Anxiety Disorder Identification with Biomarker Detection through Subspace-Enhanced Hypergraph Neural Network

Anxiety Disorder Identification and Biomarker Detection Based on Subspace-Enhanced Hypergraph Neural Network

Anxiety Disorder Identification and Biomarker Detection Based on Subspace-Enhanced Hypergraph Neural Network

Academic Background

Anxiety Disorders (ADs) are prevalent mental health issues globally, affecting approximately 7.3% of the population. Patients with anxiety disorders typically exhibit excessive fear, worry, and related behavioral abnormalities, which severely impact their social functioning and quality of life, while also placing a significant burden on families and society. Anxiety disorders can be categorized into several subtypes, such as Generalized Anxiety Disorder (GAD), Social Anxiety Disorder (SAD), Panic Disorder (PD), and Specific Phobia (SP). Although these subtypes are often diagnosed through clinical observation, it is still necessary to differentiate patients from healthy individuals using biomarkers to better identify abnormal changes in the brain.

In recent years, Deep Learning (DL) techniques have been widely applied in the diagnosis of mental disorders, particularly Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs), which have shown excellent performance in anxiety disorder classification. However, traditional deep learning models struggle to capture non-Euclidean relationships among brain regions when processing functional Magnetic Resonance Imaging (fMRI) and Electroencephalography (EEG) data. To address this issue, Hypergraph Neural Networks (HGNNs) have been proposed to capture complex structural information among brain regions.

Source of the Paper

This paper was co-authored by Yibin Tang, Jikang Ding, Ying Chen, Yuan Gao, Aimin Jiang, and Chun Wang, affiliated with the College of Information Science and Engineering at Hohai University, the School of Microelectronics and Control Engineering at Changzhou University, and the Department of Psychiatry at Nanjing Brain Hospital affiliated with Nanjing Medical University. The paper was published in 2025 in the journal Neural Networks, titled “Anxiety Disorder Identification with Biomarker Detection through Subspace-Enhanced Hypergraph Neural Network.”

Research Process

1. Data Preprocessing

The research team collected data from 179 anxiety disorder patients at the outpatient clinic of Nanjing Brain Hospital, including 48 GAD patients, 51 PD patients, 25 SAD patients, and 55 SP patients. Additionally, 108 Healthy Control (HC) participants were recruited. All participants underwent MRI scans to obtain T1-weighted anatomical images and resting-state echo-planar imaging sequences. Using DPABI, MRICron, and SPM toolboxes, the MRI data were preprocessed to obtain three types of multimodal data: Amplitude of Low-Frequency Fluctuations (ALFF), Regional Homogeneity (ReHo), and Voxel-Based Morphometry (VBM). Finally, the research team focused on 18 regions of the limbic system, with each participant’s regional multimodal data having a dimension of 18×3.

2. Feature Selection and Extraction

The research team proposed the Subspace-Enhanced Hypergraph Neural Network (SEHGNN) model and embedded it into the existing Binary Hypothesis Testing (BHT) framework, forming the SEHGNN-BHT flowchart. During the feature selection stage, the Support Vector Machine with Recursive Feature Elimination (SVM-RFE) method was used to calculate the reliability weights of each multimodal data, and the top 10 regions were selected as typical multimodal data. In the feature extraction stage, the SEHGNN model extracted high-dimensional features through the Subspace-Enhanced Hypergraph Convolution (SEHGC) operation.

3. Anxiety Disorder Classification

In the anxiety disorder prediction stage, the research team evaluated the clustering performance of high-dimensional features under different hypotheses and determined the true hypothesis by calculating and comparing variability scores, assigning the corresponding label to the test subject. To improve classification performance, the research team also introduced an ensemble learning strategy, running the SEHGNN-BHT framework multiple times and using a majority voting strategy to determine the final predicted label.

Main Results

The research results showed that the SEHGNN model performed exceptionally well in anxiety disorder classification, achieving an accuracy of 84.46%. With the ensemble learning strategy, the model’s performance further improved, reaching an accuracy of 94.1%. Additionally, the SEHGNN model successfully identified biomarkers associated with anxiety disorders, which were consistent with existing research reports, providing strong evidence for the method’s effectiveness and interpretability.

Conclusions and Significance

The SEHGNN model proposed in this study emphasizes the impact of each hyperedge and achieved a high anxiety disorder classification accuracy of 94.1% with the help of ensemble learning. The method utilizes the SEHGNN-BHT framework to process multimodal data, and experimental results demonstrated that the SEHGNN model outperformed other deep learning-based models within the limited regions of the limbic system. Furthermore, through t-tests and Multivariate Analysis of Variance (MANOVA) for biomarker detection, the biomarkers identified by the SEHGNN model aligned with recent research findings, highlighting the method’s interpretability and its potential in uncovering the developmental mechanisms of anxiety disorders.

Research Highlights

  1. High Classification Accuracy: The SEHGNN model achieved an accuracy of 94.1% in anxiety disorder classification, significantly outperforming other deep learning-based models.
  2. Biomarker Identification: The SEHGNN model successfully identified biomarkers associated with anxiety disorders, providing new perspectives for disease diagnosis and treatment.
  3. Subspace-Enhanced Hypergraph Convolution Operation: By introducing a learnable weight matrix, the SEHGNN model adaptively enhanced the impact of hyperedges, improving the effectiveness of feature extraction.
  4. Ensemble Learning Strategy: By running the SEHGNN-BHT framework multiple times and using a majority voting strategy, the research team further enhanced the model’s classification performance.

Other Valuable Information

The research team also conducted hyperparameter tuning experiments and found that selecting 10 brain regions as typical multimodal data yielded the best classification performance. Additionally, the research team quantified the contribution of each region to anxiety disorder classification through subspace signal decomposition, further validating the effectiveness of the SEHGNN model.

This study provides new tools and methods for the diagnosis and treatment of anxiety disorders, holding significant scientific value and application prospects.