Combined surface electromyography , near-infrared spectroscopy and acceleration recordings of muscle contraction : The e ® ect of motion

This is an Open Access article published by World Scienti ̄c Publishing Company. It is distributed under the terms of the Creative Commons Attribution 4.0 (CC-BY) License. Further distribution of this work is permitted, provided the original work is properly cited. Journal of Innovative Optical Health Sciences Vol. 10, No. 2 (2017) 1650056 (13 pages) #.c The Author(s) DOI: 10.1142/S1793545816500565


Introduction
Electromyography (EMG) is most commonly used in the ¯eld of neurophysiology or electrodiagnostic medicine. 1There are multiple ways of gaining data on muscle activity, including intramuscular measurements performed using a needle or a wire. 2 EMG studies can also be realized using noninvasive techniques, which is also preferable for patients, because intramuscular measurements can be painful and may potentially lead to destruction of muscle ¯bers. 3Surface EMG (sEMG) employs surface electrodes that can be attached to skin overlying the muscle. 4The signal is an interference pattern of all the active motor units (MUs) that are the smallest functional groupings in a muscle. 2,5sEMG has been used to study muscle activation, the muscle force and fatigue processes. 1,6,7he origin of the sEMG signal is the bio-electric activity of muscle ¯bers, which is due to ion channel activity (e.g., Na þ , K þ , Cl À , Ca 2þ Þ and it can be described as a zero-mean random process, where the standard deviation is comparative to the ¯ring rate and number of active MUs. 1,2The recorded signal represents the average potential of the area and appears on the skin with a limited bandwidth, ranging approximately from 10 Hz to 400 Hz. 8,9The signal amplitude tends to be extremely small as well, from some microvolts to a few millivolts peakto-peak. 6,8The electrodes of sEMG record this electrical activity in the muscle or transduce its ionic current. 10The electrodes record from a large muscle area, which means that the recorded sEMG signal can be a®ected by electrode characteristics, including physical dimension, shape, technology and constituent materials as well as the electrode placement and orientation. 1,6,9,10Moreover, electrodes have low-signal resolution, a proportionately narrow frequency band and are sensitive to movement artifacts. 1,2In addition, the subcutaneous fat layers and high skin layer conductivity cause signal blurring similar to EEG studies. 1 As the amount of subcutaneous tissue increases, the median frequency decreases. 6One important consideration is the electrode-skin impedance, as variations therein may a®ect the recorded sEMG signal.This can be due to drying of the gel, changes in skin properties or changes in the electrolyte-skin interface.An additional issue is the noise level at the electrode-skin interface, which is highly dependent on skin treatment. 10As a result, sEMG often requires some degree of skin preparation such as cleansing the skin of body oils with alcohol, inserting conductive paste or gel, and, in some cases, shaving body hair. 2,8,11In some measurements, the outermost layer of the skin is also abraded. 11n assessing muscle function, local oxygen consumption and muscle metabolism are often of great interest.Such measurements rely on using opticsbased methods, particularly near-infrared spectroscopy (NIRS). 12In addition to studies of brain function, NIRS is being increasingly utilized for the study of muscle function.Three molecules account for most of the NIRS light absorption: hemoglobin (both oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (Hb)), myoglobin (Mb) and cytochrome oxidase c. 13 Mb, an iron-containing protein found in muscle ¯bers, is structurally similar to Hb, and their NIRS spectra are very similar.Consequently, it is currently not possible to explicitly distinguish the contributions of Mb and Hb to °uctuations of deoxygenation, and deoxygenation in a muscle is commonly represented as their sum: deoxy [Hb þ Mb]. 14 Calculation methods for estimating these concentrations in NIRS are based on the modi¯ed Beer-Lambert law (MBLL).Being well-known, MBLL has been widely described in numerous publications. 14,15The general idea is that the absorption of the tested sample is proportional to the concentration of the absorber, its extinction coe±cient and the path length of light traveling through the sample.
Though the e®ects of muscle oxygenation status on resistance training adaptations, for example, are not fully understood, NIRS has become an increasingly popular choice for monitoring local muscle oxygenation.NIRS re°ects the balance of oxygen delivery to working muscles and oxygen consumption in capillary beds, 10 such measurements are commonly performed, for example, during exercise. 19,20At present, muscle activity can also be monitored in real time by portable multi-channel wireless NIRS devices. 21oth sEMG and NIRS have a high time resolution, but a rather limited spatial resolution when measuring noninvasively deep within a tissue.For example, NIRS has a depth sensitivity of approximately 1.5 cm and the spatial resolution of approximately 1 cm. 16A useful method for improving spatial accuracy involves the use of multidetector probes, 31 which allow measurements of back-scattered light at di®erent distances to be combined in order to obtain scattering and absorbance properties at di®erent tissue depths. 32In the same vein, there is a movement towards sEMG recording with two-dimensional (2D) array systems, known as high-density sEMG (HD-sEMG). 1In HD-sEMG, muscle activity is monitored with multiple closely placed electrodes that overlie a restricted area of the skin.This allows recording muscle activity in both the temporal and spatial dimension. 33HD-sEMG enables pinponting the location of innervation zones with high accuracy, estimating the velocity, length and orientation of muscle ¯ber conduction, estimating MU locations, decomposing sEMG interference signals and analyzing the properties of single MUs. 1 Multimodal studies of muscle function are becoming increasingly common, most of them based on a combination of NIRS and EMG. 18,22,23Many of these studies focus on investigating the relationship between muscle electrophysiology and oxygenation or hemodynamics [24][25][26][27][28] or on gathering more reliable information on, for example, neuromuscular and metabolic activity in relation to muscle fatigue or injury. 29,30Moreover, wireless hybrid sEMG/NIRS sensors have recently been proposed for multimodal analysis by such researchers as Guo et al. 34 Surface EMG and accelerometer (ACM) have been previously used for example for sign language recognition, 45 fatigue assessment, 46 Parkinson's disease progress assessment 47 and for monitoring daily activities of patients who have had a stroke. 48

E®ect of Motion Artifacts
Two kinds of motion artifacts may appear in measurements of muscle function: motions caused by sudden movements, such as swinging of the arm, and tremor generated by the muscle contraction mechanism. 49Particularly, in noninvasive and wearable sensors these are present and should be considered in the analysis.
In general, muscle motion may induce motion artifacts into the signal of interest, which can mislead the subsequent analysis.However, removing these artifacts may also lead to loss of useful data.Most artifact removal methods focus on motions caused by sudden movements, which are commonly present in noninvasive and wearable applications.For example, proposing ACM-based motion artifact removal (ABAMAR), Virtanen et al. 52 applied a 2D ACM to detect the presence of motion in sleeping subjects.They sought to correct the baseline shift of the NIRS signal and used a threshold, which was in the rate of acceleration change, to detect movement.On detecting a movement, a °ag is raised for a time interval Tm.However, if Tm is less than 1 s, the system sees no movement.To correct the baseline shift, the mean of the NIRS signal before and after Tm are calculated separately to obtain A before and A after .If the absolute di®erence between A before and A after exceeds 2.6 times the SD before TM, a baseline shift change is identi¯ed.The amplitude of the signal within Tm is set to A before , while that of the signal after Tm is multiplied by A before / A after .In another strand of research, Molavi et al. 41 applied wavelet transform to remove motion artifacts from muscle NIRS measurements without using an ACM.This type of wavelet analysis produces coe±cients that make motion artifacts appear as isolated large coe±cients in the wavelet domain.To counter that, coe±cients which do not belong to the original signal distribution are removed by a statistical test at a speci¯c probability level.This probability is the tuning parameter of the proposed algorithm.
E®ect of motion artifacts caused by muscle contraction itself have been studied to a lesser extent.Using state-of-the-art, technology can minimize the in°uence of some extrinsic noise sources as power line noise and cable motion artifact as well as intrinsic thermal noise originating from the electronics of the ampli¯cation system.In order to reduce the noise components of the signal, Luca et al. 50roposed to include a high-pass ¯lter to sEMG systems, which reduces baseline noise and the motion artifact from the movement without removing too much of the low-frequency components.Nevertheless, the choice of right high-pass corner frequency depends on application and muscle.Raez et al. 51 also suggest that a way to reduce motion artifact is by designing the electronic circuitry and set-up properly.Signal conditioning can also be used in motion artifact reduction.Common ways are the use of adaptive and wavelet ¯lters. 2 In our approach, we pursued to fatigue the muscles and cause muscle tremor.Surprisingly, we did not ¯nd articles about the role of muscle tremor to the motion artifacts.
In the following, we present a multimodal sensor built for performing simultaneous measurements with sEMG, NIRS and ACM.The aim was to study the e®ect of motion in isometric and isotonic exercises targeting the biceps brachii muscle.A muscle contraction is described as isometric when the muscle contracts without a change in its length, and isotonic when muscle length changes, while the contraction force remains the same. 52

The measurement device
For measurements of sEMG and acceleration, we utilized a noncommercial prototype produced by Fibrux. 53The prototype consists of a sEMG sensor and inbuilt three-dimensional (3D) acceleration sensor.The cut-o® frequencies for the sEMG sensor are 15 Hz and 250 Hz and for the ACM 10 Hz.The prototype by Fibrux was integrated with a NIRS device, built in the Optoelectronics and Measurement Techniques Unit, University of Oulu.The NIRS device utilizes a commonly known method, lock-in ampli¯cation, which is closely related to the frequency-domain NIRS technique. 35In lock-in ampli¯cation, instead of ampli¯cation of the received signal over the whole frequency band of interest, only amplitude-modulated light signals at a speci¯c narrow-band reference frequency are am-pli¯ed.As a result, a signi¯cant increase can be achieved in signal-to-noise performance.When an ampli¯ed reference signal is multiplied by the lockin reference signal and then low-pass ¯ltered, intensity changes in the remaining low-pass ¯ltered output signal correlate with attenuation changes in the received light signal.Furthermore, the lock-in ampli¯cation technique enables separation of several signals from each other.Thus, it is possible to label each wavelength of light for a speci¯c modulation frequency, illuminate all wavelengths simultaneously and then separate them in the receiver. 35ormally, this technique requires multi-distance recordings to ¯t a linear regression of distance versus changes of phase, to accurately estimate hemodynamics derived from the ¯tted slopes. 36However, mainly because of the large electrode size of sEMG sensors, we decided to use a single distance pair of optodes, located at 3 cm from each other, with two EMG electrodes placed between the optodes, as shown in Fig. 1.Although this type of single distance technique has its limitations, it is used in most NIRS-based muscle studies. 17Temporal oxygenation changes were calculated at the wavelengths of 660 nm and 830 nm.In our NIRS, sEMG and acceleration recordings, the sampling rate was 500 Hz, and synchronization between the di®erent modalities was provided by a trigger pulse.
HbO, Hb þ Mb and total deoxy t(Hb þ Mb) were calculated from raw NIRS time courses using MATLAB's HomER2 NIRS-processing package. 35 di®erential pathlength factor (DPF), needed for the calculation, is determined by the absolute optical pathlength divided by the known distance between optodes.The value of DPF is dependent on the place of measurement.In the case of the adult forearm, for example, its mean value is 3:59 AE 0:32 for both male and female subjects, while the value for the adult brain is 5:93 AE 0:43. 37On the other hand, Duncan et al. have reported separate DPF values for males and females as a function of wavelength.At 832 nm, DPF was 3:53 AE 0:55 for males and 4:34 AE 0:78 for females.At 690 nm, the DPF value for males was 3:96 AE 0:62, while that of females was 4:91 AE 0:81.Why females have a statistically signi¯cant higher forearm DPF than males is probably due to a thicker fat layer. 38Nonetheless, DPF is still a constant value which in°uences only the absolute values of Hb, HbO and HbT. 39

Subjects and performed task
The sensor was placed on the subjects' right biceps brachii (Fig. 2.) and ¯xed in place with an elastic band.Five healthy volunteers participated this study, performing three exercises (static, dynamic and isometric) targeting the biceps brachii muscle.Static and dynamic exercises were conducted on the same day, while the isometric exercise took place a week later.A static exercise was always conducted ¯rst, followed immediately by a dynamic exercise.The purpose was to fatigue the muscle in order to examine correlations between sEMG and NIRS data, and to observe the e®ect of muscle motion recorded simultaneously by an ACM.In both static and dynamic exercises, the subject was sitting in a chair with their arm on a table.No weights were used, since we wanted to study the e®ect of natural muscle power.The subjects were instructed to tense the muscle for 10 s and then relax it for 20 s.This protocol was repeated ¯ve times.In the dynamic task, the subjects did ¯ve sets of ¯ve braking contractions of biceps brachii at maximal power, with a 20 s rest between the sets.In the isometric exercise, the subjects ¯rst stood straight and then leaned against a wall while holding a customized 5 kg weight at an elbow angle of 90 for as long as possible.This isometric exercise was repeated after 1 min break.It includes an optode consisting of a ¯ber optic near-infrared (NIR) light source at a distance of 3 cm from the ¯ber optic detector, both with a ¯ber thickness of 2.5 mm; a built-in 3D ACM and two copper electrodes with a ground plate placed between the optode.Thus, NIRS and sEMG measured the average activity of the muscle from approximately the same area.This noncommercial sEMG sensor prototype was produced by Fibrux.
In our experiments, our aim was to study the motions caused by the muscle itself.Thus, the subjects were advised to avoid other movements than °exing the biceps brachii muscle.

Results
SEMG signals can be evaluated both in the time and frequency domains.Among the most common time-domain methods are root mean square (RMS) and average muscle ¯ber conduction velocity (CV).In the frequency domain, the focus of attention is on the median frequency of the power density function. 40Frequency domain measurements can be useful in fatigue assessment, as fatigue causes a shift in median frequency.
In the following Figs.3-6, we chose to survey the signals in time domain in order to demonstrate the clear correspondence between the three modalities in contrast to muscle contraction.All presented signals are measured from right biceps brachii (see Fig. 2) of Subject 4. The responses of the other subjects were similar.Shown are deoxy (Hb þ Mb), original data (blue) and upper envelograms (red) of synchronized sEMG and ACM signals, taken with 100 point smoothing.Additionally, the signals are normalized to scale 0 to 1 using Origin 2015 version from Originlab.Deoxy (Hb þ Mb) were obtained from the NIRS data using MATLAB.The spectral range of all presented envelope signals is approximately between 0-1.8 Hz, thus shown are lower signal °uctuations caused by muscle contraction and motions.In addition, higher frequency transients for EMG and ACM are included with the envelopes, except in Figs. 5 and 6.
In the performed experiments, test subjects found it di±cult to control the contraction of the muscle accurately or to repeat the contraction equally, thus such duration and the strength of contraction constantly varied.Thus, statistical analysis in this case would not be relevant.Thereby, in the following, we show only examples of signal responses from single subjects.
Figure 3 shows an e®ect of four static muscle contractions.The muscle contraction causes a simultaneous response in ACM, muscle deoxy (Hb þ Mb) concentration with the activation of MUs, visible in sEMG.
Figure 4 shows signal responses for dynamic contractions.In dynamic exercise, we can see steep spikes both in deoxy (Hb þ Mb) and sEMG signals.steeply and receives the maximum value at the same time as sEMG signal has its maximum.There are high spikes at 70 s and 117 s which are easily to spot at the beginning of the short braking contraction series.However, this motion artifact was not seen in EMG and ACM measurements.
Figures 5 and 6 show responses for isometric contractions when continuously holding up a 5 kg weight as long as possible, followed by a 1 min break before the second contraction.In the beginning, the ACM and EMG signals are quite stable but as the fatigue starts to increase, due to the long  isometric contraction, more muscle tremor is present, particularly in Fig. 5 between 160 s-200 s, and in the second contraction, Fig. 6, already between 30 s-90 s.Moreover, deoxy (Hb þ Mb) slightly °uctuates.The signal behavior varies along the test subjects, however, with the same main character as presented here.

The e®ect of motion
(a) Spikes in signal Many higher spikes in ACM, visible in static and dynamic exercises, were most likely caused by sudden muscle motions.These, however, were lesser visible in corresponding sEMG and NIRS signals which positively indicates that sEMG and NIRS are more sensitive to muscle contraction rather than its sudden motions.Interestingly, on the other hand, the dynamic experiment measurement, in Fig. 4, shows motion artifact induced to NIRS as high spikes at 70 s and 117 s which are easily to spot at the beginning of the short braking contraction series.Mostly likely, the spikes in NIRS are caused by the movements of the optode, because these are only visible in NIRS signal.Thus, we cannot utilize data from ACM to remove the artifact, for example, with the method proposed by Virtanen et al. 52 However, there is still possibility to remove this artifact using algorithm proposed by Molavi et al. 40 because it requires signal from NIRS measurement only.Utilizing this algorithm, the ¯rst spike on NIRS signal could be eliminated, as seen in Fig. 7, but the shape of the NIRS signal changed a bit.However, the second spike could be removed completely but the algorithm can attenuate it.

(b) Muscle tremor
During contraction, the muscle produces constant tremor generated by the muscle contraction mechanism. 48

EMG (dynamic)
Original data Upper envelope Fig. 7. Applying motion artifact removal to dynamic experiment.Top: original NIRS signal, middle: NIRS signal after removing motion artifact, bottom: EMG signal.
Fig. 9, the di®erence is greater, particularly in EMG.In addition, we noticed clear shift in median frequency of the power density, when compared the ¯rst contraction to the last contraction of the exercise.This was in accordance to the commonly known phenomenon that fatigue causes a shift in median frequency. 40able 1 shows correlations between acceleration signal and corresponding EMG signal during static and dynamic exercises.The average correlations for ¯ve subjects is 0.18 (static) and 0.22 (dynamic).   and Virtanen 52 can be potentially used, in particular, when the signal analysis is based on multimodal data, gathered using such presented measurement device.According to Elcadi 12 in healthy subjects, there is an inverse linear relationship between NIRS and EMG signals, meaning that as EMG amplitude increases, oxygenation decreases. 1This correlation between increasing load, EMG and VO 2 has also been found in other studies, [40][41][42][43][44] which indicates that muscle force increases as a function of load. 42hanges in EMG amplitude are mainly caused by a change in the number of active MUs and their mean ¯ring rate. 40Our results show that as the EMG amplitude increases, there is a simultaneous increase in deoxy concentration (Hb þ Mb), see Figs. 3-6.Moreover, Figs. 5 and 6 indicate that the tremor in ACM as well as in EMG, caused by the isometric contraction, increases due to fatique, while deoxy (Hb þ Mb) starts to slightly °uctuate.
Despite clear advantages o®ered by the sEMG technique, it is di±cult to integrate into routine clinical use due to a lack of standards for sensors, con¯gurations, electrode placements and recording protocols.Nonetheless, the utility of sEMG as a tool of examining neuromuscular function is obvious.Moreover, there is an ongoing e®ort to standardize and validate sEMG.A case in point in Europe is SENIAM (surface EMG for the noninvasive assessment of muscles).Also the International Society of Electromyography and Kinesiology (ISEK) has put forward recommendations regarding the use of sEMG techniques and methodologies. 3

Conclusions
A combination of these three noninvasive methods o®ers valuable insights for the study of muscle function.In addition to energy consumption data provided by NIRS, we also gain knowledge about muscle level activation using sEMG, and the e®ect of motion on these two signals.However, the exact relationship between the modalities and physiological event still need further studies.
SEMG is an intriguing research area and, combined with NIRS and ACM, it can deliver multiple bene¯ts.For example, HD-sEMG yields new information about the central nervous system and MUs.Moreover, as current clinical methods can be painful, they render large studies practically impossible.A noninvasive technique, in contrast, would increase the number of available test subjects, allowing researchers and clinicians to reap major bene¯ts.

Fig. 2 .
Fig. 2. Anatomical image of the arm of Subject 4. The measuring place was on the right biceps brachii, illustrated with white circles in the anatomical image in (a) sagittal (b) coronal and (c) axial slices.An important challenge for sEMG and NIRS is the fact that soft tissue (mostly fat), shown in white color, intervenes with the signal, especially when the active muscle is 10 mm or more below skin surface. 1

Fig. 1 .
Fig.1.Illustration of the sensor.It includes an optode consisting of a ¯ber optic near-infrared (NIR) light source at a distance of 3 cm from the ¯ber optic detector, both with a ¯ber thickness of 2.5 mm; a built-in 3D ACM and two copper electrodes with a ground plate placed between the optode.Thus, NIRS and sEMG measured the average activity of the muscle from approximately the same area.This noncommercial sEMG sensor prototype was produced by Fibrux.

Fig. 3 .
Fig. 3. Signal responses from ACM, deoxy (Hb þ Mb) and sEMG during a static exercise.There is a strong °uctuation in all the three signals caused by the four static contractions.

Fig. 4 .Fig. 5 .
Fig.4.Signals of ACM, deoxy (Hb þ Mb) and sEMG during dynamic exercise.We can see that the ACM signal also °uctuates between the dynamic contractions, and also after the third sets of contractions.This, however, is not visible in deoxy (Hb þ Mb) and sEMG signals which indicates that deoxy (Hb þ Mb) and sEMG are more sensitive to muscle contraction.

Fig. 6 .
Fig.6.The response of the second isometric contraction of Subject 4. In the second contraction, Subject 4 could hold up the 5 kg weight only for 90 s whereas in the ¯rst contraction duration was 200 s.We can see that due to the fatigue caused by the ¯rst isometric exercise, the tremor e®ect in ACM as well as in EMG is more present in the second than in the ¯rst isometric contraction.Contraction causes increase in deoxy (Hb þ Mb) and °uctuates a little more when fatique starts.

Fig. 8 .Fig. 9 .
Fig. 8. Power spectra of EMG (a) and ACM (b) during rest (red) versus static contraction (black).During contraction, the power spectra of ACM is visible up to 40 Hz and in EMG up to 150 Hz.
This is clearly visible in ACM signal in all Figs.3-6, as well as in Figs.8 and 9.However, how this a®ects the EMG and NIRS signals should be more studied.In Fig.8is shown the tremor e®ect in ACM signal and corresponding EMG signal, at spectrum range of 15 Hz-250 Hz.When compared the spectrum between rest and contraction, there is obviously a notable di®erence in the power density.Moreover, during dynamic contraction, shown in

Table 1 .
Correlation between ACC and EMG during static and dynamics exercises.