Detection of K-Complex in Sleep Eeg Signal with Matched Filter and Neural Network

Detection of K-Complex in Sleep Eeg Signal with Matched Filter and Neural Network

Prof V.V.Shete1, Mrs. Sapana Sonar2 , Ms. Ashwini.Charantimatp, Mr. Sachin Elgendelwar4 MIT College of Engineering,Pune,Maharashtra

Depart ment of Electronics and Telecommun ication

Abstract: The k-complex is a transient waveform that contributes in assessment of sleep stages. Main problems in assessments are similarity of k-complex to the other waves in EEG signal, which is buried in noise. To overcome these problems, methods based on reference signal, such as, matched filtering, neural network, etc. are used. Detection of a particular event,

1. k-complex with a known pattern in a noisy EEG signal

   is the subject of this paper.

   Keyword: – k-co mple x, matched filter, Art ificia l Neural Network (A NN).

   as, sounds, touches on the skin, and internal ones. K- comple x waveform shapes can vary severely due to the unstable structure of EEG signal. Visual recognition of k- comple x in an all night sleep EEG is time consuming and difficult task [2].

   1. Introduction:

      Electroencephalograph (EEG) is the spontaneous electrica l activ ity along the scalp. German physiologist and psychiatrist Hans Berger recorded the first human EEG in 1924[1]. These brain waves are usually recorded on a mu ltichannel recorder fro m electrica l signal which is amp lified and monitored by electrodes placed on several locations on the scalp.

      K-co mp le x is a transient wave of sleep stage II in EEG signal, wh ich has sharp positive wave followed by negative wave. In sleep analysis, the k-comp le x is one of the key features used for determination of diffe rent stages of sleep, stage II in sleep is detected by the presence of one or mo re k-co mple xes [2]. They are mo re frequent in the first cycle of sleep. It is having frequency range from 0.5 Hz to 2 Hz and amplitude > 10 0v. K-co mp le xes are roughly occurring at every 1.0 1.7 minutes and are often followed by bursts of sleep spindles. They are generated in widespread cortical locations, though they tend to predominate over the frontal parts of the brain. They occur spontaneously but also in response to external stimuli, such

      ———————————————————————————-

      Fig.1:- k-co mp le x followed by spindle.

      K-co mple x detection method has been attempted here with the help of reference signal. This reference signal is compared with input signal and check availability of reference signal in it. In this process, two methods , such as, matched filter and neural network methods have been introduced for detection of k-comp le x. Then, the results obtained from both these methods are compared based on sensitivity and specificity para meters. Compa risons of these parameters with those of the other existing methods are used for verification.

   2. Me thodolog y

      1. EEG data ac quisition

         The imple mentation process is carried out using the database available in the physonet website. The available database was in EDF format, for co mpatibility it is converted in to ascii format for the MATLA B 7.9.The raw signal is then segemented according to the Rechtschaen and Kales[ 1968] sleep stage classification standard[3]. The segmented stage II signal is filtered using 3rd order Savitzky-Go lay Filter. Savitzky-Go lay filtering can be thought of as a generalized moving average. Savitzky – Go lay s moothing filters are typically used to "smooth out" a noisy signal. This filter is a lso called a digita l s moothing polynomia l filter or a least-squares smoothing filter. The Savitzky-Go lay filtering method is often used with frequency data or with spectroscopic (peak) data.

      2. Work flow

         Signal s(t)

         The complete flo w of the imp le mantation is give below.

         Stage II e xt raction

         Filtration

         Noise n(t)

         Filter h(t)

         EEG signal

         Y(t) = sh(t)+nh(t)

         Refe rence signal

         Feature e xtraction

         Neural network

         Co mpare the results

         Matched filter

         Fig.2:- A lgoritha m for detection of k-co mple x

      3. Matche d Filter

         When template of a typical version of a signal event is available, it become possible to design a filter that is matched to the characteristics of the event. Matched filter is obtained by correlating a known signal or te mplate with an unknown signal to detect the presence of the template in the unknown signal[4][5]. This is equivalent to convolving the unknown signal with a conjugated time-reversed version of

         Fig.3:- Basic mod le for matched filter.

         To use matched filter, it is usually requried that the reference signal is known beforehand. Here te mplate signal is obtained by using amplitude and frequency defination criteriaof the k-co mple x.

         Results for matc he d filter:

         The results obtained by matched filter are shown as follows.

         the template.

         Y[n] = h[n – k] x[k]

         K= –

         If a signal that contains repetations of the event with almost the same characteristics is passed through the matched filter, the output should provide peaks at the time instance of occurance of the event. This filter is common ly used for the detection of known charateristic that are buried in noise. It ma ximizes the signal-to-noise ratio of the filterd signal[5].

      4. Artificial Neur al Network:

      Art ificia l neura l networks are nonlinear signal processing devices, built fro m e le mentary processing devices called neurons.

      A three layered feed-forward backpropagation artific ial neural network was used to classify the EEG signals, the results obtained with the network architecture of input-hidden-output nodes, show a high percentage of correct classifications .The back propagation method involves the propagation of error backwards (with no feedback involved), to update weights of the hidden and the input layers so as to minimize the least squared error. Again the outputs are calculated iteratively until the error falls to an accepted minimu m leve l. The network arch itecture used is as shown in Fig 4. A sig moid t ransfer function is used for calculating the results. Input vectors and the corresp onding target vectors were used to train the network until the network classifies the input vectors The network was trained init ia lly by the e xtracted features of the K -co mple x signal. propagation network was properly trained to classification with a high degree of correct classification. During the tra ining phase, the weights are successively adjusted based on a set of inputs and the corresponding set of desired output targets. The back-propagation algorithm needed weight adjustments in the backward sweep. Excellent results were obtained with a min imu m error of 1*e-10, a gradient 1.0 and 3000 epochs, an adaptive mo mentu m constant for speeding up the convergence. The parameters of the ANN such as acceptable min imu m error and learning rate, mo mentu m constant were set to get performance of the network over the entire set of EEG data[6][7].

      Fig 4:-feed forwa rd ANN

      Fe ature Extrac tion:

      The features are e xtracted by computing the like lihood thresholds based on amplitude and duration measurement.

      The characteristics were selected so as to reflect the visual criteria as well as possible. For the majo rity of them,

      they were e xtracted fro m significant points other possible K-co mple x. These significant points are similar to those of Bankman et a l. [8] and are illustrated in following figure-

      Fig.5:- Featuresof k-co mp le x waveform

      – x_ min and t_min correspond to the minimal va lue of the pseudo K-comp le x.

      • x_ ma x and t_ma x correspond to the ma ximal value of the EEG in the interval .

      • t_end is the last time interval of the refe rence signal.which is negative follo wed by postive.

      • t_star corresponds to the first local ma ximu m value.

      -x_ mid & t_ mid are the first value greater than 0u V met by scanning the EEG fro m le ft to right starting from t_min.

      Following features are obtained based on the above mentioned thresholds are

      The duration of the K-co mple xe is represented by:

      f3=(t_end-t_start)

      A minimu m peak to peak a mp litude is first required although the related threshold is low:

      f2=(x_max – x_min)

      Concerning the sharpness of the first negative wavecompared to the second positive wave, a relevant criteria is:

      f3=(t_end- t_mid)/ (t_mid- t_start)

      To ensure that the amplitude of the negative component is at least 50% of the positive amp litude component:

      f4=abs(x_min)/ x_max

      The sharpness of the negative wave was represented by:

      f5=abs(x_min)/( (t_mid1- t_start)* f sampling

      In order to avoid the possibility of some features dominating the classification process, the values of each feature were normalized so that the range for each feature lies in between 0 to 1. These extracted features where fed as

      the input for artific ial neutral network tra ined with back propagation algorithm.

   3. Results

      The Sensitivity and Specificity of both the classifier is as follows,

      	Sensitivity	Specificity
      M atched Filter	86.47%	67.66%
      ANN	96.06%	52.62%

      Table 1:- Results for matched filter and ANN.

   4. Conclusion

It has been shown that the k-comple x are co mprise less than five percent of time in sleep. However, this does not prevent from developing an accurate k-co mple x detection system for these 5 percent, as the signal is noisy. Therefore, the proposed methods are very effic ient for the detection of events buried in noise, even when the SNR (Signal to Noise Ratio) is very small. In the presence of noise, ANN gives more sensitivity and less specificity than Matched filter as shown in table 1.

REFERENCES

1. A. Rechtschaffen and A. Kales, A manual of standardized terminology, techniques and scoring system for sleep of hu man subjects, U.S. Pub lic hea lth service,

   U.S. Govern ment printing office, Washington D.C. 1968.

2. Didie r Henry, D. Sauter, Co mparision of detection methods: Application to k- comple x detection in sleep EEG Proc. Ann. Inter. conf. IEEE EM BS, vol. 2, pp. 1218-1219, 1994

3. Rangara j M. Rangayyan,Bio medica l signal analysis a case study approach 2004, Wiley-Ind ia Edit ion. [6]B.H.Jansen,B.M.Dawant,K.Maddahi:AI Techniques for K-co mple x Detection in Hu man Sleep EEG.University of Houston,1989.

1. IsaacN.Bankman,Vincent G.Sigillito,RobertA.wise,andP hilip L.Smith :Feature based detection of k-comp le xes in Hu man EEG using Neural Net work,IEEE 1992.

2. Automatic K-co mp le xes Detection in Sleep EEG Recordings using Like lihood Thresholds :S. Devuyst, T. Dutoit, P. Stenuit, and M. Ke rkhofs . Buenos Aires, Argentina, August 31 – September 4, 2010