i.e., the coefficients of h(t)) are updated by a correction criterion. Multiple correction criteria such as least mean squares (MLS) and recursive least squares (RLS) could be used to implement the adaptive filter. In this paper, an adaptive filter based on the RLS algorithm was developed, which can reach optimal performance, according to previous findings in [8-9].With the reference ECG noise derived from ICA analysis of multi-channel EMG signals, the adaptive filter could then be applied to each EMG channel for producing its denoised version respectively.C. Evaluation dataset descriptionSynthetic noisy EMG signals were derived from relatively clean sEMG recordings added with typical ECG recordings to examine the performance of the proposed methods in the validation study. The dataset used in this study was recorded from 2 healthy subjects (2 males, 29 and 35 years). The study was approved by Institutional Review Board of Northwestern University (Chicago, IL, USA). Written informed consent was obtained from all subjects prior to the study. There were two groups of EMG data and two groups of ECG data, each group with 8-channel recordings over a period of 25 seconds. All the evaluation dataset used in this study was recorded by a Refa 128 EMG system (TMS International BV, Enschede, Netherlands) at a sampling rate of 2000 Hz per channel. One EMG data group (EMG 1) was selected from the dataset used in our pilot study[15], where a flexible piece of 64-channel high-density EMG electrode array was used to place above the thenar muscles for EMG recordings when a healthy subject was asked to perform some voluntary muscle contractions. The other EMG data group was recorded via 8 individual electrodes placed on the left biceps brachii of another healthy male subject while the subject was continuously performing elbow flexion/extension movements. The size of each individual electrode is 10 mm in diameter while the recordingVV Figure 1. The results for ECG artifact removal applied on a typical 8-channel synthetic noisy sEMG signals: (a) 8-channel clean sEMG data;(b) 8-channel ECG artifact; (c) 8-channel synthetic noisy EMG signal; (d) The extracted independent components after ICA; (e) 8-channel denoised sEMG data.surface is 5 mm in diameter. After that, other 8 electrodes were placed above the chest muscle. Thus, the ECG interference was then obviously present in each of the 8 channels placed on left biceps brachii, from which two trails of recordings were collected respectively when the subject was asked to stay relaxed without any voluntary muscle contraction, thereby resulting in two ECG data groups.Each of the two EMG data groups could be superposed by each of the two ECG groups, generating the four groups of the synthetic signals.In addition, the real EMG signals with ECG artifact should be collected to examine the effects of real signals on the performance of the proposed methods. After the ECG recording trail, with the 16 electrodes placed on chest muscle and left biceps brachii respectively, the real noisy 8-channel EMG signal with ECG contamination was collected from the biceps brachii when the subject was asked to voluntarily perform elbow flexion/extension movements.D. Performance evaluationTo evaluate the performance of the proposed method for ECG artifact removal, conventional Butterworth high-pass filter at 30 Hz cutoff frequency (denoted as BW30) was used to compare. In order to have an objective comparison of the two ECG removal methods, signal-to-error ratio (SER) of the denoised EMG was used to evaluate the overall improvement of signal quality.III.R ESULTSFig. 1 illustrates the results of the proposed methods for ECG artifact removal applied on the typical 8-channel synthetic noisy sEMG signals. The 8-channel clean sEMG data (Fig. 1a) was superposed by an 8-channel ECG artifact at a SNR of 2dB, resulting in an 8-channel synthetic noisy EMG signal (Fig. 1c). After the ICA performed on the synthetic noisy signal, the independent components (Fig. 1d) were thenextracted. The number of extracted ICs was equal to the number of input channels (e.g., recording electrodes). The extracted ICs were normalized after centering and whitening process of ICA[11]. It is obviously that the fourth channel (IC4) in Fig. 1d is the IC that is dominated by ECG artifact. Then the selected ECG IC was used as a reference in the following adaptive filtering process. After that, the denoised version of signal in each channel is shown in Fig. 1e respectively.The SERs of the processed data after the proposed ECG removal method (ICA+AF) applied on the four groups of synthetic data with different signal-to-noise ratio (SNR) were calculated respectively and summarized in the Table 1. At the same time, the resultant SERs of the data after BW30 were also evaluated for the comparison purpose. It is found that the proposed method based on the combination of ICA and AF significantly outperforms the conventional BW30 in removing ECG artifact from multi-channel EMG recordings (paired student t-test, p<0.001, with each of the four datasets). The proposed method was further applied on a real multi-channel sEMG data with ECG contamination in this study. The experimental result was presented in Fig. 2 where the top tracing was the original ECG contaminated sEMG signal, the middle one was the extracted independent source component and the bottom one was the processed denoising sEMG signal. It is obviously that the seventh channel (IC7) in Fig. 2b is the IC that is dominated by ECG artifact. Then theVVVTime (s)Figure 2. The results for ECG artifact removal applied on a real 8-channel noisy sEMG signals: (a) 8-channel original ECG contaminated sEMG signals; (b) 8-channel extracted independent source components; (c) 8-channel denoised sEMG data.selected ECG IC was used as a reference in the adaptive filtering process.IV.D ISCUDSSION AND CONCLUSIONIn this study, a novel method for ECG artifact removal from multi-channel EMG recordings was described. The proposed method consisted of two-step procedures, namely extracting the ECG IC by applying the ICA algorithm on the multi-channel noisy recordings and then removing the ECG artifact in each channel by using adaptive filter with the extracted ECG IC employed as the reference. It was found that by applying the proposed method on both synthetic and real EMG signals, ECG artifacts could be successfully removed from multi-channel sEMG recordings. In the validation study, the proposed method yielded better performance in removing ECG artifact than the conventional BW 30 method.The traditional adaptive filters, used in some previous studies, applied on the ECG contamination removal require a synchronic series of clean ECG noise as the reference signal for adaptive calculation. One benefit of our method is that it provides an automatic way to obtain a reference input for the widely used adaptive filter by applying ICA on multi-channel EMG recordings. Compared to other approaches using adaptive filters for ECG rejection in previous studies[8][9], the proposed method in this study requires no additional recordings of ECG, thus facilitating the experimental procedure. Furthermore, according to the ICA property, the reference signal derived from the original EMG recordings is definitely synchronized with the ECG contamination. The synchronization of the reference and the targeted noise is strictly required by the adaptive filtering technique and potentially helpful in improving the performance of adaptive filters.The basic idea of ICA is to reconstruct the original independent source signals from the observation sequences by finding a linear mapping such that the demixed sequences are statistically independent. In this study, the number of extracted ICs was equal to the number of input channels (e.g., recording electrodes). This method can be work when the minimum number of electrode no less than the number of the original independent source signals.In addition, many previous studies have been reported based on adaptive filters for removing ECG from EMG recordings, where the presented noisy signal was always ECG-dominant. In this case, the adaptive filter was likely to yield satisfactory effect of suppressing ECG noise when the reference signal was good enough. However, if the ECG contamination has similar amplitude to the EMG, the noise reduction performance is more sensitive to how synchronized the reference signal is with the contaminated noise. In contrast, the ICA applied on the EMG recordings with ECG contamination is able to derive ECG-dominant IC that is suitable for adaptive filters, as demonstrated in Fig. 2. Consequently, the proposed method could also yield satisfactory performance when the ECG artifact is not dominated in the noisy EMG recordings.R EFERENCES[1]0 $ 2VNRHL DQG + +X ³0\RHOHFWULF FRQWURO V\VWHPV²AVXUYH\ ´Biomedical Signal Processing and Control, 2(4): 275-294, Oct.2007.[2]Y. Li, X. Chen, X. Zhang, K. Wang, Z. Jane Wang, ASign-Component-Based Framework for Chinese Sign Language Recognition Using Accelerometer and sEMG Data, IEEE Trans. on Biomedical Engineering, vol. 59, no. 10, pp. 2695 ± 2704, Oct. 2012.[3]X. Li, A. Suresh, P. Zhou, W. Rymer. Alterations in the SpikeAmplitude Distribution of the Surface Electromyogram Post-Stroke.IEEE Trans on Biomedical Engineering, 2012 Jun 19. [Epub ahead of print] (in press)[4]X. Zhang, P. Zhou, Sample entropy analysis of surface EMG forimproved muscle activity onset detection against spurious background spikes, Journal of Electromyography and Kinesiology, vol. 22, pp.901±907,2012.[5]H. Chan, Y. Tsai, L. Meng, T. Wu, The Removal of Ocular Artifactsfrom EEG Signals Using Adaptive Filters Based on Ocular SourceComponents, Annals of Biomedical Engineering, vol. 38, no. 11, pp.3489±3499, Nov. 2010.[6] C. J. DeLuca, L.D. Gilmore, M. Kuznetsov, S.H. Roy, Filtering thesurface EMG signal Movement artifact and baseline noisecontamination, Journal of Biomechanics, vol. 43, pp. 1573±1579,2010.[7]J.N.F. Mak, Y. Hu, K.D.K. Luk, An automated ECG-artifact removalmethod for trunk muscle surface EMG recordings, MedicalEngineering & Physics, vol. 32, pp. 840±848, 2010.[8] C. Marque, C. Bisch, R. Dantas, S. Elayoubi, V. Brosse, C. Perot,Adaptive filtering for ECG rejection from surface EMG recordings,Journal of Electromyography and Kinesiology, vol. 15, pp. 310 ± 315, 2005.[9]G. Lu, J.S. Brittain, P. Holland, J. Yianni, A.L. Green, J.F. Steina, T.Z.Aziza, S. Wanga, Removing ECG noise from surface EMG signalsusing adaptive filtering, Neuroscience Letters, vol. 462, pp. 14-19, Sep.2009.[10]M.S. Redfern, R.E. Hughes, D.B. Chaffin, High-pass filtering toremove electrocardiographic interference from torso EMG recordings, Clin Biomech, vol. 8, pp. 44 ± 48, 1993.[11]J.D. Costa Junior, D.D. Ferreira, J. Nadal, A.M.F.L. Miranda de Sá,Reducing electrocardiographic artifacts from electromyogram signals with Independent Component Analysis, 32nd Annual InternationalConference of the IEEE EMBS, 2010.[12] A. Hyvärinen, J. Karhunen, E. Oja. Independent Component Analysis,New York: John Wiley & Sons, 2001.[13]J. Yang, H. Sakai, A New Adaptive Filter Algorithm for SystemIdentification using Independent Component Analysis, IEICE Trans on Fundamentals of Electronics, Communications and Computer Sciences, vol. E90-A, no.8, pp. 1549-1554, 2007.[14]N.W. Willigenburg, A. Daffertshofer, I. Kingma, J.H. van Dieën,Removing ECG contamination from EMG recordings: A comparisonof ICA-based and other filtering procedures, Journal ofElectromyography and Kinesiology, vol. 22, no. 3, pp. 485 ± 493, June 2012.[15]P. Zhou, P.E. Barkhaus, X. Zhang, W.Z. Rymer, Characterizing thecomplexity of spontaneous motor unit patterns of amyotrophic lateral sclerosis using approximate entropy, Journal of Neural Engineering,vol. 8, no. 6, 2011.。