IEFS-GMB: Gradient Memory Bank-Guided Feature Selection Based on Information Entropy for EEG Classification of Neurological Disorders

These authors contribute equally to this work. Abstract Deep learning-based EEG classification plays a pivotal role in the automated detection of neurological disorders, offering significant advantages in diagnostic accuracy and early intervention for personalized clinical treatment. However, the performance of such classification approaches is fundamentally limited by the intrinsic low signal-to-noise ratio characteristic of EEG signals. Consequently, feature selection (FS) is essential in optimizing the EEG representations derived from neural network encoders, thereby enhancing the overall efficacy of EEG classification frameworks. Currently, few FS methods have been tailored for EEG neurological diagnosis, and most FS methods from other fields are designed for specific network architectures and lack clarity in interpretation, which restricts their direct utility in EEG classification. These authors contribute equally to this work. Consequently, these approaches may lack the robustness necessary to effectively handle data variability. To address these challenges, we introduce IEFS-GMB, a novel I nformation Entropy-based F eature Selection approach guided by a Gradient Memory Bank. This method begins by establishing a dynamic gradient memory bank that archives the sampled gradients from previous training iterations.