Machine learning based EEG classification by diagnosis: Approach to EEG morphological feature extraction

Abstract

A hypothesis that spike morphological features contain information that can be used for epilepsy type detection by machine learning methods is discussed. Investigation of approach to EEG (electroencephalogram) spike morphological feature definition in relation to machine learning based EEG classification by diagnosis is presented in this study. Two approaches of defining EEG spike morphological features are investigated: A) numerically evaluating EEG spike geometric features, e.g., upslope, downslope; B) using 300 ms of spike (without additional features extracted) for classification. Lists of spikes are used for the classification. Before start of the algorithm some basic preprocessing steps are taken: electric utility frequency (50Hz) is removed. The EEG classification by diagnosis algorithm consists of these main steps: 1) EEG spike detection by morphological filter; 2) EEG classification employing spike morphological features (employing discussed approaches) by diagnosis using machine learning based classification algorithms. Various classification algorithms (e.g., artificial neural network based classifier, AdaBoost, decision tree, random forest, extremely randomized tree, etc.) and their quality metrics are considered (e.g., accuracy, true positive rate, true negative rate, etc.) as well as results of k-fold cross-validation are investigated in this work. EEGs from children (3-17 years old) are classified in this work. The EEGs under classification are patients diagnosed with: I) benign childhood epilepsy, II) structural focal epilepsy. Current results show that best performance (87% ± 1%) is exhibited by Extremely randomized tree based EEG classifier employing spike upslope and downslope data.A hypothesis that spike morphological features contain information that can be used for epilepsy type detection by machine learning methods is discussed. Investigation of approach to EEG (electroencephalogram) spike morphological feature definition in relation to machine learning based EEG classification by diagnosis is presented in this study. Two approaches of defining EEG spike morphological features are investigated: A) numerically evaluating EEG spike geometric features, e.g., upslope, downslope; B) using 300 ms of spike (without additional features extracted) for classification. Lists of spikes are used for the classification. Before start of the algorithm some basic preprocessing steps are taken: electric utility frequency (50Hz) is removed. The EEG classification by diagnosis algorithm consists of these main steps: 1) EEG spike detection by morphological filter; 2) EEG classification employing spike morphological features (employing discussed approaches) by diagnosis using machine learning based c...