Evaluating five different adaptive decomposition methods for EEG signal seizure detection and classification

Abstract Signal processing and machine learning methods are valuable tools in epilepsy research, potentially assisting in diagnosis, seizure detection, prediction and real-time event detection during long term monitoring. Recent approaches involve the decomposition of EEG signals into different modes in a data-dependent and adaptive way, which may provide advantages over commonly used Fourier based methods when dealing with nonlinear and non-stationary data. Examples of such methods include empirical mode decomposition (EMD), extended EMD (EEMD), complete EEMD with adaptive noise (CEEMDAN), empirical wavelet transform (EWT) and variational mode decomposition (VMD). In this work, feature sets extracted from original non-decomposed signals and from the aforementioned adaptive decomposition methods are evaluated for the classification of EEG seizure data using two freely available datasets. We provide a previously unavailable common methodology for comparing the performance of these methods for EEG seizure detection, with the use of the same classifiers, parameters and spectral and time domain features. Overall, results were similar between the evaluated decomposition methods, with slightly superior values for VMD and CEEMDAN. Features extracted from the original non-decomposed signals resulted in inferior class separability, but fairly accurate predictions could still be achieved with specific classifiers. The evaluated decomposition methods are promising approaches for seizure detection, but their use should be judiciously analysed, especially in situations that require real-time processing and computational power is an issue. Another contribution of this work is the development of python packages for EWT (ewtpy) and VMD (vmdpy), already available at the PyPI repository.