Epileptic EEG Signal Classification with Marching Pursuit based on Harmony Search Method Ping Guo 1,* , Jing Wang 1,4 , X. Z. Gao 2 , and Jarno M. A. Tanskanen 3 1 Laboratory of Image Processing and Pattern Recognition, Beijing Normal University, Beijing 100875, China 2 Department of Automation and Systems Technology, Aalto University School of Electrical Engineering, Finland 3 Department of Biomedical Engineering, Tampere University of Technology, Finland 4 School of Foundational Education, Peking University Health Science Center,China pguo@ieee.org, wang_jing@bjmu.edu.cn , xiao-zhi.gao@aalto.fi, tanskanen@ieee.org Abstract— In Epilepsy EEG signal classification, the main time-frequency features can be extracted by using sparse repre- sentation with marching pursuit (MP) algorithm. However, the computational burden is so heavy that it is almost impossible to apply MP to real time signal processing. To reduce complexity of sparse representation, we propose to adopt harmony search me- thod in searching the best atoms. Because harmony search me- thod can find the best atoms in continuous time-frequency dic- tionary, the performance of epilepsy EEG signal classification is enhanced. The validity of this method is proved by experimental results. Keywords—Harmony search method; Electroencephalogram; Seizure detection; Overcomplete dictionary; Sparse representation. I. INTRODUCTION Epilepsy is a chronic neurological disorder that affects ap- proximately 1% of the world’s population. It is characterized by recurrent unprovoked seizures, which are caused by abnor- mal electrical discharges in the brain. Electroencephalogram (EEG) is an electrical signal recorded from the scalp or intra- cranial, and reflects the mass activity of neurons and their inte- ractions. EEG is used by physicians to assist diagnosing many neurological disorders, especially the epilepsy. The detection of epileptic seizures in EEG signals is very important in the diag- nosis of epilepsy. In the past decade, interpretation of EEG has been limited to only visual inspection by neurophysiologists, individuals trained to qualitatively make a distinction between normal and abnormal EEG. Unfortunately, detection of epilep- sy that needs visual inspection of long recordings of EEG is usually a time-consuming and high-cost process. Therefore, several diagnostic aid approaches for automatically detecting epileptic seizures from EEG signals have been proposed and studied in recent years. Various techniques have been developed in the literature for the detection of epileptic seizures in EEG [1-13]. All of the seizure detection schemes consist of two principal stages. In the first stage, features are extracted from raw EEG data in time domain, frequency domain, or time-frequency domain. In the second stage, the features extracted from EEG are used for * Corresponding author, phone: +861058800441; email:pguo@ieee.org. training classifiers that differentiate between the normal and epileptic EEG. A number of classifiers have been proposed and employed, including Bayesian classifier [1], support vector machine (SVM), and different kinds of artificial neural net- works (NNs) [2-6], artificial neuron-fuzzy inference system and dynamic fuzzy NN [7 8]. Besides classifiers, the perfor- mance of the epilepsy detection is mostly dependent on the features that are extracted to characterize the raw signals. Therefore, the features for classification play a critical role in the performance of any classifiers. In recent years, sparse representation is widely applied to signal and image classification. Sparse representation, first, proposed in the signal processing literature by Mallat and Zhang, is representation that account for the most or all the information of a signal by a linear combination of a small number of elementary signals called atoms [14]. Indeed, the general problem of finding a representation with the smallest number of atoms from an arbitrary dictionary has been shown to be NP-hard. This has led to considerable efforts being put into the development of many sub-optimal schemes, e.g., marching pursuit (MP), orthogonal matching pursuit and basis pursuit, all of which can reach a solution close to the optimum by relaxing some constraints of the original optimization prob- lem. The MP has been widely applied because of its simplifica- tion. However, the computational burden in the signal sparse decomposition process that iteratively builds up the signal approximation one coefficient each time is significant. Harmony Search (HS) method is inspired by the underlying principles of the harmony improvisation [15]. Similar to the genetic algorithms, particle swarm optimization, differential evolution [16] and computational swarm intelligence systems [17], the HS method is a stochastic search technique. It does not require any prior domain information beforehand, such as the gradient of the objective functions. However, different from many other population-based evolutionary approaches, it only utilizes a single search memory to evolve. Thus, the HS me- thod has the interesting characteristics of algorithm simplicity. In [1], with features extracted based on the sparse repre- sentation, Bayesian classifier are used for the epilepsy detec- tion. This method can obtain better classification accuracy when suitable parameters are set, and it is also robust to noise. However, the computational burden of this algorithm is very 2012 IEEE International Conference on Systems, Man, and Cybernetics October 14-17, 2012, COEX, Seoul, Korea 978-1-4673-1714-6/12/$31.00 ©2012 IEEE 283