Applications of machine learning techniques for the modelling of EEG data for diagnosis of epileptic seizures

Abstract

The aim of this study was to create a simple and effective epileptic seizure detector using EEG data, signal processing and machine learning techniques. This study proposes a simple and effective training set acquisition method for epileptic seizure detection. This training set acquisition method was applied and analyzed in the following three analysis methods. ‘Analysis Method 1’ uses Multilevel Wavelet Decomposition as a feature vector design process and both Support Vector Machine (SVM) and Extreme Learning Machine (ELM) as feature vector classification methods. ‘Analysis Method 2’ uses filter data as part of the feature vector design process and both SVM and ELM as feature classification methods. ‘Analysis method 3’ feeds singlechannel, multi-level wavelet decomposition feature vectors, separately, to both SVM and ELM classification methods. All analysis methods were tested with more than 185 seizures and 977 hours of electroencephalogram (EEG) data, all acquired from the ‘CHB-MIT Scalp EEG Database’. Results obtained from all analysis method prove that this simple training set acquisition is not only very effective but even better than other training methodologies used in previous studies. Results obtained from the first two analysis methods show that when classifying multi-channel data at once, the ELM classification technique performs better than the SVM classification technique. ELM’s results obtained from the first two analysis methods reach more than 99% ‘Sensitivity’ and less than 2 seconds ‘Latency’. Results obtained from the third analysis method show that the SVM classification method is able to detect all seizure instances with only one channel data being fed during the feature vector classification process. SVM’s results obtained from this analysis method reach a 100% ‘Sensitivity’ and less than 1 second ‘Latency’. Apart from providing a simpler way to detect an epileptic seizure, that is by classifying each channel data separately; this analysis method shows that a single channel data is enough to detect all the patient’s seizures. All the latter analysis methods show that future studies can be further conducted in order to improve the overall ‘Specificity’ results obtained in this study. ‘Specificity’ results can be improved by increasing the frequency ranges used during the feature vector design processes. This should improve the ‘Specificity’ results since larger frequency ranges would create spectral EEG features which better distinguish between seizure and non-seizure EEG data.