Focal/non-focal EEG records classification with MFCC based machine learning algorithms

Epilepsy, which occurs as a result of abnormal electrical spread in the brain and behaves as repetitive seizures, is accepted as an important neurological disease. According to the reports of the World Health Organization (WHO), more than 50 million people worldwide have this disease. Despite advanced drug technologies for epilepsy, 33% of patients do not respond to drug treatment and other unpredictable dimensions of the disease await them in the future. EEG, which is used in the detection of epilepsy patients, stands out as a convenient standard for the presence of epileptic seizures and other anomalies, as it has the ability to provide content for the electrical activity of the brain. In terms of location, epilepsy is generally divided into 2 types. In focal epilepsy, a specific part of the brain is affected by the disease. On the other hand, many parts of the brain are affected locally, even if they are not directly affected in non-focal epilepsy (generalized) seizures. In addition to these, EEG is also used to detect ictal and inter-ictal activities. Ictal EEG refers to the moment of seizure consisting of pins, and inter-ictal EEG refers to the time between seizures. Observational analysis of EEG recordings is both a waste of time and one of the causes of human error in disease diagnosis. An automated classification approach has become a necessity in distinguishing between different signal characteristics and assisting clinicians in decision making. In this thesis, it is aimed to classify the focal/non-focal epileptic EEG recordings within the Bern Barcelona dataset with Mel Frequency Cepstral Coefficients (MFCC) and machine learning algorithms. For this purpose, the high performance of MFCC on a difficult data set was confirmed after the classification of pre-ictal / ictal / inter-ictal EEG recordings in the Hauz Khas dataset with MFCC based machine learning algorithms. It is aimed to classify the focal/non-focal EEG recordings within the Bern Barcelona dataset with the provided MFCC supported machine learning algorithms. In the study, normalized optimum sub-feature sets were obtained by applying z-score normalization and Principal Component Analysis to MFCC feature sets. In the classification stage, 7 different conventional machine learning methods were used. As a result of the study, ictal vs. inter-ictal, ictal vs. pre-ictal vs inter-ictal vs. pre-ictal classification, 100% accuracy, sensitivity, specificity, and f1-score performance metrics were achieved for all groups with Quadratic Discriminant Analysis, Logistic Regression and Cubic SVM. For multi-way classification approach of ictal vs. pre-ictal vs. inter-ictal , only the Quadratic Discriminant Analysis algorithm was able to parse all the data correctly. Focal vs. non-focal EEG classification, all classes were correctly differentiated by Quadratic Discriminant Analysis, Logistic Regression and Cubic SVM. It is predicted that the method proposed in the thesis study and the stated inferences will become clearer by applying the current thesis study to different neurological diseases. Clinicians and physicians can apply the current methodology to their own data, compare it with their previous results and have a chance for validation.