Classification of EEG Signals for Epilepsy Detection Using PCA Analysis

Epilepsy is a disease related to the abnormal seizures caused by the central nervous system. The epileptic seizure is occurred due to malfunctioning of the electrophysiological system of the brain. The collection of EEG data is become a tedious and time consuming process so it is difficult to detect of epileptic seizures by visual observation of patient’s EEG. It needs an expert to detect epileptic activity by analyzing the entire length of recorded EEG data. In this study, we propose a method for classifying EEG signals for the detection of epilepsy using principal component analysis (PCA) and machine learning techniques. The proposed method involves pre-processing the EEG signals to reduce noise and artifacts, followed by PCA to extract the most relevant features of the signals. It reduces the dimensionality of data and it is linear combinations of the original features. Then the features obtained from the PCA components are feed to machine learning classifier. The most commonly used classifiers are support vector machine or naive bayes, random forest to classify the epileptic brain signals. In this work, the accuracy, sensitivity obtained with PCA reduction and without PCA reduction is compared. The dataset consisted of EEG recordings from multiple electrodes, with a total of over 1000 signals. We used a kfold cross-validation and data partition (33%) method to estimate the performance of the classifier. We have used K-Nearest Neighbors (kNN), Support Vector Machine (SVM), and Navie bayes algorithms by exploiting the PCA feature reduction technique in the dataset to envisage epilepsy, and the performance of classifiers are analyzed with or without using PCA technique. The study revealed that the SVM classifier demonstrated superior performance with a classification accuracy of 96.77% when utilizing the holdout technique. The proposed model exhibited higher measures of accomplishment, including sensitivity (99.34%), specificity (94.2%), and AUC (99%), compared to other classifiers when holdout cross validation was used.