Automatic spike detection in EEG by a two-stage procedure based on support vector machines
Introduction
ElectroEncephaloGram (EEG) is a one dimensional and multi-channel signal obtained from the brain. EEG is widely used clinical tool in the diagnosing, monitoring and managing of neurological disorders related to epilepsy. Epilepsy may be defined as a symptom of paroxysmal and abnormal discharges in the brain that may be induced by a variety of pathological processes of genetic or acquired origin. This disorder is often characterized by sharp recurrent and transient disturbances of mental function and/or movements of different body parts that result from excessive discharges of groups of brain cells. The presence of epileptiform activity, which is distinct from the background EEG activity, confirms the diagnosis of epilepsy, although it can be confused with other disorders producing similar seizure-like activity. During seizures, the scalp EEG of patients with epilepsy is characterized by high-amplitude synchronized periodic EEG waveforms. In between seizures, epileptiform transient waveforms, which include spikes and sharp waves, are typically observed in the scalp EEG of such patients. An EEG spike, which is different from the background activity, has a pointed peak and duration of 20– [1]. Although it may occur alone, a spike is usually followed by a slow wave, which lasts 150–, forming what is known as spike and slow wave (SSW) complex [2].
The evaluation of EEG for the detection of epilepsy generally includes visual scanning of EEG recordings for these spikes and seizures by an experienced electroencephalographer (EEGer). This process is time consuming, especially in the case of long recordings. In addition, disagreement among readers of the same record is possible due to the subjective nature of the analysis [3]. The use of ambulatory monitoring, which produces or longer continuous EEG recordings, became more common, thus further increasing the need for an efficient automated detection method. Several attempts have been made to automate the spike and seizure detection process through computer based methods. In most of these detection systems, measurements of electrographic parameters of EEG waveforms, such as sharpness, slope, duration and amplitude are compared with thresholds representative of a typical true spike and seizure.
Various methods have been used for the detection of spikes, sharp waves and other transient signals in the scalp EEG. Mostly they are based on measuring the duration and sharpness of individual waves using their second derivatives [4], [5], [6]. Another spike detection system has been developed that is sensitive to states of EEG such as active wakefulness, quite wakefulness, desyncronized EEG, phasic EEG and slow EEG [7]. In addition to these methods, mimetic techniques have been widely used for detecting epileptiform discharges. For spike detection, mimetic methods usually decompose the EEG into waves or half waves by determining the extrema of the amplitudes. Each wave is then examined for its fit to a set of predetermined criteria, e.g. duration, amplitude, slope and sharpness [6], [8]. Similarly, some filtering techniques have been proposed for spike detection [9]. All these studies have tried to find some standards for detecting spikes using objective criteria.
Artifacts should be eliminated for a robust detection of epileptiform spikes. With an EEG signal free of artifacts, a reasonably accurate detection of spikes and sharp waves is possible; however, difficulties arise with artifacts. This problem increases the number of false detections that commonly plague all automatic detection systems [10], [11]. Many studies using artificial neural networks (ANN) approach for detecting EEG spikes have been reported [8], [11], [12], [13], [14], [15]. ANN based spike detection systems basically use either of two different input representations: (1) the extracted EEG features or (2) the raw EEG signal. In the former case, the extracted spike features such as slope, duration, amplitude and sharpness are presented to the ANN for training and testing purposes [16]. The success of such a system depends on proper selection of the features, which is in some sense a trial and error procedure. In the second case, the raw EEG signal is presented to the ANN after a proper scaling and windowing [8], [17], [18]. Although it saves memory and time, the feature extraction based method is not very efficient in terms of the classification performance. On the other hand, the second method whose classification performance is superior to that of the first method requires a lot of memory and is very time consuming.
The novel support vector machine (SVM) based method proposed in this paper employs a two stage classification procedure that would display good classification performance while being efficient in terms of memory and time requirements. The former classification is realized by a modified non-linear digital filter which separates spikes and spike like non-spikes from trivial non-spikes. SVM has been examined as a post-classifier with radial basis kernel function [19], [20]. SVM has a good performance resulting in 90.3% sensitivity, 88.1% selectivity and 9.5% false detection rate for the test set. In this study, the simulations were performed using MATLAB running on a Pentium Celeron PC computer.
Section snippets
Recording and data acquisition
The EEG data used in this study were acquired from 25 epileptic patients who had been under evaluation and treatment in the Neurology Department of Dokuz Eylül University Hospital, İzmir, Turkey. 18 of these EEG records were used in training, while the remaining 7 in testing procedure. Data were obtained from a clinical EEG monitoring system, which stores continuous EEG data on its hard disk. EEG data were acquired with Ag/AgCl disk electrodes placed using the 10–20 international electrode
Results and discussion
To visualise the problem and to see the effect of C parameter, we restrict ourselves to the two features that contain the most important information about the class, namely the duration and the amplitude. As can be seen from Fig. 5, the separation of a group of spikes from a group of non-spikes is not so trivial. In the SVM classification, support vectors are represented by circles as shown in Fig. 5. Fig. 5 shows the results of an SVM classification for two different degrees of
Conclusion
In this study, we introduce a novel two stage classification procedure based on SVM for spike detection that will contribute to the clinical applications with its good efficiency and accuracy levels. The proposed approach accomplishes peak detection, feature extraction, pre-classification, and spike detection by preserving the original form of the spikes. We demonstrate that the pre-classification done by the non-linear digital filter can successfully separate the spikes and spike like
Summary
In this study, we introduce a two-stage procedure based on support vector machines for the automatic detection of epileptic spikes in a multi-channel electroencephalographic signal. In the first stage, a modified non-linear digital filter is used as a pre-classifier to classify the peaks into two subgroups: (i) spikes and spike like non-spikes (ii) trivial non-spikes.
Pre-classification procedure is performed to eliminate trivial non-spikes and also to capture spikes and spike like non-spikes.
Acknowledgements
The authors gratefully acknowledge the help of Neurology Department of the Dokuz Eylül University in the preparation of the database. The first author Nurettin Acır would also like to thank to TÜBİTAK (Turkish Scientific and Technical Research Council) Münir Birsel Fund for support.
Nurettin Acır graduated from Erciyes University, Turkey, in 1995. He took M.Sc. degree from Nigde University, Turkey, in 1998, all in Electronics Engineering. He is going on Ph.D. studies at Electrical and Electronics Engineering Department of Dokuz Eylül University, Turkey. He is now a visiting scholar at Neurosensory Engineering Lab., University of Miami, USA. His interest areas include intelligent systems, biomedical signal processing, artificial neural networks.
References (27)
- et al.
Quantification of abnormal EEG spike characteristics
Comput. Biol. Med
(1974) - et al.
Automatic recognition and quantification of interictal epileptic activity in the human scalp EEG
Electroencephalogr. Clin. Neurophysiol
(1976) - et al.
State-dependent spike detectionconcepts and preliminary results
Electroencephalogr. Clin. Neurophysiol
(1991) - et al.
Automated interictal EEG spike detection using artificial neural networks
Electroencephalogr. Clin. Neurophysiol
(1992) - et al.
Automatic spike detection via an artificial neural network using raw EEG dataeffect of data preparation and implications in the limitations of online recognition
Clin. Neurophysiol
(2000) Multifold features determine linear equation for automatic spike detection applying neural nin interictal EcoG
Clin. Neurophysiol
(1999)- et al.
Detection of spikes with artificial neural networks using raw EEG
Comput. Biomed. Res
(1998) - et al.
Automatic EEG spike detectionwhat should the computer imitate
Electroencephalogr. Clin. Neurophysiol
(1993) - et al.
A glossary of terms most commonly used by clinical electroencephalographs
Electroencephalogr. Clin. Neurophysiol
(1974) - et al.
Wavelet preprocessing for automated neural network detection of EEG spikes
IEEE Eng. Med. Biol
(1995)
Methods of Analysis of Brain Electrical and Magnetic Signals
Semiautomatic quantification of sharpness of EEG phenomenon
IEEE Trans. Biomed. Eng
A multistage system to detect epileptiform activity in the EEG
IEEE Trans. Biomed. Eng
Cited by (0)
Nurettin Acır graduated from Erciyes University, Turkey, in 1995. He took M.Sc. degree from Nigde University, Turkey, in 1998, all in Electronics Engineering. He is going on Ph.D. studies at Electrical and Electronics Engineering Department of Dokuz Eylül University, Turkey. He is now a visiting scholar at Neurosensory Engineering Lab., University of Miami, USA. His interest areas include intelligent systems, biomedical signal processing, artificial neural networks.
Cüneyt Güzeliş graduated from İstanbul Technical University, Turkey, in 1981. He took M.Sc. and Ph.D. degree from İstanbul Technical University, Turkey, in 1984 and 1988, respectively, in all electrical engineering. Between 1989 and 1991 he worked in the Department of Electrical and Computer Engineering at University of California, USA, as visiting researcher and lecturer. He is now a professor at Electrical and Electronics Engineering Department of Dokuz Eylül University, Turkey. His interest areas include neural networks, signal processing and nonlinear circuits and systems.