EEG signal processing using machine learning to detect epileptic seizures

Abstract

Epileptic seizures are characterised by abnormal synchronous electrical discharge in the brain. Due to the adverse effects that epilepsy has on people suffering with the condition, the risk of bodily harm and even premature death are a constant threat depending on the severity and quantity of seizures. Development in medicine and technology has produced several techniques used to monitor and localize seizure activity in the brain. One common technique is the use of an Electroencephalogram (EEG), which is a procedure in which metal electrodes are placed around the patient’s scalp and the electrical activity produced by the brain can be measured. Therefore, the aim of this research is to investigate different signal processing and machine learning techniques in order to be able to automatically detect, with a high accuracy, seizures from EEG recordings. In this dissertation, a seizure detection system was proposed which makes use of Discrete Wavelet Transform, in order to decompose EEG signals into a number of subband signals, and an ensamble learning technique known as classifier stacking, to combine several well performing classifiers into a single meta classifier. The results of using classifier stacking versus using a single classifier, as traditionally done, are compared and contrasted. The CHB-MIT scalp EEG dataset was used in order to obtain EEG signals containing seizure and non-seizure activity. The trained models were evaluated using three metrics being Sensitivity, Specificity, and Accuracy. The best result from all the tested models was achieved using five classifier stacking using a combination of a Support Vector Machine, Naive Bayes, K-Nearest Neighbours, Random Forest, and Multilayer Perceptron Neural Network. This model achieved an accuracy of 96%, a sensitivity of 95.35%, and a specificity of 96.65%. When comparing this result to the best result achieved from the most proficient single classifier, this model is able to reduce the number of incorrectly classified signals by 24.1%. The results obtained from this research showed that using classifier stacking produced several models with a noticeable improvement in every performance metric measured.