Comparison of three time–frequency decomposition methods for the classification of EEG signals of epilepsy patients

Abstract:
Methods of signal decomposition and pattern recognition play important role in the analysis of electroencephalogram (EEG) signals for detecting epileptic seizures. Prevalent time-frequency-based methods of signal decomposition, such as the discrete wavelet transform (DWT), empirical mode decomposition (EMD), and variational mode decomposition (VMD), deliver better performance than conventional Fourier decomposition on non-linear and non-stationary EEG signals. This study examines and compares these three time-frequency decomposition methods in terms of the classification of EEG signals. The EEG signals were decomposed into different sub-signals in different frequency bands using these methods, and five efficient features were extracted to represent different types of EEG signals. We then combined them into feature vectors and sent them to the three classifiers BLDA, RF, and SVM for classification using 10-fold cross-validation. Two EEG databases were used to evaluate the performance of the time-frequency-based decomposition methods, which contained ictal and interictal EEG signals, and focal and non-focal EEG signals. The results show that all three methods deliver fairly good classification results, especially VMD. VMD yielded the best classification on ictal and interictal signals with an accuracy of 100%, and also recorded the best classification on focal and non-focal signals with an accuracy of 87.10%.