Detection of seizure from EEG signals may help neurologist analyze the condition of epileptic patients. Variance in types of epileptic seizures provides challenge in recognizing the pattern of seizure signals from normal ones. In this paper, we discuss the classification of seizure and non-seizure conditions in epilepsy based on bandwidth features of EEG signals. These features were extracted using Empirical Mode Decomposition (EMD). This method decomposes signal into eight Intrinsic Mode Functions (IMFs) and its residue. The IMFs is then analyzed by Hilbert transform to obtain amplitude modulation bandwidth (BAM) and frequency modulation bandwidth (BFM). These features are fed as input to Support Vector Machine (SVM). We propose to combine the first four IMFs to extract bandwidth features and we implement this method on two data sets, public data set that contains signals from both extracranial and intracranial EEG, and data set OF Rumah Sakit Universitas Airlangga (RSUA) that contains signals from extracranial EEG only. The results showed that in general, the values of accuracy and specificity using combined features of first four IMFs outperformed those using features of single IMF, for all SVM kernel functions. The best accuracy and specificity on public data set were 97.3% and 98.25% respectively using RBF kernel, while the best sensitivity was gained using polynomial kernel based on only IMF1 bandwidth features by 99.0%. The experiments on RSUA data showed that the best accuracy was 96.5% and the best specificity was 99.38% when using Mexican Hat kernel based on combined features of first four IMFs, while the best sensitivity was 94% when using polynomial kernel based on only IMF1 bandwidth features.

Original languageEnglish
Pages (from-to)568-576
Number of pages9
JournalProcedia Computer Science
Publication statusPublished - 2019
Event5th Information Systems International Conference, ISICO 2019 - Surabaya, Indonesia
Duration: 23 Jul 201924 Jul 2019


Dive into the research topics of 'Epileptic seizure detection based on bandwidth features of EEG signals'. Together they form a unique fingerprint.

Cite this