How Homogenizing the Channel-wise Magnitude Can Enhance EEG Classification Model?

Huyen Ngo, Hoang Khoi Do, Duong Nguyen, Viet Dung Nguyen, Lan DangarXiv 2024 · DOI · arXiv

Abstract

Electroencephalogram (EEG) data classification is often hindered by the "dominant lead" problem, where high-magnitude channels overshadow subtle neural information. This paper proposes a two-step preprocessing pipeline: Inverted Channel-wise Magnitude Homogenization (ICWMH) and Feature Enrichment via Skip Connection (FEvSC). By transforming raw EEG signals into homogenized 2D images and highlighting neural transitions through edge detection, the method achieves substantial accuracy gains—up to 20%—over traditional deep learning architectures.

Introduction

Electroencephalogram (EEG) signals are notoriously difficult to analyze due to their temporal, non-linear, and redundant nature. A significant challenge is the "dominant lead" problem: electrodes placed at different scalp locations naturally exhibit varying signal powers. Standard Deep Learning (DL) models frequently over-focus on these high-magnitude "loud" channels, neglecting vital information in "quieter" ones. While many researchers attempt to solve this through complex neural architectures like LSTMs or specialized CNNs, this study suggests that the primary bottleneck lies in the data representation itself rather than the model architecture.

Methodology

The authors propose a two-step approach to normalize EEG data and enhance its informative features:

  1. Inverted Channel-wise Magnitude Homogenization (ICWMH): This technique normalizes the amplitude across all channels. Using "bottle-neck normalization," multi-channel signals are squeezed into a fixed range. Subsequent interpolation creates a 2D "EEG-encoded image," ensuring that no single electrode dominates the learning process based solely on voltage.
  2. Feature Enrichment via Skip Connection (FEvSC): After encoding the EEG as an image, classic computer vision techniques—specifically Canny Edge Detection—are applied. In this context, an "edge" represents a sharp transition in neural activity. A skip connection adds these edge features back to the original homogenized image, allowing the model to focus on high-frequency neural "bursts" while maintaining structural context.
The proposed ICWMH+ED pipeline, featuring magnitude homogenization and edge-based feature enrichment.

The resulting three-layer image is then processed by a standard Convolutional Neural Network (CNN) for classification.

Experimental Setup

The method was evaluated on two distinct benchmarks:

  • Perceive Lab Dataset: EEG responses from subjects viewing 2,000 unique objects across 40 ImageNet classes (128 channels, 50Hz).
  • High-Gamma Dataset (HGD): A motor-movement dataset involving four types of physical actions (128 electrodes).

Results and Discussion

The proposed ICWMH+ED method consistently outperformed established baselines across both datasets.

Performance comparison of the proposed ICWMH+ED method against various baseline models.

On the Perceive Lab dataset, the model achieved 65.78% accuracy, significantly higher than EEGChannelNet (39.58%) and standard LSTMs (23.97%). For the High-Gamma Dataset, it reached 57.18% accuracy, more than doubling the performance of standard grayscale-encoded image methods (24.54%).

Key Insights

  • Representation over Complexity: Fixing data representation can allow simpler CNNs to outperform massive state-of-the-art models.
  • Gradient Dominance: ICWMH prevents gradients from being overwhelmed by high-magnitude channels, forcing the model to learn from the entire multi-channel spectrum.
  • Edges as Transitions: Treating EEG as an image enables spatial filters to act as low-computation feature extractors for critical neural boundaries.
  • Interpolation Sensitivity: Bilinear interpolation was found to be superior to nearest-neighbor methods, indicating that smooth transitions are vital for preserving temporal logic.

Conclusion

While the ICWMH+ED pipeline offers significant performance improvements, it is sensitive to hyperparameter tuning, particularly the Canny edge detection thresholds and Gaussian blur kernel size. Future research will focus on automating threshold selection and testing the robustness of this 2D encoding on real-time, noisy streaming EEG data.

Summary updated

BibTeX
@misc{ngo2024how,
  author = {Huyen Ngo and Hoang Khoi Do and Duong Nguyen and Viet Dung Nguyen and Lan Dang},
  title = {{How Homogenizing the Channel-wise Magnitude Can Enhance EEG Classification Model?}},
  year = {2024},
  eprint = {2407.20247},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.20247},
}
Figure