Expectation Maximization Based PCA and Hessian LLE with Suitable Post Classifiers for Epilepsy Classification from EEG Signals

Abstract

Brain is a highly complicated structure which contains billions of neurons which help in maintaining the electrical charges of the system. Due to the sudden and unexpected electrical discharges occurring in the brain, epilepsy occurs and is a most commonly occurring neurological disorder next to stroke. For the easy monitoring of the activities of the brain, Electroencephalography (EEG) is widely used. For the primary purpose of assessment of brain activities, EEG serves as a valuable treasure and an indispensable tool. The various information related to the physiological state of the brain can be bought out by EEG. For the analysis of the EEG records in a visual manner, expert neurologists consume more time and to detect the epileptic seizures, only EEG signals are used widely as it contains vital information. The EEG recording contains a lot of noise and so it is difficult to isolate the seizures from other artifacts with resembles more or less similar time-frequency patterns. Various automatic detection and machine learning algorithms have been used to predict the risk of epileptic seizures in raw EEG signals. In this study, the dimensions of the EEG signals are reduced initially with the help of two approaches, namely, Expectation Maximization Based Principal Component Analysis (EM-PCA) approach and Hessian Local Linear Embedding (HLLE) approach. The dimensionally reduced values are then classified to predict the risk of epilepsy from EEG signals with the help of two post classifiers namely Weighted K Nearest Neighbour (WKNN) and Linear Support Vector Machine (L-SVM). The performance metrics are analyzed in terms of various measures like Performance Index, Accuracy, Time Delay, Specificity and Sensitivity. The study shows that the best result is obtained when HLLE is used as a dimensionality reduction technique and classified with WKNN as it gives the highest perfect classification rate with an average accuracy of 97.743%.