december 7-9, 2016. florianopolis, brazil sciencedirect · · 2017-04-01peer review under...

IFAC-PapersOnLine 49-32 (2016) 183–187

ScienceDirectScienceDirect

Available online at www.sciencedirect.com

2405-8963 © 2016, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.Peer review under responsibility of International Federation of Automatic Control.10.1016/j.ifacol.2016.12.211

© 2016, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.

Realtime EMG analysis for transcutaneouselectrical stimulation assisted gait training

in stroke patients �

T. Schauer ∗ T. Seel ∗ N. D. Bunt ∗,∗∗ P. Muller ∗

J. C. Moreno ∗∗∗

∗ Control Systems Group, Technische Universitat Berlin, Berlin,Germany (e-mail: [email protected])

∗∗ University of Twente, Enschede, The Netherlands∗∗∗ Spanish National Research Council, Cajal Institute, Madrid, Spain

Abstract: This contribution describes a method for realtime analysis of muscle activity duringapplication of Functional Electrical Stimulation (FES) to the assessed muscles. Inertial sensorsat the foot are used for realtime gait phase detection in order to synchronize the stimulationwith the gait. After detecting and muting stimulation artifacts and after extraction of Inputer-Pulse Intervals (IPIs), a non-causal high-pass filter is applied to a section of the IPI to extractthe voluntary EMG activity. The filter suppresses FES-evoked EMG activity (M-wave) andelectrode discharging artifacts. The initial filter states are chosen by an optimization procedureto minimize undesired filter transients. The obtained filtered EMG signal is then rectified andaveraged to produce a scalar measure of the volitional EMG activity over the last IPI. Thevolitional EMG activity during four different detected gait phases is calculated after everycompleted step and displayed to the stroke patients for biofeedback or to the therapist in orderto adjust the FES. The system has been initially evaluated with healthy subjects walking on atreadmill. It was demonstrated that different walking styles of an individual can be distinguishedby the EMG analysis also during active FES support.

Keywords: Signal Processing, Biofeedback, Optimization, Gait, Rehabilitation, FunctionalElectrical Stimulation, Filter Design

1. INTRODUCTION

A large part of the population in developed countries willbe affected by a stroke during their lifetime (Truelsenet al., 2006). Most people suffering from a stroke willundergo rehabilitation therapy, to regain some or all oftheir previous function. The therapy will however cometo an end at some point. A part of the stroke patientswill be allowed additional therapy after some time. Dur-ing this later periods, the rehabilitation will mainly befocused on gaining additional strength and improving thecondition of the patient. The actual movement patternsof the patient are most often kept unaltered by the ther-apy. We believe this later stage of rehabilitation wouldgreatly benefit from the possibility to have realtime bio-feedback on muscle activity, e.g. during walking, as wellas the possibility to interfere with the movement on themuscular level. This interference could either be sensorystimulation, to communicate to the patient the timing atwhich the muscle of interest should be activated (Lauferand Elboim-Gabyzon, 2011). Or it could be functionalelectrical stimulation, evoking a muscle contraction thatis beneficial to the movement pattern (Kafri and Laufer,2015). By giving the rehabilitation practitioner realtime

� The work was conducted within the research project BeMobil,which is supported by the German Federal Ministry of Educationand Research (BMBF) (FKZ16SV7069K).

feedback about volitional muscle activity during the exer-cise, a better understanding of the current gait and theeffect that the stimulation has on the volitional muscleactivation and movement will be obtained. The combinedconcept of the rehabilitation practitioner’s experience andthis technology could result in an improved rehabilitationtherapy.

2. MATERIALS AND METHODS

2.1 System Description

Two wireless Inertial Measurement Units (IMUs) andtwo wireless EMG sensors with an average transmissionlatency of 50 ms are used (MUSCLELABTM, ErgotestInnovation A/S, Norway). The IMUs are placed on theinstep of both feet. Each EMG sensor possesses two bipolarmeasurement channels and is equipped with a referenceelectrode. The EMG is measured at 1000Hz, and accel-erations and rates of turn of the IMUs are sampled at50Hz. For stimulation, a current-controlled multi-channelstimulator (RehaStim I, Hasomed, Germany) with gal-vanically isolated USB interface is used. The stimulationfrequency is set to 25Hz, and bi-phasic pulse duration pwand amplitude I can be adjusted in realtime from pulse topulse. For acquiring data and for controlling stimulation,interface blocks have been programmed in Simulink R© (TheMathworks Inc., USA) and realtime code generation is

1st IFAC Conference on Cyber-Physical & Human-SystemsDecember 7-9, 2016. Florianopolis, Brazil

Copyright@ 2016 IFAC 183









1. INTRODUCTION

















1. INTRODUCTION

















1. INTRODUCTION

















1. INTRODUCTION









184 T. Schauer et al. / IFAC-PapersOnLine 49-32 (2016) 183–187

performed by means of the Linux Target for Simulink R©

Embedded Coder R©.

2.2 Stimulation pattern

The FES is administered in synchronization with the gaitcycle. A velocity-adaptive realtime gait phase detection(GPD) (Seel et al., 2014; Muller et al., 2015) is employedto trigger the stimulation. Four gait phases (foot flat, pre-swing, swing phase, loading response) and four gait events(full contact, heel rise, toe-off, initial contract) are detectedas shown Fig. 1.

full contact initial contacttoe-offheel rise

foot flat pre-swing swing phase loadingresponse

Fig. 1. By means of a foot/shoe-mounted inertial sensor,the transitions between four distinct gait phases canbe detected in realtime.

Up to eight muscles can be stimulated dependent on thegait deficits of the stroke patient. For each muscle thebeginning and end of the stimulation interval can be set toa gait event. Additionally, the time points can be shifted(with respect to the selected gait events) forward andbackward in time by a given percentage of the estimatedtotal gait cycle duration.

2.3 EMG Signal Processing

Electromyography (EMG) can be used for multiple pur-poses in FES gait training. Figure 2 displays an exemplaryraw surface electromyography (sEMG) recording duringactive FES. When analyzing EMG signals, one has to dis-tinguish between FES-evoked EMG and patient-inducedEMG, where the latter includes both intentional (voli-tional) and unintentional muscle activity (Merletti et al.,1992). By means of online signal processing, both quanti-ties can be determined from the raw EMG also in betweenthe stimulation pulses, i.e. during active stimulation. TheFES-evoked EMG is manifested in the so-called M-wavewhich is a good measure for the amount of motor unitsrecruited by the last stimulation impulse. Recent studiesfor the upper extremities show that feedback control of theM-wave magnitude compensates the effects of muscularfatigue and maintains a desired stimulation effect (e.g.force production) (Klauer et al., 2012, 2016).

The EMG that is due to patient-induced muscle activityis much smaller than the M-wave. This rather noise-likesignal with frequency components in the range of 30 to300 Hz (De Luca and Knaflitz, 1992) can be separatedfrom the M-wave about 20 to 30 ms after each stimulationpulse by high-pass filtering or by subtraction of an esti-mated/predicted M-mave, see e.g. (Ambrosini et al., 2014).The patient’s EMG can be used to trigger the stimulationonset, to modulate the intensity profile of stimulation orsimply to monitor the effect of stimulation on the muscleactivity and motor coordination of the patient.

To enable a robust detection of stimulation pulse instantsand inter-pulse-intervals, we also stimulate the muscles

74.2 74.3 74.4 74.5 74.6 74.7 74.8

Time [s]

-2

0

2

4

6

8

10

Raw

EM

G [m

V]

74.2 74.21 74.22 74.23 74.24 74.25

-2

-1

0

1

2Stimulation artefact

M-wave

A

74.52 74.53 74.54 74.55 74.56 74.57

-2

-1

0

1

2 B

Fig. 2. EMG recording during active stimulation with astimulation frequency of 25Hz. (A) Stimulation pe-riod with almost no volitional muscle activity, (B)stimulation period with clearly visible volitional mus-cle activity.

at a sub-sensory level (I = 6mA, pw = 50µs) when nofunctional stimulation at a sensory or motor level is active.

The last 500 received EMG samples are held in buffer(covering 0.5 s) for online analysis. This buffer is period-ically investigated to find stimulation artifacts. For this,the raw EMG signal is double differentiated and rectified.Stimulation artifact are then determined by searching localmaxima (peaks) in the signal using the Matlab functionfindpeaks with appropriate setting for the function pa-rameters MinPeakDistance and MinPeakProminence. Thefirst parameter describes the minimally required distancebetween two stimulation artifacts and is set to 38ms.The second parameters describes the required minimalprominence of the peaks and is set half as large as theobserved signal variance in the buffer. The prominenceof a peak measures how much the peak stands out dueto its intrinsic height and its location relative to otherpeaks. The obtained group of detected peaks is furtherextended by reconstructing stimulation time points in-between found peaks at the stimulation frequency.

After the stimulation instants are determined, a regionaround each point is suppressed by setting the value forthis period to zero. The most recent found IPI is extractedfor further analysis. At a stimulation frequency of 25Hzand an EMG sampling frequency of 1000Hz, it contains40 EMG samples. EMGi(k), k = 1, . . . , 40, represents thek-th sample within the i-th found stimulation period.

To determine the volitional EMG activity, we first extractthe EMG from sample N1 to N2 of the IPI. This sub-interval contains beside volitional EMG activity the low-frequent tail of the M-wave and voltage transients from anyremaining charge on the electrodes after the stimulus. Toremove the latter and to extract the higher frequent partof the volitional EMG activity we apply a non-causal high-pass filtering in which the initial filter states are chosen sothat filter transients become minimal. A 6th-order elliptichigh-pass filter with a passband edge frequency of 200Hz,3 dB of ripple in the passband, and 80 dB of attenuationin the stop band is used to filter the EMG data forwardand backwards in time. To determine the optimal initialfilter state, we rewrite the entire filter process in vectorial

2016 IFAC CPHSDecember 7-9, 2016. Florianopolis, Brazil

184

T. Schauer et al. / IFAC-PapersOnLine 49-32 (2016) 183–187 185

form. Let the input vector for the filtering process be

U = [EMGi (N1) , EMGi (N1 + 1) , · · · , EMGi (N2)]T.

Filtering forward in time: The 6th order Elliptic high-pass filter is given in form of a state-space model repre-sentation (F ∈ R6x6,G ∈ R6x1,E ∈ R1x6, D ∈ R):

x(k + 1) = Fx(k) +Gu(k)

yf (k) = Ex(k) +Du(k),

where k = N1, . . . , N2 is the sample index and u(k) areelements of U . The vector of the output samples of thisforward filtering process is

Y f = [yf (N1) yf (N1 + 1) . . . yf (N2)]T

Using the Toeplitz matrix

Q =

D 0 · · · 0 0EG D · · · 0 0EFG EG · · · 0 0

......

. . ....

...EFN−3G EFN−4GG · · · D 0EFN−2G EFN−3GG · · · EG D

and the observability matrix

O =[E,EF ,EF 2, . . . ,EFN−1

]Twe can write the forward filtering problem as

Y f = QU +Ox

where x = x(N1) is the initial state of forward filtering.

Non-causal filtering: Let us introduce first the row andand column reversion operatorsR and C with the followingproperties (A,B and C are compatible matrices):

A=BC

R(A) =R(B)C

C(A) =BC(C)

BR(C) = C(B)C

C(R(BC(R(C))) = C(R(B))C

C(R(A)) =AT if A Toeplitz

The forward-backward filtering involves these steps:

(1) Filter U through (F ,G,E, D) forward in time toobtain the vector Yf ,

(2) Reverse the result Yf in time by applying the rowreversion operator,

(3) Filter the reversed sequence R(Yf ) through (F , G,E, D) again,

(4) Time-reverse the last filter output again to obtain theforward-backward filtered sequence Yfb

Using the previously introduced matrices Q and O we canwrite this process as follows

Yf =QU +Ox

Yfb =R(QR(Yf ) +Ox)

=R(QR(QU +Ox) +Ox)

=R(Q)R(Q)U +R(Q)R(O)x+R(O)x

=QTQU +QTOx+R(O)x

with the initial states x and x = x(N2) of the forward andthe backward filtering, respectively.

Determine the optimal initial states: In order to reduce

filter transients, the initial state vector xinitial = [x,x]Tis

chosen in such a way that the cost function J = Y TfbY fb

becomes minimal. To determine the optimal initial statevector, we set the first derivative of the cost function tozero

∂Y TfbY fb

∂xinitial= 0

and solve for xinitial. This yields the optimal initial state

xoptinitial =

[([QTO R(O)

]T [QTO R(O)

])T]†

·((−QTQU)T

[QTO R(O)

])T

where † is the pseudo inverse. The non-causal filter is then

Yfb =QTQU +[QTO R(O)

]xoptinitial.

The volitional muscle activity EMGVi of the interpulse

interval i is finally obtained by rectification and meanvalue calculation of the forward and backward filteredEMG:

EMGVi =

1

N2 −N1 + 1

N2∑k=N1

|yfb(k)|.

The mean volitional EMG activity during the differentdetected gait phases is calculated after every completedstep and displayed to the stroke patients for biofeedbackor for the therapist in order to adjust the FES.

3. RESULTS

In first experiments, four healthy test subjects were walk-ing on a treadmill at different speeds with and withoutFES. The subjects were asked to keep their normal walk-ing pattern. For the experiments, the setup as shown inFigures 3 and 4 is used. EMG is measured at the m. tibialisanterior (TA) and at the m. gastrocnemius (GAS) of bothlegs using AgCl electrodes (Ambu R© Neuroline 720, AmbuA/S, Denmark). The EMG electrodes are placed perpen-dicular to the muscle fibre direction on the stimulated legin order to minimize the size of the M-wave that is inducedby FES. The reference electrode is placed on top of thesubtalar joint of the same leg. On one leg stimulationelectrodes (ValuTrode R©, Axelgaard Manufacturing Co.,Ltd., USA) are placed on the m. tibialis anterior and thegastrocnemius. Stimulation intensities were I = 20mA andpw = 200µs for TA and I = 22mA, pw = 226µs for GAS.


185

186 T. Schauer et al. / IFAC-PapersOnLine 49-32 (2016) 183–187

EMG(1000 Hz)

EMG filteringand analysis

Gait Phase Detection (50 Hz)

FES pattern

generator (25 Hz)

IMU

EMG

GPs

Fig. 3. Schematic representation of the system setup.

The tibialis anterior is activated shortly before the toe off(10% of the estimated gait cycle duration) and until heelstrike to support the lifting of the foot during the swingphase. The gastrocnemius is stimulated to support push-off before the heel off (20% of the estimated gait cycleduration) until toe off.

The upper graph of Fig. 5 shows an example of therecorded raw EMG of TA with marked stimulation ar-tifacts for normal walking with 2 km/h. The lower graphdisplays the blanked and episode-wise high-pass filteredEMG signals of GAS and TA together with the calculatedenvelope (EMGV

i for both muscles). Also shown are thedetected gait phases of two gait cycles.

Fig. 6 shows mean volitional EMG-activities during thedetected gait phases within five strides for one healthysubject walking A) normally, B) with emphasized push-off(more GAS activity expected), and c) with emphasizeddorsiflexion (more TA activity expected). In all threecases FES was administrated as described above. Themean volitional EMG activities have been normalized withrespect to the observed minimal and maximal mean EMGactivities over all walking styles. The change of muscleactivity patterns for the different walking styles can beclearly seen. During normal walking, TA is active from thepre-swing phase to the loading response with a maximumat the swing phase, in which the GAS is not active. TheGAS is active in the stance and pre-swing phase andinactive in the swing phase. For the abnormal walkingpatterns B and C we see more co-contraction. As expected,GAS is most active during walking with strong push-off,and TA is most active when performing a strong foot liftduring walking.

4. DISCUSSION AND CONCLUSIONS

A combined system for functional electrical stimulationand EMG measurement has been presented that will beused in gait therapy after stroke. Preliminary tests withhealthy subjects demonstrated the feasibility of the de-veloped realtime EMG filter algorithm. The system func-tioned well for all subjects and walking styles (normalwalking, emphasized push-off, emphasized dorsiflexion) atdifferent stimulation levels ((sub)sensory, sensory and mo-tor). The observed volitional EMG profiles in the presenceof FES during normal walking correspond well to the oneswithout FES and are in accordance with the expected

Fig. 4. Experimental setup with wireless IMU and EMGunits and functional electrical stimulation at the rightleg.

161.5 162 162.5 163 163.5 164 164.5

Time [s]

-2

-1

0

1

2

3

EM

G [m

V]

Raw EMG signal of TA with peak detection

raw EMG signal

FES of TA

detected peaks

reconstructed peaks

161.5 162 162.5 163 163.5 164 164.5

Time [s]

-1

0

1

2

3

EM

G [0.0

1 V

]

Envelope construction of GAS and TA

filtered EMG of GAS

envelope of GAS

filtered EMG of TA

envelope of TA

gait phase

Fig. 5. Upper graph: Raw EMG of TA with markedstimulation artifacts. Lower graph: Filtered EMG ofGAS and TA with envelope (EMGV

i ) and gait phases(0 – stance, 1 – pre-swing, 2 – swing, 3 – loadingresponse).

activity patterns described in (Perry and Burnfield, 2010).Exaggerated volitional muscle activity during differentwalking patterns was clearly visible in the calculated EMGprofiles. The clinical application is simplified by the useof wireless sensors, the arbitrary positioning of the IMUsat the foot/shoe and adaptation of the GPD to differentwalking speeds. Future work will include tests with strokepatients investigating the clinical use of EMG biofeedbackduring FES-assisted gait therapy.


186

T. Schauer et al. / IFAC-PapersOnLine 49-32 (2016) 183–187 187

A B C0

20

40

60

80

100

No

rm.

vo

l. E

MG

act.

Vol. EMG activtiy of M. tibials anterior

StancePre-swingSwingLoading Resp.

A B CWalking styles

0

20

40

60

80

100

No

rm.

vo

l. E

MG

act.

Vol. EMG activtiy of M. gastrocnemius

Fig. 6. Normalized mean volitional EMG activities duringthe detected gait phases over five strides for threewalking conditions: A) normal walking, B) walkingwith emphasized push-off, and C) walking with em-phasized dorsiflexion (more foot lift).

ACKNOWLEDGEMENTS

The authors like to thank Ralph Dorn for programmingthe MUSCLELABTM interface in Simulink.

REFERENCES

Ambrosini, E., Ferrante, S., Schauer, T., Klauer, C.,Gaffuri, M., Ferrigno, G., and Pedrocchi, A. (2014). Amyocontrolled neuroprosthesis integrated with a passiveexoskeleton to support upper limb activities. Journal ofElectromyography and Kinesiology, 24(2), 307–317. doi:10.1016/j.jelekin.2014.01.006.

De Luca, C.F. and Knaflitz, M. (1992). Surface Elec-tromyography, What’s New? CLUT Publishers, Turino,Italy.

Kafri, M. and Laufer, Y. (2015). Therapeutic effects offunctional electrical stimulation on gait in individualspost-stroke. Annals of Biomedical Engineering, 43(2),451–466. doi:10.1007/s10439-014-1148-8.

Klauer, C., Ferrante, S., Ambrosini, E., Shiri, U., Dahne,F., Schmehl, I., Pedrocchi, A., and Schauer, T. (2016).A patient-controlled functional electrical stimulationsystem for arm weight relief. Medical Engineering &Physics. doi:10.1016/j.medengphy.2016.06.006.

Klauer, C., Raisch, J., and Schauer, T. (2012). Lin-earisation of electrically stimulated muscles by feed-back control of the muscular recruitment measured byevoked EMG. In Proc. of the 17th International IEEEConference on Methods and Models in Automation andRobotics (MMAR), 108–113. IEEE.

Laufer, Y. and Elboim-Gabyzon, M. (2011). Does sensorytranscutaneous electrical stimulation enhance motor re-covery following a stroke? a systematic review. Neurore-habilitation and neural repair, 25(9), 799–809.

Merletti, R., Knaflitz, M., and De Luca, C.J. (1992).Electrically evoked myoelectric signals. Crit Rev BiomedEng, 19(4), 293–340.

Muller, P., Seel, T., and Schauer, T. (2015). Experimentalevaluation of a novel inertial sensor based realtime gaitphase detection algorithm. In Proc. of the 5th Euro-pean Conference on Technically Assisted Rehabilitation- TAR 2015. Berlin, Germany.

Perry, J. and Burnfield, J. (2010). Gait Analysis: Normaland Pathological Function. Slack Inc.

Seel, T., Landgraf, L., Escobar, V.C., and Schauer, T.(2014). Online gait phase detection with automaticadaption to gait velocity changes using accelerometersand gyroscopes. Biomedical Engineering / Biomedizinis-che Techik, 59(s1), S795–S798. doi:10.1515/bmt-2014-5011.

Truelsen, T., Piechowski-Jozwiak, B., Bonita, R., Mathers,C., Bogousslavsky, J., and Boysen, G. (2006). Strokeincidence and prevalence in Europe: a review of availabledata. European journal of neurology, 13(6), 581–598.


187

december 7-9, 2016. florianopolis, brazil sciencedirect · · 2017-04-01peer review under...

Documents