Diagnosis of parkinson's disease

AUTOMATED DETECTION OF PD RESTING TREMOR USING PSD WITH RECURRENT NEURAL NETWORK CLASSIFIER

ABSTRACT

Diagnosis of Parkinson's disease (PD) is a challenging problem for medical community. Typically characterized by tremor, PD occurs due to the loss of dopamine in the brain's thalamic region that results in involuntary or oscillatory movement in the body. The early stage of the PD is referred as resting tremors, which appears when the muscles are relaxed. It is well known that surface EMG recording provides clinical information on the neuro-physiological characteristics of the tremors. This paper discusses the detection of resting tremors by extracting power spectral density features from EMGs. Two methods namely, PSD by Welch and Burgs are applied by configuring the order of the predictors and are then classified using a recurrent neural network model, Elman Neural Network (REN). Experiments are performed using EMG patterns and statistical measures such as mean and maximum of PSD are used to classify the normal and abnormal PD subjects. It is found that the mean value of power spectral density by Welch with recurrent neural network yields a classification accuracy of 90.1%. The proposed work need to be validated with larger datasets for real -time clinical application.+ +

Index TermsParkinson's Disease; EMG tremors; power spectral density; recurrent neural network

I. INTRODUCTION

Electromyography(EMG), the measure of electrical activity produced by skeletal muscles, is one of the major diagnostic parameter for detection of Parkinson's disease (PD) [1-3] . The brain is the master controller of the body activities and that includes the motor activities as well. The degeneration of the hypothalamus in the brain leads to very severe complications of which Parkinson's disease(PD) is the most widespread. PD is characterized by muscle rigidity, resting muscular tremor that is very rare for normal subjects, a slowing of motor action (bradykinesia) and a loss of muscular contraction that leads to loss of the entire motor activity (akinesia)[2-4] in extreme cases. Reduced motor activity in this disease makes it detectable with the help of EMG measurements from the patient.

The early indication of the PD is the resting tremor with a trembling or shaking in one of the hands. This is due to the involuntary action of the muscles. This muscle activation is well exploited by investigating the EMGs and several work have been reported in the literature for the diagnosis of PD [6-13]. The alternating properties of the EMG can be exploited properly, provided approximate features are extracted. Time and frequency domain features such as frequency spectrum estimation, amplitude and the frequency band in which maximum signal contribution have already been reported [6-13]. In this research study the importance of power spectral information is investigated by using two methods, namely, PSD by Welch and PSD by Burgs. Fig.1 shows the proposed schematic for detection of PD. The optimal features based on statistical means are then classified using Recurrent Neural network model, Elman network. Then the classifier accuracy is estimated based on the network performance in recognizing the true-false positive and negative patterns respectively.

II. Materials and Methods

A. Data source:

For experimental study the EMG are recorded at the Neurology Department, Sri Ramachandra University, Chennai, India. The subjects under the age group of 20-30 years are selected and recorded under rest and activated motion from the extensor carpi radialis muscle. Resting tremors are recorded from subjects diagnosed by physicians as PD and induced muscular contractions are recorded from normal subjects. All the EMG data are free from artifacts and external power line interference. The EMG recordings are for a restricted duration of 30 min with sampling frequency of 100 Hz. Fig.2 Shows the sample recordings

B. Feature Extraction

In order to characterize the tremor accurately, the entire EMG time series is segmented into 1s, say patterns. Furthermore this ensures stationarity of the signal and time-frequency domain parameters can be extracted subsequently. The PSD of the EMG signal is evaluated through autoregressive Burgs method and Welch method of estimation [14-16].

PSD using AR Burg:

The Burg Method block estimates the power spectral density (PSD) of the input frame by estimating the reflection co-efficients. This method fits an autoregressive (AR) model to the signal by minimizing (least squares) the forward and backward prediction errors while constraining the AR parameters to satisfy the Levinson-Durbin recursion and producing an optimal combination of both these errors. The input is the sampled EMG data representing a frame of consecutive time samples from a single-channel signal. The output is the estimate of the signal's power spectral density at NFFT equally spaced frequency points in the range [0,Fs), where Fs is the signal's sample frequency.

(1)

The burg spectrum is advantageous in the fact that it minimizes the forward and backward prediction errors in the least squares sense, with the AR coefficients constrained to satisfy the L-D recursion and it always produces a stable model.

Welch method:

It is an improved method of estimating the PSD [14-16]. This method consists of dividing the time series data into (overlapping) segments, The original data segment is split up into L data segments each containing N samples overlapping by D points.

  1. If D = N / 2, the overlap is said to be 50%
  2. If D = 0, the overlap is said to be 0%.

This method divides the data into eight segments by default with a maximum of 50% overlap between them and uses a Hamming window. Normally the windowing function affects the computation of the peridogram at the centre of the segment than at the edges, which results in a loss of information. To avoid this loss of information, overlapping of the segments is carried out. Discrete Fourier transform is applied to compute the periodogram of each segment. Then the estimated periodogram is time-averaged, in order to reduce the variances of individual power estimates. This result is called the Welch estimate. Although overlap between segments tends to introduce redundant information, this effect is diminished by the use of a nonrectangular window, which reduces the importance or weight given to the end samples of segments (the samples that overlap).

(2)

III PERFORMANCE EVALUTION

To detect the PD from the given EMG time series, neural network model is incorporated to detect the normal and PD patterns. A feedback Elman neural network model is used for the classification . In order to evaluate the statistical importance of the proposed power spectral density features, the predictor order for Welch and Burg's method are varied and mean as well as maximum value of PSD for each segment(1s duration) is calculated. This process will provide the abrupt variations of EMGs for the abnormal cases. The performance of the proposed scheme is evaluated in terms of sensitivity(SE) , specificity (SP) and classification accuracy (CA).

(3)

(4)

(5)

where the True Positives refers to correctly detected normal EMG patterns and True Negative refers to correctly detected PDs.

Due to the presence of feedback connections in Elman network from the output to the context nodes, the robust changes in values are reflected and a high accuracy of classification can be possibly achieved. Fig.3 shows the model of recurrent Elman neural network [17 ]. The network is configured optimally with hidden neurons =60, tan sigmoid and log sigmoid functions for input- hidden and hidden-output layers respectively, back propogation gradient descent momentum as learning algorithm. For training the network, 2000 EMG patterns are used and 1200 for testing the efficiency of the network. Figs 4 and 5 show the variation of PSD for normal and PD EMGs.

It can be seen from the Table 1 that the classifier performance varies based on the order of the predictor as well as type of PSD estimation algorithm. Among the different modes, PSD- Burg (max) yields the CA of 95.48%. Fig. 6 shows the classifier results obtained using PSD-Welch with Elman neural network.

IV Conclusions

This paper discusses the automated detection of resting tremor characterizing Parkinson disease using power spectral density features with recurrent Elman neural network. The muscular activations were studied using the EMG recordings and mean and maximum of power spectral densities were estimated based on the Welch and Burg's algorithms. It can be seen from the experimental results that the proposed scheme yields promising classification results. To validate the efficiency for clinical usage, attempts are being made to test with lager datasets.

ACKNOWLEDGEMENT

The research results presented here are based on data obtained from the Neurology Department, Sri Ramachandra University, Chennai, India.

REFERENCES

[1] A. Meigal, S. Rissanen, M. Tarvainen, P. Karjalainen, I. Iudina-Vassel, O. Airaksinen, M. Kankaanp, Novel parameters of surface EMG in patients with Parkinson's disease and healthy young and old controls Journal of Electromyography and Kinesiology, Volume 19, Issue 3, Pages e206-e213

[2] J Jankovic, S Fahn . Physiologic and pathologic tremors. Diagnosis, mechanisms, and management. Ann Intern Med 1980;93:460-5.

[3] W. H Oertel, S Fahn. Parkinsonism. In: T Brandt, LR Caplan, J Dichgans, HC Diener, C Kennard, editors. Neurological disorders course and treatment. San Diego: Academic Press; 2003.

p. 1021-79.

[4]J.A Robichaud , K.D Pfann, S.Leurgans , D.E Vaillancourt, C.L. Comella, D.M Corcos. Variability of EMG patterns: a potential neurophysiological marker of Parkinson's disease? Clin Neurophysiol. 120(2):390-7, 2009

[5]David E. Vaillancourt, Karl M. Newell,The dynamics of resting and postural tremor in Parkinson's disease, Clinical Neurophysiology, Volume 111, Issue 11, November 2000, Pages 2046-2056

[6]Shou-Yan Wang, Tipu Z. Aziz, John F. Stein, Xuguang Liu,Time-frequency analysis of transient neuromuscular events: dynamic changes in activity of the subthalamic nucleus and forearm muscles related to the intermittent resting tremor Journal of Neuroscience Methods, Volume 145, Issues 1-2, 30 June 2005, Pages 151-158

[7]Mehmet Engin, Serdar Demirag, Erkan Zeki Engin, Grbz elebi, Fisun Ersan, Erden Asena, Zafer olakoglu,The classification of human tremor signals using artificial neural network ,Expert Systems with Applications, Volume 33, Issue 3, October 2007, Pages 754-761

[8]A. Hossen, M. Muthuraman, J. Raethjen, G. Deuschl, U. Heute,Discrimination of Parkinsonian tremor from essential tremor by implementation of a wavelet-based soft-decision technique on EMG and accelerometer signals,Biomedical Signal Processing and Control, In Press, Corrected Proof, Available online 24 April 2010

[9] S Pan., K Warwick, J Burgess, M.N Gasson, S.Y Wang, Aziz and, T. Z., et al.(2007). Prediction of Parkinson's disease tremor onset with artificial neuralnetworks. In Proc.5th IASTED international conference on biomedical engineering (pp. 341-345).

[10] M Gasson, T Aziz., K Warwick, S Wang, J Stein. (2005). Towards a demand driven deep-brain stimulator for the treatment of movement disorders. In Proceedings of the 3rd IEE international seminar on medical applications of signal processing, London (pp. 83-86).

[11] D. Wu, K. Warwick, Z. Ma, J. G. Burgess, Song Pan, Tipu Z. Aziz. Prediction of Parkinson's disease tremor onset using radial basis function neural networks. Expert Systems with Applications, Volume 37, Issue 4, April 2010, Pages 2923-2928

[12] Lauk, M., Timmer, J., Lu cking, C.H., Honerkamp, J. andDeuschl, G., A software for recording and analysis of human tremor, Comput. Methods Programs Biomed., 60,(1999) 65-77.

[13]Abdulhamit Subasi , Mustafa Yilmaz , Hasan Riza Ozcalik ,Classification of EMG signals using wavelet neural network,Journal of Neuroscience Methods 156 (2006) 360-367

[14] Rabiner, L.R., and B. Gold. Theory and Application of Digital Signal Processing. Englewood Cliffs, NJ: Prentice Hall, 1975. Pgs. 399-419.

[15] Welch, P.D. "The Use of Fast Fourier Transform for the Estimation of Power Spectra: A Method Based on Time Averaging Over Short, Modified Periodograms." IEEE Trans. Audio Electroacoust. Vol. AU-15 (June 1967). Pgs. 70-73.

[16] Oppenheim, A.V., and R.W. Schafer. Discrete-Time Signal Processing. Englewood Cliffs, NJ: Prentice Hall, 1989. Pgs. 311-312

[17] S.Marra and F.C Morabito,F.C , A New Technique for Solar Activity Forecasting using Recurrent Elman Networks, proceedings of world academy of science, engineering and technology, vol.7,pp.68-73,2005

Please be aware that the free essay that you were just reading was not written by us. This essay, and all of the others available to view on the website, were provided to us by students in exchange for services that we offer. This relationship helps our students to get an even better deal while also contributing to the biggest free essay resource in the UK!