Cascade of classifiers to classify interictal EEGs of patients with epilepsy

• Main Author: Jiang, Zhubo
• Other Authors: Justin Dauwels
• Format: Theses and Dissertations
• Language: English
• Published: 2018
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: http://hdl.handle.net/10356/73144
• Institution: Nanyang Technological University

Epilepsy is a chronic disease influencing many people’s health worldwide. According to the study of the WHO, there are over 50 million epilepsy patients around the world. Now, electroencephalogram (EEG) is still a primary method to analyze epilepsy. Experts can detect epilepsy by visual analysis of EEGs, which record the electrical signals of the human brain. Epileptiform transients (ET) or spikes usually appear in the EEG of epileptic patients. The spikes are the main indicators for epilepsy. However, detecting epilepsy by only visual inspection may need couple of hours, and there is a lack of experts who can read EEGs. Moreover, there is no standard definition for spikes, which makes the spike detection based diagnosis of epilepsy, tedious and expert-centered. Experts do not always agree on which waveforms are spikes and which ones are not. Hence, an automated method for analysis of epileptic patients’ EEG data is of importance for management and diagnosis of epilepsy. Many methods have been applied to detect the spikes such as template matching, neural network, SVM or random forest. In this thesis, we develop an efficient classification method to eliminate most background waveforms through an effective cascade of classifiers. A cascade of winning classifiers is designed to reject most background waveform for EEG data in several consecutive stages, while prereserving most spikes. Validating a classification method needs sufficiently large data. We have used 93 epileptic patients’ EEG data from Massachusetts General Hospital, which include 18164 spikes in total. We apply the 10-step cascade of decision tree, random forest and (support vector machine) SVM separately to the data by applying cross validation. In the numerical tests of this study, on average, the cascade of decision tree rejected 98.94%% of all background in the EEG dataset while preserving 86.22% of the spikes. The cascade of SVM rejected 98.89% of all background in the EEG dataset while preserving 86.97% of the spikes. The cascade of random forest rejected 98.84% of all background in the EEG dataset while preserving 87.32% of the spikes.