Medical & Biological Engineering & Computing

https://doi.org/10.1007/s11517-019-01949-4

ORIGINAL ARTICLE

Mechanomyography-based muscle fatigue detection during electrically

elicited cycling in patients with spinal cord injury
Jannatul Naeem 1 & Nur Azah Hamzaid 1 & Md. Anamul Islam 1 & Amelia Wong Azman 2 & Manfred Bijak 1,3

Received: 7 March 2018 / Accepted: 5 January 2019

# International Federation for Medical and Biological Engineering 2019

Abstract
Patients with spinal cord injury (SCI) benefit from muscle training with functional electrical stimulation (FES). For safety reasons
and to optimize training outcome, the fatigue state of the target muscle must be monitored. Detection of muscle fatigue from mel
frequency cepstral coefficient (MFCC) feature of mechanomyographic (MMG) signal using support vector machine (SVM)
classifier is a promising new approach. Five individuals with SCI performed FES cycling exercises for 30 min. MMG signals
were recorded on the quadriceps muscle group (rectus femoris (RF), vastus lateralis (VL), vastus medialis (VM)) and categorized
into non-fatigued and fatigued muscle contractions for the first and last 10 min of the cycling session. For each subject, a total of
1800 contraction-related MMG signals were used to train the SVM classifier and another 300 signals were used for testing. The
average classification accuracy (4-fold) of non-fatigued and fatigued state was 90.7% using MFCC feature, 74.5% using root
mean square (RMS), and 88.8% with combined MFCC and RMS features. Inter-subject prediction accuracy suggested training
and testing data to be based on a particular subject or large collection of subjects to improve fatigue prediction capacity.

Keywords Functional electrical stimulation . Muscle fatigue . Spinal cord injury . Mechanomyography . Mel frequency cepstral
coefficients (MFCC)

1 Introduction fatigue is apparent due to the inverse recruitment of motor unit

[6]. It has also been reported that overloading the stimulated
Spinal cord injury (SCI) could lead to partial or complete muscle might lead to muscle damage [7]. These drawbacks
paralysis of the upper and/or lower limbs depending on the limit the FES functional capacity and should be overcome to
level and severity of the injury. SCI causes complications optimize the training and functional benefit of FES. Hence,
including pressure sores [1], muscle spasticity [2], loss of monitoring of muscle condition during the FES-evoked cy-
muscle strength, cardiovascular disease [3], and osteoporosis cling is necessary when training the muscle for a longer period
[4]. Functional electrical stimulation (FES)-evoked cycling of time.
can improve their health benefits including increased muscle Physiologically, muscle fatigue is defined as the drop of
strength, volume, insulin sensitivity, glucose metabolism, and muscle force during a continuous steady muscle contraction
endurance [5]. In all FES applications, however, fast muscle [8]. To this date, there are several methods to evaluate muscle
fatigue including torque response [9] and joint angle measure-
ment [10]. Surface electromyography (EMG) is another non-
* Nur Azah Hamzaid invasive method to evaluate FES-evoked muscle contraction
[email protected]
and fatigue. However, this technique is not well adopted prac-
tically due to the interference from the electrical stimulus and
1
Department of Biomedical Engineering, Faculty of Engineering, from motion artifacts or the surrounding noise. Incorporating
University of Malaya, 50603 Kuala Lumpur, Malaysia additional circuits can help to avoid amplifier saturation and
2
Department of Electrical and Computer Engineering, Faculty of blank stimulation artifacts. Nevertheless, the complexity of
Engineering, International Islamic University Malaysia, 53100 Kuala these circuits [11] may lead to challenging fatigue evaluation.
Lumpur, Malaysia
3
Mechanomyography (MMG), a technique that measures
Center for Medical Physics and Biomedical Engineering, Medical mechanical muscle response including muscle vibration, has
University Vienna, 1090 Vienna, Austria
been used to evaluate muscle activities [11] and muscle
 Med Biol Eng Comput

fatigue [12] in voluntary contractions [13]. Dynamic muscle subject respectively. However, the number of recruited sub-
contractions, including concentric and eccentric contractions, jects was relatively low for classification. Alves et al. [31]
produce force [14, 15]. These contractions which were the used a genetic algorithm for MMG signal feature extraction
result of the recruitment of motor unit (MU) following the based on a linear discriminant analysis (LDA) classifier to find
MU firing rate can be monitored using MMG. Several studies out the effect of accelerometer location of single-site forearm.
have reported the correlation of MU recruitment and firing It was reported that the placement of five accelerometers on
rate of the muscle fibers with MMG amplitude during motor single-site forearm achieved group accuracy nearly 73% for
nerve stimulation [16–18]. MMG parameters in time [19], all three classes of muscle actions. However, the classifiers
frequency [20], and joint time-frequency (TF) [21] domains were vulnerable to the changes of forearm position and longi-
were analyzed during isometric FES-evoked contractions. tudinal and transverse displacements of the sensors.
However, MMG responses in time and/or frequency domains While many researchers extracted features from MMG sig-
of dynamic evoked muscle force pattern were reported to be nals using genetic programming, genetic algorithms [21, 32],
non-linear [22] given the inconsistent loading and non- statistical analysis [33, 34], and wavelet transform [35], to
stationary nature of muscle contraction during FES cycling date, mel frequency cepstral coefficients (MFCC) feature have
[23]. This non-linearity may be due to several factors includ- not been introduced to perform MMG signal analysis of dy-
ing changes in the muscle fiber length, number of active motor namic muscle contraction. The most common and widely
units, firing rates, and thickness of the tissue between muscles adopted MFCC feature was in automatic speech recognition
[23, 24]. application. In speech recognition, uttered speech is consid-
In order to analyze the non-stationary MMG signals, wave- ered dynamic in nature due to its frequency changes with each
let transform (WT), short-time Fourier transform (STFT), and speech signal. Similarly, during contraction, muscle generates
Wigner-Ville transform as joint TF signal processing tech- low frequency (5–50 Hz) vibration [30]. Additionally, several
nique were proposed [17, 22, 25]. Wavelet transform has been researchers have reported that MMG frequency (mean power
used in several researches to describe non-stationary MMG frequency, median frequency) signal pattern changes when
signals produced during dynamic muscle contractions. Beck muscle is artificially stimulated [36–39]. One related work
et al. [26] proposed a new wavelet analysis method where 11 by Doulah and Fattah [40] proposed MFCC feature applica-
non-linearly scaled wavelet filter banks were used to analyze tion to classify normal and neuromuscular diseased muscles
MMG signals. The intensity of the MMG signals was pro- using EMG. The researchers employed MFCC feature with
posed to be used in statistical pattern identification of the motor unit action potential rather than direct MFCC feature of
dynamic muscle contractions. EMG signal by template matching decomposition method.
Furthermore, Ryan et al. [27] compared short-time Fourier The MFCC feature achieved a total classification accuracy
transform (STFT) with continuous wavelet transform (CWT) of up to 92.50%.
for MMG signal analysis and showed that the two were sim- We hypothesized that dynamic muscle force response of
ilar in response. Xie et al. [28] proposed several features such eccentric and concentric contraction (muscle length changes)
as STFT, short wavelet transform (SWT), wavelet packet during FES-evoked cycling can be extracted directly from the
transform (WPT), and S-transform joined with singular value muscle surface using MMG-derived MFCC feature and eval-
decomposition to classify different hand motion pattern in- uated with a support vector machine (SVM), and that the
cluding wrist flexion, extension, open, and grasp using MMG signals can be classified as Bnon-fatigued^ and
MMG signals for prosthetic control which achieved 89.7% Bfatigued.^ The proposed MFCC classification accuracy was
accuracy. compared to RMS features. Accuracy of fatigue prediction
A work by Silva and colleagues [29] acquired MMG sig- among subjects using MFCC in comparison to the generally
nals from a microphone-accelerometer sensor pair to classify adopted RMS feature was analyzed.
two activities of prosthesis to control wrist extension and wrist
flexion using RMS feature. The classification accuracy
achieved from two subjects was around 70% based on their 2 Methods
cross-validation tests. However, this work focused on RMS
feature and did not consider any time-frequency domain fea- 2.1 Participants
tures. Subsequently in another work, Saliva et al. [30] im-
proved the accuracy of RMS-based MMG signal classifica- Five individuals with SCI with American Spinal Injury
tion of muscle activity for opening and closing of the prosthe- Association Impairment Scale (ASIAIS) classification A and
sis by coupling the accelerometer-microphone sensor and B, implying no voluntary motor control [41], were recruited
MMG socket to eliminate interference of the recorded signal. from the University of Malaya Medical Centre, Kuala
The experiment was performed on two subjects and the im- Lumpur, Malaysia. Participants volunteered to participate in
proved accuracies attained were 88% and 71% for each this study (Table 1) after giving their informed consent. This
Med Biol Eng Comput

Table 1 Participants’ demography

Participants Gender Age Body mass (kg) Height (cm) Level of ASIAIS TAI (year) BMI (kg/m2)
lesion

1 M 45 82.0 172 T4 B 14 27.7

2 M 49 62.4 171 C7 B 11 21.3
3 M 28 79.6 162 T1 A 3 30.3
4 M 33 71.6 179 C6 B 13 22.3
5 F 47 72.0 165 C6 B 15 26.4
Mean 40.4 73.5 169.8 11.2 25.6
SD 9.3 7.7 6.6 4.8 3.36

ASIAIS American Spinal Injury Association Impairment Scale, TAI time after injury, F female, M male, BMI body mass index, ASIAIS A sensory and
motor complete impairment, ASIAIS B motor complete impairment

study was granted by the University of Malaya Research stimulus pulse train was 0.4 s which contained a total of 12
Ethics Committee (approval no: 1003.14 (1)). All participants stimulation pulses. However, as the muscle fatigues, the cy-
understood the study protocol. The subjects’ exclusion criteria cling speed might decrease, resulting in an increase of stimu-
were subjects with metal implanted in the stimulated limb, lus duration.
cognitive impairment or without tolerance to FES sensation, The stimulation parameters were set as follows: biphasic
severe spasticity [2] which is related to muscle tone and stiff- current amplitude adjusted up to 120 mA (peak) or the highest
ness [42], and undesirable muscle responses from quadriceps tolerable current of each participant; pulse frequency was
muscle as determined by a certified physician. Participants fixed at 30 Hz and the biphasic pulse width was set at
were asked to abstain from any FES-related exercise at least 400 μs + 400 μs (positive + negative phase). Initially, a 1-
48 h before the testing [43]. min warm up session [44] was initiated using the same stim-
ulation parameters, the current set to produce weak muscle
contractions. Electrical stimulation pulses were delivered to
2.2 FES experimental protocol the quadriceps, hamstrings, and glutei muscles via self-
adhesive electrodes (size 9 cm × 15 cm, RehaTrode,
All participants were seated in their manual wheelchair com- HASOMED, Germany). The first electrode was positioned
fortably during the FES cycling session with their feet safely between 6 and 8 cm near the patellar border and the second
secured to the pedals using physiotherapy straps. Each partic- electrode was placed approximately 1/3 of the distance be-
ipant underwent a 30-min FES cycling session on a FES cycle tween the region of inguinal line and the superior patellar
ergometer (MOTOmed Viva 2, RECK-Medizintechnik border and slightly lateral to the muscle center line to ensure
GmbH, Betzenweiler, Germany), interfaced with the stimulation coverage over the three muscle bellies of VL, RF,
RehaMove 2 FES system at a cycling speed of 40 rpm. The and VM [45] (Fig. 2). VL, RF, and VM are the main superfi-
aim of this protocol was to induce peripheral muscle fatigue cial quadriceps muscles. Each muscle behavior is different
by means of continuously repetitive muscle contraction. A [46] and can be detected using external sensors placed on
commercial electrical stimulator (Rehabstim2, HASOMED the skin.
GmbH, Magdeburg, Germany) which produces biphasic rect-
angular current-controlled stimulation pulses was synchro-
nized with the motor resisted cycling ergometer. The stimula- 2.3 MMG signal acquisition
tion (maximum 120 mA, 30 Hz, biphasic, pulse width ±
400 μs) started every time the pedal reached a crank angle In our case, the sensors were attached to the skin with double-
of 45° and ended at 135°. In a 40-rpm cycling speed, a total sided adhesive tape to over the belly of VL, RF, and VM
rotation of 360° takes 1.5 s (Fig. 1). The duration of the muscles [47]. The three MMG sensors used were

Fig. 1 Schematic diagram of

pulse train synchronized with
ergometer pedal angle position in
early non-fatigued condition, at
40 rpm
 Med Biol Eng Comput

Fig. 2 a Experimental setup and

b MMG sensor placement over
the muscles (RF rectus femoris,
VL vastus lateralis, VM vastus
medialis) represented by the black
solid circles

accelerometer based, Sonostics BPS-IIVMG transducers (fre- consensus that at higher window length, the signal may not
quency range 20–200 Hz, sensitivity 30 V/g, diameter be stationary while shorter frame may not have enough infor-
32.6 mm, thickness 12.5 mm, mass 10 g). Other researchers mation to extract significant signal feature [55, 56]. This was
described their usage for muscle assessment in [19, 48, 49]. followed by the Mel-filterbank Eq. (1) [57] designed with 26
Sensor locations were marked with a permanent marker to triangular filters uniformly spaced on the Mel scale between
ensure consistent placement. The effect of skin thickness lower and upper frequency limits. The Mel scale is defined as
and fat on MMG parameters was not considered in this study a perceptual scale of frequency when measured to its original
as it has been reported that no relationship was found between frequency.
skinfold thickness and MMG RMS (voltage)-force correlation To calculate filterbank energies (FBEs) (26 filters per
[50] and no strong correlation was reported between skinfold frame), the filterbank was applied to the magnitude spectrum
thickness as well as MMG median and peak frequency [51]. values. The 26 log filterbank energies consisted of com-
pressed FBEs were then de-correlated using the discrete co-
2.4 Signal processing sine transform (DCT) equation as given in Eq. (2) [58]. The
Cn of the MFCC feature calculated from Eq. 2 was fed into
MMG sensors were interfaced with a personal computer using Eq. 3 as the feature vector x, as expressed in Eq. 3. From the
the BACQKnowledge^ data acquisition and analysis software results, 13 out of the 26 DCT coefficients were discarded from
package (BIOPAC Inc., Santa Barbara, CA, USA). The raw our application based on common practice from the literature
signals were sampled at 2 kHz and band pass filtered (fourth- [57, 59, 60]. The reasons include the fact that fast changes in
order Butterworth) at 20–200 Hz in order to reduce additional the filterbank can degrade recognition performance and com-
noise that might have originated from motion artifacts. The putational cost [59].
signals were then processed using MATLAB (Version 2013,

f
The Mathworks, Natick, MA) for segmentation and classifi- melð f Þ ¼ 1127ln 1 þ ð1Þ
cation. They were segmented by each contraction automati- 700
rffiffiffiffiffi

cally using peak detection algorithm whereby only the propul- 2 N πi
sion phase (0.25 s) of the contraction (minimum 0.4 s) was Cn ¼ ∑ m j cos ð j−0:5Þ ð2Þ
N j−1 N
segmented. The MFCC and RMS features were then extracted
from each contraction and used for training and testing the where N is the number of filter bank and mj is the log filter
SVM. bank amplitude.
The MFCC process flow is presented in Fig. 3. The select-
2.5 MMG classification ed 13 out of the 26 DCT coefficients were used to train a SVM
classifier.
2.5.1 MFCC feature Figure 4 shows typical non-fatigued (a) and fatigued (b)
MMG signals recorded from the three localized sensors of
The first stage of MFCC feature extraction method was to RF, VL, and VM. The fatigued signals in (b) have generally
apply STFT analysis [52, 53] with window frames of 25 ms smaller amplitudes than (a) and the duration of contraction is
to the signal that was considered as stationary [54]. The power longer in (b) compared to (a). This is because, fatigued muscle
spectrum was computed for every 25 ms frame with 10 ms requires more time to complete a single contraction cycle.
forward shift throughout the 250 ms MMG signal. Typically, Significant differences in MMG signals among RF, VL, and
window lengths are in 20–40 ms range because of the VM can also be clearly observed from Fig. 4.
Med Biol Eng Comput

Fig. 3 Block diagram of the

MFCC algorithm

The MFCC generates only frequency coefficients from the 2.5.2 Support vector machine
time series of MMG signal (250 ms) as shown in Fig. 5.
Figure 6 shows the MMG recognition steps from the input In the application of biomedical signal classification, support
MMG signal to training and recognition of MMG signal. vector machine is a widely used machine learning technique
For each of the test subject, their corresponding MMG [62]. SVM consists of an optimal hyperplane with a margin
responses as described above were recorded for 30 min. that separates the two classes of data with maximum distance
Data of each subject’s MMG signals captured from RF, VL, between them [63]. Therefore, the boundary partition between
and VM muscles were separated into two groups. Out of all the two classes of information was maximized. Optimal place-
data, 75% of the total contraction signals were used as training ment of the hyperplane was dependent on the portion of the
data while the remaining 25% were used for testing. For the training data, referred to as support vectors, which lies near to
training data, the MMG signals were partitioned into two cat- the hyperplane. The optimal hyperplane is described in Eq. (3)
egories: non-fatigued muscle contractions and fatigued mus-
cle contractions. Grouping of the two muscle conditions was
wT :x þ b ¼ 0 ð3Þ
based on the assumption that the first 10 min of the cycling
session represents the responses of a non-fatigued muscle and
where w is the weight vector, x is the input vector from input
the last 10 min, a fatigued muscle [47]. From the recorded
space, and b is the bias. During the training phase, the data
data, 400 non-fatigued contraction samples and 400 fatigued
from class 1 were labeled as + 1 and data from class 2 were
contraction samples were extracted from each of the RF, VL,
labeled as − 1 (Fig. 8). At the classification detection phase,
and VM MMG responses respectively, totaling 1200 samples
the data was classified as + 1 if wT.x + b > = + 1 and when
for fatigue and non-fatigued each. Results were validated with
wT.x + b < = − 1, it was classified as − 1.
k-fold cross-validation [61] method where out of each mus-
For testing stage classification, Eq. 3 can be written as
cle’s 400 contractions set, 300 contractions were used for
training and 100 for testing. Therefore, a total of 900 non-
fatigued and 900 fatigued contractions from the three sensors
(RF, VL, VM) were used to train a SVM classifier and later f ðyÞ ¼ wT :x þ b
tested with 300 contractions. Figure 7 illustrates the contrac-
N ð4Þ
hence; wT ¼ ∑ αi k ðxi yi Þ
tion training and testing methods from one sensor. i¼1

Fig. 4 Typical normalized (to

gmax) MMG signals during one
revolution of cycling: a non-
fatigued and b fatigued muscle
contractions from sensors placed
at RF (top), VL (mid), and VM
(bottom)
 Med Biol Eng Comput

elements of the training set that lies on the boundary hy-

perplane to the two classes [65]. However, in real case
scenario, there could be more than one hyperplane. In
our case, two classes of muscle contractions of non-
fatigued (class 1) and fatigued (class 2) were considered
for the training and testing subgroups.

2.6 Muscle fatigue prediction among subjects

During electrical stimulation throughout the FES cycling

exercise among the SCI individuals, their muscle fatigue
responses may be similar yet distinct in different individ-
uals to a certain extent. This may suggest different indi-
vidual training and performance accuracy in comparison
to training and prediction algorithm using cumulative sub-
ject data. Therefore, inter-subject prediction accuracy was
generated to determine the accuracy when one subject’s
trained model was tested with the other subjects’ non-
Fig. 5 Time series of MFCC cepstrum index plot in frame index of 0.25-s
MMG signal
fatigue and fatigue contraction signals.
Furthermore, results were validated using 4-fold cross-
validation methods where all the data were used for train-
Equation 4 can be written as
ing and testing the classifier. The prediction accuracy was
f ðyÞ ¼ ∑N
i¼1 αi k ðxi yi Þ þ b ð5Þ calculated in three different models which used MFCC
feature only, RMS only, and the combined MFCC and
where yi is the tested vector of MMG signal, xi is the support RMS features. Individual muscle prediction accuracy per-
vector (Fig. 8) calculated from training data set, and αi is their formance was also analyzed.
weight and constant bias is b. K(x,y) is the kernel function
which calculates dot product of two vectors x and y in high-
dimensional feature space. The radial basis function (RBF)
kernels (Eq. 6) were considered due to its better performance 3 Results
than linear kernel.
  Over 30 min of cycling period, the cycling speed (mea-
K RBF ðx; yÞ ¼ exp −γ jjx−yjj2 ð6Þ sured with the built in speed monitor) was used as an indi-
cator of fatigue, i.e., when the speed drops significantly in
where γ is the control parameter (kernel width) and ||.|| de- comparison to the initial speed. From the experiment, the
notes the Euclidean norm. subjects’ average cycling speed usually decreased through-
In Fig. 7, SVM classification with a hyperplane that out the 30 min of training. Duration of an example con-
maximizes the separating margin between two classes is traction was 1.43 s from the first 10 min and 1.82 s from
indicated by B◊^ and BO^ and support vectors are the last 10 min (Fig. 4).

Fig. 6 Block diagram of MMG

signal training and recognition
method
Med Biol Eng Comput

Fig. 7 Selection of contractions

used for training and testing from
a cycling session

3.1 MFCC and RMS features predicted contractions 3.2 Muscle fatigue prediction accuracy
and accuracy among subjects

The predicted and expected results of all participants’ muscle An inter-subject prediction is when one subject training data
contractions using MFCC feature (Table 2) and RMS feature was used to create a model and then tested with another sub-
(Table 3) are presented as a confusion matrix [66] for the ject’s data which is expected to be non-fatigue or fatigue. The
fourth fold of repetition. Hence, confusion matrix shows the inter-subject prediction accuracy of muscle fatigue using
expected and recognized number of contractions diagonally. MFCC feature is presented in Table 4 while prediction accu-
Overall, the average prediction accuracy using the MFCC and racy based on RMS feature is shown in Table 5.
RMS feature is 92.2% and 75.9% respectively. Inter-subject correlation was assumed if high prediction
Based on the results in Table 2, subject numbers 1, 2, and 4 accuracy was achieved when test data of one subject was fed
obtained more than 95% accuracy based on MFCC feature. into training data of the other’s training data, and vice versa.
Highest accuracy was of subject 2 with a maximum number of This was demonstrated in Table 4 whereby when test data of
298 contractions for fatigued and 295 contractions as non- subject 2 was fed into subject 1 trained data it achieved highest
fatigue were predicted correctly. The same level of accuracy accuracy of 78% and 71% vice versa, using the MFCC fea-
was achieved for subject 1 and 4 at 96.3%. Subject 1 had the tures. This was again demonstrated in Table 5 when both
least prediction error of fresh contractions with two contrac- subjects 1 and 2 displayed high accuracy based on each
tions identified as fatigued contraction compared to the other other’s training and testing data using RMS feature.
subjects. The lowest accuracy of contractions was for subject
5 at an accuracy of 84%.
Table 3, on the other hand, presents the prediction accuracy 3.3 Cross-validation of the combined results of three
based on the RMS feature of the MMG signals. Similar to sensors (RF, VL, and VM)
MFCC feature, subject 2 had the highest accuracy with all
non-fatigue contractions were predicted correctly but a total For results’ validation, all the non-fatigue and fatigue contrac-
of 14 fatigue contractions failed to be predicted. The accuracy tion samples were used for training and testing. Four-fold
obtained for subject 3 was less than 60% and for subject 4 was
less than 50%. The lowest number of non-fatigue contraction
Table 2 Number of expected and predicted contractions sample in
was predicted for subject 4 which was only 163 contractions
confusion matrix and accuracy using MFCC feature. Each subject’s first
over the expected 300 non-fatigue contractions. and second row shows expected (first row: expected True, second row:
expected False) and predicted number (first row: predicted True, second
row: predicted False) of contractions

Subject Expected result Predicted contractions Accuracy (%)

Non-fatigue Fatigue Non-fatigue Fatigue

1 300 0 293 7 96.3

0 300 15 285
2 300 0 295 5 98.8
0 300 2 298
3 300 0 276 24 85.5
0 300 63 273
4 300 0 289 11 96.3
0 300 11 289
5 300 0 218 82 84.0
0 300 14 286
Fig. 8 SVM hyper plane separated by support vectors [64]
 Med Biol Eng Comput

Table 3 Number of expected and predicted contractions sample in Table 5 Predicted inter-subject accuracy using RMS feature
confusion matrix and accuracy using RMS feature. Each subject’s first
and second row shows expected and predicted number of contractions Trained subject Test subject accuracy (%)
(first row: expected and predicted True, second row: expected and
predicted False) to be detected Subject 1 Subject 2 Subject 3 Subject 4 Subject 5

Subject Expected result Predicted contractions Accuracy (%) Subject 1 97.1 48.1 48.3 85.8
Subject 2 90.6 49.5 47.8 87.1
Non-fatigue Fatigue Non-fatigue Fatigue
Subject 3 65.8 67.5 59.8 65.5
1 300 0 299 1 87.5 Subject 4 89.3 97.3 49.1 87.1
0 300 74 226 Subject 5 88.8 97.3 48.8 48.1
2 300 0 300 0 97.6
0 300 14 286
3 300 0 280 20 59.6 4 Discussion
0 300 222 78
4 300 0 163 137 48.3 Classification of non-fatigued and fatigued muscle contrac-
0 300 173 127
tions using MFCC feature was hypothesized to have a higher
5 300 0 278 22 86.5
prediction ability among subjects in comparison to the gener-
0 300 59 241
ally adopted RMS feature. Hence, the overall performance of
MFCC might be higher because the MFCC feature incorpo-
rates inherent calculations of frequency components and pow-
er spectrum of MMG signals in time and frequency domain.
cross-validation used a total of 1200 contractions (three sen- However, the average inter-subject correlation prediction
sors) of which 75% used for training and 25% for testing accuracy of MFCCs was around 50% which was a very low
(Tables 6 and 7). performance in classification measurement. Interestingly, the
Results showed that with all four repetitions in overall, the overall inter-subject accuracy based on the RMS feature re-
MFCC feature shows better performance than RMS. sulted in better performance compared to MFCC feature.
However, the mean accuracy of four repetitions for subjects However, in both MFCC and RMS features, there was insuf-
1 and 2 was higher (93.9% and 98.9%) in RMS compared to ficient prediction consistency observed among subjects.
MFCC (90.8% and 93%). The RMS of subjects 3 and 4 de- Results revealed that FES muscle fatigue classification of
picts the lowest performance compared to MFCC. However, dynamic FES cycling using MFCC feature projected better
effect of repetitions standard deviation for MFCC feature was accuracy than the RMS feature. The number of correctly iden-
below 10%. On the other hand, standard deviation of RMS tified contractions as non-fatigued and fatigued was higher in
feature reached 13.2% for subject 4 and for subject 5 it MFCC when compared to RMS. Some non-fatigued muscle
reached 19.8%. contractions overlapped with other fatigued contractions using
Table 8 demonstrates that the combined MFCC and RMS both MFCC and RMS feature. This might be due to the inac-
features reached a maximum mean accuracy of up to 93% curate assumption that the muscles were not fatigued in the
while lowest was 83% and the standard deviation reached first 10 min, when the muscles might actually be undergoing
maximum of about 12%. Results also show that the first and early muscle fatigue within the first 10 min of cycling. This is
second repetitions have higher accuracy compared to the third backed by the research findings that suggested that electrical
and fourth repetitions. muscle activation is responsible for faster muscle fatigue in
comparison to voluntary contractions [67]. This is also related

Table 6 Accuracy results for MFCC features with four repetitions

Table 4 Predicted inter-subject accuracy using MFCC feature
Subject Repetition accuracy (%) Mean ± SD
Trained subject Test subject accuracy (%)
First Second Third Fourth
Subject 1 Subject 2 Subject 3 Subject 4 Subject 5
1 93.3 88.8 85 96.3 90.8 ± 4.9
Subject 1 78 49 50 50 2 96.5 95.1 81.6 98.8 93.0 ± 7.7
Subject 2 71 49.8 51 44 3 96.3 83.6 72.6 85.5 81.7 ± 9.7
Subject 3 59.1 68.5 50 50.6 4 100 100 85 96.3 92.8 ± 7.1
Subject 4 44.1 38.5 50 57.1 5 94.1 94.3 86.8 84 89.8 ± 5.2
Subject 5 49 38 58.1 74.1 Mean ± SD 96.0 ± 2.6 92.4 ± 6.3 82.2 ± 5.6 92.1 ± 6.8
Med Biol Eng Comput

Table 7 Accuracy results for

RMS features Subject Repetition accuracy (%) Mean ± SD

First Second Third Fourth

1 97.1 100 91.3 87.5 93.9 ± 5.6

2 99.3 99.8 99.1 97.6 98.9 ± 0.9
3 59.3 66.83 57.8 59.6 60.8 ± 4.0
4 40.1 22.83 53.1 48.3 41.0 ± 13.2
5 92.5 84.3 48.6 86.5 77.9 ± 19.8
Mean ± SD 77.66 ± 26.5 74.752 ± 32 69.98 ± 23.4 75.9 ± 20.8

to the inability to modulate firing frequency or recruitment In many research works, muscle contraction classification
pattern of motor units [68] and the Binverse recruitment.^ and prosthetic control based on MMG signals have been in-
Moreover, the captured muscle responses might contain mo- vestigated using different feature extraction methods, such as
tion artifacts [69] since subjects who are in the sitting position the RMS, wavelet transform, SWT, and STFT, and genetic
on the wheelchair were not completely fixed and their limbs algorithms [21, 28, 30, 31].
were moving during cycling. Furthermore, different muscle Two types of muscle contractions for wrist extension and
properties of RF, VL, and VM and variations in the placement flexion were investigated by Saliva et al. which used RMS
of MMG sensors may have also affected the signals. Table 9 feature while Xie et al. studied the STFT, SWT, WPT, and
shows each of the three individual VM, VL, and RF muscle S-transform features to classify hand motion pattern [28,
performance accuracy of MFCC and RMS feature. The mean 30]. However, in their work, the classification accuracy
accuracy of MFCC feature is higher than RMS feature in RF, achieved was below 90%. Alves et al. [31] reported classifi-
VL, and VM muscles. When compared to MFCC on subjects cation accuracy of three class movement of MMG signal wrist
1 and 5, RMS feature shows higher accuracy in the three flexion, wrist extension, and semi-pronation of single-site
muscle groups. The standard deviation accuracy of three mus- forearm based on sensor placement was about 73%. Authors
cles in each subject was higher in MFCC, yet the overall mean conveyed the degradation of accuracy might be influenced by
accuracy of RMS was lower than MFCC. It is interesting to several factors including sensor placement, classifier com-
highlight that inter-individual and intra-individual sensor for plexity, training method, signal feature, and muscle architec-
each patient has different accuracy to detect muscle fatigue ture. Another recent research [21] has implemented wavelet
due to the geometry of the each muscle structure [70]. Using transform and modified pseudo-wavelet by using SVM clas-
the MFCC feature, it could be suggested that one sensor on RF sifier to investigate non-fatigue and fatigue contractions and
muscle can quantify the whole quadriceps muscle assessment. achieved approximately 80.63% accuracy, though during the
There have been several studies that documented the experiment very small number of trials (73 trials) were used
change in MMG mean power frequency [18, 36, 39] and for training compared to this study. Madeleine et al. [71] sug-
median frequency [38] over the stimulation time due to the gested that linear and nonlinear analyses of MMG signal of
inverse recruitment of motor units. These were the basis of wrist extensor could be assessed using average rectified values
which MFCC feature can retrieve frequency components from (ARV) of the MMG output or RMS linear feature with varia-
the muscle contractions. tions in sensor load, location, contractions type, and time.
Their results show that higher ARV value was observed in
load variations compared to variations in location and contrac-
Table 8 Accuracy for combined MFCC and RMS features
tion type, while variance ration in percentage of recurrence
and percentage of determinism is 22.8% and 0.1% respective-
Subject Repetition accuracy (%) Mean ± SD ly. Variations in location revealed that ARV was lowest with
31.2 ms−2 and, 9.9% recurrence and 43.6% determinism,
First Second Third Fourth
while varying the time revealed ARV of 89.4 ms−2, with
1 90.6 96.1 83.8 80.8 87.8 ± 6.8 27% recurrence and 6.6% determinism. However, this re-
2 100 99.8 75.6 98.8 93.5 ± 11.9 search focused on ARV or RMS features only. Sarlabous
3 97 85 73 78.5 83.1 ± 10.3 et al. [72] used the dog model to quantify stochastic nature
4 96.5 95.1 81.1 95.3 92.0 ± 7.2 of MMG signals to estimate muscle force using ARV or RMS
5 93.6 95.1 81.5 80.1 87.5 ± 7.8 parameters based on the Lempel-Ziv algorithm. Both studies
Mean ± SD 95.5 ± 3.5 94.2 ± 5.5 79 ± 4.5 86.7 ± 9.5 emphasized that non-linear analyses are found promising
when analyzing muscle fatigue or muscle force. In this study,
 Med Biol Eng Comput

Table 9 Individual three muscles performance accuracy of RF, VL, and VM

Subject MFCC feature accuracy% Mean ± SD RMS feature Accuracy % Mean ± SD

RF VL VM RF VL VM

1 77 90 66.5 77.8 ± 11.7 94 100 99.5 97.8 ± 3.3

2 100 98 99 99 ± 1 97 98.5 100 98.5 ± 1.5
3 94 78 62.3 78.1 ± 15.8 61 63 56 60 ± 3.6
4 98 99.5 89 95.5 ± 5.6 12 75 42 43 ± 31.5
5 88 86.5 65 79.8 ± 12.8 92 98.5 90 93.5 ± 4.4
Mean ± SD 91.4 ± 9.2 90.4 ± 8.7 76.3 ± 16.5 71.2 ± 36.1 87 ± 16.9 77.5 ± 26.7

RMS and MFCC features of the MMG signal were classified and sessions. Thus, increasing the repetition of trials in the
using SVM classifier with the RBF kernel to map the data in long run would positively influence the accuracy of the results
higher dimensions for non-linear MMG data separation. [31]. However, the effect of different window length of MMG
The MFCC feature is a representation of the short-term signal on accuracy could be an interesting topic for future
power spectrum of a signal, based on a linear cosine transform researcher.
of a logarithmic power spectrum on a non-linear Mel scale of
frequency. Therefore, non-linear MMG signal feature during
cycling was deemed to be more suitable for MFCC than clas-
sical STFT in dynamic muscle contraction classification. 5 Conclusions
Moreover, the computational costs of short wavelet transform
or wavelet packet transform are higher than MFCC which This study is the first one to demonstrate the adoption of
plays important role in real-time applications. MFCC feature, which had been primarily applied in the
The total accuracy for MFCC feature of MMG signal BAutomatic Speech Recognition^ domain, for MMG classifi-
achieved of up to 96% accuracy for non-fatigue and fa- cation of fatigued and non-fatigued contractions throughout
tigue classification for the first fold repetition and average dynamic FES cycling. MFCC feature showed better accuracy,
accuracy of 4-fold was 90.7%. Combined MFCC and up to 90.7% in comparison to RMS feature with an accuracy
RMS features with an average accuracy of 88.8% did of 74.5%. Thus, the proposed features can be used in muscle
not show any significant improvement of accuracy. fatigue prediction in such dynamic and cyclical evoked mus-
Studies on FES and fatigue related exercise can benefit cle contraction as long as the system is trained with data from
from our findings by implementing MFCC feature extrac- the monitored subject. Inter-subject prediction accuracy is in-
tion of the MMG signal. consistent and has low accuracy, indicating the need to have a
For fatigue detection correlation among subjects, the re- larger pool of training data. Further investigations will help to
sults illustrate good correlation between subjects 1 and 2 in better understand the nature of the MMG signals and influenc-
both MFCC and RMS. However, generally, higher accuracy ing factors such as physiological properties and physical mi-
was found in RMS feature adoption. The relationship between lieu. The method introduced in this study could be implement-
these two subjects could be a similarity in muscular behavior ed in FES systems to monitor the state of muscle fatigue to
during stimulation [73]. These results may suggest the possi- increase patient safety and to optimize patient training by
bility of using an identified similarly performing muscle, as in adapting the FES parameters during electrically evoked con-
subjects 1 and 2, to pool into the other subjects’ data to im- tractions in individuals with motor complete SCI.
prove its fatigue prediction learning and ability, or to use one
subject’s trained algorithm to predict another’s. RMS feature Funding The study was financially supported by the University of
Malaya Research Grant (UMRG), grant no. RP035A-15HTM.
showed better correlated accuracy among the subjects but the
results were not consistently high in all subjects. Therefore,
more investigation of muscle responses is required during Compliance with ethical standards
stimulation to find out the correlation among subjects.
This study was granted by the University of Malaya Research Ethics
A new method of MFCC feature extraction for MMG sig- Committee (approval no: 1003.14 (1)).
nal classification during FES cycling has been successfully
implemented using SVM classifier. The outcomes of this
study however were limited by the number of trials within Publisher’s note Springer Nature remains neutral with regard to jurisdic-
each subject due to challenges in multiple subject recruitment tional claims in published maps and institutional affiliations.
Med Biol Eng Comput

References 19. Ibitoye M, et al. (2016) Torque and mechanomyogram relation-

ships during electrically-evoked isometric quadriceps contractions
in persons with spinal cord injury. Vol. 38
1. Liu LQ, Moody J, Traynor M, Dyson S, Gall A (2014) A systematic
review of electrical stimulation for pressure ulcer prevention and 20. Ryan ED, Cramer JT, Egan AD, Hartman MJ, Herda TJ (2008)
treatment in people with spinal cord injuries. J Spinal Cord Med Time and frequency domain responses of the mechanomyogram
37(6):703–718 and electromyogram during isometric ramp contractions: a compar-
ison of the short-time Fourier and continuous wavelet transforms. J
2. Elbasiouny SM, Moroz D, Bakr MM, Mushahwar VK (2010)
Electromyogr Kinesiol 18(1):54–67
Management of spasticity after spinal cord injury: current tech-
niques and future directions. Neurorehabil Neural Repair 24(1): 21. Al-Mulla MR, Sepulveda F (2014) Novel pseudo-wavelet function
23–33 for MMG signal extraction during dynamic fatiguing contractions.
Sensors 14(6):9489–9504
3. Furlan JC, Fehlings MG (2008) Cardiovascular complications after
acute spinal cord injury: pathophysiology, diagnosis, and manage- 22. Hong-Bo X, Yong-Ping Z, Jing-Yi G (2009) Classification of the
ment. Neurosurg Focus 25(5):E13 mechanomyogram signal using a wavelet packet transform and
singular value decomposition for multifunction prosthesis control.
4. Tan CO, Battaglino RA, Morse LR (2013) Spinal cord injury and
Physiol Meas 30(5):441
osteoporosis: causes, mechanisms, and rehabilitation strategies. Int
J Phys Med Rehabil 1:127 23. Bonato P, Roy SH, Knaflitz M, de Luca CJ (2001) Time-frequency
parameters of the surface myoelectric signal for assessing muscle
5. Davis GM, Hamzaid NA, Fornusek C (2008) Cardiorespiratory,
fatigue during cyclic dynamic contractions. IEEE Trans Biomed
metabolic, and biomechanical responses during functional electrical
Eng 48(7):745–753
stimulation leg exercise: health and fitness benefits. Artif Organs
32(8):625–629 24. Cramer JT, Housh TJ, Weir JP, Johnson GO, Coburn JW, Beck TW
(2005) The acute effects of static stretching on peak torque, mean
6. Chou L-W, Binder-Macleod SA (2007) The effects of stimulation
power output, electromyography, and mechanomyography. Eur J
frequency and fatigue on the force-intensity relationship for human
Appl Physiol 93(5):530–539
skeletal muscle. Clin Neurophysiol 118(6):1387–1396
25. Akataki K, Mita K, Watakabe M (2004) Electromyographic and
7. Fouré A, Nosaka K, Wegrzyk J, Duhamel G, le Troter A, Boudinet
mechanomyographic estimation of motor unit activation strategy
H, Mattei JP, Vilmen C, Jubeau M, Bendahan D, Gondin J (2014)
in voluntary force production. Electromyogr Clin Neurophysiol
Time course of central and peripheral alterations after isometric
44(8):489–496
neuromuscular electrical stimulation-induced muscle damage.
PLoS One 9(9):e107298 26. Beck TW, Housh TJ, Fry AC, Cramer JT, Weir JP, Schilling BK,
Falvo MJ, Moore CA (2009) A wavelet-based analysis of surface
8. Wigmore DM, Befroy DE, Lanza IR, Kent-Braun JA (2008)
mechanomyographic signals from the quadriceps femoris. Muscle
Contraction frequency modulates muscle fatigue and the rate of
Nerve 39(3):355–363
myoglobin desaturation during incremental contractions in humans.
Applied physiology, nutrition, and metabolism = Physiologie 27. Ryan ED, Beck TW, Herda TJ, Hartman MJ, Stout JR, Housh TJ,
appliquee, nutrition et metabolisme 33(5):915–921 Cramer JT (2008) Mechanomyographic amplitude and mean power
frequency responses during isometric ramp vs. step muscle actions.
9. Kent-Braun JA, Callahan DM, Fay JL, Foulis SA, Buonaccorsi JP
J Neurosci Methods 168(2):293–305
(2014) Muscle weakness, fatigue, and torque variability: effects of
age and mobility status. Muscle Nerve 49(2):209–217 28. Xie H-B, Zheng Y-P, Guo J-Y (2009) Classification of the
mechanomyogram signal using a wavelet packet transform and
10. Weir JP, McDonough AL, Hill VJ (1996) The effects of joint angle
singular value decomposition for multifunction prosthesis control.
on electromyographic indices of fatigue. Eur J Appl Physiol Occup
Physiol Meas 30(5):441–457
Physiol 73(3):387–392
29. Silva J, Heim W, and Chau T (2004) MMG-based classification of
11. Islam MA, Sundaraj K, Ahmad RB, Ahamed NU (2013)
muscle activity for prosthesis control. In The 26th Annual
Mechanomyogram for muscle function assessment: a review.
International Conference of the IEEE Engineering in Medicine
PLoS One 8(3):e58902
and Biology Society
12. Tarata MT (2003) Mechanomyography versus electromyography,
30 . Silv a J, Heim W, Ch au T (20 05) A self-con tained,
in monitoring the muscular fatigue. BioMed Eng OnLine 2:3–3
mechanomyography-driven externally powered prosthesis. Arch
13. Xiaogang H, William ZR, Nina LS (2014) Motor unit firing rate Phys Med Rehabil 86(10):2066–2070
patterns during voluntary muscle force generation: a simulation
31. Alves N, Sejdic E, Sahota B, Chau T (2010) The effect of acceler-
study. J Neural Eng 11(2):026015
ometer location on the classification of single-site forearm
14. Ibitoye MO et al (2016) Estimation of electrically-evoked knee mechanomyograms. Biomed Eng Online 9(1):23
torque from mechanomyography using support vector regression.
32. Kattan A, et al. (2009) Detecting localised muscle fatigue during
Sensors 16(7):1115
isometric contraction using genetic programming. in IJCCI.
15. Mohamad NZ et al (2017) Mechanomyography and torque during
33. Al-Mulla MR, et al. (2009) Statistical class separation using sEMG
FES-evoked muscle contractions to fatigue in individuals with spi-
features towards automated muscle fatigue detection and
nal cord injury. Sensors 17(7):1627
prediction. In 2009 2nd International Congress on Image and
16. Orizio C, Liberati D, Locatelli C, de Grandis D, Veicsteinas A Signal Processing.
(1996) Surface mechanomyogram reflects muscle fibres twitches
34. Al-Mulla MR, Sepulveda F (2010) Novel feature modelling the
summation. J Biomech 29(4):475–481
prediction and detection of sEMG muscle fatigue towards an auto-
17. Barry DT, Cole NM (1990) Muscle sounds are emitted at the reso- mated wearable system. Sensors 10(5):4838–4854
nant frequencies of skeletal muscle. IEEE Trans Biomed Eng 37(5):
35. Beck TW, Housh TJ, Johnson GO, Weir JP, Cramer JT, Coburn JW,
525–531
Malek MH (2005) Comparison of Fourier and wavelet transform
18. Orizio C, Solomonow M, Baratta R, Veicsteinas A (1992) Influence procedures for examining the mechanomyographic and electro-
of motor units recruitment and firing rate on the soundmyogram and myographic frequency domain responses during fatiguing
EMG characteristics in cat gastrocnemius. J Electromyogr Kinesiol isokinetic muscle actions of the biceps brachii. J Electromyogr
2(4):232–241 Kinesiol 15(2):190–199
 Med Biol Eng Comput

36. Peters EJD, Fuglevand AJ (1999) Cessation of human motor unit 54. Roberto Gil-Pita, B.l.-G., Manuel Rosa-Zurera, Tailored MFCCs
discharge during sustained maximal voluntary contraction. for Sound Environment Classification in Hearing Aids, in
Neurosci Lett 274(1):66–70 Advanced Computer and Communication Engineering
37. Orizio C (1992) Soundmyogram and EMG cross-spectrum during Technology: Proceedings of ICOCOE 2015. Vol. 362. Springer,
exhausting isometric contractions in humans. J Electromyogr 2015., M.A.O. Hamzah Asyrani Sulaiman, Mohd Fairuz Iskandar
Kinesiol 2(3):141–149 Othman, Yahaya Abd Rahim, Naim Che Pee, Editor. 2015
38. Kouzaki M, Shinohara M, Fukunaga T (1999) Non-uniform me- 55. Wei H, et al. (2006) An efficient MFCC extraction method in
chanical activity of quadriceps muscle during fatigue by repeated speech recognition. In 2006 IEEE International Symposium on
maximal voluntary contraction in humans. Eur J Appl Physiol Circuits and Systems.
Occup Physiol 80(1):9–15 56. Shang-Ming L, et al. (2001) Improved MFCC feature extraction by
39. Esposito F, Orizio C, Veicsteinas A (1998) Electromyogram and PCA-optimized filter-bank for speech recognition. in IEEE
mechanomyogram changes in fresh and fatigued muscle during Workshop on Automatic Speech Recognition and Understanding,
sustained contraction in men. Eur J Appl Physiol Occup Physiol 2001. ASRU '01.
78(6):494–501 57. Gupta S et al (2013) Feature extraction using MFCC. Signal Image
40. Doulah ABMSU and Fattah SA (2014) Neuromuscular disease Processing 4(4):101
classification based on mel frequency cepstrum of motor unit action 58. Young S, et al. (2002) The HTK book Cambridge university engi-
potential. In 2014 International Conference on Electrical neering department, 3: p. 175
Engineering and Information & Communication Technology 59. Wahyuni ES (2017) Arabic speech recognition using MFCC fea-
41. Kirshblum SC, Burns SP, Biering-Sorensen F, Donovan W, Graves ture extraction and ANN classification. In 2017 2nd International
DE, Jha A, Johansen M, Jones L, Krassioukov A, Mulcahey MJ, conferences on Information Technology, Information Systems and
Schmidt-Read M, Waring W (2011) International standards for neu- Electrical Engineering (ICITISEE)
rological classification of spinal cord injury (revised 2011). J Spinal
60. Jong-Hwan L, et al. (2000) Speech feature extraction using inde-
Cord Med 34(6):535–546
pendent component analysis. In 2000 IEEE International
42. Moon S-H, Choi J-H, Park S-E (2017) The effects of functional Conference on Acoustics, Speech, and Signal Processing.
electrical stimulation on muscle tone and stiffness of stroke patients. Proceedings (Cat. No.00CH37100)
J Phys Ther Sci 29(2):238–241
61. Xu Q and Liang Y.-Z (2001) Monte Carlo Cross Validation 56:1–11
43. Bickel CS, Slade JM, VanHiel LR, Warren GL, Dudley GA (2004)
62. Subasi A (2013) Classification of EMG signals using PSO opti-
Variable-frequency-train stimulation of skeletal muscle after spinal
mized SVM for diagnosis of neuromuscular disorders. Comput
cord injury. J Rehabil Res Dev 41(1):33–40
Biol Med 43(5):576–586
44. Fazio C (2014) Functional electrical stimulation for incomplete
spinal cord injury. Proceedings (Baylor University. Medical 63. Cutajar M et al. (2013) Hardware-based support vector machine for
Center), 27(4):353–355 phoneme classification. in Eurocon 2013
45. Szecsi J, Straube A, Fornusek C (2014) A biomechanical cause of 64. Hashem EM, Mabrouk MS (2014) A study of support vector ma-
low power production during FES cycling of subjects with SCI. J chine algorithm for liver disease diagnosis. Am J Intell Syst 4(1):9–
NeuroEngineering Rehabil 11(1):123 14
46. Ouamer M, Boiteux M, Petitjean M, Travens L, Salès A (1999) 65. Akay MF (2009) Support vector machines combined with feature
Acoustic myography during voluntary isometric contraction reveals selection for breast cancer diagnosis. Expert Syst Appl 36(2, Part
non-propagative lateral vibration. J Biomech 32(12):1279–1285 2):3240–3247
47. Islam MA, Hamzaid NA, Ibitoye MO, Hasnan N, Wahab AKA, 66. Khezri M, Jahed M (2007) Real-time intelligent pattern recognition
Davis GM (2018) Mechanomyography responses characterize al- algorithm for surface EMG signals. Biomed Eng Online 6(1):45
tered muscle function during electrical stimulation-evoked cycling 67. Marsolais EB, Edwards BG (1988) Energy costs of walking and
in individuals with spinal cord injury. In: Elsevier editorial standing with functional neuromuscular stimulation and long leg
system(tm) for clinical, vol 58, pp 21–27 braces. Arch Phys Med Rehabil 69(4):243–249
48. Padmanathan Y, et al. (2014) Assessment of muscle performance 68. Bickel CS, Gregory CM, Dean JC (2011) Motor unit recruitment
using vibromyography (VMG) and electromyography(EMG). In during neuromuscular electrical stimulation: a critical appraisal. Eur
2014 IEEE 19th International Functional Electrical Stimulation J Appl Physiol 111(10):2399–2407
Society Annual Conference (IFESS). 69. Ibitoye MO, et al. (2014) Mechanomyographic parameter extrac-
49. Herzog W, Zhang YT, Vaz MA, Guimaraes ACS, Janssen C (1994) tion methods: an appraisal for clinical applications. Sensors
Assessment of muscular fatigue using vibromyography. Muscle (Basel, Switzerland) 14(12):22940–22970
Nerve 17(10):1156–1161 70. Orizio C (1993) Muscle sound: bases for the. Crit Rev Biomed Eng
50. Cooper MA, Herda TJ, Vardiman JP, Gallagher PM, Fry AC (2014) 21(3):201–243
Relationships between skinfold thickness and electromyographic 71. Madeleine P, Hansen EA, Samani A (2014) Linear and nonlinear
and mechanomyographic amplitude recorded during voluntary analyses of multi-channel mechanomyographic recordings reveal
and non-voluntary muscle actions. J Electromyogr Kinesiol 24(2): heterogeneous activation of wrist extensors in presence of delayed
207–213 onset muscle soreness. Med Eng Phys 36(12):1656–1664
51. Jaskólska A, Brzenczek W, Kisiel-Sajewicz K, Kawczyński A, 72. Sarlabous L, Torres A, Fiz JA, Morera J, Jané R (2013) Index for
Marusiak J, Jaskólski A (2004) The effect of skinfold on frequency estimation of muscle force from mechanomyography based on the
of human muscle mechanomyogram. J Electromyogr Kinesiol Lempel–Ziv algorithm. J Electromyogr Kinesiol 23(3):548–557
14(2):217–225 73. Vromans M, Faghri PD (2018) Functional electrical stimulation-
52. Irino T, et al. (2002) Evaluation of a speech recognition/generation induced muscular fatigue: effect of fiber composition and stimula-
method based on HMM and straight. in INTERSPEECH tion frequency on rate of fatigue development. J Electromyogr
53. Obuchi Y (2004) Delta-cepstrum normalization for robust speech Kinesiol 38(Supplement C):67–72
recognition. In: Proc. ICA
Med Biol Eng Comput

Jannatul Naeem is currently pur- Dr. Amelia Wong Azman is an

suing PhD at the University of Assoc. Prof. at the Department
Malaya and received his B.Sc. de- of Electrical and Computer
gree in Electrical and Electronic E n g i n e e r i ng , I n t e r n a t i o n a l
Engineering in 2009 and Msc in Islamic University Malaysia. Her
Communication Engineering at recent research interest is on the
the International Islamic use of ES for rehabilitation pur-
University of Malaysia, 2014 poses.

Dr. Nur Azah Hamzaid is a se- Dr. Manfred Bijak is Assistant

. nior lecturer at the University of Professor at the Medical
Malaya, Malaysia. She obtained University of Vienna, Austria.
her Bachelor in Mechatronics His area of research is
Engineering (Hons) in 2006 and Biomedical Engineering with a
then pursued her PhD in focus on functional electrical
Rehabilitation Engineering at the stimulation and development of
University of Sydney. implantable stimulators.

Md. Anamul Islam pursued the

PhD degree in Biomedical
Electronic Engineering. He is cur-
rently working at the department
of physical therapy, College of
Staten Island-City University of
New York, New York. His re-
search interests are on neuro-
physiology and spinal cord injury.