Feature relevance analysis for preictal electroencephalography classification

Abstract. Epilepsy is a disorder that a ects the central nervous system and it is characterized by un-predicted disruption of normal brain activity that leads to recurrent seizures. Additionally, epilepsy is one of the most common neurological diseases worldwide according to World Health organization (WHO), a ecting 1 % of world population. Around 30 % of this popu-lation however su er from drug resistant epilepsy and must live with eventual seizures which a ict their quality of life. Electroencephalogram test is currently the main tool to diagnose epilepsy which enables neurologists to determine the states of an epileptic brain. This test can be contrasted with video-telemetry and other biomarkers such as Magnetic Resonance Imaging (MRI) to ndthe seizure focus. These recordings nevertheless could take hours and sometimes days which could be burdensome to analyze. As a result, computational techniques have been proposed to detect and predict seizures. Most of these techniques have therefore implemented feature extraction using methods of amplitude, frequency and multivariate analysis in order to reduce the input space, and then classi cation methods mainly based on machine learning techniques. So far the results that have been obtained have shown similar rates of classi cation. Therefore in this thesis, a methodology for anticipating a seizure was proposed. This metho- dology used a database that belongs to Neurocentro from Pereira and two research groups: Gaunal group and GCYPDS, both from Universidad Nacional de Colombia. This database, contains the information (reports and evaluations) of a set of patients with di eret types of epilepsies but people who su ered recurrent seizures. This methodology is developed by means of the selection of the most relevant features using a metric of the maximum separability of the classes which are the epileptic brain states (preictal, ictal and interictal). This approach allowed us to improve the classi cation performance using di erent types of classiers in order to identify the zones likely to be related with the moment prior to a seizure onset. In addition, this methodology highlights the performance of Partial Least Squares techniques that relate information which is as important for brain activity as to the epileptic brain states. Two frameworks were implemented in order to see di erences between two distinct domains such as the raw data (time domain) and frequency related representation. Finally, classi cation rates were measured from the classi ers carried out for preictal and interictal classi cation. Particularly, the highest performances were given by the partial least squares regression and in the parietal and occipital lobes which are the lobes related to somatosensory information (sensitivity of 88 ; 24 %) and visual information, respectively. These results make sense according to the subjective premonitory or aura symptoms manifested before a seizure in most of the epileptic patients

• Autores: Salazar Londoño, Maria Clara
• Fecha de publicación: 2017
• Institución: Universidad Nacional de Colombia
• Repositorio: Universidad Nacional de Colombia
• Idioma: spa
• OAI Identifier: oai:repositorio.unal.edu.co:unal/63764
• Acceso en línea: https://repositorio.unal.edu.co/handle/unal/63764; http://bdigital.unal.edu.co/64291/
• Palabra clave: 62 Ingeniería y operaciones afines / Engineering; Epilepsy; Feature extraction; Ictal; Metrics; Principal Component Analysis (PCA),; Partial Least Squares (PLS); Epilepsia - Métodos de simulación; Extracción de características; Selección de características; Interictal; Análisis de componentes principales (PCA); Mínimos cuadrados parciales (PLS); Preictal
• Rights: openAccess
• License: Atribución-NoComercial 4.0 Internacional