This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.
Essential tremors (ET) are slow progressive neurological disorder that reduces muscular movements and involuntary muscular contractions. The further complications of ET may lead to Parkinson's disease, and hence it is important to detect at the early onset. This study deals with the identification of the presence of essential tremors in the EMG of the patient and the linear algebraic decomposition of the signal is treated as feature selection parameter to distinguish Essential tremors from the normal EMG. Levinson's recursive decomposition and AR Burg decomposition algorithms are applied to extract the correlation features for calculating autoregressive coefficients and two non-linear classifier models, Support Vector Machines and Multi Layer Perceptron neural networks are applied. A medical decision support system is developed by integrating the features and classifiers and optimality is arrived by estimating the evaluation parameters, such as sensitivity, specificity and classification accuracy. The experimental study shows that the AR Burg feature with MLP yields a classification accuracy of 98%. This feature network combination can effectively be applied real time for the early detection of ET from any EMG signal.
Index Terms- essential EMG tremors; automated classification; MLP; SVM
Electromyography (EMG) is the estimation of the electrical activity of the muscles and this provides a quantitative measure of the muscular contraction. Essential tremors (ET) are caused due to genetic mutation and transferred to future generations through Autosomal Dominant Transmission . Though ET originates in the Central Nervous System (CNS) it can visually observed by the involuntary muscular contraction in the subject's limb muscles. Postural tremor of the outstretched arms, intentional tremor of the arms and rest tremor in the arms are also very common in ET . The Essential tremor has a frequency range of 4-12 Hz and it is observed mostly when the affected muscle is under work. Stress, when experienced physically or mentally may worsen the tremors [3-5].
Essential tremors and Parkinson's disease differ in the fact that ET starts at the Cerebellum of the brain whereas PD originates due to the degeneration of the Hypothalamus. Clinical studies have shown that patients affected with ET may gradually develop PD as the risk of Parkinsonism is greater with patients with ET. ET is generally postural and action tremors and they originate bilaterally and the tremor duration is only 1-2 sec, which is very short when compared with the resting tremors in PD that lasts for 8-9 seconds .
Researchers so far had conducted many tests to successfully differentiate PD and ET, but it still proves to be a serious matter of clinical research. They have successfully classified the various stages of the Essential tremor as definite Essential tremor, Probable essential tremor and Possible ET with the progress of tremors from head and neck in the first stage to arms in the second and the third stage is the tremors present during action and rest or continuous tremors in the arms  which can then progress to any part of the body. This research study suggests a decision support system for detection of Essential tremors. Fig.1 shows the schematic of automated classification.
The features are extracted using the autoregressive Burg spectral estimation and Levinson's-Durbin algorithm. The AR model is a statistical analysis tool applied for linear prediction of the signal under consideration and Levinson's-Durbin recursion is an algebraic tool for the signal decomposition and analysis. Both the above mentioned methods focus on the reduction of the forward and backward projection errors and improvising the accuracy of discrimination of the signals they work on.
Fig.1 Basis of automated classification
MATERIALS AND METHODS
The electromyogram signals are recorded using Nihon Kohden MEB-2200 EMG system at a sampling frequency of 100 samples per second. The signals are obtained from the Neurology department of Sri Ramachandra Medical University and Research Center, Chennai, India. The Essential Tremor EMG is recorded from various patients of age group 20-40 and the normal EMG from subjects of same age group under rest and activated muscular contraction. The tremor signals are recorded from abnormal subjects, diagnosed manually as patients with ET and earlier stages of ET under rest conditions (resting tremors) and tremors during muscular contractions. Normal EMG is obtained from induced muscular contractions and muscle at rest. The obtained EMG data are ensured to be free of noise and any motion artifacts or power line interference. The signals are obtained for 30 minute duration under continuous monitoring to avoid any external noise interference. Fig.2 illustrates the sample recordings.
Fig.2 Sample recordings of EMG obtained from (a) normal/healthy subject (b) subject with ET
The obtained EMG signals are first segmented into 1 sec data. The continuous non-stationary EMG signals are segmented in order to ensure the stationarity of the signals as well as the feature extraction accuracy increases with short segment duration. Signal decomposition algorithms like the Burg spectral estimation using autoregressive model and the Levinson's- Durbin algorithm are applied to extract the classification features from the given signal.
AR Burg model:
The autoregressive model of the signal x is given by the equation:
Where are the parameters of the model with c as its constant and is the error component.
The AR model computes the parameters of the signal first and then computes the spectral estimate from the extracted parameters. Since this method is easier to estimate the spectrum of the signal AR model is the widespread algorithm. The AR model of order p is given by,
Where the coefficients of the signal are denoted by a (k) and the white noise by w (n). w (n) is an estimate of . To make the output to be stable using the AR method factors like model order, length of the segment under valuation and the stationarity of the signal are essential[8-19].
This method clearly concentrates on estimating the reflection coefficients of the signal and to minimize the forward and backward projection errors. The reflection coefficient is estimated from:
The projection errors are given by:
…… (4) Forward prediction error
…….. (5) Backward projection error
The sum of the forward and backward projection errors are gives the total least squares error.
Levinson's- Durbin algorithm:
This method is algebraic method to analyze the signal proposed by Levinson and further improvised by Durbin. This is a recursive algorithm and it is most widely applied because of its ease to be understood and for signals of short duration can be analyzed faster than any other algorithm. This method is mainly applied for Toeplitz matrices or a constant diagonal matrix. This method progresses by estimating the forward and backward vectors. The forward vectors are used to estimate the values of the backward vectors, which then help in deriving a solution to the problem [19-20].
….. Forward vector. (6)
…… backward vector. (7)
The end solutions for these vectors are given by:
Forward vector … (8) and the backward vectors are obtained from the forward vectors as … (9). The solution to the given time series or the EMG signal is also obtained in the same recursive manner.
Fig.3 Feature extraction using AR burg
Fig.3 figure indicates the feature values of the Essential Tremor and the normal EMG signal extracted using AR Burg algorithm. The plot clearly indicates that the patient with ET have higher AR Burg feature values than those of the normal. It also indicates that the values of ET are highly random and show clear abnormality when compared to the normal subject's features. Fig.4 shows the plot of the features extracted using Levinsons algorithm.
Fig. 4 Signals after parametric evaluation
Figure 4 plots the feature values extracted using the Levinson's-Durbin recursion algorithm. This algorithm also very distinctly discriminates the essential tremor from normal EMG signal. These features are then fed into artificial neural network, Multi Layer Perceptron and Support Vector Machines for classification and performance evaluation. This evaluation is carried out in order to highlight the feature that gives better classification accuracy and ease in computation.
Artificial Neural network is a complex network of interconnected nodes or neurons that perform computational tasks. Simple classification from individual neurons makes it a powerful discriminating tool. The network, when fed with an input signal, processes it, adapts itself to the learning algorithm and then accurately discriminates the input . The adaptability of the network depends on the training algorithm and the number of hidden neurons that executes the computational task.
The architecture of any neural network consists of an input layer, where the data that is to be discriminated is fed into the network and the last layer is the Output layer, which gives the classification results. The layer in between the input layer and the output layer is the hidden layer that is fully occupied by a complex set of interconnected neurons. The number of neurons in this hidden layer determines the classification accuracy of the network. Small number of neurons can effectively classify only small and simple data sets while too many neurons will make the network topology too complex and will levy a major setback in terms of computational complexity. It is very essential to design a neural network with optimal configuration for successful simple classification with enhanced accuracy. The number of hidden layers and the number of neurons in the hidden layer can be drawn to a conclusion only by trial and error basis. The methods to compute this number of hidden neurons have been successfully described in many research articles [22-23].
Fig.5 Neural Network Architecture.
Support vector machines are another set of linear classifiers. Linear classifiers are the neural networks that discriminates the data sets with the input of linear combination of multiple features. They are also called as margin classifiers as they describe each data in a data set as points and classify those points into a plane in which they are trained to belong to.
SVM denotes all the data as points in space and separates these data sets into planes in space denoted as Hyperplanes. These hyperplanes are the classification barrier between the two major types of vectors (when only two patters are employed). Classification accuracy of the SVM is described the best as it provides a distinct band gap between the two vectors. The classification errors are denoted by the placement of misclassified vectors in the opposite planes .
In general, the SVM creates a single or collection of hyperplanes in an infinite space, which are applied for classification and regression. Better classification is accomplished with a hyperplane, which is farthest to the nearest training vectors or datasets.
Unlike the usual neural network models the SVM does not control the network complexity by compromising on the number of features it can process. The normal neural nets apply back-propagation algorithm and converges to an optimal local minima and therefore SVM proves to be advantageous in this regard as it focuses on high level of accuracy with optimal configurations and it also deviates from the traditional approach as it does not have the local minima approach. SVM select their own model sizes with their selection of the support vectors thereby adaptability increases .
For the classification of the Essential tremors from the normal EMG data sets the algebraic and the linear recursion features extracted through AR Burg and Levinson's -Durbin algorithm are applied. The numbers of patterns applied for training are 1600 for the Multi layer percptron network and the best optimal configuration is drawn into conclusion. After the training is accomplished the neural network efficiency, classification accuracy and the computational complexity are evaluated using 1000 test patterns.
Fig. 6 Classification of the ET vs Normal EMG using Multi Layer Perceptron
The above figure indicates the performance of the Multi Layer Perceptron for the classification of Essential tremors which provided a classification accuracy of 98.1%.
The features were also examined using the support vector machines and the performance of the same was evaluated.
Fig.7 Classification using SVM
Table 1 Classification of patients with ET against Normal EMG
SVM (in %)
MLP (in %)
Subject 1 with ET vs. Normal
Subject 1 with ET vs. Normal
Subject 2 with ET vs. Normal
Subject 2 with ET vs. Normal
Subject 3 with ET vs. Normal
Subject 3 with ET vs. Normal
In the Table 1 the classification of various subjects with Essential tremors are tested against the normal EMG. It is observed that the best feature for classification is found to be the AR Burg method of feature extraction when evaluated using the Multi layer perceptron neural network. The optimal configuration was arrived with various trials and best method is observed to give a classification accuracy of 98 %.
A medical decision support system for the detection of essential tremors from EMG recordings has been proposed. From the above experimental study, it becomes more evident that it is very essential to identify the ET and its onset as the later stages of ET may consequently lead to Parkinson's disease. This evaluation results also yield a solution for real time implementation of the feature network combination to diagnose essential tremor and provide a scope for treatment and earlier recovery for the patients suffering from ET. We draw to a conclusion that the AR Burg feature extraction procedure with the Multi Layer Perceptron model for classification is ideal to be implanted for automated diagnosis of Essential Tremor and its stages and provide better healthcare with prevention from further complications.