Biomedical Signal Processing and Control 18 (2015) 334359

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

jo ur nal homep age: www.elsev ier .com/ locate /bspc

Review

Current state of digital signal processing in myoelectric interfaces andrelated

Maria Haa Department ob Brain Researcc Brain and MinEspoo, Finland

a r t i c l

Article history:Received 18 JuReceived in re27 December Accepted 11 F

Keywords:Surface electroMyoelectric inClassicationFeature extracPattern recognition

Contents

1. Introd2. Acqui

2.1.

2.2. 2.3.

3. Decod3.1.

3.2.

CorresponTel.: +358 50 3

E-mail add1 Tel.: +358 2 Tel.: +358

http://dx.doi.o1746-8094/ uction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335sition system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336sEMG electrodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3362.1.1. Electrode types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3362.1.2. Size and shape of electrodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3362.1.3. Inter-electrode distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3372.1.4. Placement of electrodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3372.1.5. Number of electrodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338Filtering and sampling rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338Preprocessing algorithms for classication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338ing myoelectric information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3393.1.1. Windowing technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3393.1.2. Segment length . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3393.1.3. State of the sEMG signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3403.1.4. The effect of segmentation strategies on the delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3413.2.1. Time domain features. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3413.2.2. Frequency domain features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342

ding author at: Department of Electrical Engineering and Automation, School of Electrical Engineering, Aalto University, PO Box 15500, 00076 Aalto, Finland.823008.resses: maria.hakonen@aalto. (M. Hakonen), harri.piitulainen@aalto. (H. Piitulainen), arto.visala@aalto. (A. Visala).50 568 0654.50 5755936.

rg/10.1016/j.bspc.2015.02.0092015 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). applications

konena,c,, Harri Piitulainenb,1, Arto Visalaa,2

f Electrical Engineering and Automation, School of Electrical Engineering, Aalto University, PO Box 15500, 00076 Aalto, Finlandh Unit, Department of Neuroscience and Biomedical Engineering, 15100, 00076 Aalto, Espoo, Finlandd laboratory, Department of Neuroscience and Biomedical Engineering, School of Science, Aalto University, PO Box 15100, 00076 Aalto,

e i n f o

ly 2014vised form2014ebruary 2015

myographyterface

tion

a b s t r a c t

This review discusses the critical issues and recommended practices from the perspective of myoelectricinterfaces. The major benets and challenges of myoelectric interfaces are evaluated. The article aimsto ll gaps left by previous reviews and identify avenues for future research. Recommendations aregiven, for example, for electrode placement, sampling rate, segmentation, and classiers. Four groupsof applications where myoelectric interfaces have been adopted are identied: assistive technology,rehabilitation technology, input devices, and silent speech interfaces. The state-of-the-art applicationsin each of these groups are presented.

2015 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-NDlicense (http://creativecommons.org/licenses/by-nc-nd/4.0/).

M. Hakonen et al. / Biomedical Signal Processing and Control 18 (2015) 334359 335

3.2.3. Time-frequency domain features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3433.2.4. Spatial domain features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343

3.3. Myoelectric control strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3433.3.1. Pattern recognition-based control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344

. . . . . 4. Challe . . . . . .

4.1. . . . . . . 4.2. . . . . . 4.3. . . . . . 4.4. . . . . . .4.5. . . . . . .

5. Applic . . . . . .5.1. . . . . .

. . . . . . . . . .

. . . . . .5.2. . . . . . .

. . . . . . . . . .

5.3. . . . . . . . . . .

. . . . . 5.4. . . . . .

6. Concl . . . . . Ackno . . . . . Refer . . . . .

1. Introdu

The westrated sigmarket segness, healthlacks an effpopular inpand portablcannot easPortable inpally, for peoinjury, quaddystrophy, inadequateing them towheelchair85% reportwhom a pocontrol it. Oculties withtasks. Such eaccommodamade to ovewell as camlography (E

EMG intecle activity device. Thecommunicaand direct ethe user is muscles. EMdevice relyiused to meaelectrodes amuscular Ethe skin inwhere a nee

overedgeces. Cnditicy inectroasive, surfl, wh

EMGtram

ial abings,e iEM

EMG sEM3.3.2. Non-pattern recognition-based control . . . . . . . . . . . . . . . . . . . . . . . . . .nges and future trends . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .Number of control commands. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Simultaneous and proportional control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .Variation in limb posture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .Variation in contraction force . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .Interface integrity with time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .ations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .Assistive technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5.1.1. Prostheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5.1.2. Electric power wheelchairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5.1.3. Assistive robots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .Rehabilitative technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5.2.1. Exoskeletons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5.2.2. Serious games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .Input devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5.3.1. Armbands for mobile devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5.3.2. Muscle computer interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .Silent speech recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

usion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .wledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

ences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

ction

arable and mobile technology market has demon-nicant growth and adoption in various end-userments, in particular telecommunication, tness, well-care, and medical monitoring. However, the technologyective method to communicate with devices. Currentlyut methods such as touch screens, small keyboards,e controllers are impractical in situations where handsily be used to directly manipulate an input device.ut devices are also difcult to carry around. Addition-ple with severe physical disabilities such as spinal cordriplegia, hemiplegia, Parkinsons disease, or muscularthe traditional user interfaces currently available are. Fehr et al. [1] surveyed 200 practicing clinicians, ask-

provide information about their patients with power

tissue knowlinterfatory coaccuraface elnoninvtrodeschannefor theever, inpotentrecordBecaustext of

The

s relying on conventional controllers. Of respondents,ed evaluating some number of patients annually forwer wheelchair is not an option because they cannotf the patients, 40% with power wheelchairs had dif-

steering tasks and 59% needed assistance with suchxamples indicate the need for new controller interfacesting the abilities of the patients. Attempts have beenrcome these problems by using voice commands [2], asera- [3], electroencephalography (EEG)- [4], electroocu-OC)- [5] or electromyography (EMG)-based control [6].rface classies the voluntary-contraction-related mus-and associates it to the desired function of the given

EMG interface could offer an intuitive and easy way oftion that relieves the user from portable control devicesye contact to the device. The only requirement is thatable to activate some of his or her voluntary skeletalG interfaces generate control commands for a given

ng on information content of EMG signals. The methodssure these signals include surface EMG (sEMG) wherere placed on the skin over the measured muscle, intra-MG (iEMG) where the electrodes are inserted throughto the muscle tissue and percutaneous EMG (pEMG)dle or wire is inserted under the skin and subcutaneous

action poteface electrowith contramuscle conThe concepand the rsdeveloped itoward sEMfaces have people withpeople [25traditional in the eld

The sEMother manrequire lesscontrol witmyoelectricling the devsEMG signacontrol, sEMto ambientto the user. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353

the aponeurosis of the muscle. According to our best, pEMG measurements have not been used in the EMGomparative studies have found that, at least in labora-ons, intramuscular and surface recordings yield similar

classifying hand and forearm movements [7,8]. Sur-des are advantageous because they are inexpensive and. In contrast to relatively selective intramuscular elec-

ace electrodes detect activity from many muscles on oneich makes it possible to acquire sufcient information

interface with smaller number of electrodes [8]. How-uscular recordings may be benecial because of theirility to overcome some of the major problems of surface

such as electrode shifts and skin impedance changes.G [713] and have seldom been investigated in the con-

interfaces, this study deals only with sEMG recordings.G signal is a superposition of individual motor unit

ntials (MUAPs) within the pick-up range of the sur-des. As sEMG amplitude and frequency content changesction-force level [14,15], it is possible to associate thetractions of the user to control the device concerned.t of sEMG interface was introduced in the 1940s [16],t sEMG application, a myoelectric prosthetic arm, wasn 1960 [17]. In the recent years, the interest has grownG interfaces. It has been noted that myoelectric inter-a huge potential in applications designed not only for

disabilities [1824] but also in applications for healthy30]. The numerous benets of the sEMG interface overinput devices have inspired patents [31,32], especiallyof mobile technology.G interface is suggested to offer many benets overmachine control methods. The sEMG control may

attention from the user than EEG based controls or theh eye movements. In contrast to visual-based control,

control allows the user to look around while control-ice. Compared to many other biosignals, such as EEG,ls have relatively high signal-to-noise ratio. Unlike voiceG control has only a minimal delay, is not sensitive

sound perturbations, does not cause embarrassment, or disrupt the environment. The sEMG interfaces can

336 M. Hakonen et al. / Biomedical Signal Processing and Control 18 (2015) 334359

offer an alternative interface that require only minimal motor skillsfrom the user, which is a signicant improvement for people withimpaired motor skills. However, to get accustomed with the inter-face, some training is always required. The electrodes can be leftunder the cltile electrodMyonteck Lcontrol comels of staticinterface dspecial conc

sEMG sithe body [3[37], the shapplied in tabilities canstill able tocontrol the possible to to commancontrol becgrasp postuing action w

This artiissues relatsEMG interof-art applipreviously and Merletthe themesing optimalpreprocessicommands tions are suthe type of csEMG inter

2. Acquisit

The objeto provide acle contractwith the deor thresholThe acquisichannels, centrates onsystem: sEMrate, and pr

2.1. sEMG e

In sEMGity is detecto control ationed motelectrode(smembraneneuromusctrodes mea

The mosfaces incluTypically, osignal is thface electro

muscle [46]. Monopolar electrode conguration measures a dif-ference between the electrode on active site (the muscle) and acommon reference electrode on non-active site (typically on bonyarea) [47]. Laplacian conguration uses typically one central sur-

ectroly sho

also grid

placs subize ambeare imtime d recmplen-Invy arces. G e

ElectG sigonlynge ualitationlectrationitionsEMGon anmatil anrable

may mat]. Podes)emeas nol elef chas perharaG m

ts [4e (A

conshlornducals. Sd cades [

Size aour bnsidis macontsis orom

2 cmcatiol [58signaothes or even be imbedded into them [33,34]. Some tex-es, e.g. smart shorts measuring EMG activity (Mbody,td., Kuopio, Finland), are commercially available. Theirmands can be given by subtle motions or different lev-

(isometric) muscle contractions. Therefore, the sEMGoes not mark the user as handicapped or in need ofessions.

gnals can be measured from the supercial muscles of5]. Muscles of the upper limbs [18,36], the lower limbsoulder or the head, face, and neck [3840] have beenhe sEMG interfaces. Thus, a person with mobility dis-

control the device with the muscle(s) that he/she is contract. For example, a person with quadriplegia candevice with contractions of facial muscles [19], and it isrecord the sEMG signal from the amputees stump skind the prosthesis [18,36,41,42]. This allows an intuitiveause it allows the amputee to command for example are of prosthesis simply by performing the correspond-ith his/her residual muscles.cle aims to discuss and conclude the most importanted to sEMG signal measurement and processing fromface point of view as well as give a review of the state-cations for sEMG control. Related reviews have beenpublished by Oskoei and Hu [43], Miscera et al. [21]ti et al. [44]. This review is organized as follows. First,

related to the acquisition system are discussed, includ- electrode type and conguration, sampling, lters, andng algorithms. Second, the steps in decoding of sEMGare studied. Based on recent literature, recommenda-ggested for segmentation strategy, feature selection andlassier. The nal chapter introduces applications withface.

ion system

ctive of the acquisition system and signal processing is high quality sEMG signal(s) where the posture or mus-ion specic information can be extracted and associatedsired control command using classiers, proportionald algorithms, onset analysis, or nite state machines.tion electronics of the sEMG interface consists of sEMGlters, ampliers, and an A/D converter. This section con-

the most essential issues in the design of an acquisitionG electrodes, cut-off frequencies of lters, sampling

eprocessing algorithms.

lectrodes

interfaces, voluntary-contraction-related muscle activ-ted when the user contracts his/her muscles in order

device. The sEMG signal is composed of superposi-or unit action potentials propagating underneath the) along the active muscle bers (along their excitablethe sarcolemma) starting from innervation zones (i.e.

ular junctions) toward the tendon regions [45]. The elec-suring the sEMG signal form an sEMG channel.t common electrode derivations used in sEMG inter-de bipolar, monopolar, and Laplacian conguration.ne channel bipolar derivation is applied, in which sEMGe voltage difference between a pair of recording sur-des aligned along the length of the skin surface of the

face elrecentest hasa densetions is

Thitype, sand nuissues tional detailefor exathe Nobut theinterfathe sEM

2.1.1. sEM

Commor spohigh qpreparskinepreparIn adduse in irritatiinamtive gecompatrodes

Theior [56electrodisplacwhereAg/AgCow otrode ithese cfor sEMartifacchloridtrodessilver cby a comaterition, anelectro

2.1.2. To

who costudy based the balated f2 cm classicontrosEMG de and number of surrounding electrodes, and has alsown promise in sEMG interfaces [48,49]. Research inter-

increased toward high-density sEMG (HD-sEMG) where of surface electrodes allowing various electrode deriva-ed on a restricted skin area [50,51].section aims to give recommendations for electrodend inter-electrode distance, as well as the placementr of electrodes in the context of sEMG interfaces. Theseportant in regard to classication accuracy, computa-

and production costs of the sEMG control system. Moreommendations for sEMG measurements can be found

from the website of the Surface ElectroMyoGraphy forasive Assessment of Muscles (SEMIAM) project [35,52],e not necessarily always optimal for customized sEMGA review by Merletti et al. [46] discusses more deeplylectrode system and amplier technology.

rode typesnals can be measured both with wet and dry electrodes.

used wet electrodes require conductive electrolyte gelbetween the electrode and the skin, but can providey sEMG signals. The wet electrodes often require skin

(e.g. shaving and skin abrasion), which can reduceode impedance and motion artifacts [53]. However, the

is somewhat time consuming and requires expertise., the wet electrodes may not be optimal for long-term

interfaces since the conductive gel may dry, can caused discomfort, and is potential cause of skin allergy and

on [54]. Modern dry electrodes do not require conduc-d skin preparation, and still can reach signal quality

to wet electrodes [55]. For this reason, the dry elec- be more applicable for sEMG interfaces [46].erial of the electrode affects its electrochemical behav-larizable electrodes (e.g. gold, platinum and iridium

are characterized by capacitive behavior because onlynt current passes between the skin and electroden-poralizable electrodes (e.g. galvanized and sinteredctrodes) behave like resistors since they allow a freerge across the electrodeskin interface [56]. No elec-fectly non-polarizable or polarizable but approximatescteristics. Polarizable electrodes are not recommendedeasurements because of their high sensitivity to motion6,56]. Commonly used non-polarizable silversilverg/AgCl) electrodes are highly stabile. The Ag/AgCl elec-ist of a silver metal surface plated with a thin layer ofide [53,54]. Some polymers and fabric of threads coatedtive layer have also both shown promise as an electrodeuch dry electrodes are ideally suited for textile integra-n yield sEMG signal quality comparable to wet Ag/AgCl57].

nd shape of electrodesest knowledge, Young et al. [58] are the only authorsered an optimal electrode size for sEMG interfaces. Theirde with bipolar in the context of pattern recognition-rol strategy that decides the control commands onf the posture-specic values of feature vectors calcu-multiple EMG signals. The electrode sizes (1 cm 1 cm,

and 3 cm 3 cm) were shown not to signicantly affectn accuracy and completion rates in target achievement]. However, the benet of large electrodes was that thels acquired with them were signicantly less sensitive

M. Hakonen et al. / Biomedical Signal Processing and Control 18 (2015) 334359 337

to the changes of sEMG recording site when shifted up to 2 cmin perpendicular to the muscle bers [58]. This was because theelectrode with the largest pickup volume was possibly recording aportion of the same source as in the non-shifted condition. An alter-native stratclassicatiothe trainingis time consbe moved tof the classi

No recominterfaces wshape of thfound cleartrode shapefrom takingevidence thspectral diplar, rectangIn special siquency of sEvelocity, tracle bers) bwith larger

2.1.3. Inter-Bipolar E

are more tolarger the inthe two elethe pick-upcic the amdistance voMost studietion of SENselective retance betwBased on min supercifrom typicais always d(typically wfound that Iconsistentlyommendedincreases thet is a smvolume [23be a spatialeffect of ele

IED alsospace is relof surface pgenerating in order to aused, and resystems.

2.1.4. PlaceThe sign

placing the The electroered volumtissue and opass lter smamplitude a

frequency content of the sEMG signal are affected by the distancebetween the sEMG electrodes and the sources.

It is recommended that bipolar sEMG channels should be placedon the propagating part of the muscle, between the endplate area,

ervatis typate buscllimb]. Espes ofhysi

l dispbstapeciad rept-by-tribu7,69,indivIZs w

occic cousclmetelecters a

avoiized.ost owers. Hhat mtudientedulatid thete sils yinnelrans

to bt electronetationesentve (mls arther

shiftssEMMG ve toto ones extribuarin

l musevelsch the an

tudiedes h

equch wegy suggested to reduce the effect of electrode shifts onn accuracy is to include exemplars of possible shifts in

session of the classier [59,60]. However, this approachuming and troublesome because the electrodes have too expected displacement locations during the traininger.mendations for the sEMG-electrode shape for sEMGere found in the literature, and in many studies thee electrodes was not reported. Also SENIAM has not

and objective criteria for a recommendation for elec- and expected no major inuence on the sEMG signal

different electrode shapes [61]. However, there isat particular electrode types may be associated withs in the sEMG power spectrum [62]. Typically, circu-ular or bar electrodes are used in the sEMG interfaces.tuations, such as when determining mean spectral fre-MG signal or estimating mean muscle ber conductionnsversal (with respect to longitudinal axis of the mus-ar electrodes may be more appropriate than electrodeslongitudinal dimension, e.g., square electrodes [63].

electrode distanceMG channels are preferred in sEMG interfaces as theylerant to noise than for example monopolar ones. Theter-electrode distance (IED) is, i.e. the distance betweenctrode poles that form a bipolar channel, the wider

volume sampled and the higher but less spatially spe-plitude of the signal. A rough estimate of an electrodelume is a sphere with radius equal to the IED [64].s of sEMG interfaces have followed the recommenda-IAM and used an IED of 20 mm, which yields relativelycordings. The optimal IED also depends on the dis-een the recording electrodes and the source muscle.odeled sEMG signals, sEMG amplitude is reduced lessal muscles than in the deeper ones if IED is reducedl 20 mm to 10 mm [63]. Nevertheless, the sEMG signalominated by supercial sources, i.e. MUAPs, closest toithin 1020 mm) the electrode [23]. Young et al. [64]EDs larger than commonly recommended 20 mm were

more tolerant against electrode shifts and thus rec- the use of IEDs of up to 40 mm. Although larger IEDe likelihood of crosstalk from nearby muscles, the ben-aller electrode shift relative to the electrode detection]. A potential conguration for sEMG interfaces couldly selective concentric-ring electrode reducing both thectrode shift and crosstalk [65].

is critical to HD-sEMG systems. The sampling rate inated to IED, and too long IED may result the samplingotentials at the rate below Nyqvist frequency, and thusspatial aliasing. A recent study [66] demonstrated thatvoid spatial aliasing maximum IED of 10 mm should becommended the IEDs to be below 10 mm for HD-sEMG

ment of electrodesal-to-noise ratio of sEMG signals can be improved byelectrodes as close to the sEMG signal source as possible.des are separated from the muscle of interest by a lay-e conductor composed of subcutaneous tissue (adiposether soft tissues), and the skin, acting as a spatial low-oothing the detected MUAPs and thus decreasing theirnd frequency content [46]. Thus, both amplitude and

i.e. innthe IZ propagnate mupper-IZs [67site sidof the pa smalmay suand esrate ansubjecthe discles [6for all major IZs canin stat[72]. Mits geosEMG cle btotallyminim

In mare follcle beshifts tbeen snel orieby calcsite anopposichannelar chausing tprovedwithouthe eleThe beinformare prselectichanne

Anotrode in the HD-sEsensitipared providtial disload-shseveraeffort lin whifrom th

In selectrocles orapproaion zone (IZ), and tendon region. In fusiform muscles,ically a relatively narrow band from where the MUAPsidirectionally toward the tendons [67]. However, pen-

es with a more complicated structure, including many muscles, have more complex and diffusively localizedecially, if bipolar sEMG electrodes are placed on oppo-

the IZ, there may occur a substantial level of cancelationological sEMG signal, reducing its amplitude [68]. Thus,lacement of the sEMG electrodes with respect to the IZntially reduce the amplitude of the bipolar sEMG signal,lly its low-frequency components [68]. To get accu-eatable measurements, the IZs should be identied onsubject basis, using an electrode array [69]. Although,tion of IZs has been reported for several human mus-70], general distributions of IZs are not necessarily valididuals because of a large variability in the location ofithin and between subjects [70]. Additionally, a shift inur with dynamic changes in joint angle [71] and evennditions when isometric contraction level is increasede undergoes substantial three-dimensional changes inry, especially during dynamic contractions, and thusrodes are shifted with respect to the underlying mus-nd IZs [73]. Due to these issues, it is not possible tod the effect of IZs, but their effects should be optimally

studies of sEMG interfaces, common recommendationsd, and the bipolar electrodes are placed parallel to mus-owever, this placement is very sensitive to electrodeay occur in real use of sEMG interfaces [59]. There haves on untraditional electrode orientation where a chan-

perpendicularly with respect to muscle bers is formedng differential between an electrode from one control

corresponding electrode from the control site on thede of the arm [58,64,74]. Although, these transverseeld poorer classication accuracy than traditional bipo-s placed longitudinally with respect to muscle bers,verse channels in addition to traditional channels hase useful for ensuring low classication error with andctrode shift [64]. This result is achieved by comparingde congurations with the same number of channels.

of using transverse channels is that they provide global (muscle group), useful especially when electrode shifts, while solely longitudinal channels result relativelyuscle-specic) recordings. Transverse and longitudinale illustrated in Fig. 1.

possible approach to reduce the effect of IZs and elec- would be use of HD-sEMG that has shown promiseG based control in the last years [34,50,51,75,76]. Asscans almost the whole muscle skin surface it is less

issue of electrode locations and re-placements com-e channel bipolar conguration. In addition, HD-sEMGtraction of features from EMGs that depend on the spa-tion of the MUAPs in the same muscle, and on theg between muscles if the electrode grid extents overcles. This may aid in differentiation between tasks and. The effect of IZs may also be reduced using HD-sEMGe IZ channels can be detected and thus be discardedalyses [77].s on pattern recognition-based sEMG interfaces, bipolarave been placed either with reference to specic mus-

idistantly over the muscles of interest. The untargetedould be more preferable in sEMG interfaces because

338 M. Hakonen et al. / Biomedical Signal Processing and Control 18 (2015) 334359

Fig. 1. Longituelectrodes.

[2011] IEEE.

it is simpleincrease claisons of inlocal intramglobal surfawhen amplinvolved incation accelectrodes iconsistent dsion. Thus, tdisplaceme

2.1.5. NumThe opt

mostly bee[7,8,64,78,7cation accuchannels [7shown to brecommendhalf were lonels [64]. Hdepend on iincreased tfor four suuntil increaor seven. Fosignicantlduring a typvidually for

When ingave the beatively far provide mogross hand electrode suon exor diradialis longwhen study

placed approximately on exor digitorum prefunds and extensor dig-itorum communis [78,79].

The drawback of small number of electrodes is that the fault one electrode can cause signicant degradation in clas-on aineed [8relyiram)capag whsEMtatio

to btrainpmeMG. MsEMGow aes [3

lterin

amp makly 10ing e. Hig

sEM by aablesl. [8ces bainhic

class0 Hz y suba higtiont (at

domoni

consdinally (channels L1 and L2) and transversely (T1 and T2) oriented

Reprinted, with permission, from Ref. [58].

r to implement. The targeting of specic muscles mayssication accuracy [8], but not always [7]. Compar-

tramuscular and surface electrodes have shown thatuscular measurements do not outweigh the relativelyce recordings in sEMG classication [7,8]. It seems thatitude information is available from most of the muscles

motion, the classier is capable to yield high classi-uracy. More important than the initial placement ofs that the information content of the measurements isuring the use of the sEMG interface and training ses-he sEMG classication system should be robust againstnts of electrodes.

ber of electrodesimal number of electrodes for sEMG interfaces hasn studied by placing the electrodes on the forearm9]. In laboratory conditions, the increase in classi-racy has become saturated with three to four bipolar

of onlysicatibad skremovsEMG variogto be traininof HD-compushownin the develoHD-sEin HD-and allpostur

2.2. Fi

Thewhichtypicalrecordversionused inmainlytheir cLi et ainterfatrum munits, wment 2010healthuse of the mocurren

Theto harmcan be,8]. A relatively small number of electrodes have beene sufcient also when electrode shifts are present: theed subset is four to six bipolar channels, of which onengitudinal channels and the other half transverse chan-owever, the optimal number of electrodes may alsondividual anatomy of the subject. Andrews et al. [78,79]he number of bipolar channels from one to eight, andbjects the classication accuracy remained very poorsed sharply when the number of channels was ver eight subjects, the classication accuracy increased

y up to three electrodes. The classication was testeding task with the classication system optimized indi-

each subject.crementally adding new channels, the channels thatst classication accuracy for small subsets were rel-from each other, as would be expected because theyst new information to the classier [7,8,78]. Whereand forearm postures have been studied, the optimalbset usually includes electrodes placed approximatelygitorum supercials, exor capri ulnaris, extensor caprius or brevis, and extensor capri ulnaris [7,8], whereasing nger movements, the selected electrodes were

avoid aliasiest frequensampling this commonlower than mation for athe lower scomputatioLi et al. [81tionally mothe respectin healthy storing methe samplinfrequency b

2.3. Prepro

The algoinclude clapendent coby class-spccuracy. The malfunctioning channels (e.g. due tolectrode contact) can be automatically detected and0], but re-training of the classier is still needed. HD-ng on spatial-domain information (e.g. experimental

could solve this problem since HD-sEMG has shownble to maintain high performance even without re-en some electrodes are omitted [76]. The drawbacksG based systems are increased production costs andnal demands. However, the drop in performance hase relatively small when the number of electrodes useding of the classier is reduced from 96 to 24 [76]. Thent of more powerful microprocessors enables the use of

oreover, it has been shown that e-textiles can be used systems, and thus are easy to apply, are non-obtrusive

classication accuracy of 90% for nine hand and wrist4].

g and sampling rate

litude of sEMG signal is typically well below 10 mV,es it sensitive to artifacts. Therefore, it is amplied05000-fold with amplier, preferably as close to thelectrodes as possible, and ltered prior to an A/D con-h-pass cut-off ranging between 5 and 20 Hz is typicallyG studies to eliminate slow signal variations causedrtifact motion due to the movement of electrodes and

and typically ranging between 0 and 20 Hz. However,1,82] suggest a high-pass cut-off at 60 Hz for sEMGecause the lower frequency components of sEMG spec-ly contain information on ring rates of active motorh may not make a signicant contribution to the move-ication [61]. They found that sEMG power betweenonly slightly improves the classication accuracy forjects (by 0.25%) and transradial amputees (by 1.6%). Theh-pass lter at 60 Hz more than effectively attenuates

artifacts as well as power interferences by alternatingthe frequency of 50 Hz in Europe).inant energy (about 95%) of an sEMG signal is limitedcs up to 400500 Hz, i.e. the components over 500 Hzidered as noise [83]. According to the Nyqvist rule, tong, the sampling rate should be equal to twice the high-cy of interest contained within the signal. Therefore, thee rate of 1000 Hz with a low-pass lter at 400450 Hzly used in sEMG recordings. However, sampling rates1000 Hz may still preserve sufcient physiological infor-ccurate classication of the movements. One benet of

ampling rate is simultaneously reduced processing andnal complexity for the controller of the sEMG interface.,82] found that sampling at 500 Hz could be computa-re optimal than typical 1000 Hz for sEMG interfaces, asive classication accuracy was decreased only by 0.8%subjects and by 2.2% in amputees but halved both themory and data processing time [81]. To avoid aliasing,g rate of 500 Hz requires a low-pass lter with cut-offelow 250 Hz.

cessing algorithms for classication

rithms used to preprocess sEMG data for classicationsswise principal component analysis (cPCA) and inde-mponent analysis (ICA). In cPCA, sEMG data are rotatedecic projection matrixes to spatially decorrelate the

M. Hakonen et al. / Biomedical Signal Processing and Control 18 (2015) 334359 339

Fig. 2. Schemcontrol metho

[2007] IEEE.

data beforeapproximatin a way tunclasses bettanalysis winaffecting ththat, when for both int

ICA can faces. ICA emixture of Ganesh et arithms for issEMG. Temthe best peTimemy et increases cl88% to 93%.low contracabout non-g

3. Decodin

sEMG cgroups: patbased stratcommands

Pattern rfrom segmesier for thinclude proUnlike pattto-one mapand no needto decode mstages are das well as pclassicatio

3.1. Segmen

In sEMGreal-time. Tnamely winthree issuestate of EM

These parameters affect both the classication accuracy andresponse time of the system.

Windre awingque, d feasing makn in uiringmech ises b

delaarge d lon

Segmge daack ists. Tcy ane dtext

eal-t oneon ercatio58]. Ecommostheand 4w lenrequheormatel bet

acceatus useand

teccy byreviocy ofsing atic diagrams of pattern and non-pattern recognition-based sEMGds.

Adapted with permission, from Ref. [43].

features are extracted [84]. The projection matrixese some nonlinear low-dimensional data manifold. Thises the data, allowing the classier to identify the motioner. Hargrove et al. [85] found that, by using cPCA, thedow length can be cut from 256 ms to 128 ms without

e classication accuracy. Their other study [86] showedcPCA is used, classication errors reduced signicantlyact-limbed and amputee subjects.be used to reduce the crosstalk effect in sEMG inter-stimates the set of independent sEMG signals from agiven signals by estimating an un-mixing matrix [87].l. [88] compared the performance of different ICA algo-ometric hand gesture identication using four channelporal Decorrelation Source Separation (TDSEP) yieldedrformance of 1 s duration for an analysis window. Al-al. [89] found that the FastICA preprocessing techniqueassication accuracy for different window lengths from

The sEMG data has been shown to be super Gaussian attion levels [90], which matches the FastICA assumptionaussianity.

g myoelectric information

ontrol strategies can be separated into two maintern recognition-based and non-pattern recognition-egies [21,43]. The phases required to decode motorin the two systems are illustrated in Fig. 2.ecognition-based approaches calculate feature vectorsnts of the signal and use them as an input to a clas-e prediction of postures. Non-pattern-based methods

3.1.1. The

windotechniysis anproceswhich be seefor acqeach seapproaproductrollerWith lto avoi

3.1.2. Lar

drawbdata seaccuraand timthe conin all rlimb tosicaticlassilable [users the pr50 ms windopling f

In tto estiintervacationif the f(MV) iFigs. 4 cessingaccuran 1 paccuradecreaportional, onset analysis, and nite state machines.ern recognition-based methods, they do not allow one-ping, but their advantages are simple implementation

of training. Next sections describe each stage requireduscle contractions to the device control commands. Theata segmentation, feature extraction, feature reduction,attern recognition- and non-pattern recognition-basedn.

tation

interfaces, the sEMG data needs to be analyzed inherefore, the analysis is performed on time segments,dows, epochs or segments. This section discusses thes, i.e. windowing technique, segment length and theG signal, that need to be considered in segmentation.

tion with thIt should bemore votesthe nal deaccuracy wsmall windshort windspace. The ing the clasa scare resotion, no greis a notablesessions [93ture vectorclassicatioof the featuowing techniquere two alternative windowing techniques: adjacent

and overlapped windowing [43,91,92]. In the formercustom-length consecutive segments are used for anal-ture extraction. Because of high-speed processors, thetime is usually less than the duration of time segment,es the processor idle for a certain amount of time, as canFig. 3 (left). Overlapped windowing uses the idle timeg more data to be processed. As Fig. 3 (right) illustrates,nt overlaps the previous one. The overlapped window

more appropriate in sEMG control systems because itetter classication accuracy and a more constant con-y and reduces the length of the maximum delay [92].segments, overlapped windowing is necessary in orderg latency in real time operation [93].

ent lengthta windows increase classication accuracy, but the

that more time is required to collect and process largerhus, a trade-off has to be done between classicationd real-time constrains. The effect of classication errorselay on controllability has mostly been investigated in

of upper limb prostheses [58,91,94,95] but is essentialime applications. A test where the user moved a virtual

of six predetermined target positions showed that clas-rors less than 10% yield controllable systems whereasn errors over 35% yield systems that are not control-stimates of the delay (i.e. the time between onsets ofand and actuation of the device) which does not makesis feel unresponsiveness to the user range between00 ms [91,92,9599]. This delay range corresponds togths of 50400 samples and 25200 samples for sam-

encies of 1000 Hz and 500 Hz, respectively.y, a segment of t 200 ms contains enough information

motion states of a limb because that is the minimumween distinct contractions [93]. However, high classi-uracy is also possible with segments less than 200 msres have been selected carefully and if majority votingd as a postprocessing mechanism, as can be seen in

5. In sEMG signal processing, MV is a common postpro-hnique that aims to increase the overall classication

analyzing the current class decision along with theus class decisions [92]. As is apparent from Fig. 4, the

the unprocessed decision stream degrades rapidly bysegment length, but MV effectively prevents degrada-e help of more decisions available with short segments.

noticed that MV also increases the delay because the is used the more windows need to be processed beforecision. No appreciable difference exists in classicationhether a large window with small number of votes or aow with a large number of votes is used [91]. However,ows are recommended because they need less storagesaving in storage space is important when implement-sier as an embedded system where memory is usuallyurce [92]. When MV is used with overlapped segmenta-at improvement on performance is apparent, but there

decrease in the discrepancy of accuracy over different]. The optimal window length also depends on the fea-

calculated from windows in pattern recognition-basedn [93], as shown in Fig. 5. A more detailed descriptionres is given in Section 3.2.

340 M. Hakonen et al. / Biomedical Signal Processing and Control 18 (2015) 334359

Fig. 3. Windo ion (d1, d2, d3) can be produced. In adjacent windowing (left), the classieris idle majoritwindow slidesdecision imme

Figure is modi

Fig. 4. [2003] IEEE.

3.1.3. StateThe sEM

alternative [95,100]. Thmotor unit contractionated from ais under a cthe featureed more acontractionstate signalthe data retem responnot as profodrawback oing from clacontractionwing techniques. Each analysis window takes nite time to process (t) before a decis

y of time because the processing time is much less than window length. Overlapped wi

along at relatively small increments (inc). Setting the amount of overlap equal to the diately when the previous decision has been completed.

ed version from Ref. [91], available: http://www.rehab.research.va.gov/jour/11/486/pdf

Classication error vs. window length. Td = delay due to MV. Reprinted, with permission, from Ref. [92].

of the sEMG signalG classication studies have usually considered twostates of sEMG signal: a transient state and a steady statee transient state emanates from burst of simultaneousactivity, as a muscle goes from rest to some voluntary

level or when the contraction force is dynamically devi- steady contraction level. In the steady state, a muscleonstant (isometric) contraction. It has been shown thats extracted during isometric contraction can be classi-ccurately than the features extracted during dynamic

[100], and therefore most studies have used steady-s to generate the training data [86,93,103]. In addition,corded during isometric contraction allows faster sys-se because the degradation of classication accuracy isund when the window length is decreased. The mainf using transient sEMG signals is that it prohibits switch-ss to class in an effective or intuitive manner because a

must initiate from rest [95]. This severely impedes the

Fig. 5. Classilapped segmeof 300 ms andment sizes. Mlength; ZC = zeorder autoreg

[2008] IEEE

coordinatiodoms (DOFundetermincontractionswitching bvoting. Howstate signalthat investidynamic ponicantly imanalysis (L(SVM) class

In addititinguished of reasons, potentiatiomaintain gindows (right) increase frequency of class decisions because analysisprocessing time allows the controller to begin processing next class

/farrell486.pdf.

cation accuracy for some single features and feature vectors. Over-ntation with an increment of 200 ms was applied for segment lengths

500 ms while disjoint segmentation was applied for the other seg-AV = mean absolute value; RMS = root mean square; WL = waveformro crossing; AR2 = 2nd order autoregressive coefcients; AR6 = 6th

ressive coefcients.

. Reprinted, with permission, from Ref. [93].

n of complex tasks involving multiple degrees of free-s). When steady-state data is used, an sEMG signal is ined state during transition between different levels ofs. Therefore, the classication errors that occur whenetween classes should be reduced by using majorityever, an approach to classify both transient and steady-s concurrently has shown to outperform the methodsgate these two states separately [8,104]. Including somertions of sEMG signal during the learning process sig-proved the performance of both a linear discriminant

DA) [8,104] classier and a support vector machineier compared to the static training [104].on, muscle fatigued and non-fatigued states can be dis-[101,102]. EMG signal varies over time due to numbere.g., motor learning, muscle fatigue, and post-activityn. The muscle fatigue has been dened as inability toven task demands [101,102], and can be of both central

M. Hakonen et al. / Biomedical Signal Processing and Control 18 (2015) 334359 341

and peripheral origin [101,102]. Muscle fatigue can develop withina few seconds or a few minutes depending on the task [105]. Musclefatigue alters the task or posture specic time and spectral domainproperties of the sEMG signal and thus is a challenge for accurateclassicatiohas been shmedian freqmal contracsquare (RMeffect of musupervisingtures are difatigue [108

3.1.4. The eFarrel [9

age and besdifferent sein Table 1. Ttraction clasaccurate estion of the overlap, andTnew is relaanalysis wialgorithms tion, and thunder the dwindow len

3.2. Feature

Raw sEMture vectorthe signal m[21,43]. Becers also peof the systeaccording t

1) time dom2) frequenc3) time-sca4) spatial d

Table 2 sbe calculatedescriptionbe found in

Feature processing accuracy issier [7,125feature spaputational cclusters witthus minimthe ability oa noisy envset should bwith reason

Althoughoptimal f[93,10911made deepticularly fro

is little consensus between these studies because of signicantdifferences in the study details: most notably, the number of pos-tures classied, posture types and durations, data acquisition andthe classication system used, the number of subjects and the data

s. Co accued

ness uggencluions,in hi

has be b]. The

gro

TimefeatG siajor

matire be anmatinto fods, (2ds, ae sace oeature ha

for setudithe Lbsol

nd sites ad W

nt [7requr of tude lchant ares in mpuhownyer

alsoudgiis fe,8,60) [12o pr

for11,1omped input

del scatio9] acatio

fun pern of sEMG signals in humandevice interfaces [106]. Itown that spectral variables, such as the mean (MDF) anduency (MNF), are reduced during sustained submaxi-tions whereas amplitude variables, such as root meanS), increase [106]. Methods suggested to minimize thescle fatigue include adaptive classiers [107], on-line

mechanisms [90] and sEMG feature vectors (sEMG fea-scussed in Section 3.2) that are robust against muscle].

ffect of segmentation strategies on the delay1,94] presented equations to estimate the worst, aver-t case delays as well as delay ranges in the context ofgmentation strategies. These equations are presentedhe equations assume the transition between the con-ses to be instantaneous but have still shown to produce

timates of the controller delay [91]. The delay is a func-window length, processing time, amount of window

the number of majority votes used. The window shiftted to processing time , which is determined by thendow length Ta, processor, memory, type of features,used to extract features and perform pattern recogni-e number of sEMG channels. Thus, the only parametersesigners direct control for a given feature set are thegth and the number of majority votes.

s

G signals are mapped into smaller-dimension fea-s because features describe the information content ofore efciently than random and complex raw signal

ause of the smaller size of the feature vectors, classi-rform faster, which improves the real-time propertiesm. EMG features can be grouped into four categorieso the domain where they are calculated:

ain (TD) features,y or spectral domain (FD) features,le or time-frequency domain (TFD) features andomain (SD) features.

hows some features of each category. SD features cand only from HD-sEMG conriqurations. A more detailed

of features, especially that of TD and FD features, can Refs. [109,110].selection is the most important step in sEMG signalbecause the effect of the feature set on classication

even greater than the effect of the type of the clas-,126]. Three properties determine the quality of the

ce: maximum class separability, robustness, and com-omplexity [110]. A high quality feature space results inh maximum class separability or a minimum overlap,izing the misclassication rate. Robustness describesf the feature space to preserve cluster separability in

ironment. The computational complexity of the featuree low so that the related procedure can be implementedable hardware and in real-time.

several studies have been made to ndeatures for classication of sEMG signals1,118,121,122,127,128], few of these studies havely quantitative comparisons of their qualities, par-m the viewpoint of redundancy. In addition, there

set sizehigherperformrobustbeen sthis coconditmaintasystemit may[78,79feature

3.2.1. TD

in sEMTheir mmathetures ato noismathetures imethomethofrom thformanSome f

ThevectorsEMG stion is mean a(ZC), aestimatude ansegmetude, fnumbeamplitsignal ture sechangeand cobeen smultila

It iswith Htor). ThLDA [7(GMMare twcation[1091more cdescribthe comthe moclassi4th [13classidensityfeaturemparative studies have shown that TD features achieveracy for the LDA classier, whereas TFD features out-them for the SVM classier [104,129]. Considering theof LDA over SVM, TD features classied with LDA havested as optimal for sEMG classication [130]. However,sion bases on studies made in low-noise laboratory

and more study is needed to nd the features capable togh classication accuracy in real use. The classicationalso been shown to be subject dependent, and thereforeenecial to tailor a classication system to each subject

following subsections describe the feasibility of eachup in sEMG classication in more detail.

domain featuresures are the most commonly used feature groupgnal classication [36,42,81,93,113,114,125,131,132].

advantage is that they are fast to calculate because nocal transformation is needed. However, because TD fea-ased on signal amplitude, they are relatively sensitived artifacts. Based on observations of scatter plots andcal properties, Phinyomark et al. [133] divided TD fea-ur main types: (1) energy and complexity information) frequency information methods, (3) prediction modelnd (4) time-dependence methods. Adding the featuresme category in the feature vector may increase the per-nly slightly, due to small difference in feature space.res from each group are presented in Table 2.ve been attempts to determine an optimal TD featureveral classiers. Examples of TD feature vectors used in

es are presented in Table 3. The most common combina-DA classier with Hudgins feature vector consisting ofute value (MAV), waveform length (WL), zero crossinggnal slope changes (SSC) [42,93,95,100,122,134]. MAVn average of absolute value of the EMG signal ampli-L, the cumulative length of the waveform over the time4,110,133]. WL is a combined measure of signal ampli-ency and duration describing signal complexity. ZC is aimes that amplitude values of the EMG signal cross zeroevel and SSC the number of times that slope of the EMGges sign [74,109,110,135]. The benets of Hudgins fea-

relatively high classication accuracy, stability againstsegment length, low discrepancy over several sessions,tational simplicity [93]. In addition to LDA, it has also

to be an optimal TD feature vector at least for SVM andperception (MLP) [93].

common to combine autoregressive (AR) coefcientsns features (often referred to as the TDAR feature vec-ature vector has shown high classication accuracy with,64,89,136] MLP [137] and Gaussian mixture models5]. Autoregressive (AR) and Cepstral (CC) coefcientsediction model features used in sEMG signal classi-

which AR coefcients are more commonly applied15,118]. AR and CC coefcients are computationallylex than other TD features and need longer segments, as

Section 3.1.2. The larger the model order, the greateration time needed to determine the coefcients. Thus,hould be kept as simple as possible without sacricingn accuracy. Previous studies have suggested 3rd [138],nd 6th [132] AR order model as an optimal model forn of movements from sEMG signals. The probabilityction of sEMG amplitude has also shown promise as a TDmitting sEMG classication [140]. In a simulation study

342 M. Hakonen et al. / Biomedical Signal Processing and Control 18 (2015) 334359

Table 1Equations to estimate worst-case, average, and best-case controller delay as well as difference between best- and worst-case controller delays.

Classier type Worst-case delay Average delay Best-case delay Difference between best and worst case

No overlap, no majority voting D = 13 Ta + D=Ta+ D = 12 Ta + TaNo overlap, with majority voting D =

(n2 + 1

)Ta + D =

(n+12

)Ta + D =

(n2

)Ta + Ta

Overlap, no majority voting D = 12 Ta + Tnew + D = 12 Ta + n2 Tnew + D = 12 Ta + TnewOverlap, with majority voting D = 12 Ta +

(n+12

)Tnew + D = 12 Ta +

(n2

)Tnew + D = 12 Ta +

(n+12

)Tnew + Tnew

Reproduced from Ref. [75], available: http://www.rehab.research.va.gov/jour/11/486/pdf/farrell486.pdf.Note: General recommendation is that Tnew is approximately an order of magnitude less than Ta , meaning that users should experience more consistent delay with overlappedwindows. D = maximum delay between users intended movement and controller producing correct output class, n = number of majority votes, Ta = analysis window length, = processing time, Tnew = amount of window overlap.

that classied sEMG data according to three force levels the coreshape model outperformed higher order statistic combinations inall simulations due to the precise PDF shape screening embeddedin its formalism.

3.2.2. Frequency domain featuresFrequency domain (FD) features can be used to estimate mus-

cle fatigue [141144], force production [145] and changes in motorunit recruitment and ring patterns [43]. FD features are calcu-lated from power spectral density (PSD), which can be estimatedusing Periodogram or parametric methods [146]. PDS is mainlydetermined by the ring rate of the recruited motor units in thelow-frequency range (below 40 Hz), whereas the morphology of

their MUAPs traveling along the respective muscle bers mainlydetermines the high-frequency range (above 40 Hz) [147].

Phinyomark et al. [109] studied the properties of thirty-sevenTD and FD features and found that TD features were superior toFD features. In addition to their poorer classication accuracy, FDfeatures are computationally more complex than TD features. How-ever, combining FD features with successful TD features may yieldmore robust classication than a feature vector consisting solelyof TD features. Phinyomark et al. suggested mean frequency (MNF)as a potential candidate to combine with TD features because itsdiscriminant pattern in the feature space is different from thatof TD features [120]. Phinyomark et al. [121] modied the MNFand median frequency (MDF) features in order to increase the

Table 2Some typical features used in sEMG classication. Notations:xi = ith signal sample in a segment, N = number of samples in a segment, fj = frequency of the spectrum at frequencybin j; Pj = the EMG power spectrum at frequency bin j; M = length of the frequency bin, Aj = amplitude.

Mathematical denition References

Time domain featuresEnergy and complexity information methods

Mean absolute value MAV = 1N Ni=1|xi| [74,111]Integrated EMG IEMG = N

i=1|xi| [112,113]Variance VAR = 1N1 Ni=1x2i [111,112]Root mean square RMS =

1N

Ni=1x

2i

[8,110,114]

Waveform length WL = N1i=1 |xi+1 xi| [74,113,115]

Log detector LOGDET = e1/NNi=1 log(|xi |) [115]Frequency information methods

Zero crossing [74,116]Wilson amplitude WAMP = N1

t=1 [f (|xn xn+1|)] [111,115,117]

Slope sign changeSSC = N1

t=2 [f (|(xi xi1) (xi xi+1)|)]where f (x) =

{1, if x throshold0, otherwise

[74,93]

Prediction model methodsAutoreg t at tim

Cepstral um co

Time-depeMean ab mentHistogra paced

Frequency dMean freqMedian freModied m

Time-frequShort time tion;

Continuou ter; a

Discrete w coefc

Stationary

Wavelet p

Spatial domExperimenressive coefcients xn = Pi=1ai,nxni , P = model order, ai,n = ith AR coefcien coefcients cn = an nk=1

(1 kn

)akcnk , c1 = a1 , cn = nth cepstr

ndence methodssolute value slope MAVSi = MAVi+1 MAVi ; i = 1, . . ., K 1; K = number of segm of EMG HEMG divides elements in the EMG signal into equally s

elements for each segment.omain featuresuency MNF = M

j=1fjPj/Mj=1Pj

quency MDFj=1 Pj = Mj=MDFPj = 12 Mj=1Pj

ean frequency MMNF = Mj=1fjAj/

Mj=1fjAj

ency domain features Fourier transform STFT(k, m) = N1

r=1 x(r)g(r k)ej2mi/N ; g = window funcs wavelet transform WTx(, a) = 1a

x(t)

(ta

)dt; t = translation parame

functionavelet transform DWT splits the signal into an approximation and detail low- and high-pass lters. The approximation coefcients areapproximation and detail coefcients. By repeating the proceslower resolution components.

wavelet transform SWT does not decimate the signal at each stage, avoiding the and WPT.

acket transform WPT is a generalized version of DWT that is applied to both lohigh-pass results (details).

ain featurestal periodogram (h) = 12n(h)

n(h)i=1 [x(zi) x(zi + h)]

2; h = distance vector, x(zi) = mpairs h units apart in the direction of the vector he instant n [7,115]

efcient, ai = AR coefcient [112,115]

s covering the signal [109,118] segments and returns number of signal [109,111,119]

[120][120][121]

k = time sample; m = frequency bins [95,100]

= scale parameter; = mother wavelet [95,100,122]

ients by passing it through complementary [95,123,124]

further split into a second-levels, one signal is broken down into many

problem of nonlinear distortion of the DWT [95]

w-pass results (approximations) and [95,98,100,122]

easurement at location zi . n(h) = number of [76]

M. Hakonen et al. / Biomedical Signal Processing and Control 18 (2015) 334359 343

Table 3TD feature vectors used in sEMG interfaces. H = healthy subject, A = amputee subject. 1 = classication accuracy under electrode location shift, 2 = classication accuracyunder muscle contraction level change, 3 = classication accuracy with muscle fatigue. CKML =Cascaded-kernel learning machine, SE = sample entropy, AR6 = 6th order ARcoefcients.

Feature vect Classe

MAV, WL, ZC 6 MAV, WL, ZC 10 MAV, WL, ZC 6 MAV, WAMP 12 MAV, WAMP 4 MAV, WL, AR 4 WL, LOGDET 4 SE, CC, RMS, 11 IEMG, WL, V 11 AR6, MAV 11 AR6, ZC 11 AR6, SSC 11 AR6, WL 11 AR6, RMS 6 AR, HIST 8

robustness to the amplthe variatiothe toleranied MNF results shoregarding wgested as a for a more pFD featurescoordinatiotion force [1

3.2.3. Time-Time-fre

cation inclucontinuouswavelet trpacket trantransform (increase boCWT, DWT frequency rband but pohigh frequeefcient thture in sEMof its abilitfrequency bare computimplementeto meet theappropriateare used [9the robustnbecause bybe restricte

TFD fearequires dimspeed and ture projecprincipal co[100,103,12wavelet coefeatures cancients or sEMG analy

of thmarted of di

conseconlse pd sEM58,1e whlternent

Spati-sEMmpoings.on behe spn m

obsdiffe

theD-sEw strposiodograblet longvariae expor Classier Classication accuracy (%)

, SSC SVM 96 , SSC, AR6 LDA 97 , SSC, AR6, RMS GMM 97 , VAR, WL ANN 98 , AR, CC LDA 701, 782, 873

, CC LDA 701, 782, 883

, AR, CC LDA 701, 782, 883

WL LDA 98 AR, ZC, SSC, WAMP GRA 96

LDA 98H, 79A LDA 97H, 75A LDA 97H, 74ALDA 98H, 79A SVM 96 CKLM 93A, 97H

property of these features by applying mean and medianitude spectrum instead of the power spectrum becausen of the amplitude spectrum is smaller. They comparedce of sixteen traditional TD and FD features, and mod-and MDF features against white Gaussian noise. Thewed that a modied MNF is the most robust featurehite Gaussian noise. Therefore, modied MNF was sug-highly potential feature to augment the other featuresowerful and robust feature vector. Recently, two novel

derived from discrete Fourier transform and musclen has shown robustness against variations in contrac-48].

frequency domain featuresquency domain (TFD) features used in sEMG classi-de short time Fourier transform (STFT) [95,100,122],

Wavelet transform (CWT) [95,100,122], discreteansform (DWT) [95,123,124,133,149152], waveletsform (WPT) [95,98,100,122] and stationary waveletSWT) [95]. The drawback of STFT is that it cannotth time and frequency resolution simultaneously [153].and SWT overcome this deciency by providing goodesolution with poor time resolution in low frequencyor frequency resolution with good time resolution inncy band [153]. Since DWT is computationally morean CWT, it has become the most common TFD fea-G interfaces. WPT has also gained interest because

y to provide the frequency information in both lowand and high frequency band. Although TFD featuresationally more complex than TD features, they can bed with fast algorithms that have shown to be capable

real-time requirements in sEMG classication when

be onePhinyocompucients vectorlevel rMyopustructe[1561removtional componal.

3.2.4. HD

only terecordentiatiabout tbetweeby thevated well asSince Hvery fethis putal varcompaagainssmall late th dimensional reduction and segmentation techniques2,98,122,152,154]. Wavelet transforms may improveess of the system compared to TD and FD features

using subsets of wavelet coefcients the analysis cand only to interesting frequency bands.tures yield a high-dimensional feature vector thatensionality reduction transformation to increase the

accuracy of the classication. The most popular fea-tion algorithms in sEMG signal classication are themponent analysis (PCA) and uncorrelated LDA (ULDA)9,155]. However, the feature extraction applied forfcient is usually a TD technique [133,150,156,157]. TD

be calculated either directly from the wavelet coef-the reconstructed sEMG signal [150]. Most studies ofsis have concluded the Daubechies wavelet family to

points whicvariodogramdifferent duexcluding tputation ofwithout re-

3.3. Myoele

Two conrecognitionpattern recthe featurerecognitionparing a vas Bipolar electrodes Subjects Reference

4 11H [93]3 12H [7]4 12H [125]

32 1H [117]2 8H [115]2 8H [115]2 8H [115]4 4H [130]7 12H [113]4 5H, 5A [132]4 5H, 5A [132]4 5H, 5A [132]4 5H, 5A [132]4 11H [93]3 2A, 1H [119]

e most suitable for sEMG signal analysis [150,158,159].k et al. [120] compared several TD and FD featuresof the sEMG signal reconstructed using DWT coef-fferent levels. Useful feature vectors included a featureisting of ZC, WAMP, and MAV computed of the second-structed sEMG signal with the Db7 wavelet and theercentage rate (MYOP) feature of the rst-level recon-G signal with the Db8 wavelet [120]. Phinyomark et al.

60] have also suggested DWT as an optimal method toite Gaussian noise (WGN) from sEMG signals. Conven-s cannot effectively remove WGN because its frequencys fall in the energy band of the physiological sEMG sig-

al domain featuresG measurements have made it possible to extract notral and spectral but also spatial information from sEMG

Spatial domain (SD) features would improve the differ-tween postures and force levels providing informationatial distribution of the MUAPs and on the load-sharing

uscles. The relevance of spatial information is supportedervation that distinct regions of the muscle are acti-rentially depending on the position of joint [161] as

duration [162] and strength [163] of the contraction.MG has only recently adopted in the sEMG interfaces,udies have investigated and designed SD features fores [75,76]. An example of SD feature is the experimen-ram [76]. It has shown to yield classication accuracy

to the classical TD features but to be more robustitudinal and transverse electrode shifts [76]. Moreover,tions in the number of measurements used to calcu-erimental variodogram do not change the number of

h form it, and consequently the feature space used in

method [76]. Thus, the number of electrodes can bering testing and training of the classier which allowshe electrodes with poor signal quality from the com-

the variogram during the use of the sEMG interfacetraining the classier.

ctric control strategy

trol strategies used in sEMG interfaces are pattern-based and non-pattern recognition-based control. Inognition-based control, a classier is used to map

vector to a desired command, whereas non-pattern-based control decides the control commands by com-lue of a single feature to predetermined threshold(s).

344 M. Hakonen et al. / Biomedical Signal Processing and Control 18 (2015) 334359

Pattern classication-based approach is usually preferred in sEMGinterfaces because it allows more versatile control scheme thannon-pattern one. This subsection describes these sEMG controlapproaches.

3.3.1. PattePattern

that the claduced in thof a given lated of the commands provided impared to coextending ttrol commadecades [16and compaA detailed sEMG intercomparativand a sufcclassicatioappropriatelinear probsimple to itions, such machines (S[168,176,17become a g

Howeveto classify sor noise. It ble to mainones becauKaufmann eclassier wtuating sEMsEMG data racy for LDAwhen trainet al. [58] foclassier wplex classiclassier ha

3.3.2. Non-Non-pat

main limitcontrol comapproach m[26], wheerequire fewprostheses.based meth[181183], control, theor force ofexoskeletonpattern recomethods to

Onset anor double-tsEMG signadepends onmethod is

to noise makes it suitable only for coarse ON/OFF detection. Anadaptive threshold method, where the threshold is determined onthe basis of the average power of previous sEMG contraction, hasalso been developed [179]. Thus, the threshold can automatically

itselfon. Tnt coreshnvel

usedmean

sine it rcomped umb

s onsresho

deteold m

it is Ald-thre

and tion

ite stumbtes uansithairn inSM aed foeen sf-artequiFSMch [1stop -lev

llen

of ths to m

use,l comad to

for y tha

desces a

umbe

ougtternle us

comssicstudof 1ed incy torn recognition-based controlrecognition-based approach relies on the assumptionssier is capable to recognize the input values intro-e training session and assign each input value to oneset of classes. Input values are feature vectors calcu-sEMG signal, and classes correspond to different controlthat are sent to the device. Pattern recognition hasportant improvements in sEMG control when com-nventional non-pattern recognition-based control byhe number DOFs and increasing the intuitiveness of con-nds. sEMG pattern classication has been studied for4,165], and several classiers have been investigated

red for use in sEMG control [93,114,126,134,166170].description of the most common classiers used infaces is given in Refs. [43,118]. However, a number ofe studies agree that with an appropriate feature setient number of channels, most classiers have similarn accuracy [7,125,126,170]. The implication is that an

feature representation makes the classication task alem. The trend seems to be toward classiers that aremplement, fast to train, and meet real-time contrac-as LDA [8,50,58,64,91,92,130,171174], support vectorVM) [136,166,175], and hidden Markov models (HMM)7], of which LDA is the most commonly used and haseneral recommendation for sEMG interfaces.r, few studies have compared the ability of the classiersEMG signals in long-term use and with additive artifactscan be assumed that linear classiers are more capa-tain high classication accuracy compared to nonlinearse of their better capability to generalize sEMG data.t al. [178] demonstrated this by showing that The LDAas the most robust against the long-term effect of uc-G signals when compared to ve state-of-art classiers.was recorded from 21 days, and the classication accu-

was 82.37% when trained with recent data and 78.73%ed with data collected only during the rst day. Youngund that the LDA classier also outperforms The MLPhen electrode shifts are present. However, more com-ers may be appropriate in long-term use where a propers to classify novel patterns during online training.

pattern recognition-based controltern-based controllers have simple structure, but theation is that they allow only restricted number ofmands [43]. However, non-pattern recognition-baseday provide an intuitive interface for navigation menuslchairs [22] and assistive robots [179] that usuallyer commands than, for example, multifunction control

The methods included in non-pattern recognition-ods are proportional control [20,180], onset analysisand nite state machines [22,184186]. In proportional

strength of muscle contraction determines the speed a device. Proportional control has been adopted ins [20] and prostheses [180], where it is used withgnition-based or other non-pattern recognition-based

control speed or force of an assistive device.alysis is performed either by using single-threshold

hreshold methods. In the single-threshold method, thel is compared with an amplitude threshold whose value

the mean power of the background noise [43]. Thisfast and simple to implement, but its high sensitivity

adjust sicaticonstaated thtime-emonlysignal

ThebecausTo ovedeveloxed ndetectthe thhigherthreshis thatXu anddoublestabiledescrip

Finnite nThe staand trwheelcvides aof an Fdesignalso bstate-oeffort ring an approaate or the low

4. Cha

Onefaces iIn realcontromay lenizablestrategsectioninterfa

4.1. N

Althnon-paversatimandsthe clasEMG imum includaccura when contraction power varies, so as to avoid misclas-o perform a specic function, the user must produce antraction to keep the sEMG amplitude above the associ-old [43]. The single-threshold method can be based on aoped signal or instant signal value. In the rst case, com-

algorithms are signal mean value, low-pass-ltered value, or MarpleHorvat and Gilbey algorithm [187].gle-threshold method is generally unsatisfactorystrongly depends on the choice of the threshold [43].e this problem, the double-threshold method was[22,182]. It applies single-threshold detection to aer of consecutive values of an auxiliary variable andet when a certain number of auxiliary values crossld. The double-threshold method (see Fig. 6) yieldsction probability, which makes it superior to the single-ethod [43]. However, the drawback of this method

complex and computationally expensive. Therefore,er [183] developed an improved method based on theshold method. The improved method is more sensitive,efcient with decreased computational cost. A detailed

of the method can be found in Ref. [183].ate machine (FSM) based controller is described by aer of states, transitions between them, and commands.sually represent predened commands for the device,ion roles are associated with the signal features. In

[22,186] and assistive robot [188] control, FSM pro-tuitive, easy interface with high accuracy. An examplend the double-threshold-based signal-onset detectionr the sEMG wheelchair is presented in Fig. 6. FSM hastudied in upper limb prostheses [184,185], and the

commercial prostheses rely on FSM-based control. Thered to control a prosthesis can be reduced by combin-

higher-level controller with pattern recognition-based85]. In this control strategy, the patient must only initi-a movement, and using information from local sensors,el controller will control the movement.

ges and future trends

e major challenges related to the design of sEMG inter-aintain high classication accuracy in long-term use.

the muscle contractions, i.e. the classes associated tomands, are performed in a variety of conditions, which

differences in signal properties making them unrecog-the classier. Another challenge is to develop a controlt allows simultaneous movement of multiple DOFs. Thiscribes the main challenges related to the design of sEMGnd introduces potential approaches to overcome them.

r of control commands

h the pattern recognition-based method overcomes the based one with the possibility to implement a moreer interface, increasing the number of control com-plicates the classication problem, which may reduceation accuracy. The number of classes investigated inies usually remains fewer than twelve. When a max-2 classes are examined, each additional output classto the classication problem will cause the classication

drop by 0.26% [8].

M. Hakonen et al. / Biomedical Signal Processing and Control 18 (2015) 334359 345

Fig. 6. EWP co le-thrMAV signals. P h signmotion. If any ransitor activating t e singthan ATH as in d comto the sEMG si ides o

[2010] IEEE.

For humstudies haacquired fr[190,191], vUtilizing mber of classsignicantlySign Langua96.3% for twsEMG sensowere classitence accursEMG data HMM. The gesture segintensity ofsegmented and accelersEMG signacients and MsEMG signa1 kHz, and sback on theelectrodes oimi, palmarand brachio

4.2. Simulta

Pattern rat a time. Twhere a usments to peused to appmotions (i.etional paral

The sing[196] consiseach combiBraker et althe classes discrete mo

odig theposis traio thes in

two outpun. Yoy tht dis

comonecretelect

clasentVM

that accOFs

discntrol using a nite state machine with the double-threshold method. (a) The doubTH and ATH are the primary threshold and auxiliary threshold, respectively. If bot

signal is less than PTH but above ATH , as in (2), (4), (6), and (8) that is regarded as a to contraction. In this case, the activation is reserved with the previous state. In th(7). (b) State transition diagram for EPW control. (c) Denitions of motion states angnal from the electrode placed on the levator scapulae muscle on the left and right s

Adapted, with permission, from Ref. [22].

andevice interface purposes, increasing number ofve combined sEMG information with informationom other sources, such as images [189], accelerationoice [192], EOG signals [193], and EEG signals [194].ultiple sensor modalities allows increasing the num-es without compromising the classication accuracy. Zhang et al. [195] managed to classify 72 Chinesege (CSL) words with average accuracies of 95.3% ando subjects by merging the information acquired by vers and a three-axis accelerometer. 40 CSL sentencesed with an overall word accuracy of 93.1% and a sen-

acy of 72.5%. The effective fusion of accelerometer andwas achieved using a decision tree and multistreamstart and end points of sEMG signals of meaningfulments were detected automatically on the basis of the

the sEMG signals. These active segments were furtherinto analysis windows of 250 ms with 125 ms overlapometer signals were segmented synchronously with thels. The feature vector consisting of 4th order AR coef-AV was calculated from each analysis window of the

ls. The sEMG signals were ltered between 20 Hz andampled at 1 kHz. The accelerometer was placed on the

[174] mtrainintwo opclass itrast tmotionwhen their odecisiostrateger thaand alla compone disclass semotionimplemusing Sshowscationthree D6.6% on forearm near the wrist and ve bipolar silver sEMGn the following forearm muscles: extensor digiti min-

is longus, extensor carpi ulnaris, extensor carpi radialis,radialis.

neous and proportional control

ecognition-based classiers can activate only one DOFhis is a signicant drawback, especially in prostheseser has to make a combination of sequential move-rform a coordinated task. The strategies that have beenly pattern recognition-based classication to combined. activate multiple DOFs) are single, parallel and condi-lel classication (see Fig. 7).le classication approach proposed by Davidge et al.ts of one classier in which each discrete (one DOF) andned motion (two DOFs) were labeled as separate classes.. [197] introduced a parallel classication that predictswith separate classiers each of which is trained withtions and a separate subset of channels. Young et al.

gle classiestrategy hadfor all classigies is thatLimiting traapproach, bto over 80%cially if theand the elerecognitioncontrol.

Regressirecognitiondecide for afor each DOtional contis relativelyperceptronversus kinebetween sEeshold method is used to recognize shoulder elevation motions fromals are less than ATH , as in (1), (5), and (9), that is recognized as noion state in which one or both muscles are relaxing from contractionle-contraction state, only one signal exceeds PTH , and another is lessmands by shoulder elevation motions. Left EMG and right EMG referf the shoulder, respectively.

ed this method, using one classier for each DOF and classiers to discriminate between three classes, theng motion classes of a DOF and no motion. Each motionned using data from its discrete motion and, in con-

previous parallel approach, also using the combinedwhich it participates. The output is a combined actionf the parallel classiers have active motion classes as

t. This approach uses all channels for every classicationung et al. [174] also developed a conditional parallel

at classies each motion class with a separate classi-criminates between its designated discrete movementbined movements that have this discrete movement asnt. Thus, each classier makes a priori assumption that

motion class is active. The nal output is the movemented by both of the conditional classiers that contain thes. Each of the three classication strategies has beened using LDA classiers and the parallel strategy alsoclassiers [198]. A comparison of LDA implementations

the conditional parallel classier yields the best classi-uracy of the three approaches [174]. In classication of, the conditional parallel strategy yielded error rates ofrete and 14.1% on combined motions, whereas for sin-

r the corresponding values were 9.4% and 14.1%. Parallel

the poorest performance, suggesti