DETECTING LOCALISED MUSCLE FATIGUE

DURING ISOMETRIC CONTRACTION

USING GENETIC PROGRAMMING

Ahmed Kattan, Mohammed Al-Mulla, Francisco Sepulveda and Riccardo Poli

School of Computer Science and Electronic Engineering, University of Essex, Wivenhoe, U.K.

Keywords: Muscles fatigue, Transition-to-Fatigue, Genetic Programming, EMG, K-means.

Abstract: We propose the use of Genetic Programming (GP) to generate new features to predict localised muscles

fatigue from pre-filtered surface EMG signals. In a training phase, GP evolves programs with multiple

components. One component analyses statistical features extracted from EMG to divide the signals into

blocks. The blocks’ labels are decided based on the number of zero crossings. These blocks are then

projected onto a two-dimensional Euclidean space via two further (evolved) program components. K-means

clustering is applied to group similar data blocks. Each cluster is then labelled into one of three types

(Fatigue, Transition-to-Fatigue and Non-Fatigue) according to the dominant label among its members.

Once a program is evolved that achieves good classification, it can be used on unseen signals without

requiring any further evolution. During normal operation the data are again divided into blocks by the first

component of the program. The blocks are again projected onto a two-dimensional Euclidean space by the

two other components of the program. Finally blocks are labelled according to the k-nearest neighbours.

The system alerts the user of possible approaching fatigue once it detects a Transition-to-Fatigue. In

experimentation with the proposed technique, the system provides very encouraging results.

1 INTRODUCTION

The electro-myo-gram (EMG) is a test used to record

the electrical activity of muscles (Lieber 2002).

Muscles produce an electrical potential that is non-

linearly related to the amount of force produced in a

muscle. Analyzing these signals and associating

them with muscle state has been an area of active

research in the biomedical community for many

decades. For example, Sony (Dubost & Tanaka

September 2002) has presented a hardware system

for musical applications that is controlled by EMG

signals. The system is able to recognise different

gestures and associate them with particular

commands to the machine. Detecting muscle fatigue,

however, is still an extremely challenging task.

Atieh et al. (Atieh et al. December 2005) tried to

design more comfortable car seats by trying to

identify and classify EMG signals using data mining

techniques and statistical analysis to determine

localised muscle fatigue. Artificial neural networks

have been used to detect muscle activity by Moshoua

et al. (Moshoua, Hostensa & Papaioanno July 2005).

In that work, wavelet coefficients were proposed as

features for identifying muscle fatigue. Song and

collaborators (Song, Jung & Zeungnam 2006)

proposed an EMG pattern classifier of muscular

fatigue. The adaptation process of hyperboxes of

fuzzy Min-Max neural networks was shown to

significantly improve recognition performance.

In this paper we investigate the idea of

predicting localised muscle fatigue by identifying a

transition state which resides between the non-

fatigue and the fatigue stages within the EMG signal.

Genetic Programming (GP) has been used to

automate this process (Poli, Langdon & McPhee

2008).

As we will illustrate in the following sections, the

proposed approach was able to give an early warning

before the onset of fatigue in most of the

experimental cases we looked at. Therefore, this

approach shows some potential for application

domains such as ergonomics, sports physiology, and

physiotherapy.

292

Kattan A., Al-Mulla M., Sepulveda F. and Poli R. (2009).

DETECTING LOCALISED MUSCLE FATIGUE DURING ISOMETRIC CONTRACTION USING GENETIC PROGRAMMING.

In Proceedings of the International Joint Conference on Computational Intelligence, pages 292-295

DOI: 10.5220/0002315402920295

 SciTePress

2 THE METHODOLOGY

We try to spot regularities within the EMG data and

to associate them to one of three classes: i) Non-

Fatigue, ii) Transition-to-Fatigue, and iii) Fatigue.

Each class indicates the state of the muscle at a

particulate time.

The system works in two main stages: i)

Training, where the system learns to match different

signals’ characteristics with different classes, and ii)

Testing, where the system applies what it has learnt

to classify unseen data.

In the training phase, the system processes

filtered EMG signals and performs two major

functions: i) Segmentation of the signals based on

their statistical features, and ii) Classification of the

identified segments based on their types (i.e., Non-

Fatigue, Transition-to-Fatigue, or Fatigue). For these

tasks, GP has been supplied with a language that

allows it to extract statistical features from EMG.

Table 1 reports the primitive set of the system.

Table 1: Primitive Set.

Primitive Set Input

Median, Mean, Avg_dev, Std,

Variance, Signal size, Skew,

Kurtosis, Entropy, Zero crossings

Vector of real

number

+, -, /, *, Sin, Cos, Sqrt

Real Number

The system starts by randomly initialising a

population of individuals using the ramped half-and-

half method (Poli, Langdon & McPhee 2008). Each

individual has a multi-tree representation. In

particular, each individual is composed of one

splitter tree, and two feature-extraction trees. (Multi-

tree representations of this kind are common in GP,

and have been used, for example, for data

classification in (Estevez & Pablo 25-28 September

2007) ).

2.1 Splitter Tree

The main job of splitter trees is to split the EMG

signals in the training set into meaningful segments,

where by “meaningful” we mean that each segment

indicates the state of a muscle at a particular time.

The system moves a sliding window of size L

over the given EMG signal with steps of S samples.

At each step the splitter tree is evaluated. This

corresponds to applying a function, f

splitter

, to the data

within the window. The output of the program is a

single number, λ, which is an abstract representation

of the features of the signal in the window. The

system splits the signal at a particular position if the

difference between the λ’s in two consecutive

windows is more than a predefined threshold θ. The

threshold θ has been selected arbitrarily (θ = 10).

Once the data have been divided into blocks, the

system labels each block with one of the three

identified classes based on the number of zero

crossings in the raw EMG signal, i.e., the number of

times the signal crosses the zero-amplitude line

(details are in section 2.3). A good splitter tree

should be able to place block boundaries at the

transitions between three types of muscle states: i)

Non-Fatigue, ii) Transition-to-Fatigue, and iii)

Fatigue.

Preliminary tests showed that an average EMG

signal in our set has 50% of non-fatigue, 10%

transition-to-fatigue and the remaining 40% is

fatigue. Thus, the splitter tree can be considered to

be good if it divides the signal into the three types of

blocks with both meaningful proportions (i.e.,

fatigue > non-fatigue > transition-to-fatigue) and

meaningful sequence (non-fatigue should appear

before transition-to-fatigue and fatigue). Splitter

trees that violate these conditions are discouraged by

penalizing their fitness value (see section 2.4).

2.2 Feature-Extraction Tree

The main job of the two feature-extraction trees in

our GP representation is to extract features using the

primitives in Table 1 from the blocks identified by

the splitter tree and to project them onto a two

dimensional Euclidian space, where their

classification can later take place.

We used a standard pattern classification

approach on the outputs produced by the two

feature-extraction trees to discover regularities in the

training signals. In principle, any classification

method can be used with our approach. Here, we

decided to use K-means clustering to organise blocks

(as represented by their two composite features) into

groups. With this algorithm, objects within a cluster

are similar to each other but dissimilar from objects

in other clusters. The advantage with this approach is

that the experimenter doesn’t need to label the

training set. Also, the approach does not impose any

constrains on the shape of the clusters. Once the

training set is clustered, we can use the clusters

found by K-means to perform classification of

unseen data.

In the testing phase, unseen data go through the

three components of the evolved solution. Blocks are

produced by the splitter tree and then projected onto

a two-dimensional Euclidean space by the two

DETECTING LOCALISED MUSCLE FATIGUE DURING ISOMETRIC CONTRACTION USING GENETIC

PROGRAMMING

293

feature-extraction trees. Then, they are classified

based on the majority class labels of their k-nearest

neighbours. We use a weighted majority voting,

where each nearest neighbour is weighted based on

its distance from the newly projected data point.

More specifically the weight is w = 1 / distance (x

,), where x

is the nearest neighbour and z

is the

newly projected data point.

Once the system detects a transition-to-fatigue, it

alerts the user about a possible approaching fatigue.

2.3 Labelling the Training Set

The approach described above is based on an

unsupervised learning model. In our case the given

EMG signals in the training set are unlabelled. Here,

we used the zero crossings to recognise the state of

the muscle.

There are several ways to recognise muscle state

from the EMG signal. Some Authors (Mannion et al.

1997), (Finsterer August 2001) argue in favour of

the idea of counting the number of times the

amplitude of the signal crosses the zero line based on

the fact that a more active muscle will generate more

action potentials, which overall causes more zero

crossings in the signal. However, at the onset of

fatigue the zero crossings will drop drastically due to

the reduced conduction of electrical current in the

muscle.

Hence, our system decides the labels of the

blocks found by the splitter tree in the training set

based on the number of zero crossings of the EMG

signal. Before the system starts the evolution

process, it scans the training set signals and divides

them into blocks of predefined length. The number

of zero crossings from each block is stored in a

sorted vector that does not allow duplicate elements.

Then, the system divides this vector into three parts.

The lower 40% is taken to represent fatigue, the

middle 10% represents transition-to-fatigue, and the

higher 50% is non-fatigue. These three propositions

(10%, 40% and 50%) were selected based on the

preliminary tests with the EMG signals (as we

mentioned in Section 2.1). The numbers of zero-

crossings of the blocks in these three groups are then

used to identify three intervals. Later, the output of

the splitter trees is classified into one of these three

groups based on the interval in which the number of

zero crossings in it falls.

Despite the simplicity of this labelling technique,

the labels that it assigned to the output of the splitter

trees were consistent and were judged by an expert

to indicate the muscle state in most of the cases.

However, noise in some parts of the EMG signal can

result in wrong labels. Sometimes when the EMG

signal is noisy, the labels tend to not be in a

meaningful sequence and do not reflect the actual

state of the muscle. To solve this problem, we

carefully selected a training set with the least noisy

signals. Also, each fragment of signal in the training

set was manually screened to ensure that it had the

correct label.

2.4 Fitness Measurement

The calculation of the fitness is divided into two

parts. Each part contributes with equal weight to the

total fitness. The fitness contribution of the splitter

tree is measured as follows.

Splitter tree fitness is measured by assessing the

amount of help provided to the feature extraction

trees in projecting the segments into tightly grouped

and well separated clusters, plus a penalty if

required. More formally, the quality of the splitter

tree can be expressed as follows. Let f

Feature-extraction

the fitness of the feature-extraction trees, and µ a

penalty values:

Splitter

= f

feature-extraction

+ µ.

where µ is added if, for example, the splitter does

not respect the non-fatigue/transition-to-

fatigue/fatigue sequence. The value of µ is fixed.

The second part of the individual’s fitness is the

classification accuracy (with K-means) provided by

the feature-extraction trees. After performing the

clustering using K-means we evaluate the accuracy

of the clustering by measuring cluster homogeneity

and separation. The homogeneity of the clusters is

calculated as follows.

The system counts the members of each cluster.

Note that, each data point in the cluster represents a

block of the signal. Since we already know (from the

previous step) the label for each block, we label the

clusters according to the dominant members. The

fitness function rates the homogeneity of clusters in

terms of the proportion of data points – blocks – that

are labelled as the muscle’s state that labels the

cluster. The system prevents the labelling of

different clusters with the same label even in cases

where the proportions in two or more clusters are

equal.

The Davis Bouldin Index (DBI) (Bezdek & Pal

Jun 1998) was used to measure clusters separation.

DBI is a measure of the nearness of the clusters’

members to their centroids and the distance between

clusters’ centroids. A small DBI index indicates well

separated and grouped clusters. Therefore, we add

the negation of the DBI index to the total feature

extraction fitness in order to encourage evolution to

IJCCI 2009 - International Joint Conference on Computational Intelligence

294

separate clusters (i.e., minimise the DBI). It should

be noted that the DBI here is treated as a penalty

value, the lower the DBI the lower penalty applied to

the fitness.

Thus, the fitness of feature extraction trees is as

follows.Let H be a function that calculates the

homogeneity of a cluster and let CL

be the i

cluster.

Furthermore, let K be the total number of clusters

(three clusters in our case: fatigue, transition-fatigue

and non-fatigue). Then,

Homogeneity

DBI

)H(CL

−

∑

The total fitness of the individual is:

f = (f

feature-extraction

/2) + (f

Splitter

/2)

3 EXPERIMENTS

Experiments have been conducted in order to

investigate the performance of the proposed

technique. The aim of these experiments is to

measure the prediction accuracy with different EMG

signals.

3.1 EMG Recording

The data was collected from three healthy subjects

(aged 23-25, non-smoker, athletic background). The

local ethical committee approved the experiment’s

design. The three participants were willing to reach

physical fatigue state but not psychological fatigue.

Selection criteria were used to minimise the

differences between the subjects, which would

facilitate the analysis and comparison of the

readings. The participants had comparable physical

muscle strength. It was also preferred that the

subjects had an athletic background with a similar

muscle mass, which would facilitate the correlation

of the results. It was also important that the

volunteers were non-smokers, as it is known that

smoking affects physical abilities, which again could

lead to inaccurate readings of muscle fatigue in a

participant who smokes.

The participants were seated on a chair. Each

participant was asked to hold a weight training bar -

dumbbell – until their muscle fatigued. The steps in

our test bed set up are the following: 1) A bipolar

pair of electrodes was placed on the right arm's

biceps muscle for EMG recording. 2) The force

gauge was perpendicular to the dumbbell to ensure

that the force gauge was taking the correct reading.

3) A strap wass handed to participants to enable the

force gauge reading. 4) A protractor was used to

ensure a 90 degree angle of the elbow for the initial

setup. 5) A dumbbell was handed to the participant.

6) A laser was embedded in the dumbbell to give

visual guidance of the elbow angle. 7) The elbow

position was padded, so the participant was

comfortable.

The myoelectric signal was recorded using two

channels; Double Differential (DD) recording

equipment at 1000Hz sampling rate with active

electrodes on the biceps branchii during two

isometric dumbbell exercises, with 30% Maximum

Voluntary Contraction “MVC” and 80% “MVC”

respectively. For each of the three participants 6

trials were carried out, providing 18 trials in total.

3.2 GP Setup

Of the 18 EMG signals (trials) acquired, we used 3

trials for the training set, one trail from each subject.

In this way we hoped we would allow GP to find

common features for the three different participants

and build a general prediction model. The

experiments that are presented here were done using

the following parameters: population of size 100,

maximum number of generations 30, crossover with

probability of 90%, mutation with probability 5%,

reproduction with probability 5%, tournament

selection of size 5 and maximum tree depth of 10.

The performance of our approach has been

measured through 18 independent runs, each of

which trains the system and uses the output of the

training to predict the muscle state of 15 EMG

signals (5 signals for each participant). The aim is to

obtain a good prediction for each participant and a

reasonably general prediction algorithm that

performs well on average for all participants. Each

GP run results in one splitter tree and two feature

extraction trees. Since during offline tests it is

possible to know the actual label for each block by

counting its number of zero crossings, we compared

the GP predictions against the actual labels and

counted the proportion of times this was correct (hit

rate).

Table 2 reports the best achieved hit rate for each

test signal, as well the average hit rate in all runs.

Also, the worst hit rates are presented to show the

algorithm performance in its worst case. Moreover,

standard deviation for each signal is presented to

illustrate the system's robustness. The last row of

Table 2 reports the approximate average processing

time for both training and testing.

DETECTING LOCALISED MUSCLE FATIGUE DURING ISOMETRIC CONTRACTION USING GENETIC

PROGRAMMING

295

Table 2: Summary of performance of 18 different GP runs.

Signal Avg. Hit Best Hit Worst Hit Std.

60.36% 77.27% 37.84% 8.99

57.54% 77.36% 38.89% 9.03

60.69% 85.71% 39.39% 11.24

62.40% 100% 25.00% 19.31

57.96% 81.82% 9.09% 17.92

56.62% 90% 39.29% 13.10

62.73% 82.35% 41.18% 12.98

41.67% 100% 0.00% 28.58

65.08% 100% 0.00% 23.78

58.24% 100% 25.00% 20.00

67.08% 100% 42.31% 14.73

62.28% 78.79% 43.59% 8.27

66.05% 85.71% 40.00% 11.76

56.75% 100% 0.0% 16.56

65.34% 100% 16.56% 42.84

Average Testing Time

Average Training Time

1 min/test

signal

18 Hours

Figure 1 summarise the runs’ information. We

measured the quality of each run by calculating the

average hit rate for all test signals (test set). This

reflects the accuracy of the evolved programs when

dealing with signals from different participants.

More specifically, the first column in the histogram

in Figure 1 shows the average of the runs’ qualities

(i.e., the average of the averages). The second

column in the figure shows the quality (i.e., average

hit rate for all test signals) of the best evolved

predictor. The third column shows the quality of the

worst evolved predictor. Finally, the last column

shows the average standard deviation for all runs.

The low standard deviation with the reasonable

average prediction accuracy indicates that our

system is likely to produce accurate models within a

few runs.

Figure 1: Summary of 18 runs.

Table 3 reports the performance of the best

evolved predictor as well as the performance of the

worst, signal by signal. It should be noticed that the

achieved results are promising, especially

considering that the system is predicting muscle

fatigue for three different individuals.

Table 3: Best GP test run vs. worst GP test run.

Signal/ results Best run Worst run

66.67% 37.84%

66.66% 53.12%

75.47% 48.91%

25% 50%

66.67% 40%

90% 40.74%

81.25% 65.52%

50% 33.33%

100% 50%

100% 42.86%

100% 42.31%

78.79% 43.84%

75% 56.25%

80% 69.23%

100% 42.857%

4 CONCLUSIONS

Our results are encouraging, in the sense that a good

prediction has been achieved and further significant

improvements could be obtained.

Although, this approach has achieved good

prediction rates (100% in some cases), it suffers

from a major disadvantage. The proposed labeling

mechanism – being based on zero crossings – is not

very reliable

(as described in section 2.3). This is

due to the fact that the number zero crossings is

affected by noise. Thus, the labels of the signal’s

blocks might not be accurate. This has the potential

to hamper or prevent learning in the system. We

avoided this problem altogether by carefully

selecting the least noisy signals for the training set

and suitably preprocessing the signals. Nevertheless,

there remains the possibility that some errors in the

labeling still occur.

There are many directions for us to further

improve the performance of this technique. For

example, a simple extension in the set of statistical

functions available in the primitive set might

IJCCI 2009 - International Joint Conference on Computational Intelligence

296

improve the classification accuracy. Also, the use of

more sophisticated technique to label the EMG

blocks (e.g., fuzzy classification) might improve the

system’s reliability and flexibility.

REFERENCES

Atieh, M, Younes, R, Khalil, M & Akdag, H December

2005, 'Classification of the Car Seats by Detecting the

Muscular Fatigue in the EMG Signal', International

Journal of Computational Cognition.

Bezdek, JC & Pal, NR Jun 1998, 'Some new indexes of

cluster validity', IEEE Transactions on Systems, Man,

and Cybernetics, Part B: Cybernetics, vol 28, no. 3,

pp. 301-315.

Dubost, G & Tanaka, A September 2002, 'A Wireless

NetWork based Biosensor Interface for Music', Proc.

of Intern. Computer Music Conf.

Estevez, NB & Pablo, A 25-28 September 2007, 'Genetic

Programming-Based Clustering Using an Information

Theoretic Fitness Measure', IEEE Congress on

Evolutionary Computation, IEEE Press.

Finsterer, J August 2001, 'EMG-interference pattern

analysis J', Journal of Electromyography and

Kinesiology, vol 11, pp. 231-246(16).

Lieber, RL 2002, Skeletal muscle structure, function &

plasticity: the physiological basis of rehabilitation,

2nd edn, Lippincott Williams & Wilkins.

Mannion, AF, Connolly, B, Wood, K & Dolan, P 1997,

'The use of surface EMG power spectral analysis in the

evaluation of back muscle function', Journal of

Rehabilitation, vol 34, pp. 427-39.

Moshoua, D, Hostensa, I & Papaioanno, G July 2005,

'Dynamic muscle fatigue detection using self-

organizing maps', Applied Soft Computing, Elsevier.

Poli, R, Langdon, WB & McPhee, N 2008, A field guide to

genetic programming, http://lulu.com.

Song, J-H, Jung, J-W & Zeungnam, B 2006, 'Robust EMG

Pattern Recognition to Muscular Fatigue Effect for

Human-Machine Interaction', Springer.

DETECTING LOCALISED MUSCLE FATIGUE DURING ISOMETRIC CONTRACTION USING GENETIC

PROGRAMMING

297