A Classification Approach for Focal/Non-focal EEG Detection Using CepstralAnalysis

Abstract

Electroencephalogram (EEG) is a convenient neuroimaging technique due to its non-invasive setup,practical usage, and high temporal resolution. EEG allows to detect brain electrical activity to diagnoseneurological disorders. Epilepsy is a crucial neurologic disorder that is reasoned from occurrence of suddenand repeated seizures. The goal of this paper is to classify the focal (epileptogenic area) and non-focal(non-epileptogenic area) EEG records with cepstral coefficients and machine learning algorithms.Analysis is carried out using publicly available Bern-Barcelona EEG dataset. Mel Frequency CepstralCoefficients (MFCC) are calculated from EEG epochs. Feature sets are normalized with z-score anddimension reduction is realized using Principal Component Analysis. Fine Tree, Quadratic DiscriminantAnalysis, Logistic Regression, Gaussian Naïve Bayes, Cubic Support Vector Machine, weighted k-nearestneighbors, and Bagged Trees are applied for classification stage. A value of k=10 is used for crossvalidation. All focal and non-focal EEG pairs are perfectly classified with acc., sen., spe., and F1-score of100% and AUC with 1 via. Quadratic Discriminant Analysis, Logistic Regression, Cubic SVM andWeighted k-NN. Proposed work recommends MFCCs as a single marker and this provides lesscomputation workload, practicality, and direct processing of focal / non-focal EEG time series. Proposedmethodology in this paper serves one of the highest achievements to literature and can assist neurologistand physicians to validate their diagnosis.