EMG Signal Feature Extraction, Normalization and Classification for Pain and Normal Muscles Using Genetic Algorithm and Support Vector Machine

Electromyography (EMG) is a diagnostic method by which specialists evaluate the functional state of skeletal muscles and peripheral nerve endings. The assessment is based on the level of their electrical activity. To conduct EMG, an electromyography is used - an apparatus that enhances and records the bio potentials of the neuromuscular system [1]. Modern computer devices record even the minimum values of electrical impulses, automatically read the amplitude and frequency of periods, and also perform their spectral analysis. The device consists of a complete computer system capable of recording certain signals (bio potentials) of muscle tissue [2]. Using the device, bio potentials are strengthened, which helps doctors to determine the degree of damage to muscle tissue without a surgical diagnostic operation. Diodes are attached to the computer system that record deviations from the norm. Using the apparatus, the signal is amplified, and an image is displayed on the screen that displays the state of the muscle tissue and peripheral nerves of the body area under study [3]. Modern devices display the image directly on the monitor, but the old generation electromyography captures the received pulses on paper. It has been observed that exist several techniques to process complex EMG signal that is assisted by EMG classification using either of the Artificial Neural Networks (ANN), Multi-Layer Perceptron (MLP), Support Vector Machines (SVM), Linear Discriminant Analysis (LDA) and K-Nearest Neighbor (KNN). Researchers have put their effort in order to classify the EMG data and have identified the major issues in the classification as

a) Preprocessing and Data Selection

b) Training and Classification

Following aspects have been discovered in order to classify an EMG signal.

1. The signal contains many artefacts for the same type of disease as the signal varies due to the small movement of human during the recording process of signal.
2. The presence of noise also increases the complexity and hence difficult to classify the signal.
3. Optimal set of data for each class will lead to better classification accuracy.
4. The machine learning schemes required signal with optimal and refined data so that training and then classification can be performed in a better way [4].

To solve the above defined problems, a new system has been designed using signal Attribute selection, feature optimization with classification techniques.

1.png

 Figure 1. (a) Non- invasive (b) Invasive techniques

EMG is used to detect the muscular information based on the contraction and expansion period of muscles through the electrodes, which are placed on the surface of the human body. The process of gathering information can be invasive or non- invasive. The process of conducting non-invasive surface EMG is performed using electrodes whereas the invasive EMG required a needle, which is being inserted into the patient body for collecting the muscle related information. Both the processes are shown in Figure 1.

The EMG model can be simply represented by Eq. (1).

$y(n)=\sum_{t=0}^{N-1} h(t) e(n-t)+x(n)$    (1)   

y(n)$\rightarrow$ Modelled EMG signal.

e(n)$\rightarrow$ Fringing impulse.

h(t)$\rightarrow$ Motor Unit Action Potentials (MUAPs), which is used to provide significant source of information that is useful for diagnosis of neuromuscular disorders.

x(n)$\rightarrow$ While Gaussian Noise.

n $\rightarrow$ Number of motor unit.

This research deals with the Attribute selection and classification of the EMG signal into two categories that is either pain or normal activity of muscles.

The entire work is shown in Figure 2, which consists of three main parts such as attribute row selection using genetic Algorithm (GA), Similarity measure using Cosine Similarity, application of critic method to normalize features and classification using Support Vector machine (SVM).

3.1 Dataset

The dataset is collected from https://www.kaggle.com/nccvector/electromyography-emg-dataset link. The considered dataset contains pain and normal muscular data of 1000 row for each category with seven different un-named attributes. The electric potential observed is as like presented in Table 1.

Table 1. Dataset

Where, t is the time interval for time ranging from t1 to t8 measured in miliseconds.

2.png

Figure 2. Flow of proposed work

3.2 Attribute selection

Attribute selection of EMG signal is performed on each recorded movement data per session using the electrodes placed on the human body. The amplitude variation in the EMG is high when any movement is occurs otherwise the signals are in rest and amplitude variation reduces. EMG signal, each voltage-time amplitude value need Attribute selection to find out the relevant features for painful data or normal data to achieve better classification accuracy. The benchmark for the selection and rejection is relative to the value which is being used. It means, the selection and rejection will depend upon the other relative values available in that class of other patients.

In this research, attribute selection of the uploaded EMG signal is performed using nature inspired Genetic Algorithm. GA selects rows among the available dataset as per the designed fitness function represented by Eq. (2).

Fitness function$=1 \quad(1-e) \times F s>F t$

0 Otherwise            (2)

where, Fs = Current Attributes,

            Ft = All Attributes Row Values.

Each row has been tested using Eq. (2). If the row satisfied the fitness function, then categorize it either pain or normal. Otherwise, repeat the process for next row. Using this process, the unwanted signal has been removed and the row also obtained in reduced and desired form. The process of GA is shown in Figure 3.

3.png

Figure 3. GA process

Step 1: Initialization of Population (Rows): Initialize, population string (raw EMG data) known as chromosomes. To resolve problem that is to select EMG signal, which is being recorded during motion and to reject the EMG signal, which is recorded at rest position has been performed based on the designed fitness function.

Step 2: Selection: In this step, the selection of appropriate signal, which is higher than fitness value is eliminated and those less than or equal to fitness value are selected. Those values of rows that have lowest value are known as parent and contributed to the generation of new member named as children.

Step 3: Mutation: helps to search with the best row selection based on mutation threshold.

Step 4: Termination: The Attribute selection process is terminated while the desired rows are selected and categorized as pain and normal signal [17, 18].

Feature selection using GA

3.3 Cosine similarity

Cosine similarity is being applied on the selected rows, which return a single similarity index for each row. Suppose, the selected rows after GA of 700 has been obtained from the available 1000 rows for both pain and normal EMG signals. Therefore, similarity has been measured for each row with the remaining 699 rows. In this way a single or average value has been obtained using Eq. (3).

Avg $\operatorname{sim}=\frac{\sum_{i=1}^{n} \cos _{\operatorname{sim}}}{n}$            (3)

where, n is total number of similarities

After obtaining the average value, 20% of value has been added and subtracted to the obtained average similarity value as denoted by $b_{1}$ and $b_{2}$ [19]. Mathematically, can be represented by Eq. (4);

$B_{1}=\operatorname{Avg} \operatorname{Sim}+\frac{A v g \operatorname{sim} N o.}{100}$

$B_{2}=\operatorname{Avg} \operatorname{sim}-\frac{\operatorname{Avg} \operatorname{Sim} \mathrm{No} .}{100}$            (4)

If the value of $B_{1}$ and $B_{2}$ lies between the similarity value and those values are used for the training and classification purpose using SVM approach. Otherwise drop the data [20].

Cosine similarity

3.4 Critical method

Critic method is used to normalize the feature vector attained from the application of Genetic Algorithm [21]. Critic method acts on three elements as follows:

1. The current state of value

2. The maximum value of the section

3. The minimum value of the section

Following pseudo code is applied in order to implement the critic method.

Application of Critic

The critic method takes the current attribute value and checks it with the average value of other members which is called the selection value in this case. If the attribute value is less than the threshold value, it checks for the maximum and minimum value of this attribute in all the available rows of the category. The current attribute value is subtracted from the min value and is divided by a total of max and min value. The outcome is replaced by the current attribute value.

3.5 Support Vector Machine (SVM)

SVM is a machine learning approach used to classify EMG signal as pain and normal. Using this approach, a hyperplane has been constructed to distinguished two different data as normal and pain in this case. To train SVM, let the input train data be as: $\left(a_{1} b_{1}\right),\left(a_{2} b_{2}\right), \ldots,\left(a_{m} b_{m}\right) \epsilon P^{N} \times\{-1,+1\}$.

$a_{i} \rightarrow$ input value,

$b_{i} \rightarrow$ Assigned class to which input belongs {-1,+1}.

In case, if the input data is not separated linearly, then a transform of ($\varphi: P^{N} \rightarrow P^{M}$) has been used with a new feature space represented by $P^{M}$.

Using this as a function, the obtained hyperplane can be separated as per the Eq. (5);

$\omega \times \varphi(a)+b=0$

$\omega \in P^{M}$ and $b \in P$           (5)

The training can be said best with optimal hyperplane and a minimum error. In case, if the signals are too close or overlap with each other, then a kernel function is used to separate that data. The kernel function might be Radial Basic Function (RBF), polynomial, linear, a Gaussian etc. [1, 22].

The training and testing using SVM is shown in Figure 4.

Classification using SVM

4a.png

4b.png

Figure 4. Training and testing of EMG Signal using SVM