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

Liang Zhang, Hanyang Dong, Jia-Hong Gao, Yi Sun, Kuntao Xiao, Wanli Yang, Zhao Lv, Shurong ShengarXiv 2025 · DOI · arXiv

Abstract

Electroencephalography (EEG) is a critical tool for diagnosing neurological disorders, yet its low signal-to-noise ratio often hinders automated classification. This paper introduces IEFS-GMB, a plug-and-play feature selection module that leverages a Gradient Memory Bank (GMB) and Information Entropy-based Feature Selection (IEFS) to enhance model performance and interpretability. By utilizing historical gradient information and uncertainty-based filtering, IEFS-GMB achieves significant accuracy improvements across multiple clinical datasets while providing clinicians with meaningful visualizations of diagnostic features.

Introduction

Electroencephalography (EEG) provides a non-invasive window into brain activity, making it indispensable for identifying conditions such as epilepsy, Parkinson’s, and Alzheimer’s disease. However, the clinical utility of EEG is often limited by its inherent noise. Muscle movements, eye blinks, and electrode artifacts can produce waveforms that closely resemble the electrical spikes associated with neurological pathology.

The Challenge of EEG Noise

Automated EEG analysis faces a significant hurdle: distinguishing between true diagnostic signals and transient artifacts. As shown in Figure 1, the morphological similarity between noise and actual spikes makes it difficult for standard deep learning models to maintain high precision without sophisticated feature selection.

EEG waveforms illustrating the visual similarity between noise artifacts (blue) and diagnostic spikes (red).

Existing Feature Selection (FS) methods often struggle with architecture rigidity, typically being optimized for specific network types like CNNs. Furthermore, they often lack "memory," focusing only on the current training batch, which makes them sensitive to the high variability of EEG data. Finally, the "black box" nature of many deep learning models remains a barrier to clinical adoption, where interpretability is paramount.

The IEFS-GMB Framework

To address these issues, the authors propose IEFS-GMB, which consists of two core components:

  • Gradient Memory Bank (GMB): This module maintains a dynamic queue of gradients from previous training iterations. By sampling relevant historical gradients using cosine similarity and applying a temporal decay rate, the model gains a robust, long-term perspective on feature importance, rather than reacting to the noise of a single batch.
  • Information Entropy-based Feature Selection (IEFS): This module calculates the importance of local features by analyzing the accumulated gradients. It uses information entropy as a proxy for confidence; lower entropy indicates higher prediction certainty, allowing the model to amplify diagnostic signals and suppress noise.

Because it operates on feature representations within the network, IEFS-GMB is "plug-and-play" and can be inserted into almost any modern EEG encoder.

Experimental Evaluation

The IEFS-GMB module was tested across four diverse, publicly available datasets: TUEV (spike detection), TUSZ (epilepsy diagnosis), APAVA (Alzheimer’s), and TDBRAIN (Parkinson’s). It was integrated into state-of-the-art encoders like SpikeNet and TimeXer to evaluate its generalizability.

Results and Interpretability

The integration of IEFS-GMB led to accuracy improvements ranging from 0.64% to 6.45% across the benchmark datasets, outperforming leading feature selection methods such as SCConv and TAL.

Beyond raw performance, the method offers significant clinical value through its interpretability. The module can generate "heatmaps" that map weights back to the raw EEG signal. In practice, the model successfully identified specific clinical markers, such as the "canine tooth" waveform characteristic of epileptic spikes, providing a transparent rationale for its classifications. Despite the added memory component, the module remains lightweight and efficient, often achieving peak performance with fewer parameters than competing methods.

Conclusion and Future Work

IEFS-GMB represents a significant step forward in making deep learning for EEG both more accurate and more interpretable. By treating uncertainty as a filter and leveraging historical training context, the module effectively navigates the noisy landscape of neurological data. Future work will focus on optimizing computational overhead and adapting the framework for real-time clinical monitoring of streaming EEG data.

Summary updated

BibTeX
@misc{zhang2025iefs,
  author = {Liang Zhang and Hanyang Dong and Jia-Hong Gao and Yi Sun and Kuntao Xiao and Wanli Yang and Zhao Lv and Shurong Sheng},
  title = {{IEFS-GMB: Gradient Memory Bank-Guided Feature Selection Based on Information Entropy for EEG Classification of Neurological Disorders}},
  year = {2025},
  eprint = {2509.15259},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.15259},
}
Figure