LETTER TO EDITOR
|Year : 2004 | Volume
| Issue : 3 | Page : 399--400
An artificial neural network to detect eeg seizures
Rakesh K Sinha, Amit K Ray, Navin K Agrawal
School of Biomedical Engineering, Institute of Technology, Banaras Hindu University, Varanasi - 221 005, India
Rakesh K Sinha
School of Biomedical Engineering, Institute of Technology, Banaras Hindu University, Varanasi - 221 005
|How to cite this article:|
Sinha RK, Ray AK, Agrawal NK. An artificial neural network to detect eeg seizures.Neurol India 2004;52:399-400
|How to cite this URL:|
Sinha RK, Ray AK, Agrawal NK. An artificial neural network to detect eeg seizures. Neurol India [serial online] 2004 [cited 2020 Nov 24 ];52:399-400
Available from: https://www.neurologyindia.com/text.asp?2004/52/3/399/12757
Electroencephalography (EEG) in its conventional form is not routinely useful in the diagnosis of any specific epileptic seizure discharge as it produces a huge amount of normal data intermixed with relatively rare clinically significant events. Several computer algorithms and programs for the detection of seizures have been developed, but these methods are found insufficient to recognize the exceptions and minimize the number of false detections. To overcome these problems, artificial neural network (ANN) has been used for the classification of seizures in EEG., It has been suggested that instead of analyzing only the frequency and amplitude changes, obtaining power spectra by fast Fourier Transform (FFT) of EEG provides more information.
The experiments were carried out with male Charles Foster rats weighing between 200-250 grams. EEG recording electrodes were implanted on the rat's head under Urethane (Sigma, USA) anesthesia (1.5 gm/kg i.p). Three stainless steel screw electrodes of 1 mm diameter soldered with flexible radio wires were implanted extradurally. Two electrodes for EEG recordings were placed on the frontal and occipital part of the skull and one reference electrode was placed on the anteriormost part of the skull. To induce epileptic seizures, 50 units of Benzyl Penicillin in a volume of 10ml was injected 2 mm below the cortical surface of the parietal region of the brain. The intracerebral injection of the above-mentioned dose of benzyl penicillin induces spontaneous seizure initiation after 3-5 minutes of injection and observed its continuous presence as spike patterns in EEG up to an average of 30 minutes. The single-channel bipolar EEG signals were recorded with standard amplifier setting just before and continuously for 45 minutes through an electroencephalograph (EEG-8, Madicare, India) from the time of injection of penicillin to the end of seizure patterns. The EEG signals were not only recorded on papers, but also on two minutes separate data files to the computer hard disk after digitization of the traces at 256 Hz. The data were collected using the Visual Lab-M software (ADLink Technology Inc., Taiwan).
A three-layered feed forward back propagation program written in C++ programming language was used for the detection of seizures. It has been described earlier that single hidden layer neural networks are universal approxomator and universal classifier; thus the network contains only one hidden layer. One-second epochs of seizure and normal patterns were taken and their FFT or Power Spectra calculated separately after filtering the raw data. Selected digital FFT values of seizure and normal patterns were used for the training of the network. The number of neurons in the input layer is fixed by input data from digital values of FFT between 10 Hz to 30 Hz. The patterns of seizure and normal EEG and the selection of bandwidth used for the training and testing of ANN are presented in [Figure:1]. The proposed network has 40 neurons in the input layer for both training and testing purpose. The output layer has only one unit. The firing (high value) of output neuron confirms the seizure otherwise normal EEG patterns. The schematic architecture of ANN is shown in [Figure:2]. 100 data sets, 50 sets each from seizure and normal EEG, were prepared through this method and arranged randomly in a separate file named as 'TRAINING.DAT'. The error tolerance and the learning rate parameters were assigned as 0.001 and 0.5 to start the network and it was trained for 1 million cycles. Once the simulator reaches the error tolerance that was specified or the maximum number of iterations, assigned earlier, the simulator saves the state of the network, by saving all its weights in a file called 'WEIGHT.DAT'. In test mode, the network was provided for a set of test data, which was prepared and stored in a separate data file named as 'TEST.DAT'. The method of the preparation of the test data file was similar to that of preparation of training data file but without assigning the output and hence the test data file contains only input patterns. When the file was applied to the trained network, the 'OUTPUT.DAT' file was generated, which contains the output from the network for all the input patterns. The network goes through a cycle of operation in this mode, covering all the patterns in the test data file. The output of the network classifies the test patterns to seizure or normal pattern.
The performance of the network was evaluated in terms of percentage recognition rate of seizures that were marked by visual inspection and were correctly detected by the network. The percentage recognition rate was calculated for variation in number of sets used for training.
Number of correctly
Performance of ANN (%) = ___________________ x 100
Total number of
The results of the seizure and normal events detected by the network compared with those detected manually are summarized in [Table:1]. Manually detected events were taken as standard and agreement percentage represents the percentage of epochs in which ANN-detected seizure or normal events agreed with manually detected ones.
The present study proposed an approach for the computerized recognition and discrimination of EEG seizure patterns following intracerebral injection of penicillin by using an ANN. The purpose was to use processed and filtered signals to extract more information and to get higher decision accuracy. The most difficult aspect of the EEG analysis is the pattern recognition and data reduction. FFT is one of the most popular approaches that permits the presentation of large data in a comprehensive manner and by selection of the component for further processing results in significant data reduction. This technique is considered as a superior method in its computational ability and is well accepted for data reduction for long-term EEG.
Features calculated from the FFT such as relative power in various frequency bands and then using an ANN to generate a single number that indicates the degree of which the event is a seizure was used previously to classify seizure patterns. The best result owing feature extraction from FFT for the training and the automated testing of ANN for the detection of seizure patterns represent very good agreement with the human manual scoring. Instead of the features from the FFT of the EEG signals, we have used the selected frequency band of digital values of the FFT from one-second epochs of the EEG signals for the training and testing of the ANN. Following training, the present network was tested with different test patterns (both normal and seizure). This detector was observed to perform very well with a high performance of the agreement of 97.33% of the correct recognition of seizure and normal patterns compared to the manual scoring.
The success of the ANN in a specific application involves the optimization of the network structure and the parameters. One hidden layer was used based on several previous studies, which showed that one hidden layer resulted in the same performance as two or more hidden layers. However, conflicting results were reported in the literature on the number of hidden nodes. The selection of the number of hidden nodes in this study was based on experiments, which showed that 12 hidden nodes were most effective in this ANN architecture for the training and classification of seizure and normal EEG patterns.
|1||Zapata-Ferrer A, Maya LR, González AG, Pantaleón Garcia MC, Nasab NM, Valencia RH, et al: Detecting the onset of epileptic seizures. IEEE Eng Med Biol 1999:78-83.|
|2||Jandó G, Seigel RM, Horváth Z, Buzsáki G: Pattern recognition of the electroencephalogram by artificial neural networks. Electroencephalography and Clin Neurophysiol 1993;86:100-9.|
|3||Webber WRS, Lesser RP, Richardson RT, Wilson K: An approach to seizure detection using an artificial neural network. Electroencephalography and Clin Neurophysiol 1996;98:250-72.|
|4||Sarbadhikari SN: A Neural network confirms that physical exercise reverses EEG changes in depressed rats. Med Eng Physics 1995;17:579-82.|
|5||Rao V, Rao H : C++ neural networks and fuzzy logic. 1st Ed. New Delhi: BPB Publications 1996.|
|6||Hassoun HM: Fundamentals of artificial neural networks. New Delhi: Printice-Hall of India Private Limited 1998.|
|7||Kulkarni PK, Kumar V, Verma HK: Diagnostic acceptability of FFT-based ECG data compression. J Med Eng Technol 1997;21:185-9.|
|8||Al-Nashash HAM: A dynamic Fourier series for the compression of ECG using FFT and adaptive coefficient estimate. Med Eng Physics 1995;17:197-203.|
|9||Sirne RO, Isaacson SI, D'Attellis CE: A data-reduction process for long-term EEGs. IEEE Eng Med Biol 1999:56-61.|
|10||Zurada, JM: Introduction to artificial neural network systems. St. Paul (MN): West Publishing Company 1997.|