Some applications of machine learning for biomedical signal processing

Abstract

Flexible analytic wavelet transform (FAWT) is suitable for the study of oscillatory signals like electroencephalogram (EEG) signals with versatile features such as shift in-variance, tunable oscillatory properties and exible time-frequency domain cov- ering. In this thesis, we propose two automated methods for the classi cation of epileptic EEG signals using FAWT for decomposition of the EEG signals into sub- bands and suitable features were extracted. The obtained features are given as input to twin support vector machine (TSVM), least squares TSVM (LS-TSVM) and ro- bust energy-based least squares twin support vector machines (RELS-TSVM) for classi cation. The proposed methods have been implemented on publicly available Bonn University EEG database [1] and the accuracy of RELS-TSVM was found to be better as compared to TSVM and LS-TSVM and is comparable to other exist- ing methods with a maximum accuracy of 100% for the classi cation of seizure and non-seizure EEG signals and 98:33% for the classi cation of seizure and seizure-free EEG signals.