EURASIP Journal on Applied Signal Processing 2001:2, 67– 81© 2001 Hindawi Publishing Corporation

Onset Detection in Surface Electromyographic Signals:A Systematic Comparison of Methods

Gerhard StaudeInstitut für Mathematik und Datenverarbeitung, Universität der Bundeswehr München, GermanyEmail: gerhard.staude@unibw-muenchen.de

Claus FlacheneckerInstitut für Mathematik und Datenverarbeitung, Universität der Bundeswehr München, GermanyEmail: claus.flachenecker@humphrey.com

Martin DaumerInstitut für Medizinische Statistik und Epidemiologie, Technische Universität München, GermanyEmail: martin.daumer@imse.med.tu-muenchen.de

Werner Wolf

Institut für Mathematik und Datenverarbeitung, Universität der Bundeswehr München, GermanyEmail: Werner.Wolf@unibw-muenchen.de

Received 26 July 2000 and in revised form 13 May 2001

Various methods to determine the onset of the electromyographic activity which occurs in response to a stimulus have beendiscussed in the literature over the last decade. Due to the stochastic characteristic of the surface electromyogram (SEMG), onsetdetection is a challenging task, especially in weak SEMG responses. The performance of the onset detection methods were tested,mostly by comparing their automated onset estimations to the manually determined onsets found by well-trained SEMG examiners.But a systematic comparison between methods, which reveals the benefits and the drawbacks of each method compared to the otherones and shows the specific dependence of the detection accuracy on signal parameters, is still lacking. In this paper, several classicalthreshold-based approaches as well as some statistically optimized algorithms were tested on large samples of simulated SEMG datawith well-known signal parameters. Rating between methods is performed by comparing their performance to that of a statisticallyoptimal maximum likelihood estimator which serves as reference method. In addition, performance was evaluated on real SEMGdata obtained in a reaction time experiment. Results indicate that detection behavior strongly depends on SEMG parameters, suchas onset rise time, signal-to-noise ratio or background activity level. It is shown that some of the threshold-based signal-power-estimation procedures are very sensitive to signal parameters, whereas statistically optimized algorithms are generally more robust.

Keywords and phrases: EMG, electromyography, onset detection method, performance, comparison.

1. INTRODUCTION

Analysis of electromyographic signals recorded from the skinover the muscles by surface electrodes (SEMG) represents animportant tool in a variety of applications like neurologicaldiagnosis, neuromuscular and psychomotor research, sportsmedicine, prosthetics, or rehabilitation. Processing of SEMG(and other biosignals like EEG, ENG, ECG, etc.) can be con-sidered a special field of applied signal processing, with themain focus on an appropriate application of theoretical con-cepts to extract specific information from small and oftennoisy biosignals.

The instantaneous SEMG signal energy serves as ameasure of the current level of muscle activation [1], whereasthe determination of the exact onset and offset times of amuscle contraction is useful in studies of motor control andperformance [2, 3, 4, 5, 6, 7, 8, 9, 10]. The latter is of par-ticular importance in reaction time (RT) experiments, andnumerous methods were developed for SEMG onset detec-tion. Although a comparison of results obtained with vari-ous approaches was presented [11, 12, 13], an objective andcomprehensive assessment of their accuracy in SEMG onsetdetection is still lacking. Due to the fact that the underly-ing signal parameters including the “true” onset and offset

68 EURASIP Journal on Applied Signal Processing

times are unknown in real SEMG recordings, a rating by anobjective error measure is not possible. Therefore, the auto-matically determined onsets are usually compared to resultsobtained by visual inspection of the data by some well-trainedSEMG examiner. But a profound performance test should bebased upon a relevant statistical analysis which requires alarge amount of SEMG data with properly preset signal pa-rameters. This analysis is difficult or even impossible with realdata due to the poor parameter reproducibility of successiveSEMG recordings, and due to inter- and intra-rater variabilityintroduced by the manually determined reference onset esti-mates. To overcome this problem, simulated SEMG data wereused in this study for assessing the sensitivity of various detec-tion methods to changes in signal parameters. Also, the struc-ture of the various existing SEMG onset detection methodswas analyzed revealing several common components whichare identical in a logical way, but are formulated individuallyby different authors. As basis for comparison, some funda-mental algorithmic principles formulated in [12] were usedto express each method by these terms.

Detection algorithms introduced earlier by Hodges andBui [13], Bonato et al. [14], Lidierth [15], Abbink et al. [2],and two new implementations of the model-based detectorproposed by Staude et al. [12] are included in this study. Thealgorithmic layout of the methods is compared, and the de-pendence of onset detection accuracy on signal parameters isdemonstrated. The use of a statistically optimal estimator as areference for the upper performance limit allows an absoluteperformance ranking of each method. In addition, the meth-ods were tested on real SEMG data obtained in a reactiontime situation.

2. ONSET DETECTION ALGORITHMS

2.1. General scheme of onset detection

As suggested in [12], the comparison of algorithmic struc-tures is conducted by using the SEMG signal model and gen-eral scheme of event (onset) detection shown in Figure 1. Thedigitized SEMG signal may be represented by a real-valued se-quence (xk)k≥1, where xk denotes the signal amplitude at aparticular discrete time instant k; thus a single SEMG record(xk)k≥1 represents a sample observation of a discrete ran-dom process (a zero mean discrete white Gaussian noise pro-cess) (Wk)k≥1 exciting a linear system with transfer functionH(z). (Wk)k≥1 reflects the discharge timing and recruitmentof the elementary signal sources (i.e., the motor units) in-volved; H(z) describes the shape of the discharges (i.e., theaction potentials) as well as the specific bioelectrical transferfunction between generator and recording site. Measurementnoise is neglected. H(z) is modeled by an all-pole represen-tation (autoregressive (AR) filter) of order p

H(z) = 11+ a1z−1 + a2z−2 + · · · + apz−p

, (1)

where a1, a2, . . . , ap are the AR coefficients and z denotesthe complex frequency of the z-transform. The time domainrepresentation of this signal model is

xk = −p∑i=1

aixk−i +wk. (2)

As a particular feature of the signal model, the varianceσ2(k)of the white noise excitation depends on the current state ofthe muscle activity which is modulated by descending com-mands from supraspinal levels as well as by feedback fromperipheral receptors [16]. Within this framework, the indica-tion for the response onset to be detected is an abrupt changein the variance profileσ2(k), that is, it changes abruptly froma pattern σ2

0 (k) to a new pattern σ21 (k, t0) ≠ σ2

0 (k) at timeinstant t0.

Computerized onset detection requires to determinethe unknown change time t0 as accurate as possible. Mostdetection algorithms consist of up to three basic processingstages (see Figure 1):

• signal conditioning,• detection unit,• post-processor.

In an initial step, the observed SEMG signal passes througha signal conditioning unit in order to enhance the spectralcontent of the measured signal carrying information aboutthe onset of muscle activation. Frequently, a lowpass filter isapplied to reduce high frequency noise. More sophisticatedly,an adaptive whitening filter can be used which eliminatesthe spectral color that was introduced by the bioelectricalchannel H(z) but which is absolutely irrelevant for detectionof changes in the excitation signal. The observed SEMG signal(xk)k≥1 is processed by a filter with transfer function

Hw(z) = 1+ b1z−1 + b2z−2 + · · · + bqz−q (3)

with the filter orderq and the filter coefficientsbi tuned to theparameters of the shaping filter H(z). Obviously, for q = pand bi = ai, the whitening filter represents the ideal inversefilter Hw(z) = H−1(z) with respect to the transfer functionH(z). Therefore, with the parameters of H(z) exactly known,the excitation (wk)k≥1 can completely be reconstructed. Usu-ally, these parameters are unknown but can be estimated fromthe measured signal by some least-squares technique, for ex-ample [17].

In the next processing stage, a test function g(y1,y2, . . . , yk) is computed from the (pre-conditioned) signal(yk)k≥1 which serves as an indicator for the response on-set. The test function uses some or even all of the pastsamples to create an intermediate signal which is moni-tored by the decision rule in order to determine whethera change in the muscle activation pattern has occurred(alarm time ta). Some onset detection methods directly useta as an estimate t0 for the SEMG onset time t0. Someother methods use ta as a pure indicator for the exis-tence of an onset and estimate t0 through additional post-processing.

The decision rules of all methods will be written instopping rule notation,

ta = min{k ≥ 1 : g

(y1, y2, . . . , yk

) ≥ h}, (4)

Onset detection methods 69

ykH

g , ,...

Alarmtime

Changetime

Signalconditioning

Detection unit

Postprocessor

Test function Decisionrule

~

Signal model Detector

ˆ0

xk

(z)

(k)

0)(k,t

(k)

a

ykwk

k=t0

21

20

2

1y

2yk,

t

t

)y(

σ

σ

σ

Figure 1: Scheme of event detection in surface electromyographic signal (SEMG). The digitized SEMG xk is modeled by a white Gaussiannoise signal with dynamic varianceσ 2(k) exciting a linear system with transfer functionH(z). At t0, the signal variance changes from restingactivity σ 2

0 (k) to σ 21 (k, t0) of the active state. Most detectors employ up to three processing stages (signal conditioning, detection unit, and

post-processor) in order to detect an event (alarm time ta) and to determine an estimate for the unknown change time t0.

where h is an appropriate threshold. Each time a new datapoint yk is available, the current value gk of the test functionis computed from the samples y1, y2, . . . , yk and comparedto the threshold h. As long as gk < h, this procedure is re-peated sample by sample. At the first time instant ta, whengk ≥ h, the procedure is stopped, and an event alarm is given.

Estimation of the exact change time t0 by the post-processor usually starts after event alarm was given at ta, andthe estimate t0 of the unknown t0 is computed from thesamples y1, y2, . . . , yta+∆. Some methods employ a separatesignal conditioning unit which computes the input sequencefor the post-processor from the raw SEMG x1, x2, . . . , xta+∆.Most algorithms require an additional amount ∆ of samplesafter the alarm time, thus ta + ∆ represents the earliest timeinstant when the final onset estimate t0 is available. The (op-tional) change time estimation procedure is written as

t0 = f(y1, y2, . . . , yta+∆

), (5)

where f is an arbitrary function that computes the changetime from y1, y2, . . . , yta+∆.

In the following two sections, commonly used detectionmethods are described within the framework of this basiccomputational structure shown in Figure 1. The summary ofthis analysis is presented in Table 1, together with the param-eter values chosen for their performance evaluation.

2.2. Selected threshold-based onsetdetection methods

Hodges and Bui

The detection algorithm proposed by Hodges and Bui [13] isa representative for a large class of detection methods, known

as finite moving average (FMA) algorithms (cf. [5, 8, 18, 19]).All these methods employ a fixed-size sliding test window,and the output at time k is computed as the weightedsum of the W signal samples yk−W+1, yk−W+2, . . . , yk con-tained within the window. Within the framework of Figure 1,the Hodges method first conducts some signal conditioning;SEMG data samples are rectified and subsequently lowpassfiltered. From the pre-processed signal (yk)k≥1, the test func-tion g(y1, y2, . . . , yk) at time k is computed as the mean ofyk−W+1, yk−W+2, . . . , yk, and muscle activity onset is identi-fied at the instant when the test function exceeds the baselineactivity level by a specified multiple h of standard deviations.The baseline activity is adaptively determined by averagingthe initial M samples of (yk)k≥1. The algorithm can be sum-marized by the following decision rule:

ta = min{k ≥ W : gk ≥ h

},

gk = 1σ0

(yk − µ0

),

yk = 1W

k∑i=k−W+1

yk,

t0 = ta −W + 1,

(6)

where yk denotes the rectified and lowpass filtered SEMGsignal, and µ0 and σ0 are the mean and standard deviationof the M initial samples of (yk)k≥1, respectively. After thestopping rule signaled an alarm, the first index of the slidingwindow is used as an estimate for the onset time t0, which isa very simple post-processor. The Hodges method was imple-mented with the following parameters: h = 2.5, sixth-orderdigital butterworth lowpass filter with 50 Hz cutoff frequency,W = 50, M = 200.

70 EURASIP Journal on Applied Signal Processing

Table 1: Characteristic properties of all examined onset detection methods.

Hod

ges

Bon

ato

Lidi

erth

Abb

ink

AG

LRst

ep

AG

LRra

mp

Est

Opt

Signal conditioning

Pre-whitening × × × ×Data sample rectification × × ×Data sample squaring × × × ×Lowpass filtering × ×Test function/decision rule

Moving average + simple threshold × × ×χ2 test variable + double threshold ×Likelihood ratio test × × ×Post-processor/onset time estimation

Any threshold crossing ×Threshold & duration-based selection of alarm time × ×Search for maximum in (additional) test function × × × ×

Bonato et al.

The signal conditioning stage of the Bonato method [14] con-sists of an adaptive whitening filter, with the filter parametersestimated from the current SEMG data. The test functionis determined by two successive samples of (yk)k≥1 accord-ing to

gk = 1

σ20

(y2k−1 +y2

k), (7)

where σ0 is the standard deviation of the M initial samples of(yk)k≥1. Note that gk is evaluated only for odd values of k.The decision rule is

ta = min{k = 1,3,5, . . . : gk ≥ h

}(8)

and the post-processor checks the resulting alarms for theirrelevance. It accepts the onset decision only, if both followingrules apply:

• at least n out of m successive samples must exceed thethreshold in order to indicate the current muscle stateto be an active state,

• such an active state must last for at least T1 samples.

If this check holds true, the earliest beginning of this activestate epoch is taken as an estimate for the onset time t0.

As an important property, the method assumes that thepre-whitened SEMG (yk)k≥1 before the response onset isrepresented by a sequence of statistically independent zeromean Gaussian random variables with equal variance. Thus,with σ2

0 estimated from the measured sequence, the statisticalproperties of the test function gk can be explicitly specified,which facilitates a more sophisticated setting of the parame-ters h, n, m, and T1. Further, the use of a double-thresholdscheme in the post-processor provides a higher degree offreedom for selecting detection parameters. However, due tothe implicit maximum operation of the “n out of m” cri-terion the accuracy with which the response onset can be

determined is limited by the sizem of the test window. More-over, the down-sampling operation reduces time resolutionwhich may also lead to a less accurate onset estimation. Bon-ato was implemented with the parameters h = 7.74, n = 1,m = 5, T1 = 50, and M = 200.

Lidierth

Lidierth [15] proposed a full-wave rectification of the rawSEMG as signal conditioning only. The test function and de-cision rule of the Lidierth detection unit are identical to theHodges approach, but increased performance is achieved bythe extended post-processing rules:

• SEMG onset detection is accepted, if the test functionexceeds the threshold h to the active level for at leastT1 samples, and

• during T1, the test function may repeatedly fall be-low the threshold, but each time not longer thanT2 samples.

Note that these post-processing rules represent a specific caseof the post-processor suggested by Bonato; with the param-eters of the Bonato “n out of m” criterion set to n = 1and m = T2, both post-processing rules are identical, thusLidierth is a composite of Bonato and Hodges. Lidierth wasimplemented with parameters T1 = 90, T2 = 15, h = 3, andM = 200.

Abbink et al.

Again, the “new” method copies the signal conditioningand detection unit from already known methods, andadds some specific fine tuning by extending the post-processor function: Abbink [2] uses the Hodges signalconditioning and detection unit approach, only changingthe cutoff frequency to 3 Hz and the window lengthW to 1.

The post-processor stage of Abbink uses a less smooth ver-sion (y′k)k≥1of SEMG, which is obtained by applying a but-

Onset detection methods 71

terworth lowpass with 30 Hz cutoff frequency to the rectifiedSEMG (i.e., the post-processor does his own signal condi-tioning). Looking backwards from ta, the following rules areapplied:

t0 = arg maxN≤j≤ta

{nlow(j)+nhigh(j)

},

nlow(j) =j∑

i=j−N+1

Ind

{1σ0

(y′i − µ0

)< h2

},

nhigh(j) =j+N∑i=j+1

Ind

{1σ0

(y′i − µ0

)> h2

},

Ind(x) =1 if x is true,

0 otherwise,

(9)

wherenlow(j) denotes the number of normalized amplitudessmaller than the threshold h2 in the N samples directly pre-ceding the onset time candidate j, andnhigh(j) is the numberof amplitudes larger thanh2 in theN samples directly follow-ing j. Mean µ0 and standard deviation σ0 are estimated fromthe initial M samples of (y′k)k≥1, respectively. The followingparameters were used for the simulations: h = 3, h2 = 3,N = 200, and M = 200.

2.3. Detectors based upon statisticallyoptimal decision

If the signal generating process is (partially) known, detectorsbased on statistically optimal decision rules can be used. Inthis section, three algorithms are presented which are derivedfrom the model-based detection concept of [12]. In orderto determine an estimate of t0, these methods evaluate thestatistical properties of the measured SEMG signal beforeand after a possible change in model parameters. Only basicprocessing steps are described here (see [12] for details). Thespecific implementations of the decision rules for the SEMGsignal model (see Figure 1 left) are given in the appendix.

For signal conditioning, all 3 methods employ an adap-tive whitening filter for the reduction of irrelevant informa-tion. Assuming optimal performance of the filter, the (re-constructed) excitation signal (yk)k≥1 = (wk)k≥1 representsa single realization of a statistically independent Gaussianrandom process (Yk)k≥1. The statistical properties of an in-dividual random variable Yk can be fully described by itsprobability density function (PDF)

pσ(yk) = 1√2πσ2(k)

e−y2k /2σ

2(k), (10)

which depends upon the (deterministic) variance profileσ2(k). The model-based decision rules for this problemdepend on the a priori knowledge about the variance pro-files σ2

0 (k) and σ21 (k, t0) before and after the change to be

detected.

Optimal estimator (EstOpt)

If the variance profilesσ20 (k) andσ2

1 (k, t0) are exactly known(except the onset time t0), a statistically optimal decision rule

can be designed according to

ta = min{k ≥ 1 : gk ≥ h

},

gk = max1≤j≤k

Skj ,

Skj =12

k∑i=j

[(σ−2

0 (i)− σ−21 (i, j)

)y2i + ln

σ20 (i)

σ21 (i, j)

],

t0 = arg max1≤j≤ta

Staj .

(11)

This detector referred to as EstOpt algorithm computes themaximum likelihood (ML) estimate t0 of the unknown onsettime t0 provided that all remaining parameters of the signalmodel are exactly known. The EstOpt decision rule is suitedfor arbitrary but known dynamic variance profilesσ2

0 (k) andσ2

1 (k, j), thus it shows optimal performance among all pos-sible algorithms for sequential detection in this application.Therefore, the EstOpt detector may serve as a reference forperformance rating. (Essentially, the test compares the log-likelihood ratio

Skj =k∑i=j

lnpσ1

(yi, j

)pσ0

(yi) (12)

between the two distributions before and after a possiblechange at time j with a threshold h. Since the exact changetime t0 is unknown, it is replaced by its ML estimate, that is, amaximum operator selects the largest value of the test func-tion with respect to all possible change times 1 ≤ j ≤ k. Eachtime a new data point yk is available, a test window with theupper bound k fixed at the current observation and with itslower bound j comprises all the observationsyj,yj+1, . . . , ykafter the hypothetical change time j. For each candidate j, thelog-likelihood ratio Skj is computed from the observations

within the test window. If the maximum of Skj with respect toall hypothetical change times 1 ≤ j ≤ k exceeds the thresh-old h, data acquisition stops and event alarm is given (alarmtime ta). The time j at which the maximum value is obtainedserves as the maximum likelihood estimate t0 of the unknownchange time t0.)

Approximated generalized likelihood-ratio detectors(AGLRstep, AGLRramp)If the variance profilesσ2

0 (k) andσ21 (k, j) are not known like

in EstOpt, they can be replaced by estimates. For this condi-tion, the approximated generalized likelihood ratio (AGLR)detector [20] is an appropriate tool to formulate the stoppingrule. First, the variance profiles are assumed to depend uponseveral unknown parameters, that is,

σ20 (k) = σ2

0

(k, θ0

),

σ21 (k, j) = σ2

1

(k, j, θ1

).

(13)

The unknown parameter vector θ0, true up to t0, is estimatedfrom the firstM observations of the SEMG by ML techniques,resulting in

72 EURASIP Journal on Applied Signal Processing

θ0 = arg supθ0

( M∑i=1

lnpσ0

(yi, θ0

)), (14)

which is kept fixed throughout the remaining detection pro-cedure. Next, a sliding window of fixed size W is continu-ously shifted along the data sequence. For each location k ofthe window, the ML estimate θ1 of the unknown parametervector θ1 after change is determined from the W data pointscovered by the window, and the corresponding log-likelihoodratio Skk−W+1 is computed and compared with a threshold h.Finally, after a change has been indicated, the exact changetime is estimated by an ML procedure from all possible can-didates j ≤ ta. The AGLR decision rule can be summarized as

ta = min{k ≥ W : gk ≥ h

},

gk = Skk−W+1,

Skj = supθ1

12

k∑i=j

[(σ−2

0

(i, θ0

)− σ−21

(i, j, θ1

))y2i

+ lnσ2

0

(i, θ0

)σ2

1

(i, j, θ1

)],

t0 = arg maxW≤j≤ta

(Sta+∆j

),

(15)

where ∆ is an appropriate dead zone which ensures that aminimum number of observations is available for parameterestimation. Note that the maximization with respect toθ1 hasto be repeatedly performed for each combination (j, k) to betested. Therefore, the computational effort is dominated bythe number of operations required for computing θ1.

In this paper, two implementations of the AGLR decisionrule are used which differ in the complexity of σ2

1 (k, j). TheAGLRstep algorithm assumes a step-like variance profile withconstant but different variances before and after change. Thisallows for a very efficient implementation of the method butcompletely disregards the available information about the dy-namic profile of the change itself. The AGLRramp algorithmexplicitly takes this information into account by assuming a“ramp and hold”change profile. The unknown parameters θ1

of the transition are estimated from the actual observations.Thus, comparison of the AGLRstep and AGLRramp methodsshows the advantage of using available information about thechange dynamics in the detection process. Their exact im-plementations for the present signal model are given in theappendix.

3. EVALUATION OF DETECTION PERFORMANCE ONSIMULATED SEMG RECORDINGS

3.1. Signal generation

All methods were tested for robustness and efficiency on aset of simulated SEMG traces which were generated by us-ing the signal model shown in Figure 1(left). The coefficientsof the shaping filter were determined from a representativeset of real SEMG data from a previous study [21] using

standard techniques of least squares parameter estimation[17]. Model order was set to p = 8. A total number of 16000segments, each consisting of 1000 data points (i.e., 1000 msrecord length) were generated. Figure 2 depicts the principleand definitions. The response to be detected was modeled bya rapid change in the variance σ2(k) from a pattern

σ20 (k) = σ2

noise (16)

to a new pattern

σ21

(k, t0

) = σ2noise + σ2

signalu(k, t0

), (17)

where t0 denotes the onset of the voluntary muscle activa-tion, u(k, t0) describes the dynamic profile of the change,and σ2

noise and σ2signal are appropriate constants. The resting

activity σ2noise 0 results from spontaneous firing of motor

units which occurs even if the muscle is not voluntarily acti-vated. (Note that the usage of the term “noise” is technicallybut not biologically motivated.) σ2

signal denotes the magni-tude of the activation pattern. The dynamic change profilewas approximated by a unit ramp and hold transition

u(k, t0

) =

0 when k ≺ t0,(k− t0

)τ

when t0 ≤ k ≤ t0 + τ,

1 when k t0 + τ,

(18)

where τ denotes the duration of the ramp. This finite slopeof the change profile reflects the limited dynamics within theneuromuscular system [16], and τ was varied as it can beobserved in real data. The onset t0 of voluntary muscle acti-vation was randomly varied between 400 ms and 600 ms, rel-ative to the beginning of the trace. Change magnitude σ2

signal

was always set to 1 while background activityσ2noise was varied

to create variable signal-to-noise ratios (SNR)

SNR = 10 · log10

σ2signal

σ2noise

[dB]. (19)

Several sets of simulated SEMG traces, each consisting of 4000trials, were used for testing the onset detection algorithms.The assessment of the onset detection accuracy was madefor the following range of parameters (referenced as MixedTrials):

• ramp duration τ uniformly distributed between 5 msand 30 ms,

• signal-to-noise ratio SNR uniformly distributed be-tween 6 dB and 12 dB.

In particular, dependence on SNR and ramp duration werestudied. Sensitivity to SNR was investigated with the MixedSNR Trials data set:

• fixed ramp duration τ of 20 samples,• SNR uniformly distributed between 6 dB and 12 dB.

Performance for extremely small SNR was evaluated with theFixed SNR Trials data set:

Onset detection methods 73

+

ramp duration

signal

0

2noise

2noise

2

t

k

2 ( )k

τ

σσ

σ

σ

(a) Power envelope of modeled SEMG signals. After a period of background activity the SEMGpower linearly rises to the active state within τ. Parameters were systematically varied duringsimulation.

0 200 400 600 800 1000

0

1

2

3

0

σ ( )

x

k

k

t

−1

−2

−3

ms( )

(b) Sample trace of simulated SEMG.

Figure 2: Simulated SEMG.

• fixed ramp duration τ of 20 samples,• fixed SNR of either 3 dB or 6 dB.

Sensitivity to ramp duration was investigated with the MixedRamp Trials data set:

• ramp duration τ uniformly distributed between 5 msand 30 ms,

• fixed SNR of 10 dB.

All simulations were performed using MATLAB* (version5.3, The MathWorks, Natick, MA) on an IBM-compatible PC.

3.2. Results of onset detection analysis

The aim of this paper is to objectively compare the perfor-mance of commonly used onset detection methods. It shouldbe considered that several factors could influence detectionperformance:

• algorithmic delay of onset detection,• percentage of missed responses and false alarm rate,• detection error bias, that is, average deviation between

estimated and true onset,• reproducibility for signals with the same parameters,

• sensitivity of all attributes to SNR and onset ramp du-ration.

The first two factors refer to the detection power of the al-gorithm, that is, the ability of a method to signal an eventas early as possible (small delay of alarm time ta) com-bined with a false alarm rate as low as possible. Actually,due to the relatively high SNR, the majority of the simulatedSEMG responses could be detected within a reasonable timerange around the real onset, and the detection probability ofall algorithms was nearly 100%. We, therefore, focus on theremaining 3 factors which are related to the accuracy of theestimated change times t0.

Measures of detection performance

Figure 3 shows the probability density functions (PDFs) ofthe onset error ε = t0 − t0 for all methods (Mixed Trials).Two diagrams with different abscissa scales are used for im-proved visibility. The optimal estimator EstOpt indicating theupper performance limit provides a rather narrow distribu-tion which is almost symmetrical to the origin. The AGLR-ramp and AGLRstep methods are showing increasing asym-metry and width of the distributions. While the PDF of the

74 EURASIP Journal on Applied Signal Processing

EstOpt

−25 −20 −15 −10 −5 0 5 10 15 20 250

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

AGLRstep

AGLRramp

PD

F

Onset error (ms)

(a)

0 10 20 30 400

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

Onset error (ms)

Bonato

EstOpt

Abbink

Lidierth

Hodges

50

PD

F

−−−−− 1020304050

(b)

Figure 3: Probability density functions (PDFs) of onset estimationerrors for the Mixed Trials data set (SNR 6–12 dB, ramp duration5–30 ms). Note the different abscissa scaling in (a) and (b).

AGLRramp method is still centered near the origin, the PDFof AGLRstep is shifted towards higher onset errors. In the per-formance ranking,Bonato and Lidierth follow with reasonableresults. Finally, Abbink shows worse results because the PDFspreads to larger onset error values, and Hodges ranks low-est because of a wide-spread distribution with additionallystrong (negative) bias.

Table 2 depicts mean and standard deviation of onsetestimation errors and the percentage of detected onsets withan absolute error less than 100 ms for each method. The100 ms limit was chosen because it comprises at least 99%of all onset estimates computed from the Mixed SEMG Trials.Note that, due to the very thin tail of the PDF, variations inthe limit will result in only minor variations of the result-ing probability of detected onset. Analysis of the mean er-rors in Table 2 shows that the AGLRramp algorithm provides

Table 2: Statistics for onset estimates based on SEMG data set“Mixed Trials.”

MethodDetected Mean Error

onsets error STD

EstOpt 100.0% 0.6 ms 3.6 ms

AGLRramp 99.7% 0.2 ms 5.4 ms

AGLRstep 99.8% 4.2 ms 5.0 ms

Bonato 99.9% 4.0 ms 7.5 ms

Lidierth 98.9% 0.1 ms 11.1 ms

Abbink 99.6% 8.8 ms 10.4 ms

Hodges 99.9% −7.1 ms 11.8 ms

estimates very close to the “true” onset. AGLRstep, however,yields a positive bias of 4.2 ms reflecting the fact that the step-like change profile assumed by AGLRstep only represents anapproximation for the more gradual “true” change profile.The improvement in bias obtained by using the more realis-tic change pattern of the AGLRramp algorithm, however, is atthe expense of a slightly increased standard deviation. Thisis because estimation of the additional parameter introducesadditional uncertainty to the final detection result.

The accuracy function

Pa = P(∣∣t0 − t0

∣∣ ≤ a)

(20)

shown in Figure 4 can be used as another measure of de-tection performance. The accuracy function denotes the per-centage of responses that were detected with an absolute errorsmaller than or equal to a maximum tolerated error a for dif-ferent values of a. For largea, the diagram reflects the generalability of a method to detect an onset, whereas a particularaccuracy requirement of a method can specifically be assessedat small values a. Thus, the more a curve approaches the up-per left corner, the higher is the onset detection quality. As amajor result, the accuracy functions in Figure 4 confirm theranking already obtained from the PDFs shown in Figure 3.

Effect of signal to noise ratio

Figure 5 depicts the dependence of mean and standard de-viation of the onset error on SNR (Mixed SNR Trials). Allmethods show a systematic degradation of detection perfor-mance for smaller SNR. Except the Hodges approach (sensi-tivity to SNR of approximately 1 ms/dB), all methods showa similar relative SNR sensitivity of the standard deviationof approx. 0.5 ms/dB. Generally, the overall variability is sig-nificantly smaller for the methods employing a whiteningfilter (EstOpt, AGLRstep, AGLRramp, Bonato) and decreaseswith increasing level of a priori knowledge used. Again, theAGLRramp provides the smallest bias among all detectorsshowing a zero mean estimation error close to the EstOpt ref-erence (see Figure 5b). Most importantly, the AGLRramp de-tector is the only approach which is nearly insensitive to SNR.All other methods show biased results with an SNR sensitivityof 0.3–0.8 ms/dB. The poor results of the Hodges approach(sensitivity 3.3 ms/dB) demonstrate the general problem ofany simple adaptive threshold detector; the threshold level is

Onset detection methods 75

0 5 10 15 20 250

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

P

Tolerated onset error a

Bonato

EstOpt

Hodges

Lidierth

Abbink

AGLRstep

AGLRramp

a

(ms)

Figure 4: Accuracy functions of onset estimates (Mixed Trials dataset). The ordinate value denotes the probability that the toleratederror a given by the abscissa value is not exceeded. The further acurve is bent to the upper-left corner, the better is the onset detectionquality.

usually adapted to the background activity (noise) and thusvaries with SNR. The higher the noise level (decreasing SNR),the larger the threshold and, consequently, the later the onsetwill be detected.

Degradation of performance with smaller SNR is partic-ularly obvious from the accuracy functions in Figure 6 (FixedSNR Trials). For SNR = 6 dB, all methods provided accept-able results (see Figure 6a) but decreasing signal quality toSNR = 3 dB promptly reveals their specific quality and ro-bustness (see Figure 6b). The optimal reference EstOpt stilldetects 93% (SNR 6 dB) and 82% (SNR 3 dB) of all onsetswith an accuracy of 10 ms. AGLRramp and AGLRstep followclosely with a near-zero rate of undetected onsets. Even for the3 dB trials, more than 98% of onsets are detected with an abso-lute error less than 50 ms. But all the purely threshold-basedmethods (Abbink, Bonato, Lidierth, Hodges) cannot competewith likelihood-based methods; especially for the 3 dB situa-tion, their performance deteriorates extremely.

Effect of change dynamics

Figure 7 illustrates the dependence of mean onset error onramp duration τ (Mixed Ramp Trials). AGLRramp is de-signed to compensate change dynamics, and, thus results areclosest to those of EstOpt. But Bonato, Lidierth, and AGLRstepshow a similar small sensitivity to ramp duration, only Abbinkand Hodges are rather sensitive. Variability of onset estimates(around mean onset error) was not affected by variations inramp duration.

Generally, changing SNR predominately affects the onseterror PDF width and the percentage of undetected (missed)onsets, whereas increasing ramp duration leads to a delayedonset detection (onset error bias).

6 7 8 9 10 11 122

4

6

8

10

12

14

Stan

dard

dev

iati

on o

f on

set

erro

r (m

s)

EstOpt

Lidierth

AGLRstep

Bonato

AGLRramp

Abbink

Hodges

(a) Standard deviation of onset error.

6 7 8 9 10 11 12

0

5

10

15

Mea

n o

nse

t er

ror

(ms)

SNR (dB)

LidierthEstOpt

AGLRramp

Abbink

AGLRstep

Hodges

Bonato

−

−

−5

10

15

(b) Mean onset error.

Figure 5: Dependence of onset estimation error on signal-to-noiseratio (SNR) for the Mixed SNR Trials data set (SNR 6–12 dB, rampduration 20 ms).

4. EVALUATION OF DETECTION PERFORMANCEON REAL SEMG RECORDINGS

4.1. Data acquisition

All methods were additionally tested on real SEMG recordsobtained in a reaction time situation. Briefly, five subjectswithout any sign of neurological deficits were examined. Theyperformed rapid stimulus initiated transient lateral abduc-tions of the index finger in response to a bright visual “go”signal displayed at random times. During the experiment,the tip of the index finger was fixed to a lever which restrictedmovement to the horizontal plane. SEMG activity of the firstdorsal interosseus (FDI) muscle was recorded using an activesurface electrode (Liberty Mutual MYO 111). SEMG signalswere digitized for a period of 6 s around the stimulus onsetwith a sampling rate of 1000/s. After short practice, each sub-

76 EURASIP Journal on Applied Signal Processing

0 10 20 30 40 500

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Bonato

EstOptAGLRramp

Hodges

Lidierth

Abbink

AGLRstep

Pa

(a) SNR = 6 dB.

Tolerated onset error a (ms)

0 10 20 30 40 500

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Bonato

EstOptAGLRramp

Hodges

Lidierth

Abbink

AGLRstep

Pa

(b) SNR = 3 dB.

Figure 6: Dependence of accuracy functions on SNR for constantramp duration of 20 ms.

ject produced 120 single responses. The recorded data werevisually inspected and only those responses which were def-initely initiated between 100 and 400 ms after stimulus pre-sentation were included to the test data set comprising a totalnumber of 587 responses.

As a major handicap of real data, the true response onsett0 is unknown and, therefore, the onset error ε = t0 − t0 asused before cannot be determined. But performance of themethods can be assessed on a relative basis; to this, the onseterror ε = t0−tref was computed as the deviation between theonset estimates t0 provided by the tested methods and theonset estimates tref visually determined by a human expert.In order to maximize the percentage of correct detection ofthe methods, their decision thresholds were set individuallyby taking the smallest value h for which the number of de-tected responses with alarm times −50 ≤ ta − tref ≤ 200 mswas maximum. The resulting thresholds are summarized inTable 3.

5 10 15 20 25 30

0

5

10

15

Mea

n o

nse

t er

ror

(ms)

Ramp duration (ms)

EstOpt Lidierth

Abbink AGLRstep

Hodges

Bonato

AGLRramp−5

−10

−15

−20

Figure 7: Dependence of mean onset error on ramp duration. Eval-uation of Mixed Ramp Trials data set (SNR of 10 dB, ramp duration5–30 ms).

Table 3: Statistics for real SEMG data set.

Method Threshold hDetected Mean Error

onsets error STD

AGLRramp 200 100.0% 0.4 ms 3.6 msAGLRstep 200 100.0% 0.5 ms 3.5 msBonato 20 99.8 % −0.8 ms 5.5 msLidierth 3 99.5 % −2.3 ms 6.9 msAbbink 30 98.6 % −0.6 ms 9.8 msHodges 5 99.0 % −7.5 ms 9.3 ms

Figure 8 depicts two representative SEMG recordings.Onset estimates obtained with each method are indicated bymarkers. If the SEMG signal shows a large SNR and a steepchange profile, all methods provide onset estimates close tothe expert reference (see Figure 8a). If SNR decreases, on-set estimates across methods will become more variable (seeFigure 8b). The performance of the methods in real data on-set detection is depicted by histograms of the onset error(see Figure 9a) showing the deviations of the computerizedmethods from the expert’s reference and by the correspon-ding accuracy functions (see Figure 9b). Mean errors andstandard deviations are summarized in Table 3. Consistently,the majority of SEMG recordings of this particular paradigmhad large SNR and steep slopes like the sample recordingin Figure 8a. This is reflected by the very narrow error his-tograms (see Figure 9a) and the rapidly increasing accuracyfunctions (see Figure 9b) indicating satisfactory performancefor the majority of methods tested. Only the Hodges methodprovided onset estimates which were significantly smaller(7.5 ms on average) than the expert’s decision, as indicatedby the skewed error distribution (see Figure 9a). The rankingof methods is consistent with the ranking already obtainedin the simulation study attesting highest performance to theGLR-based methods followed by the Bonato approach. Note

Onset detection methods 77

Bonato

AGLRramp

Expert

Hodges

Lidierth

Abbink

AGLRstep

k

xk

(a)

Bonato

AGLRramp

Expert

Hodges

Lidierth

AbbinkAGLRstep

100 ms

xk

k

(b)

Figure 8: Raw SEMG signals obtained in a reaction time experiment:two sample traces (a) and (b) with different SNR. Onset estimatesof each method are indicated by markers. Dotted lines indicate theexpert’s reference.

the small difference between the AGLRramp and AGLRstepalgorithm indicating steep response profiles close to the step-like pattern of assumed by the AGLRstep method.

5. DISCUSSION

All methods perform well in detecting the onset time in goodquality SEMG signals; thus, the difficulty to decide on theappropriate method appears with low SNR when SEMG sig-nals are disturbed for some reason. In any case, the choice isalso dependent on implementation issues like method com-plexity, real-time implementation capability, and the requiredCPU performance.

Simple threshold-based methods are very popular be-cause of their intuitive and easy implementable structure. In

0 5 10 15 20 250

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

AGLRstepAGLRramp

PD

F

Onset error (ms)

Bonato

Abbink

Lidierth

Hodges

−5−10−15−20− 25

(a) Probability density functions of onset estimation errors.

0 5 10 15 20 250

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

P

Tolerated onset error a (ms)

Bonato

AGLRramp

Hodges

Lidierth

Abbink

AGLRstep

a

(b) Accuracy functions.

Figure 9: Analysis of deviations between estimated change timesand expert’s reference obtained on real SEMG signals with differentcomputerized detectors.

most cases, SEMG signals have a good SNR, so these methodsare well applicable with the main limitation that SNR shouldbe larger than 10 dB. But SEMG of small and deep musclesas well as SEMG recorded in patients with neuromusculardiseases may not fit into this SNR requirement, which callsfor application of more sophisticated approaches.

Generally, purely threshold-based methods represent atradeoff between detection precision and detection proba-bility. A relatively low threshold level results in early onsetdetection but also causes more false alarms. High thresholdlevels, however, will usually lead to delayed or even missedonset detection. Although this delay can be partly compen-sated by subtracting an appropriate constant value from the

78 EURASIP Journal on Applied Signal Processing

alarm time, the large variability of resulting change time esti-mates and their systematic dependence on change dynamicsconfine the application of simple threshold-based methods ifSNR is low and/or highly accurate change time estimation isrequired.

Analysis of simulated as well as real SEMG data showedthat appropriate signal conditioning of the measured rawSEMG is a prerequisite for high detection performance. Low-pass filtering the (full-wave-rectified) SEMG before com-puting the test function substantially reduces the risk offalse alarms. But it causes a very smooth incline of thepre-processed signal near the onset and thus will lead toincreased variability of estimated change times. Simple-threshold detectors like Hodges are known for their resultsbeing very sensitive to the parameters of the lowpass filterused [13]. The Abbink approach suffers from this dilemma,too, despite of the more sophisticated change time estimationprocedure.

Application of an adaptive pre-whitening filter proved tobe superior to the less specific lowpass filter. Simulations withwhite Gaussian SEMG signals have shown that Bonato andLidierth almost have the same onset error distribution,which is a consequence of their resembling detection strate-gies. Analysis of the more realistic colored SEMG data (seeFigure 4) as well as analysis of real SEMG signals (seeFigure 8) demonstrated the distinct performance gain of Bon-ato due to the use of an appropriate pre-whitening filter. Gen-erally, the implicit highpass characteristic of the whitening fil-ter preserves or even improves change dynamics as desired butaccentuates the higher frequencies of possibly superimposedadditive noise [22]. Therefore, we can expect good perfor-mance of the whitening filter provided that the noise is smallcompared to the variance profile σ2(k).

If higher robustness to changes in signal parameters andhigh detection precision is required, a pre-whitening filtertogether with a statistically optimized decision rule is thefirst choice. Consistently, the two model-based approachesAGLRstep and AGLRramp provided higher detection powerand more accurate change time estimates than the unspe-cific detectors both for the simulated and real SEMG signals.Particularly, these methods were most robust with respect tovariations in signal properties such as SNR and change dy-namics. However, model-based approaches are usually asso-ciated with higher computational efforts, particularly, whenthe dynamic of the change is not known and has to be esti-mated from the data. Results show only minor loss of per-formance for the computationally more efficient AGLRstepalgorithm compared to the more specific but more complexAGLRramp algorithm. In order to reduce the computationaleffort of such more complex parametric detectors, a step-likestopping rule according to (31) may be combined with theML change time estimator as shown in (37). This allows foran efficient implementation of the detection procedure, sincethe time-consuming estimation of the onset shape profile isonly initiated once a change has been indicated.

An important aspect of computerized onset detectionis post-processing. Generally, combining an arbitrary stop-ping rule with a post-processor testing multiple alarm/change

times for their plausibility can improve detection perfor-mance. Particularly, the detection threshold can be reducedsince false alarms can be partly compensated by the post-processor. The use of an adequate post-processor is par-ticularly important when the SEMG signal contains mul-tiple changes indicating different levels of muscle activa-tion. In this case, the detection unit produces a sequenceof change times indicating possible transitions between dif-ferent levels of muscle activation. A post-processor thengroups the segments according to their variance mergingsegments with similar variance together. Combinations ofGLR-based decision rules and post-processor have been suc-cessfully applied for the detection of pre-motor silent peri-ods in SEMG signals [22] and for the detection and classi-fication of events in the uterine EMG [23]. But any of thepost-processors presented is also dependent on signal con-ditioning and alarm time generation. This means that anyimpairment caused by signal conditioning will neverthelessaffect detection.

6. EPILOGUE: DESIGN, PRESENTATION, ANDCOMPARISON OF ALGORITHMS

“An algorithm must be seen to be understood and the bestway to learn how an algorithm works is to play with it.”(Donald Knuth, in “The Art of Computer Programming,”[24]).

Every algorithm has embedded parameters, and it worksonly properly if these parameters are appropriately chosen.“Good” parameters are often determined by trial and errorusing simulated data. Despite the importance of finding andchoosing the right parameters, usually little emphasis is de-voted to this step by papers about algorithms, presumablybecause there is no clear mathematical theory behind it (i.e.,scientifically accepted).

The presentation of algorithms to the scientific commu-nity is usually done by writing about

(1) the practical and mathematical background,(2) a useful class of models,(3) a description of the algorithm(s),(4) a theoretical analysis of the behavior in some limiting

cases,(5) a table of numbers showing the performance of the al-

gorithm tested on simulated data, while keeping someparameters in the signal model and the algorithmfixed,

(6) as in (4), but also with information about competingalgorithms,

(7) a plot of a signal together with the correspondingoutput of the algorithms in question.

What typically is missing is a tool

(a) to check and interactively change the signal and tofind good parameters for the algorithms for other ap-plications,

(b) to compare different algorithms with arbitrary com-binations of parameters,

Onset detection methods 79

(c) to check the behavior of the algorithm(s) with real

data.

These aspects are even more valid in the case of on-linealgorithms. In the statistical community, new ways to presentresults using the internet were emphasized [25], and severalgroups are currently working on software tools which allowa user with any standard browser to visualize and change sig-nals and algorithms. To our experience, systematically butinductive trying of all the parameters in both the modeland the algorithm yields quickly an idea about interestingneighborhoods in parameter space, quality, and stability of analgorithm. Software tools appropriate for this purpose shouldbe easy to use and should not require a comprehensive instal-lation procedure: neither many researchers nor referees dohave the time to reprogram a new algorithm in their ownfavorite programming environment to evaluate new signalprocessing procedures. To illustrate this point, an onset de-tection program designed in MATLAB* will be made avail-able to the readers by the authors, which allows to visual-ize the model described above, to change its parameters andto see the output of some selected algorithms. Using suchtools simplifies to assess an algorithm or method, but requiresto own a MATLAB* license up to now. But there are somedevelopments to overcome this problem (e.g., compiler ver-sion) in near future, which will allow the software package tobe made available through the internet.

APPENDIX

In this section, the implementation of the AGLRstep andAGLRramp decision rules for the present signal model areshortly described.

The structure of the tests depends upon the variance pro-files before and after the response onset t0 which, accordingto the process model, are

σ20 (k) = σ2

noise, (21)

σ21

(k, t0

) = σ2noise + σ2

signalu(k, t0

), (22)

respectively. The variance pattern σ20 (k) before change (i.e.,

response onset) depends upon a single parameter

σ20

(k, θ0

) = θ0 = constant, (23)

which is equal to the unknown variance σ2noise. The ML

estimate of θ0 can be obtained from the initial M data pointsaccording to

θ0 = arg supθ0

M∑i=1

−12

(lnθ0 +

y2i

θ0

). (24)

Differentiation with respect to θ0 and solving the resultinglikelihood equation

12

1

θ20

M∑i=1

y2i −

Mθ0

= 0 (25)

leads to

σ20

(k, θ0

) = θ0 = 1M

M∑i=1

y2i . (26)

Thus, the estimated variance profile before change is equal tothe average signal energy within the reference window.

Determination of the variance pattern σ21 (k, j) after

change is more complicated since it depends upon the dy-namic change profile u(k, j) which, generally, is not known.This issue is addressed in the next paragraphs.

AGLRstep detector

The AGLRstep detector simply assumes an approximatelyconstant variance profile of unknown magnitude θ1 afterchange, that is,

σ21

(k, j, θ1

) = θ1 = const. (27)

In this case, the log-likelihood ratio in (14) can be written as

Skj = supθ1

12

k∑i=j

[(θ−1

0 − θ−11

)y2i + ln

θ0

θ1

], (28)

where θ0 is the estimated variance before change accordingto (26). Maximization with respect to θ1 is explicitly possibleby replacing θ1 by its ML estimate

θ1(j, k) = 1k− j + 1

k∑i=j

y2i , (29)

which leads to

Skj =k− j + 1

2

(θ1(j, k)θ0

− lnθ1(j, k)θ0

− 1

). (30)

Thus, the AGLRstep decision rule can be summarized as

ta = min{k ≥ W : gk ≥ h

},

gk = Skk−W+1,

Skj =k− j + 1

2

(ρ(j, k)− ln ρ(j, k)− 1

),

t0 = arg max Sta+∆j ,

(31)

where

ρ(j, k) = θ1(j, k)θ0

=(1/(k− j + 1)

)∑ki=j y2

i

(1/M)∑Mi=1 y

2i

(32)

denotes the ratio between the estimated variances before(reference window) and after change (test window), respec-tively. This decision rule can be implemented in a completelyrecursive manner. The AGLRstep algorithm was implementedwith parameters W = 25, h = 10, ∆ = 100, and M = 200.

80 EURASIP Journal on Applied Signal Processing

AGLRramp detector

The AGLRramp detector uses a more complex test functionwhich also takes the dynamic change profile into account. Itassumes that the shape u(k, j) of the additive change in thevariance profile is known but its exact magnitude σ2

signal isunknown, that is,

σ21

(k, j, θ1

) = θ0 + θ1u(k, j). (33)

Then, the log-likelihood ratio in (14) can be rewritten as

Skj = supθ1

12

k∑i=j

[(1

θ0− 1

θ1u(i, j)+ θ0

)y2i

+ lnθ0

θ1u(i, j)+ θ0

],

(34)

where the unknown magnitude θ1 is the only parameter tobe determined. Differentiating the log-likelihood ratio withrespect to θ1 results in a likelihood equation which cannot beexplicitly solved. Therefore,θ1 is estimated from the observedsequence according to

θ1(j, k) =∑ki=j

(y2i − θ0

)∑ki=j u(i, j)

(35)

by exploiting the statistical independence of the data se-quence. The resulting AGLR decision rule for the detection ofan additive change with known dynamic profile but unknownmagnitude is

ta = min{k ≥ W : gk ≥ h

},

gk = Skk−W+1,

Skj =12

k∑i=j

[(s

1

θ0− 1

θ1u(i, j)+ θ0

)y2i

+ lnθ0

θ1u(i, j)+ θ0

],

θ1(j, k) =∑ki=j

(y2i − θ0

)∑ki=j u(i, j)

,

t0 = arg maxW≤j≤ta

Sta+∆j ,

(36)

where θ1(j, k) is individually determined for each combina-tion (j, k) according to (35).

If the exact profile is also unknown, the test may in-clude more maximizations which determine the ML esti-mates of the unknown parameters, too. Particularly, u(k, j)may be replaced by a set of N template profiles u(k, j)(n),n = 1,2, . . . , N, together with an additional maximizationwhich selects the most likely template according to

Skj = max1≤n≤N

12

k∑i=j

[(1

θ0− 1

θ(n)1 u(i, j)(n) + θ0

)y2i

+ lnθ0

θ(n)1 u(i, j)(n) + θ1

].

(37)

The AGLRramp algorithm was implemented with parametersW = 25, h = 10, ∆ = 100, M = 200, and a set of N = 8shaping functions with ramp durations τ = 5,10, . . . ,40 ms.

REFERENCES

[1] E. A. Clancy and N. Hogan, “Probability density of the surfaceelectromyogram and its relation to amplitude detectors,” IEEETrans. Biomed. Eng., vol. 46, pp. 730–739, 1999.

[2] J. H. Abbink, A. van der Bilt, and H. W. van der Glas, “De-tection of onset and termination of muscle activity in surfaceelectromyograms,” J. Oral Rhabil, vol. 25, pp. 365–369, 1998.

[3] S. V. Adamovich, M. F. Levin, and A. G. Feldman, “Merging dif-ferent motor patterns: coordination between rhythmical anddiscrete single-joint movements,” Exp. Brain Res., vol. 99, pp.325–337, 1994.

[4] P. Brodin, T. S. Miles, and K. S. Turker, “Simple reaction-time responses to mechanical and electrical stimuli in humanmasseter muscle,” Arch. Oral Biol., vol. 38, pp. 221–226, 1993.

[5] S. A. V. M. Haagh, W. A. C. Spijkers, B. Van den Boogaart,and A. Van Boxtel, “Fractioned reaction time as a function ofresponse force,” Acta. Psychol., vol. 66, pp. 21–35, 1987.

[6] J. K. Leader 3rd, J. R. Boston, and C. A. Moore, “A data depen-dent computer algorithm for the detection of muscle activityonset and offset from EMG recordings,” Electroenceph. Clin.Neurophysiol., vol. 109, pp. 119–123, 1998.

[7] S. Micera, A. M. Sabatini, and P. Dario, “An algorithm fordetecting the onset of muscle contraction by EMG signal pro-cessing,” Med. Eng. Phys., vol. 20, pp. 211–215, 1998.

[8] J. Nilsson, M. Panizza, and P. Arieti, “Computer-aided deter-mination of the silent period,” J. Clin. Neurophysiol, vol. 14,pp. 136–143, 1997.

[9] R. D. Rafal, A. Winhoff, J. H. Friedman, and E. Bernstein, “Pro-gramming and execution of sequential movements in Parkin-son’s disease,” J. Neurol. Neurosurg. Psychiatry, vol. 50, pp.1267–1273, 1987.

[10] K. Takada and K. Yashiro, “Automatic measurement of on/offperiods of EMG activity,” Medinfo, vol. 8, pp. 751–754, 1995.

[11] G. J. M. van Boxtel, L. H. D. Geerats, M. M. C. Van den Berg-Lenssen, and C. H. M. Brunia, “Detection of EMG onset inERP research,” Psychophysiology, vol. 30, pp. 405–412, 1993.

[12] G. Staude and W. Wolf, “Objective motor response onset de-tection in surface myoelectric signals,” Med. Eng. Phys., vol. 21,pp. 449–467, 1999.

[13] P. W. Hodges and B. H. Bui, “A comparison of computer-basedmethods for determination of onset of muscle contraction us-ing electromyography,” Electroenceph. Clin. Neurophysiol., vol.101, pp. 511–519, 1996.

[14] P. Bonato, T. D’Alessio, and M. Knaflitz, “A statistical methodfor the measurement of muscle activation intervals from sur-face myoelectric signal during gait,” IEEE Trans. Biomed. Eng.,vol. 45, pp. 287–298, 1998.

[15] M. Lidierth, “A computer based method for automated mea-surement of the periods of muscular activity from an EMGand its application to locomotor EMGs,” Electroenceph. Clin.Neurophysiol., vol. 64, pp. 378–380, 1986.

Onset detection methods 81

[16] G. Staude, R. Dengler, and W. Wolf, “The discontinuous natureof motor execution part I: a model concept for single-musclemultiple-task coordination,” Biol. Cybern., vol. 82, pp. 23–33,2000.

[17] L. Ljung, System Identification: Theory for the User, Prentice-Hall, New Jersey, 1987.

[18] G. Barrett, H. Shibasaki, and R. Neshige, “A computer-assistedmethod for averaging movement-related cortical potentialswith respect to EMG onset,” Electroenceph. Clin. Neurophysiol.,vol. 60, pp. 276–281, 1985.

[19] H. G. Choi, J. C. Principe, A. A. Hutchison, and J. A. Woz-niak, “Multiresolution segmentation of respiratory elec-tromyographic signals,” IEEE Trans. Biomed. Eng., vol. 41, pp.257–266, 1994.

[20] G. Staude, W. Wolf, and U. Appel, “Automatic event detectionin surface EMG of rhythmically activated muscles,” An Int.Conf. IEEE BME Soc., vol. 17, pp. 1351–1352, 1995.

[21] G. Staude and W. Wolf, “Voluntary motor reactions: doesstimulus appearance prolong the actual tremor period?,” J.Electromyogr. Kinesiol., vol. 9, pp. 277–281, 1999.

[22] G. Staude, V. Kafka, and W. Wolf, “Determination of premotorsilent periods from surface myoelectric signals,” Biomed. Tech.,vol. 45, no. 2, pp. 228–232, 2000.

[23] M. Khalil and J. Duchêne, “Uterine EMG analysis: a dynamicapproach for change detection and classification,” IEEE Trans.Biomed. Eng., vol. 47, pp. 748–756, 2000.

[24] D. E. Knuth, The Art of Computer Programming, Addison-Wesley, Massachusetts, 1998.

[25] W. West, T. Ogden, and A. Rossini, “Statistical tools on theWWW,” The American Statistician, vol. 52, pp. 257–262, 1998.

Gerhard Staude received the M.S.E.E. in1990 from the Technical University Munich,Germany, and the Ph.D. degree in electricalengineering in 1996 from the University ofthe Armed Forces at Munich, Germany. From1996 to 1997 he was a member of the interdis-ciplinary graduate college “Sensory Interac-tion in Biological and Technical Systems” ofthe Deutsche Forschungsgemeinschaft. Cur-rently, he is affiliated with the Department ofElectrical Engineering at the University of the Armed Forces, wherehe heads the Neuromuscular system research group at the Instituteof Mathematics and Computer Science. He is author of more than30 refereed journal papers in the fields of digital signal processingand analysis of the human motor system. His major research inter-ests include parameter estimation and change detection in motor-related biosignals, and basic research in the human motor systemwith special focus on biological oscillators (e.g., physiological andpathological tremors) and movement coordination.

Claus Flachenecker received his M.S.E.Ein 1994. He continued as a Ph.D. studentat BMW AG, Munich, in cooperation withthe Technical University Munich, Lehrstuhlfür Elektrische Meßtechnik. This work com-prised intense cooperation with the Euro-pean STAUMECS and German ASAM stan-dardization groups, that define plug & playintegration standards for product testing en-vironments. After receiving his Ph.D. degreein 1998, he was affiliated with the University of the Armed Forces atMunich to research biosignal related topics, set-up a driving simula-tor environment with incorporated physiological driver monitoring

and create intelligent force-feedback controllers to be used in force-feedback pedals for automobiles. Since end of 2000, he is working forCarl Zeiss Ophthalmic, Inc. (California, USA), as a systems designengineer. His responsibilities now include the design and develop-ment of diagnostic medical devices for the eye-care industry.

Martin Daumer received the diploma inPhysics in 1990 and the Ph.D. in mathe-matics both from the Ludwig-Maximilians-University Munich, Germany in 1995. In1987 he was a summer student at CERN,Geneva, and in 1992 he spent one year witha stipend of the DAAD at the Departmentfor Mathematics at the Rutgers University,New Jersey. From 1993–1996 he was a mem-ber and later post-doc of the interdisciplinarygraduate college “Mathematics in its relationship with Physics” fromthe Deutsche Forschungsgemeinschaft DFG. Since 1996 he is head-ing the Research group “Online Monitoring in Medicine” in cooper-ation with the Sonderforschungsbereich 386 “Statistical Analysis ofDiscrete Structures” at the Institute for Medical Statistics and Epi-demiology of the Technical University of Munich. He has publishedmore than 40 articles about topics in quantum physics, scatteringtheory, probability theory, biosignal processing and holds patentsfor methods and devices for change-point detection and internet-based monitoring of biosignals. In 1999 he founded the IT-companyTrium Analysis Online, which is playing a central role in various na-tional (BMBF) and international research projects, for example, theSylvia Lawry Centre for Multiple Sclerosis Research.

Werner Wolf received his M.S.E.E. in 1970and his Ph.D. in 1978, both from the Tech-nical University Munich, Germany. Now, heis a member of the Faculty of Electrical En-gineering of the University of the ArmedForces at Munich. From 1970 to 1978, hewas engaged in an interdisciplinary researchproject “Visual Perception” of the Institute ofCommunication Systems, Technical Univer-sity Munich, with special focus on the func-tional role of saccadic eye movements in vision. In 1978, he movedto the Institute of Mathematics and Computer Science now head-ing the Biosignal Processing Lab. His main interests are processingof biosignals (e.g., EP, EEG, EMG) and include basic research insensorimotor systems (eye movements, hand and arm movements,etc.). Much of his research concerns pathophysiological aspects andis performed in cooperation with clinical departments. He is authorof more than 100 scientific publications, and a member of the IEEESociety of Engineering in Medicine and Biology, the IEEE Society ofSignal Processing, the American Society of Neuroscience, and severalnational engineering societies.

Photograph © Turisme de Barcelona / J. Trullàs

Preliminary call for papers

The 2011 European Signal Processing Conference (EUSIPCO 2011) is thenineteenth in a series of conferences promoted by the European Association forSignal Processing (EURASIP, www.eurasip.org). This year edition will take placein Barcelona, capital city of Catalonia (Spain), and will be jointly organized by theCentre Tecnològic de Telecomunicacions de Catalunya (CTTC) and theUniversitat Politècnica de Catalunya (UPC).EUSIPCO 2011 will focus on key aspects of signal processing theory and

li ti li t d b l A t f b i i ill b b d lit

Organizing Committee

Honorary ChairMiguel A. Lagunas (CTTC)

General ChairAna I. Pérez Neira (UPC)

General Vice ChairCarles Antón Haro (CTTC)

Technical Program ChairXavier Mestre (CTTC)

Technical Program Co Chairsapplications as listed below. Acceptance of submissions will be based on quality,relevance and originality. Accepted papers will be published in the EUSIPCOproceedings and presented during the conference. Paper submissions, proposalsfor tutorials and proposals for special sessions are invited in, but not limited to,the following areas of interest.

Areas of Interest

• Audio and electro acoustics.• Design, implementation, and applications of signal processing systems.

l d l d d

Technical Program Co ChairsJavier Hernando (UPC)Montserrat Pardàs (UPC)

Plenary TalksFerran Marqués (UPC)Yonina Eldar (Technion)

Special SessionsIgnacio Santamaría (Unversidadde Cantabria)Mats Bengtsson (KTH)

FinancesMontserrat Nájar (UPC)• Multimedia signal processing and coding.

• Image and multidimensional signal processing.• Signal detection and estimation.• Sensor array and multi channel signal processing.• Sensor fusion in networked systems.• Signal processing for communications.• Medical imaging and image analysis.• Non stationary, non linear and non Gaussian signal processing.

Submissions

Montserrat Nájar (UPC)

TutorialsDaniel P. Palomar(Hong Kong UST)Beatrice Pesquet Popescu (ENST)

PublicityStephan Pfletschinger (CTTC)Mònica Navarro (CTTC)

PublicationsAntonio Pascual (UPC)Carles Fernández (CTTC)

I d i l Li i & E hibiSubmissions

Procedures to submit a paper and proposals for special sessions and tutorials willbe detailed at www.eusipco2011.org. Submitted papers must be camera ready, nomore than 5 pages long, and conforming to the standard specified on theEUSIPCO 2011 web site. First authors who are registered students can participatein the best student paper competition.

Important Deadlines:

P l f i l i 15 D 2010

Industrial Liaison & ExhibitsAngeliki Alexiou(University of Piraeus)Albert Sitjà (CTTC)

International LiaisonJu Liu (Shandong University China)Jinhong Yuan (UNSW Australia)Tamas Sziranyi (SZTAKI Hungary)Rich Stern (CMU USA)Ricardo L. de Queiroz (UNB Brazil)

Webpage: www.eusipco2011.org

Proposals for special sessions 15 Dec 2010Proposals for tutorials 18 Feb 2011Electronic submission of full papers 21 Feb 2011Notification of acceptance 23 May 2011Submission of camera ready papers 6 Jun 2011