GCTNet: A Graph Convolutional Transformer Network for Major Depressive Disorder Detection Based on EEG Signals

Computer Science Engineering Electrical Engineering Information Science Artificial Intelligence Neurology Clinical Psychology Medical Imaging Medicine Psychiatry
Major Depressive Disorder EEG Deep Learning Phase Locking Value Contrastive Learning

GCTNet: Graph Convolution Transformer Network for Detecting Major Depressive Disorder Based on EEG Signals

Research Background

Major Depressive Disorder (MDD) is a prevalent mental illness characterized by significant and persistent low mood, affecting over 350 million people worldwide. MDD is one of the leading causes of suicide, resulting in approximately 800,000 deaths annually. Currently, the diagnosis of MDD mainly relies on patients’ self-reports and clinical judgments by professionals. However, the subjectivity in the diagnostic process can lead to low consistency among different doctors, potentially causing inaccurate diagnoses. Studies have found that the general accuracy of non-specialist diagnosis of MDD is only 47.3%. Therefore, exploring objective and reliable physiological indicators and employing effective methods to timely identify MDD is crucial for promoting early diagnosis and intervention.

Source of the Paper

This paper was co-authored by Yuwen Wang, Yudan Peng, Mingxiu Han, Xinyi Liu, Haijun Niu from the Beijing Advanced Innovation Center for Biomedical Engineering, and Jian Cheng from the School of Computer Science and Engineering, and Suhua Chang from Peking University Sixth Hospital. The paper was published in the Journal of Neural Engineering.

Research Process and Methods

This paper proposes a Graph Convolution Transformer Network (GCTNet) for detecting MDD based on electroencephalogram (EEG) signals. The research process includes data acquisition, preprocessing, model construction, and performance evaluation.

  1. Data Acquisition and Preprocessing: The research data is divided into two datasets: self-collected and public datasets. The self-collected dataset contains 85 subjects, while the public dataset contains 64 subjects. EEG recordings of all participants were conducted under eyes-closed resting-state conditions. The EEG signals were preprocessed through steps like filtering, artifact removal, and down-sampling, and finally segmented into 5-second samples.

  2. Model Construction: The model is mainly composed of three modules: Residual Graph Convolution Network (resGCN) module, Transformer module, and Readout block. The resGCN module extracts spatial information features from the EEG data, the Transformer module extracts temporal features of the EEG signal sequence, and the Readout block integrates the outputs of the two modules for classification. Figure 2 shows the overall architecture of GCTNet.

    • resGCN Module: Constructs the adjacency matrix of the EEG graph using Phase Locking Value (PLV), extracts node representations through multiple layers of graph convolution, and enhances sensitivity with residual connections.
    • Transformer Module: Contains convolution modules, positional embeddings, learnable cls_token, and multiple layers of Transformer encoders. The Transformer captures sequential dependencies of EEG signals.
    • Readout Block: Integrates the features from the first two blocks, and outputs them to the classification and projection space, computing the Contrastive Cross-Entropy (CCE) loss.

  3. Contrastive Cross-Entropy (CCE) Loss: Combines supervised contrastive (SupCon) loss with traditional cross-entropy (CE) loss, significantly improving the identification of MDD by optimizing feature representation and classification performance.

Research Results

The research conducted detailed evaluations on the self-collected dataset (Dataset I