Computers in Biology and Medicine 38 (2008) 837 844

www.intl.elsevierhealth.com/journals/cobm

A new method of simulating surface electromyograms using probability

density functions
I. Ashida , S. Kawakami, Y. Miyaoka
Department of Health and Nutrition, Niigata University of Health and Welfare, School of Health Sciences, 1398 Shimami-cho, Niigata 950-3198, Japan

Received 21 November 2006; accepted 5 May 2008

Abstract
A new method based on probability density functions (inverse Gaussian distributions) to simulate surface electromyograms (sEMGs) was
developed for the specic use of the TP technique, which discriminates sEMG activity patterns. First, four prototypes with different activity
patterns were generated from inverse Gaussian distributions by changing their geometrical parameters, which were derived from actual recorded
sEMGs. Then, four simulated sEMGs were produced on the basis of the prototypes. The validity of the simulation method in relation to the
TP technique was statistically examined and veried. The new simulation method will be useful.
2008 Elsevier Ltd. All rights reserved.

Keywords: Surface electromyogram; Computer simulation; Integration; Probability density function; Activity pattern analysis

1. Introduction In addition, analysis of actual sEMGs often begins with in-

tegration of the original action potentials because integration
Skeletal muscle bers generate multiple action potentials of sEMGs with appropriate time constants is important for
in association with movements that can be measured by sur- measuring various parameters (e.g., duration, mean amplitude,
face electromyograms (sEMGs) [1,2], which are recorded peak amplitude, and integrated area under the curve [35]).
using electrodes placed on the skin and mucosa. sEMGs are Integrated sEMGs can also be used to interpret the activity pat-
commonly employed in basic science research (e.g., physiol- terns of the target muscles, but interpretation has been mostly
ogy and kinesiology) and are often used in clinical settings. conducted by visual observation only. We recently developed
Advanced computers and computer software have facilitated a TP technique for the quantitative evaluation of activity
sEMG simulation research. patterns observed in integrated sEMGs [7]. The TP technique
Sophisticated computerized simulation methods have been is briey summarized in the following steps: (1) each nal
developed [36]. sEMGs that have been simulated using many cumulative value of an EMG is divided into 10 or 20 equal
available methods generally compile individual action poten- sections (10%, 20%, . . . , 100% or 5%, 10%, . . . , 100% of the
tials as they occur during biological events [6]. These methods, nal value); (2) the 10 or 20 sectioned values are allocated
which were based on simulated action potentials, appear to be serially to a standardized time scale; and (3) the allocated 10
suitable for precise computer analysis of biological processes or 20 points are designated on the standardized time scale
(e.g., activation of individual action potentials from muscle as TP (T10 , T20 , . . . , T100 or T5 , T10 , . . . , T100 ). We used this
bers, volume conduction in muscular tissues, and surface elec- technique to accurately discriminate sEMG activity patterns
trode recording). Because many parameters are involved and that were recorded during different chewing and swallowing be-
proprietary software is required [36], the conventional meth- haviors [79]. In a previous paper [7], e.g., we discriminated the
ods tend to be cumbersome for simulating sEMGs. activity patterns of the suprahyoid muscles during swallowing
of tasted and tasteless foods by the TP technique. Our previous
Corresponding author. Tel./fax: +81 25 257 4702. results generated using this technique [79] suggest that the TP
E-mail address: [email protected] (I. Ashida). technique is useful in evaluating actual patterns of sEMGs, but
0010-4825/$ - see front matter 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compbiomed.2008.05.001
838 I. Ashida et al. / Computers in Biology and Medicine 38 (2008) 837 844

the theoretical basis of the technique has not yet been estab- converted to discrete variables by digital sampling; a set of 200
lished. In addition, we need a simulation method that can be digital data points was sampled from the standardized distri-
modied to the form of distributions. Because probability dis- butions with equal intervals of 5 103 s (Fig. 1c). The data
tribution functions may cause characteristic forms by changing points sampled (si , i = 1, 2, . . . , 200; Fig. 1d) served as data
values of the function used (e.g., [10]), in the present study we sources for the three prototypes.
focused on the properties of probability distribution functions.
In the research eld of electromyography, Stashuk and Naphan 2.1.3. simuEMGs amplitudes
[11] applied the probability distribution functions to indexes to Amplitudes of the sampled data points were determined by
classify motor unit action potentials in simulated signals, and assigning random numbers to individual data points (ai , i = 1,
Hamid Nawab et al. [12] improved resolutions in decompos- 2, . . . , 200; Fig. 1d). The random numbers were obtained from
ing sEMGs by modifying them into constituent motor unit ac- a normal distribution with a mean of 0.0 and SDs of si in
tion potential trains using the probability distribution functions. the exchanged distributions. The distribution of the random
However, application of the probability distribution functions numbers resulted in negative, as well as, positive amplitude
to simulations of sEMGs has not been reported. In this study, values for the sampled data points (ai ).
we aimed to establish a simulation method specic for sEMGs All of the data points with determined amplitudes were plot-
to elucidate the theoretical basis for the TP technique. ted on a scattergram (Fig. 1e), and nally, the data points were
connected consecutively with lines to complete simuEMGs
2. Methods (Fig. 1f).
2.1. Production of sEMG data
2.2. simuEMGs data analysis
We produced simulated sEMGs through the following three
steps (also see Table 1): (1) prototypes for simulated elec- The simuEMGs derived from the three probability den-
tromyograms (simuEMGs) were rst dened; (2) data sources sity functions of the inverse Gaussian distribution were in-
were then sampled from the prototypes; and (3) amplitudes for tegrated with a time constant of 0.05 s (int.simuEMGs). The
simuEMGs were calculated and allocated to the individual data simuEMGs were compared to the prototypes for degree of
sources sampled. tness (degt ) with Pearsons correlation coefcients between
the int.simuEMGs and the corresponding prototypes. The
2.1.1. Prototypes for simuEMGs procedure to determine amplitudes (see Section 2.1.3) was
Three kinds of statistical distributions derived from probabil- iterated in the three prototypes individually, and 11 stages
ity density functions of the inverse Gaussian distribution [10] ranging from 5 to 100 were selected as the number of iter-
ations (Ni , i = 5, 10, 20, . . . , 100). The means and standard
f (x; , ) = (/2)1/2 x 2/3 exp((x )2 /22 x),
errors of means (S.E.M.s) of the degt were calculated at each
x > 0, (1) iteration stage. Areas under of the curve of the int.simuEMGs
were used as prototypes for simuEMGs with three activity and the nal values of cumulative simuEMGs were adjusted to
patterns. We used three prototypes (proA , proB , and proC ) equate the three prototypes: proA ; proB ; and proC . Then, the
of simuEMG to illustrate our simulation method and three TP technique [79] was applied to the individual simuEMGs
sets of  and , corresponding to the geometrical param- to test whether activity patterns of the three prototypes were
eters of the inverse Gaussian function, for our prototypes: discriminated by the technique.
{0.3, 3.0}, {0.5, 5.0}, and {0.7, 7.0}, respectively. The proce- The TP technique was developed to quantitatively evaluate
dures to produce the three simuEMGs are based on the rst sEMG activity patterns. The technique is outlined as (1) divi-
prototype (Fig. 1), which is similar to the other two prototypes. sion of the nal cumulative value of an sEMG equally into 10
Because values on the horizontal axes (x-values) of these or 20 sections; (2) allocation of the 10 or 20 sectioned values to
distributions were mathematically innite (Fig. 1a), the a standardized time scale serially; and (3) designation of the al-
x-values were exchanged to nite values for simulation. After located values as TP (T10 , T20 , . . . , T100 or T5 , T10 , . . . , T100 ).
the exchange, the x-values of each distribution were standard- According to this technique, each TP value designates a relative
ized from 0.0 to 1.0 to compare simuEMGs based on dif- location of the EMG activity on a standardized time scale, i.e.,
ferent prototypes. Values on the vertical axes (y-values) were a T50 value indicates the standardized time for half of the nal
adjusted from 0.0 to the maximum to compare distributions cumulative EMG activity, and a smaller TP value suggests that
with different magnitudes (Fig. 1b). the area under the curve of the integrated EMG activity shows
a negatively skewed (skewed to right) distribution.
2.1.2. simuEMG data sources
sEMGs that were actually recorded from electrodes on the 2.3. Production of numerical data
skin of human subjects were composed of discrete variables
due to the all-or-none law of action potentials arising from 2.3.1. Actual recorded data
individual muscle bers. In contrast, probability density func- Actual recorded sEMG (actualEMG) data were used
tions consisted of continuous variables, according to their def- to demonstrate the usefulness of our simulation method.
inition (Eq. (1)). Consequently, the continuous variables were The actualEMGs were obtained from the anterior tongue
 I. Ashida et al. / Computers in Biology and Medicine 38 (2008) 837 844 839

Table 1
Processes of articially simulated and actually recorded surface EMGs

Process Procedure to simulate surface EMGs Events occurring during surface EMG recording

1 Selection of PDF models (Figs. 1a and 2a) Generation of muscle action potentials
2 Standardization of both x- and y-values (Fig. 1b) Conduction to surface electrodes
3 Digital sampling from PDF models and conversion from continuous to discrete variables (Fig. 1c) Amplication of surface EMG signals
4 Determination of amplitudes of data points (Fig. 1d and e) Integration of amplied EMG signals
5 Connection of serial data points (Figs. 1f and 2b) Evaluation of patterns of EMGs

Fig. 1. Electromyogram (EMG) simulations based on probability density functions. (a) A probability density function (the inverse Gaussian distribution with
geometrical parameters of {0.3 and 3.0}) was selected as a simulation prototype. (b) The values on both the time (x) and amplitude (y) axes of the prototype
were standardized as 1.0. (c) Sampling of data points from the standardized prototype at a rate of 200 Hz. (d) Amplitudes were assigned to the sampled data
points (e.g., the amplitude of the 22nd data point (a22 ) was assigned from numbers of a normal distribution (mean = 0.0, standard deviation (s22 ) = 0.17)).
(e) A scattergram was produced with the amplitude-assigned data points. (f) A simulated EMG was completed by consecutively connecting points on the
scattergram. See text for details.

of healthy young adults to determine whether body pos- amplied (AD Instruments-Japan, Nagoya, Japan), ltered
ture altered the duration of the oral phase of swallowing (wide bandpass, 105000 Hz), fully rectied, and integrated
[13]. Briey, an sEMG of the anterior tongue was recorded (time constant, 0.03 s). A total of 216 experimental trials
by small surface electrodes. The electrodes were located (2 test foods 9 subjects 4 body positions 3 sessions)
20 mm posterior to the tongue tip. The sEMG signals were was conducted as in the previous study [13], and 100 records
840 I. Ashida et al. / Computers in Biology and Medicine 38 (2008) 837 844

of swallowing were randomly extracted from the 216 trials to 2nd, and 3rd models of int.simuEMGs derived from the three
use as the actualEMGs. Each of the actualEMG records was prototypes (proA , proB , and proC ) showed characteristic shapes
classied into one of four groups according to its T50 value (i.e., negatively skewed (skewed to right, proA ), symmetrical
[7]. Means and SDs of the three sEMG parameters (peak am- (proB ), and positively skewed (skewed to left, proC ) distri-
plitude, time to peak, and cumulative activity) were calculated butions, respectively). Table 2 summarizes peak values, stan-
in each group and used to compare geometrical properties of dardized peak time, and TP values calculated from the three
the actualEMGs. int.simuEMGs models. The KruskalWallis test revealed sig-
nicant differences in the standardized peak time among the
2.3.2. Simulation data based on the actual recorded data 1st, 2nd, and 3rd models (P < 0.01), but not in the peak values.
The simuEMGs were generated using four prototypes with TP values differed greatly among the three models (Table 2);
different geometrical parameters, which were determined by e.g., the median T20 value in the 1st model was less than one-
the means of the time to peak in the four groups of actualEMGs. third of the 2nd model, and was less than one-fth of the 3rd
The simuEMGs were integrated with a time constant of 0.05 s model. The KruskalWallis test detected signicant differences
(int.simuEMGs). The simuEMGs were compared to the ac- in TP values among the proA , proB , and proC (P < 0.1).
tualEMGs with respect to time to peak, cumulative activity, and Fig. 3 shows changes in means (a) and S.E.M.s (b) of the
T50 value. degt when Ni s were increased from 5 to 100. The means var-
ied slightly among the three simuEMGs but were basically con-
2.3.3. Statistical examinations sistent through the iterations. One-way ANOVA found no sta-
The following four kinds of statistical examinations were tistically signicant differences in the means at the nal stage
used: (1) The Wilcoxons test was used to examine differences of iteration among the proA , proB , and proC . The Wilcoxons
in the degt from the rst (Ni =5) and nal stages (Ni =100). (2) tests also found no differences between the degt with N5 and
One-way analysis of variance (ANOVA) was used to examine N100 . In contrast, the S.E.M.s of the degt decreased monoton-
differences in the degt among the three int.simuEMGs at the ically with increase of Ni in all of the three simuEMGs (e.g.,
nal stage of iteration (Ni = 100). (3) Trend analysis with the S.E.M.s of the simuEMG of proB fell from 2.1 102
regression coefcients was used to examine whether or not to 0.3 102 ; Fig. 3b). Trend analysis using regression
S.E.M.s of the degt regressed signicantly against numbers
of iterations. Finally, (4) the KruskalWallis test was used to Table 2
examine TP value differences among the three models. Characteristic values of the three simulation models
The level of statistical signicance was set at P < 0.05 in all Model Area Cumulative TP values
statistical examinations.
T20 T50 T80
3. Results 1st 0.36 54.1 0.11 0.22 0.39
2nd 0.36 54.1 0.38 0.50 0.62
3.1. Simulation data based on the inverse Gaussian functions 3rd 0.36 54.1 0.61 0.78 0.89

Model: simulation models (see Fig. 2); Area: area calculated from integrated
Fig. 2 depicts sample data of three different standardized simuEMG; Cumulative: nal cumulative values of simuEMG; T20 , T50 , and
prototypes (a), simuEMGs (b), and int.simuEMGs (c). The 1st, T80 , TP values at 2nd, 5th, and 8th deciles, respectively. See text for details.

Fig. 2. Three prototypes with their simulated electromyograms (EMGs) and integrated EMGs. (a) Three types of inverse Gaussian distributions used as
prototypes (proA , proB , and proC ) for EMG simulations. (b) Simulated EMGs based on the three prototypes in (a). (c) Integrated curves of the simulated
EMGs in b. Note that peaks of the integrated EMGs are located in the initial (proA ), middle (proB ), and late (proC ) periods on the time scale.
 I. Ashida et al. / Computers in Biology and Medicine 38 (2008) 837 844 841

Fig. 3. Degree of tness and number of iterations in three simulation prototypes. (a) Changes in means of the degree of tness under 11 stages of iterations
(ranging from 5 to 100) in three simulation prototypes. (b) Changes in standard errors of means (S.E.M.s) of the degree of tness during the iterations. Note
that the S.E.M.s decrease with increasing the number of iterations, but the means do not change noticeably.

were signicantly increased (P < 0.05) caused by increased

T50 values. The areas between the curve of int.actualEMGs
and the x-axis in the four groups were 86.9 21.7 (rG1 ),
81.0 18.4 (rG2 ), 90.3 22.6 (rG3 ), and 90.4 16.8 (rG4 ),
respectively, and there were no signicant differences in the
areas under the curve among the four groups.

3.3. Simulated data based on the actual recorded data

Similar to the actual data, Fig. 5 Shows the standardized

peak time (a), the areas under the curve (b), and the T50
values (c) of four int.simuEMGs (sG1 , sG2 , sG3 , and sG4 )
with the geometrical parameters derived from the four en-
velopes of int.actualEMGs. Means and SDs of the standardized
peak time in individual data were 0.28 0.06 (sG1 ), 0.51
Fig. 4. Area under the curve of the integrated electromyograms (EMGs) in 0.08 (sG2 ), 0.62 0.06 (sG3 ), and 0.75 0.06 (sG4 ), re-
actual recorded data. Enveloped curves of integrated EMGs were produced spectively, and all four mean peak times differed signicantly
from the actual recorded data of four individual groups (rG1 , rG2 , rG3 , and from each other (P s < 0.01). The four areas calculated
rG4 ) classied by T50 values. Arrow heads indicate the individual peak time were 1.9 0.7 (sG1 ), 2.1 0.6 (sG2 ), 2.4 0.7 (sG3 ), and
(standardized time) of the four areas under the curves. See text for details.
2.3 0.6 (sG4 ), and the areas were not different among the
groups. The T50 values calculated in four int.simuEMGs were
0.32 0.01 (sG1 ), 0.39 0.01 (sG2 ), 0.50 0.01 (sG3 ), and
coefcients revealed that the S.E.M.s regressed signicantly 0.66 0.01 (sG4 ), respectively All four T50 values differed
against Ni s in all of the three simuEMGs (P s < 0.01), sug- signicantly from each other (P s < 0.01).
gesting that the S.E.M.s of the degt decreased with increasing
Ni s of the three simuEMGs. 4. Discussion

3.2. Actual recorded data We aimed to establish a simulation method specic for
sEMGs to elucidate the theoretical basis for the TP tech-
Fig. 4 depicts the four areas under the curve for the nique that we developed to analyze sEMG activity patterns.
int.actualEMGs for the four groups (rG1 , rG2 , rG3 , and The present study used the inverse Gaussian distributions of
rG4 ) classied by their T50 values. Means and SDs of probability density functions to simulate sEMGs (Figs. 1 and
the standardized peak times in individual groups were 2). This simulation method is advantageous over other meth-
0.34 0.12 (rG1 ), 0.39 0.18 (rG2 ), 0.51 0.18 (rG3 ), ods, rst and foremost because of its simplicity and ease of
and 0.65 0.20 (rG4 ), respectively, and the peak time means use. As summarized in Table 1, the method relies on only six
842 I. Ashida et al. / Computers in Biology and Medicine 38 (2008) 837 844

Fig. 5. Peak time, areas, and T50 values for the simulation data. (a) Shown are the standardized peak time of four kinds of simulation data (sG1 , sG2 , sG3 ,
and sG4 ) with the four sets of geometrical parameters derived from actual recorded data. (b) Shown are areas encompassed by the curves (see Fig. 4) and
x-axis of the simulation data. (c) Shown are T50 values to evaluate simulation data activity patterns. Values indicate means plus SD. Horizontal segments in
(a) and (c) denote pairs of bars with signicant differences ( P < 0.01) among the four simulation data sets. See text for details.

processes, each consisting of a few procedures (see Section 2). prototypes. Probability density functions for prototypes require
Conventional sEMG simulation methods based on physiologic exibility to alter the shapes of probability density function
models are more complicated due to the assembly of articial curves. The inverse Gaussian functions we used fulll the
action potentials of single muscle bers [6]. Specically, an requirement by changing only two parameters (Fig. 2a). The
average of 17 parameters are employed to simulate EMGs by changes actually resulted in three distributions with different
assembling articial action potentials. In contrast, our method distribution skewness (positively, symmetrical, and negatively
uses only four parameters (geometrical parameters for inverse skewed) and kurtosis (platykurtic, normal, and leptokurtic)
Gaussian functions, number of digital sampling points, sources (Fig. 4). Other candidates for simulations, such as Gaussian,
for amplitude assignment, and number of iterations) to sim- Chi-squared, and Gamma distributions, had the following dis-
ulate sEMGs. In addition, popular personal computers (e.g., advantages: (1) the Gaussian distributions showed symmetrical
Microsoft Windows XP operating system) and common soft- shapes only and (2) the Chi-squared and Gamma distributions
ware (e.g., Microsoft Excel ) are sufcient to perform calcu- showed negatively skewed shapes only. Thus, we concluded
lations with our method; in contrast, other simulation systems that the inverse Gaussian functions are the most suitable pro-
for biological signals require highly specic computers and/or totypes for the simulation method that we sought to develop in
proprietary software (e.g., [14]). Consequently, we concluded this study. The geometrical parameters that we derived from ac-
that the simulation method using probability density statistical tualEMGs gave characteristic shapes to the prototypes (Fig. 2),
functions can be easily applied to produce sEMG simulation and the TP technique clearly detected shape differences among
data. the produced simuEMGs (Table 2 and Fig. 5c). The present
We achieved the original aim (i.e., establishment of a sim- results strongly suggest that the method we used accurately
ulation method to analyze sEMG activity patterns) with this simulates integrated sEMG curves to analyze their activity
easy method using probability density function models as patterns.
 I. Ashida et al. / Computers in Biology and Medicine 38 (2008) 837 844 843

We used T50 values to classify actualEMGs in the present 5. Summary

study (Fig. 4). There was no special reason for using T50 values,
and other TP values, such as T20 and T80 , were also available. We recently reported an analytical tool, TP technique, to
The median T50 values ranged from 0.22 (1st model) to 0.78 discriminate surface electromyogram (sEMG) activity patterns.
(3rd model) and the range calculated was 0.56 (Table 2), while In the present study, we aimed to develop a new method to sim-
the median T20 and T80 values ranged from 0.11 (1st model) to ulate sEMGs by using probability density functions to establish
0.61 (3rd model) and from 0.39 (1st model) to 0.89 (3rd model), the theoretical basis of the TP technique. We produced simu-
respectively, and the ranges calculated were 0.50 for the two TP lated EMGs with different activity patterns by the following six
values (Table 2). This comparison suggests that the selection of procedures: (1) generation of prototypes for simulated EMGs
TP values does not affect the classication of actualEMGs. We by using probability density functions of the inverse Gaus-
adopted the standardized peak time, which was calculated in sian distributions with different sets of geometrical parameters;
the four actualEMG groups, as the determinant of the geometri- (2) standardization of both values on the time (x) and amplitude
cal parameters for the inverse Gaussian functions (Fig. 5). The (y) axes to 1.0 in the individual prototypes; (3) sampling of
adopted determinant, corresponding to the standardized peak digital data points to generate simulated EMGs from the stan-
time, appears to closely reect the impression based on visual dardized prototypes at a rate of 200 Hz; (4) assignment of am-
observation about the similarity of shapes between the proto- plitudes of simulated EMGs to individual data points sampled
types and integrative actualEMGs curves (i.e., int.actualEMGs; from normally distributed (with a mean of 0.0 and a variance of
Fig. 4). However, other choices for the determinants of the ge- the sampled values from the prototype used) random numbers;
ometrical parameters are available (e.g., maximum amplitudes (5) production of a scattergram for each prototype using the as-
of int.actualEMGs, durations of muscle activities, and angles signed amplitude data points; and (6) completion of simulated
of the peak of the integrative curves). Evaluating the inuences EMGs by connecting data points consecutively on the scatter-
of changes from the peak time to other kinds of determinants grams. We statistically examined the simulated EMGs to eval-
on geometrical parameters for the inverse Gaussian functions uate the validity of the simulation method in relation to the TP
will be necessary in the future. technique by analysis of the concordances between integrated
We further need to consider at least three technical prob- curves of EMGs simulated and prototypes used. We veried
lems: (1) digital sampling; (2) number of sampling points; and that all of the simulation data showed statistically signicant
(3) number of iterations. First, we sampled digital data points concordances (P s < 0.01) with the corresponding prototypes.
(Fig. 1c) from the standardized distributions to determine the In addition, results of the KruskalWallis test on the T20 , T50 ,
amplitudes of individual data points. Digital sampling func- and T80 values among the three simulation data sets revealed the
tionally corresponds to the differential of probability density TP values clearly discriminated differences in activity patterns
functions with variables denoting time or subdivided contin- of the three prototypes (i.e., simulated sEMGs). We also ex-
uous functions. Additionally, digital sampling by our method amined the consistency between the produced simulation data
converted continuous variables to discrete variables. Second, and EMGs that were actually recorded. We categorized 100
we selected a xed number of 200 as the number of digital sEMGs (recorded from the anterior tongue during swallowing
sampling points in each prototype after performing preliminary in healthy humans) into four groups with T50 values, and we
tests (see Section 2.1.2). The preliminary tests showed that the calculated the area under the curves of the integrated EMGs
degree of tness at the nal stage of iteration in the three pro- for the four groups. Four simulation data sets were constructed
totypes (i.e., proA , proB , and proC ; Fig. 2) were 0.80 at according to the geometrical parameters derived from the four
100 points, 0.90 at 200 points, and 0.95 at 400 points, curves. Statistically signicant differences (P s < 0.01) were
respectively. Larger digital sampling points increase the degt shown in peak time and T50 values among the four simulation
but require more computation time and labor. The decision to data sets, suggesting that the simulation data reect the orig-
use 200 digital sampling points was based on consideration of inal shapes of the actual recorded EMGs. In addition to the
the advantage (i.e., an increase in the degt ) and disadvantages results that demonstrate the validity of the simulation method,
(i.e., more time and labor for computation) of such higher sam- we discussed the simplicity of the method and its easy applica-
pling numbers as 400 and 800 points. Third, we used 11 stages tion with inexpensive computing systems. The four geometrical
of iterations, ranging between 5 and 100, to examine changes parameters used in our method were much less cumbersome
in means and S.E.M.s of the degt (Fig. 3). An increase in Ni than previously reported methods, in which approximately 17
hardly changed the means of the three models (Fig. 3a) but parameters were used. Furthermore, commonly available soft-
reduced the S.E.M.s monotonically (Fig. 3b). Here, we also ware (e.g., MicrosoftExcel ) was sufcient to calculate the
weighed the advantage (i.e., reduction in S.E.M.s) and disad- simulation data. In conclusion, the simple and easy method we
vantages (i.e., more computation time and labor) and concluded reported in the present study provides the theoretical basis for
that 50 iterations were probably optimal for the simulations in the TP technique for analysis of sEMG activity patterns.
the present study.
In conclusion, we developed a simulation method based on Acknowledgments
probability density curves with inverse Gaussian functions to
establish a theoretical basis to analyze activity patterns in sur- This study was supported in part by Grants-in-Aid for
face electromyograms using the TP technique. Younger Researchers (to I.A.) and for Scientic Research
844 I. Ashida et al. / Computers in Biology and Medicine 38 (2008) 837 844

(to Y.M.) from the Niigata University of Health and Welfare chewing of agars in healthy young males, J. Texture Stud. 38 (2007)
and by a Grant-in-Aid for Scientic Research (No. 19500667 8799.
to Y.M.) from the Ministry of Education, Science, and Culture [9] Y. Miyaoka, I. Ashida, S. Kawakami, S. Miyaoka, Differentiation of
activity patterns in the suprahyoid muscles during swallowing of umami
of Japan. tasting foods, J. Sensory Stud. 21 (2006) 283572.
[10] M.C.K. Tweedie, Statistical properties of inverse Gaussian distributions
Conict of interest statement I, Ann. Math. Stat. 28 (1957) 362377.
[11] D.W. Stashuk, R.K. Naphan, Probabilistic inference-based classication
We, Ichiro Ashida, Shin-ya Kawakami, and Yozo Miyaoka, applied to myoelectric signal decomposition, IEEE Trans. Biomed. Eng.
39 (1992) 346355.
declare that we have no any nancial and personal relation- [12] S. Hamid Nawab, R. Wotiz, C.J. De Luca, Improved resolution of pulse
ships with other people or organizations that could inappro- superpositions in a knowledge-based system EMG decomposition, in:
priately inuence our manuscript entitled, A New Method for Conference of the Proceedings of IEEE Engineering in Medicine and
Simulation of Surface Electromyograms Using Probability Den- Biology Society, vol. 1, 2004, pp. 6971.
sity Functions. [13] D. Inagaki, Y. Miyaoka, I. Ashida, K. Ueda, Y. Yamada, Inuences of
body posture on duration of oral swallowing in normal young adults, J.
Oral Rehabil. 34 (2007) 414421.
References
[14] L. Karlsson, B. Hammarberg, E. Stalberg, An application of a muscle
model to study electromyographic signals, Comput. Methods Programs
[1] G.D. Klasser, J.P. Okeson, The clinical usefulness of surface Biomed. 71 (2003) 225233.
electromyography in the diagnosis and treatment of temporomandibular
disorders, J. Am. Dent. Assoc. 137 (2006) 763771.
[2] A.D. Taylor, R. Bronks, A.L. Bryant, The relationship between
Ichiro Ashida is a Lecturer at Niigata University of Health and Welfare in
electromyography and work intensity revisited: a brief review with the Department of Health and Nutrition. He was awarded his D.Agr. from
references to lacticacidosis and hyperammonia, Electromyogr. Clin. Niigata University Graduate School of Science and Technology, Japan, in
Neurophysiol. 37 (1997) 387398. 1999. He was an Assistant Professor at Niigata University in the Department
[3] F. Buchthal, Z. Kamieniecka, The diagnostic yield of quantied of Agriculture. His main research interest is analytical physiology and its
electromyography and quantied muscle biopsy in neuromuscular practical application to evaluation of foods by analyzing of muscle activity
disorders, Muscle Nerve 5 (1982) 265280. pattern.
[4] V.F. Ferrario, C. Sforza, A. Colombo, V. Ciusa, An electromyographic
investigation of masticatory muscles symmetry in normo-occlusion Shin-Ya Kawakami is an Assistant Professor at Niigata University of Health
subjects, J. Oral Rehabil. 27 (2000) 3340. and Welfare in the Department of Health and Nutrition. He was awarded his
[5] R.L. Joynt, R.F. Erlandson, M. Rourke, Computerized synthesis of M.Agr. from Niigata University Graduate School of Science and Technology,
electromyographic interference patterns, Arch. Phys. Med. Rehabil. 69 Japan, in 2003. His research interest is neural regulation of ingestion.
(1988) 517523.
[6] R.S. Person, M.S. Libkind, Simulation of electromyograms showing Yozo Miyaoka is a Professor at Niigata University of Health and Welfare
interference patterns, Electroencephalogr. Clin. Neurophysiol. 28 (1970) in the Department of Health and Nutrition. He was awarded his Ph.D.
from Niigata University School of Dentistry, Japan, in 1990. He was a
625632.
postdoctoral fellow at the Pennsylvania State University School of Medicine,
[7] Y. Miyaoka, I. Ashida, D. Inagaki, S. Kawakami, Differentiation of USA, from 1990 to 1992. He was an Assistant Professor at Niigata University
activity patterns in the suprahyoid muscles during swallowing of foods in the Department of Oral Physiology and Associate Professor at Yonezawa
with ve taste qualities, J. Sens. Stud. 20 (2005) 473483. Womens College in the Department of Health and Nutrition. His research
[8] I. Ashida, H. Iwamori, S. Kawakami, Y. Miyaoka, A. Murayama, interest is physiology of oral functions, especially neural mechanisms of taste
Analysis of physiological parameters of masseter muscle activity during and swallowing.