By Rasmy, Mohamed (Researcher) Helwan University
This study investigates the impact of different preprocessing techniques on EEG signal classification using a CNN-BiLSTM model. Using the inherently noisy BCI Competition 2008–Graz dataset A, we applied Independent Component Analysis (ICA) to all files for artifact removal. Additionally, Artifact Subspace Reconstruction (ASR) was applied selectively: once across all files and another time only for the noisiest subjects. The effect of epoching before applying ICA and ASR was
also analyzed. While ASR showed limited impact when applied only to the noisiest subjects, its application across the entire dataset slightly improved performance. A refined CNN-BiLSTM architecture was developed to handle non-stationary signals more effectively. We also explored the role of bias initialization in BiLSTM units, though it was not the primary focus of this study. Our results highlight the importance of optimal preprocessing, careful epoching, and model design in improving EEG classification performance for brain-computer interface (BCI) tasks.
