MGSN: Depression EEG lightweight detection based on multiscale DGCN and SNN for multichannel topology

Xin Chen, Youyong Kong, Hongli Chang, Yuan Gao, Zidong Liu, Jean-Louis Coatrieux, Huazhong ShuBiomedical Signal Processing and Control 2024 · DOI

Introduction

Clinical diagnosis of depression currently relies heavily on subjective assessments, such as patient self-reports and physician experience, which can lead to diagnostic uncertainty. While Electroencephalography (EEG) offers a promising source of objective biomarkers, analyzing it is difficult due to the complex, dynamic interactions between different brain regions. Most existing deep learning methods either ignore these spatial relationships or utilize "deep" architectures that are computationally expensive and prone to overfitting on small medical datasets.

Researchers have developed MGSN, a novel framework that addresses these challenges by combining multiscale graph convolutions with energy-efficient Spiking Neural Networks (SNNs). By capturing the complex topology of brain activity across different scales and fusing responses to emotional stimuli, the model achieves high diagnostic accuracy while remaining significantly more lightweight than traditional models.

Methodology

The Multiscale Dynamic Graph Convolutional Network and Spiking Neural Network (MGSN) architecture consists of three main innovations designed to be both biologically plausible and computationally efficient.

1. Topology Information Transfer (TIT) Block

Instead of stacking many deep layers, the TIT block uses a "shallow but wide" approach. It employs multiple parallel diffusion branches to analyze different "neighborhoods" of brain activity (1-hop, 2-hop, and $k$-hop neighbors). This allows the model to capture both local and global brain connectivity patterns simultaneously without the overhead of deep sequential layers.

2. Spiking Neural Networks (SNNs)

Unlike traditional neural networks that use continuous floating-point numbers, SNNs process information as discrete binary "spikes" (0s and 1s). This mimics the "all-or-none" firing of biological neurons, making the model highly energy-efficient and better suited for spatiotemporal data like EEG.

Figure 4: The implementation idea showing the transition from EEG topology to SNN spikes encoding and depression detection.

3. Multi-Emotional Stimulus Fusion

The framework utilizes EEG data recorded while patients viewed both positive (happy) and negative (sad/fear) faces. Fusing these features creates a more robust diagnostic profile than relying on a single emotional state.

Figure 5: The complete MGSN architecture for detection under single and multi-emotional stimuli, featuring the TIT block and SNN decoder.

Experimental Results

Dataset and Preprocessing

The study utilized the MODMA (Multimodal Open Dataset for Mental Disorder Analysis), involving 24 depressed patients and 29 healthy individuals. 128-channel EEG signals were recorded during emotional-neutral face pair stimulus experiments. Signals were cleaned using Independent Component Analysis (ICA) to remove artifacts like eye blinks and head movements.

Performance Metrics

  • Accuracy: The model reached a peak classification accuracy of 90.51% when fusing "Sad-Neutral" and "Happy-Neutral" stimuli.
  • Model Complexity: MGSN is remarkably lightweight, featuring only 3.24 million parameters. In comparison, other state-of-the-art models like GAT (164M) or LGRNN (172M) are orders of magnitude larger.
  • Speed: The training time was approximately 364 seconds, significantly faster than the 900–1600 seconds required by competing architectures.

Key Insights

  • Negative Stimuli Sensitivity: EEG responses to negative stimuli (sadness and fear) yielded higher classification accuracy than positive stimuli. This suggests that negative events induce more distinct brain responses in depressed patients that are conducive to diagnosis.
  • Optimal Connectivity: Experiments with different $k$-values (the receptive field of the graph) revealed that $k=3$ was optimal. Looking too locally misses the big picture, while looking too globally introduces redundant noise.
  • Biological Markers: Visualization of the network's output confirmed distinct abnormalities in the frontal and temporal lobes of depressed patients, aligning with established neuroscientific research.

Conclusion and Future Work

MGSN demonstrates that a biologically-inspired, lightweight approach can outperform massive deep learning models in medical diagnostics. While the results are promising, future work will focus on validating the model on larger, more diverse cohorts to ensure generalizability and exploring more direct SNN training methods to replace current surrogate gradient approximations. The ultimate goal is a modular, scalable framework to support real-world clinical decision-making.

Summary updated

BibTeX
@article{chen2024mgsn,
  author = {Xin Chen and Youyong Kong and Hongli Chang and Yuan Gao and Zidong Liu and Jean-Louis Coatrieux and Huazhong Shu},
  title = {{MGSN: Depression EEG lightweight detection based on multiscale DGCN and SNN for multichannel topology}},
  journal = {Biomedical Signal Processing and Control},
  year = {2024},
  doi = {10.1016/j.bspc.2024.106051},
}
Figure