Waveform detection with RBF network Application to automated EEG analysis

Transcription

1 Neurocomputing 20 (1998) 1 13 Waveform detection with RBF network Application to automated EEG analysis Antti Saastamoinen *, Timo Pietilä, Alpo Värri, Mikko Lehtokangas, Jukka Saarinen Department of Clinical Neurophysiology, Tampere University Hospital, P.O. Box 2000, FIN Tampere, Finland Department of Neurology, Tampere University Hospital, P.O. Box 2000, FIN Tampere, Finland Signal Processing Laboratory, Tampere University of Technology, P.O. Box 553, FIN Tampere, Finland Received 30 December 1996; revised 30 December 1997; accepted 24 February 1998 Abstract Automated detection of different waveforms in physiological signals has been one of the most intensively studied applications of signal processing in the clinical medicine. During recent years an increasing amount of neural network based methods have been proposed. In this paper we present a radial basis function (RBF) network based method for automated detection of different interference waveforms in epileptic EEG. This kind of artefact detector is especially useful as a preprocessing system in combination with different kinds of automated EEG analyzers to improve their general applicability. The results show that our neural network based classifier successfully detects artefacts at the rate of over 75% while the correct classification rate for normal segments is as high as about 95% Elsevier Science B.V. All rights reserved. Keywords: RBF networks; Automated EEG analysis; Artefact detection; Waveform detection 1. Introduction Due to its capability to reflect both the normal and abnormal electrical activity of the brain, the electroencephalogram (EEG) has been found to be a very powerful tool in the field of neurology and clinical neurophysiology. The diagnosis and treatment * Corresponding author. asaastamoine@tays.fi /98/$ see front matter 1998 Elsevier Science B.V. All rights reserved. PII S (98)

2 2 A. Saastamoinen et al./neurocomputing 20 (1998) 1 13 planning of different neurological diseases is strongly based on the multi-channel, long-term recordings of EEG that are carefully inspected by a specialized physician. However, as a consequence of the vast amount of data this process is generally very time-consuming. For this reason, many different automated methods have been proposed to facilitate the diagnosis and to reduce the time needed for the visual inspection [5,6]. These methods are mainly intended for the automated detection of epileptiform activity in the long-term EEG recordings, automated sleep stage scoring and the diagnostics of different sleep disorders. Routine EEG is measured on the scalp by using so-called electrode system [2]. These recording sessions typically take about 30 min and their primary purpose in epilepsy diagnostics is to record EEG during suspected epileptic attacks. Because these attacks might not happen very often, some activating procedures such as photic stimulation, hyperventilation or sleep deprivation are often used for attempting to provoke the onset of an epileptic attack. However, in most cases remarkably longer recordings are needed for the recording of typical seizures. In addition to the extreme complexity of the signal morphologies, one of the main problems in the automated EEG analysis is the summation of different kinds of interference waveforms, i.e. artefacts, to the original EEG during the recording sessions. The most important causes for these interferences are the movements of the patient during recording session and the normal electrical activity of the heart, muscles and eyes. In case of visual inspection, these artefacts can be often easily discarded. However, during the automated analysis these signal patterns often cause serious misclassifications thus reducing the clinical usability of the automated analysing systems. In this study our main emphasis has been to create a reliable method for artefact detection to improve the diagnostic accuracy and applicability of the systems proposed earlier [8,7,11,4]. In spite of the general importance of the subject and quite exhaustive search for articles concerning the applications of neural networks for artefact detection, rather limited amount of contributions in this field were found. Generally, different artefact processing methods have been based on the application of conventional signal processing methods such as digital highpass filtering for reducing high-frequency contaminations, different correlation based subtraction techniques for correcting interferences caused by eye movements and different rule-based classifiers [1]. However, Wu et al. [13] have proposed a comparative research using neural network approach. In their paper they present a method utilizing autoregressive spectrum estimates and cross correlation studies for feature extraction and several Multilayer Perceptron Networks for different artefact types for the data classification. Although our system utilizes only one Radial Basis Function Network with a remarkably simpler network architecture our results seem also to be clinically promising. 2. Materials and methods The EEG data used in this work has been recorded at the intensive monitoring laboratory at Tampere University Hospital, Tampere, Finland, and at Vaajasalo

3 Epilepsy Unit, Kuopio, Finland. EEG recordings were digitized at the sampling frequency of 200 Hz and the artefactual parts in these recordings were scored visually before the application of the data. So far we have concentrated on the detection of muscle, movement and saturation artefacts because they seem to be the most frequent interferences to occur in the recordings for the diagnostics of epilepsy. Although the electrical activity of the eyes (Electro-oculography, EOG) produces also interferences in the EEG, these are not considered because they can normally be found only on a few channels. Moreover, we already have a reliable method for detecting EOG activity [12,10] Studied artefact types A. Saastamoinen et al./neurocomputing 20 (1998) In the following, we briefly discuss the general properties of muscle, movement and saturation artefacts. The main emphasis is to show those properties on which the detection of these artefacts can be based on. Some examples of these types of artefacts are also given. Electromyography (EMG) artefacts are originated in the EEG due to the increased muscle activity in the head area, e.g. during chewing or frowning. As it can be seen in Fig. 1, EMG artefacts consist of higher amplitudes and frequencies than the preceding activity. Thus, EEG with EMG contamination includes more power on high frequencies and the variance of the signal is higher than in case of pure spontaneous EEG. Fig. 1 shows some example signal waveforms containing EMG contamination. Movement artefacts are associated with the skeletal movements which often cause electrodes and measurement cables to move. Depending on the nature of the movements they can be fairly recognizable or more subdued. Movements of the measurement cables often produce very strong step-like artefacts with voltages saturated to maximum or minimum values of the possible voltage range. Tremor produces a rhythmical pattern of various frequencies due to the electrode movements. Because muscle activity is associated with skeletal movements, movement artefacts can contain also EMG-activity. In Fig. 2 an example signal part on movement artefacts is depicted. It also includes some amount of EMG activity due to the connection to muscle activity. Saturation artefacts can be divided into several subgroups based on their amplitude levels. For all these types slightly different approaches are used. Zero-level saturation simply means that the signal amplitudes stuck into zero level. Like the low-amplitude Fig. 1. Example EEG tracings containing EMG-activity.

4 4 A. Saastamoinen et al./neurocomputing 20 (1998) 1 13 Fig. 2. Example signal patterns containing movement artefacts. Fig. 3. Example signal patterns containing signal saturation. saturation this can be caused by some electrode problems, for example due to loose electrode connection or broken measurement cable. For this reason saturation artefacts should be detected, although they do not cause misclassifications in other signal analysis systems. Fig. 3 shows an example on a signal saturated close to zero level. As an example on the complexity of the classification problem, Fig. 4 shows some example signal waveforms containing epileptic activity. Typically, these epileptiform patterns contain sharp spikes and slower waves, either as combinations like in the Fig. 4, or separately. For the system to be clinically applicable, these kind of waveforms must not be classified as artefacts. It is easy to see that the significance of the preprocessing is crucial and that the typical properties of epileptiform activity must be considered, when the feature extraction stage is designed Feature extraction Generally, when neural network based approach is applied to waveform detection, some kind of application-specific preprocessing stage precedes the neural network classifier. The purpose of this stage is to calculate some measures that reflect the temporal properties of the data. These measures or features can then be applied to categorize the measurement data to different classes depending on whether different signal segments contain some waveform or not. Fig. 5 shows the block diagram of our artefact detection system. The meanings of different blocks are discussed in details below. Before feature extraction we first segment the recorded EEG to fixed length segments of 100 samples (0.5 s). To ensure reasonable time resolution the segmentation window is moved in steps of 25 samples (0.125 s). For each segment, several

5 A. Saastamoinen et al./neurocomputing 20 (1998) Fig. 4. Example signal pattern containing epileptic activity. Fig. 5. Block diagram of artefact detection system. For simplicity, only four hidden units are shown. features based on the local power and amplitude values of the EEG are then calculated. These features are used as inputs for the radial basis function network that has been trained using the orthogonal least squares (OLS) algorithm Movement artefacts As it can be seen in Fig. 2, the movements of the patient typically generate very strong, step-like interferences to the recorded EEG. When considering the general properties of movement artefacts, it is possible to derive a measure to detect the prevalence of this kind of waveform in a segment based on the amplitude distribution of the recorded EEG. Starting from the definition of the histogram a nonlinear filter was developed for the detection of movement artefacts. Let us consider an EEG segment with 100 samples. Sorting the amplitude values to increasing order and forming bars from the groups of ten successive samples we can form a 10-bar histogram that describes the amplitude distribution of segments. Generally, if we scale the heights of the histogram in a way, that the total area of the histogram is equal to l, we get an estimate for the probability density function. Let

6 6 A. Saastamoinen et al./neurocomputing 20 (1998) 1 13 s be the spread of the ith bar of the histogram and let n be the number of bars in the estimate. Then the height h of the ith bar is h " 1 ns. (1) The height estimates have an interesting property of giving highest values for those bars with smallest amplitude distribution, thus providing a good measure for the concentration of the signal amplitude. However, if all the amplitude values inside a bar are the same, the height of the bar will be infinite. This condition can be avoided by adding a nonzero bias to the spreads s according to the equation h " 1 n(s #bias). (2) This means that by setting the bias equal to 1/n we actually limit the maximum height of a bar in case of zero spread to h "1. Let now w be the mean of the amplitude values inside the ith bar of the histogram. Using these mean values as weight coefficients we can define a nonlinear filter with output y" w h " w ns #1. (3) According to the connection to histograms of the segments we call this filter a weighted histogram filter (WHF). It is easy to notice that the filter describes well the concentration properties of the data inside segments. If the spreads of the bars are wide, the output of the filter will be close to zero. In case of movement artefact the amplitude values tend to concentrate in the vicinity of some high positive or negative amplitude value with spreads very close or equal to zero. Thus, the heights of the bars are close to unity and the output tends to the sum of mean amplitudes which will be very high. On the other hand, although the amplitude values of normal EEG patterns concentrate close to zero level leading possibly to heights close to 1, the output of the WHF filter is still small due to the small mean amplitudes w. In some rare cases of twosided steps with positive and negative mean amplitudes close to each other, the filter output may be close to zero. This can be avoided by using the absolute values of mean amplitudes as weights thus forcing the output to be nonnegative Saturation artefacts One measure that effectively reflects the prevalence of zero-level saturation can be defined in terms of signal energy. Let denote the length of the segment and let x denote the ith amplitude value in the segment. Using a simple energy estimate E" x (4)

7 in the following estimator D A. Saastamoinen et al./neurocomputing 20 (1998) D" 1 (1#E), (5) we get a simple measure for the zero-level saturation or dead-electrode condition. Much in the same way we can define an estimator for low-amplitude saturation based on the sorted amplitude vector determined during the computation of the WHF output. Picking the minimum and maximum values inside a segment and using the difference of them as a range estimate we can define the saturation estimator S as 1 S" [1#(x!x )]. (6) Although this estimate responds well for saturation at all levels, we call it a lowamplitude saturation estimator, because this special type of saturation cannot be detected by the WHF filter due to the low mean amplitudes. Alternatively, we could define this estimator by first subtracting the mean from the segment and applying then Eq. (5). Thus, we could define the low amplitude saturation estimator also as follows: 1 S" [1# x!μ ]. (7) In the experiments described in this paper, Eq. (6) was used for computing measures for the low-amplitude saturation because it is computationally more efficient. This is because we have already sorted the data vector x during the computation of the WHF output Muscle artefacts and high-frequency contaminations The most characteristic property of muscle artefacts is their higher frequency content when compared to the normal EEG activity. Thus the most obvious way to determine the amount of EMG contamination is to filter the segment with some high pass filter and then compute the energy of the filter output. One computationally very efficient estimate can be obtained simply by computing the energy of the derivative of the segments, because the derivative in fact is a simple high-pass filter. Thus we can define our estimator as H" 1 x!x. (8) 2 In addition to the EMG contamination, this estimator efficiently reflects the other high-frequency contaminations as well. These include, e.g. the 50/60 Hz mains interferences and some irregular interference waveforms that can be found on EEG during electrode and measurement cable problems.

8 8 A. Saastamoinen et al./neurocomputing 20 (1998) Training of the radial basis function network After feature extraction, calculated features are fed to the RBF network. In this project we have used gaussian activation functions for the nonlinear mapping performed by the hidden layer. Thus, the outputs of the classifier are calculated according to equation y "w # w exp x!x d, (9) where y is the output value for feature vector x, w is a constant bias term, x and d are the centers and widths for jth hidden neuron in the network, respectively, q is the total number of hidden neurons in the network and the coefficients w are the weight coefficients in the connections from hidden neurons to the output neuron. Eq. (9) shows that the output y for ith feature vector can be obtained as a weighted sum of the activation values of hidden neurons added with a constant bias term. For the training of the network, well-known orthogonal least squares (OLS) training method was applied [3]. It is based on the idea of selecting the neurons that are installed to the network from a large set of candidate neurons based on LMS optimization procedure. In this project we first used our training data to determine the centers and widths for 300 candidate neurons by applying K-means clustering algorithm and two-nearest-neighbor statistics, respectively [9]. The initial cluster centers for the K-means algorithm were randomly selected in such a way that their distribution in the four-dimensional input space of our classifier was uniform. When the centers and widths had been selected, the parameters of the candidate neurons were frozen and 25 most efficient candidates were selected to be installed to the network by applying the orthogonal least squares (OLS) algorithm. In the OLS training algorithm the best possible hidden neurons are selected one at a time from a large set of candidate neurons based on their ability to reduce the variance of the error between the calculated and desired output values [3]. In addition to that, the connection weights w between the hidden layer and output layer are determined. The final classification results were formed by hard-limiting the network output obtained through Eq. (9) using a threshold value determined by ROC analysis. Threshold value was selected in such a way that the detection rate of the classifier for the validation data would be as high as possible with a minimum loss of clinically relevant EEG data. Those patterns that caused the network output to be smaller than the threshold were classified to contain artefactual waveforms whereas those patterns for which the output was higher were considered to be either normal or epileptic EEG waveforms. 3. Results In this project we used a data set containing in total a bit more than signal patterns from which about one half contained artefactual waveforms. To estimate the

9 A. Saastamoinen et al./neurocomputing 20 (1998) performance of the RBF classifier after training phase we used the well-known N-fold cross-validation method. In this method the whole data set is divided into N separate data sets. During each validation run, one of these sets is used as a validation set while the others are used for the training. Because we used four-fold cross validation, each training and validation set contained about and signal patterns, respectively. According to the general practice of cross-validation studies, sensitivity and specificity were used as performance measures. Sensitivity (true positive rate, TPR) gives the probability of the correct classification for positive (artefactual) signal patterns and it is defined according to the equation TPR" TP TP#FN, (10) where TP is the total amount of correctly classified artefactual signal patterns and FN is the amount of artefactual signal patterns that were misclassified as negatives. On the other hand, specificity (true negative rate, TNR) tells the probability of the correct classification in case of negative (normal or epileptic) signal patterns. It can be computed using the equation TNR" TN TN#FP, (11) where TN is the total amount of correctly classified negative signal patterns and FP is the amount of negative signal patterns that were misclassified as artefacts. During the experiments, RBF networks with different numbers of hidden neurons were tested. The sufficient number of hidden neurons was found to be around 25. The performance of the classifier with 25 hidden neurons was then examined by performing four-fold cross-validation. These results show that the original EEG activity is correctly classified at the rate of approximately 95% while the rate of correct classification for the artefacts is over 75%. The sensitivities and specificities for threshold level 0.5 are summarized in Table 1. Results of the fourfold cross-validation study for the network with 25 hidden neurons are shown in terms of average ROC curves and average performance curves in the Fig. 6. ROC curve depicts the correct classification rate of negative patterns Table 1 Sensitivities and specificities of the radial basis function classifier Performance Sensitivity Specificity (TPR) (TNR) Validation set % 94.79% Validation set % 94.84% Validation set % 94.45% Validation set % 94.55% Average 76.42% 94.65%

10 10 A. Saastamoinen et al./neurocomputing 20 (1998) 1 13 Fig. 6. Average ROC-curves and performance curves for the RBF-classifier. ROC-curve (top) shows the dependency between the sensitivity and specificity of the classifier. Performance curve (bottom) reflects the performance of the RBF classifier as a function of the threshold level. Solid line shows the dependency of the artefact detection rate on the selected threshold value. Corresponding curve for the correct classification of normal segments is shown by dashdot line. (either normal or epileptic EEG patterns) as a function of the correct artefact detection rate. Performance curve shows the dependence of the sensitivities and specificities of the classifier on the threshold level used to hard limit the output values of the RBF network. Thus, it can be effectively used for the selection of the threshold value.

11 A. Saastamoinen et al./neurocomputing 20 (1998) Discussion Setting the threshold a bit higher even higher amount of artefacts can be detected correctly. However, this happens at the cost of lower classification accuracy for true EEG. Because the correct classification of the original EEG is clinically more important than the detection of all possible artefacts, lower threshold value was selected to ensure sufficient specificity for our artefact detector. This holds especially in the case of epileptic signal segments. This is due to the fact that the analyzer classifying too many epileptic patterns as artefacts is clinically useless. To test the real performance of the classifier in the presence of epileptic activity, the training and testing data contained also a good collection of epileptic waveforms. Because almost all the normal segments were correctly classified and the detection rate for artefactual patterns is over 75%, our results seem to be promising. However, more thorough validation of the system is needed before the system can be applied clinically. Acknowledgements This study has been financially supported by the Academy of Finland, Technology Development Centre of Finland and the European Union Healthcare Telematics Project (European Neurological Network, ENN). This support is gratefully acknowledged. The authors also want to thank the reviewers for their valuable comments on the manuscript. References [1] J.S. Barlow, Artefact processing (rejection and minimization) in EEG data processing, in: F.H. Lopes da Silva, W. Storm van Leeuven, A. Rémond (Eds.), Handbook of Electroencephalography and Clinical Neurophysiology, Vol. 2, Clinical Applications of Computer Analysis of EEG and other Neurophysiological Signals, Elsevier, Amsterdam, 1986, pp [2] J.D. Bronzino, Biomedical Engineering and Instrumentation: Basic Concepts and Applications, PWS Engineering, Boston, [3] S. Chen, C.F. Cowan, P.M. Grant, Orthogonal least squares learning algorithm for radial basis function networks, IEEE Trans. Neural Networks 2 (2) (1991) [4] P. Elo, J. Saarinen, A. Värri, H. Nieminen, K. Kaski, Classification of epileptic EEG by using self-organized maps, in: I. Aleksander, J. Taylor (Eds.), Artificial Neural Networks, 2nd edn., Proc. Int. Conf. on Artificial Neural Networks (ICANN 92), Brighton, UK, September 1992, Elsevier, Amsterdam, 1992, pp [5] J. Gotman, Computer analysis of the EEG in epilepsy, in: F.H. Lopes da Silva, W. Storm van Leeuven, A. Rémond (Eds.), Handbook of Electroencephalography and Clinical Neurophysiology, Vol. 2, Clinical Applications of Computer Analysis of EEG and other Neurophysiological Signals, Elsevier, Amsterdam, 1986, pp [6] J. Gotman, Computer analysis during intensive monitoring of epileptic patients, in: R.J. Gumnit (Ed.), Advances in Neurology, Vol. 46, Intensive Neurodiagnostic Monitoring, Raven Press, New York, 1986, pp [7] V. Krajca, S. Petranek, I. Patakova, A. Värri, Automatic identification of significant graphoelements in multichannel EEG recordings by adaptive segmentation and fuzzy clustering, Int. J. Biomed. Comput. 28 (1991)

12 12 A. Saastamoinen et al./neurocomputing 20 (1998) 1 13 [8] T. Pietilä, S. Vapaakoski, U. Nousiainen, A. Värri, H. Frey, V. Häkkinen, Y. Neuvo, Evaluation of a computerized system for recognition of epileptic activity during longterm EEG monitoring, Electroencephal. Clin. Neurophysio. 90 (1994) [9] J.T. Tou, R.C. Gonzales, Pattern Recognition Principles, Addison-Wesley, London, [10] A. Värri, K. Hirvonen, V. Häkkinen, J. Hasan, P. Loula, Nonlinear eye movement detection method for drowsiness studies, Int. J. Biomed. Comput., in press [11] A. Värri, K. Hirvonen, J. Hasan, P. Loula, V. Häkkinen, A computerized analysis system for vigilance studies, Comput. Methods Programs Biomed. 39 (1992) [12] A. Värri, B. Kemp, A.C. Rosa, K.D. Nielsen, J. Gade, T. Penzel, J. Hasan, K. Hirvonen, V. Ha kkinen, A.C. Kamphuizen, M.S. Mourtazev, Multi-centre comparison of five eye movement detection algorithms, J. Sleep Res. 4 (1995) [13] J. Wu, E.C. Ifeachor, E.M. Allen, N.R. Hudson, A neural network based artefact detection system for EEG signal processing, Proc. Conf. on Neural Networks and Expert Systems in Medicine and Healthcare, Plymouth, UK, 1994, pp Antti Saastamoinen was born in Tampere, Finland, on July He studied digital signal processing and biomedical engineering in the Department of Electrical Engineering at Tampere University of Technology, where he received an M.Sc. degree in He is currently working for the Department of Clinical Neurophysiology at Tampere University Hospital. His research interests include processing of physiological signals and the implementation of biomedical signal processing systems. Timo Pietilä was born in He studied medicine in Tampere University He has been working at the Departments of Neurology and Clinical Neurophysiology at Tampere University Hospital since His areas of interest are epilepsy, EEG and computer based analysis of EEG. Alpo Va rri was born in Turku, Finland, on December He received the Master of Science, Licentiate of Technology and Doctor of Technology degrees from Tampere University of Technology, Tampere, Finland, in 1986, 1988 and 1992, respectively. He is currently working as laboratory manager in the Signal Processing Laboratory at Tampere University of Technology. His research interests include biomedical signal processing and microcomputer programming.

13 A. Saastamoinen et al./neurocomputing 20 (1998) Mikko Lehtokangas studied analog and digital electronics and applied mathematics in the Electrical Engineering Department at Tampere University of Technology where he received the M.Sc. and Dr.Tech. degrees in 1993 and 1995, respectively. Currently he is working as a senior researcher in Signal Processing Laboratory at Tampere University of Technology. His main research interests are nonlinear adaptive architectures and algorithms, and their application. Jukka Saarinen was born in Finland on 11 July, He studied computer architecture, digital techniques, telecommunications and software engineering in the Department of Electrical Engineering at Tampere University of Technology, where he received an M.Sc. degree in 1986, a Licentiate of Technology degree in 1989 and a Doctor of Technology degree in Currently, he is a Professor of Computer Engineering at Tampere University of Technology. His research interests are parallel processing, neural networks, fuzzy logic and pattern recognition.