Comparison of EWT and OVMD Techniques for Left and Right Hand Motor Imagery Classification in EEG Signals

P. Zych, P. Śniatała (Poznan Univ. of Techn., Poland)

Brain-computer interfaces (BCIs) utilizing event-related desynchronization/synchronization (ERD/ERS) are vital for stroke-related neurorehabilitation. To improve outcomes, systems must adapt to individual patient characteristics. This study evaluates orthogonalized variational mode decomposition (OVMD) as an alternative to empirical wavelet transform (EWT). Using BCI Competition IV Dataset 2b, two processing pipelines were analyzed to test these techniques within the filter bank common spatial pattern (FBCSP) framework and an instantaneous feature extraction pipeline. The results indicate differences in performance based on the feature domain. EWT significantly outperformed OVMD in the FBCSP pipeline (78% versus 60% accuracy), yet OVMD proved superior for instantaneous amplitude (IA) and frequency (IF) features due to more stable mode separation. While some subjects reached 96% accuracy, significant inter-subject variability remains. Although standard parameters yield satisfactory results, the integration of advanced signal decomposition serves as a powerful tool for BCI systems where individual adaptation is critical. The superior performance of linear classifiers suggests motor imagery (MI) classes are well-separated in the extracted feature space. Ultimately, these findings highlight the need for further parameter optimization to develop truly individualized and adaptive BCI systems for rehabilitation support.