Analysis of Classifiers for Epileptic Seizure Detection: ANFIS and SVM Gopika Gopan K * , Eldho S Kollialil † , Harsha A and Liza Annie Joseph Rajagiri School of Engineering and Technology Rajagiri Valley, Kakkanad, Kochi, Kerala, India. Email: * gopika.gopan.k@gmail.com, † eldhoskollialil@gmail.com Abstract—Epilepsy is a neurological disorder characterized by recurrent, abnormal and synchronous neural activity (seizures) in the brain resulting in characteristic abnormality in the elec- troencephalographic pattern. An automated detection of epileptic seizures provides assistance to Neurologists in the diagnosis and timely treatment of epileptic patients. Adaptive Neuro-Fuzzy Inference System (ANFIS) and Support Vector Machine (SVM) are two effective classifiers that can be used in an automated epilepsy detection system. This paper attempts to analyze the performance and effectiveness of ANFIS and SVM for epileptic seizure detection taking into consideration various signal features. The ANFIS classifier gives 100% efficiency for two-level classi- fication with Mean Teager Energy and Mean Energy using Pi- membership function for input node. The quadratic programming based SVM gave 100% accuracy with interquartile range being the feature for three-level classification of the EEG data. Keywords—Electroencephalography, Epileptic seizures, Wavelet decomposition, Support Vector Machines, Adaptive Neuro-Fuzzy Classifier I. I NTRODUCTION Epileptic seizure is an abnormal synchronization of neu- ral activity resulting in characteristic discharges in the elec- troencephalographic pattern (EEG) [1]. Spikes, polyspikes, spike-and-wave complex, sharp waves are the commonly seen rhythms in electroencephalographic pattern indicative of epileptic seizure [2,3]. Spikes lasts less than 70 millisec- onds while sharp waves happen over 80-200 milliseconds. Polyspikes are series of spikes happening quickly and spike waves are very fast waves followed by slow waves happening three times per second. Two periods of abnormal activity (inter-ictal and ictal) in- dicative of epileptic seizures can be observed in electroen- cephalographic pattern. EEG of ictal period (at the time of epileptic seizure) indicates the abnormality in the form of mainly continuous spike and sharp wave complexes observed over a duration longer than the average duration of these abnormalities during inter-ictal periods (intervals between two consecutive seizures) which show occasional transient wave- forms like isolated spikes, spike trains, sharp waves or spike- wave complexes [4]. Identification of the presence of these abnormalities from the given EEG data forms the base of any automated epileptic seizure detection system. Since EEG rarely get ictal readings, the diagnosis rely on inter-ictal EEG (most commonly obtained from patients) with seizure activated with photo stimulation, hyperventilation and other methods. Time-frequency analysis of epileptic seizures utilizing short- time Fourier transforms [5-7] was a commonly used approach for epileptic seizure detection. Approaches involving wavelet transforms [8], Lyapunov exponents, principal component analysis [9-11], and combination of wavelet analysis and Lya- punov exponents [9] were also present in the literature. Neural networks [6,9,12-14] were also used in some approaches due to its learning ability but it suffered from large noise and sensitivity issues. This paper intends to analyze the performance and effec- tiveness of two classifiers in detection of epileptic seizures, namely, Adaptive Neuro-Fuzzy Inference System (ANFIS) and Support Vector Machine (SVM). The EEG signals undergo five-level wavelet decomposition to extract the desired fre- quency band. Various features are calculated for fifth detail coefficients (which give the desired frequency range). The calculated features are then given as inputs to ANFIS and SVM and their performances are analyzed. The paper consists of four sections. Section I gives an outline of the concepts involved in this work. Section II describes the datasets used in this work and the methodologies involved with the subsections giving description of each method uti- lized in this work. Section III gives the observation and the results obtained after MATLAB simulation. Section IV draws conclusions about the outcome of the research undertaken. II. MATERIALS AND METHODS The EEG data used for this research were acquired from a reputed Neurology Clinic and datasets of normal and epileptic EEGs of 250 patients each were created. Each dataset contains a total recording of 40 minutes (≈ 0.4 minute recording of each patient) sampled at a sampling frequency of greater than or equal to 180 Hz as per Nyquist sampling criterion. Out of the total 500 single channel EEG datasets (both epileptic and normal), 400 EEG data were used for training the classifier and remaining 100 were used for testing. Fig. 1 shows the block diagram of the approach utilized in this work. EEG signals are initially decomposed using discrete wavelet transform to obtain the various brain rhythms. The fifth detail coefficients gave the appropriate frequency range for the given datasets and features were calculated using these coefficients. The features were given as input to two different classifiers, ANFIS and SVM, to evaluate their performance and the classification effectiveness of each featurs. A. Discrete Wavelet Transform Electroencephalography, being non-stationary, cannot be analyzed using Fourier transform. It requires a technique