| Abstract [eng] |
In this paper, an analysis of artificial neural network (ANN) effectivenes, when used as a tool to analyse electroencephalograms (EEG), is presented. Main target of this analysis is a 40-200 ms. long EEG spike. EEG spikes are usually used in epilepsy diagnositcs. This paper analyses EEG spike information obtained from real life patients diagnosed with Rolandic epilepsy. Five different types of classifiers were chose for EEG spike detection: 2D Convolutional Ne- ural Networks (2D CNN), Logistic Regression classifier, Decision Tree classifier, Support vector machine classifer as well as AdaBoost metaalgorithm. Signal analysis was conducted on data, that was extracted from 18 EEG channels and was used unprocessed as well as preprocessed using wavelet transform and results optimisation was done only by changing various ANN parameters. 2D Convolutional Neural Network (2D CNN) archieved 0.973 accuracy, 0.9605 sensitivity and 0.3592 specificity values. Logistic regresion model archieved 0.7088 accuracy, 0.7415 sensitivity and 0.3703 specificity values. Decision Tree model archieved 0.8535 accuracy, 0.9064 sensitivity and 0.3074 specificity values. Support Vector machine classifier archieved 0.883 accuracy, 0.9408 sensitivity and 0.2851 specificity values. AdaBoost metaalgorithm archieved 0.8650 accuracy, 0.9179 sensitivity and 0.3185 specificity values. Python programming language and several open-source libraries for machine learning such as TensorFlow, Keras as well as Scikit-Learn were used to archieve the results described in the paper. EEG were provided in the European Data Format (EDF) which were converted into CSV and divided into one spike length (maximum length of 200 ms. was chosen) intervals which were used for training of the neural network models. As seen from this paper 2D CNN returned the best results of 0.973 accuracy, 0.9605 sensitivity and 0.3592 specificity when used with preprocessed data. |
