A Novel Method for Depression Detection Based on Time-Frequency EEG: TFAGL

Academic Background

Major Depressive Disorder (MDD) is a common mental illness worldwide, characterized by symptoms such as sadness, guilt, and low self-esteem, accompanied by loss of interest, diminished enthusiasm for life, and disruptions in sleep or appetite. According to statistics from the World Health Organization (WHO), over 246 million people are affected by depression, with approximately 30-35% of severely depressed individuals attempting suicide each year, resulting in about 2-15% of suicides. Therefore, by 2024, depression is expected to become the leading cause of debilitating diseases.

Currently, clinical diagnosis of depression primarily relies on doctor-patient conversations and questionnaire surveys, which are susceptible to interference from patients’ subjective awareness and doctors’ expertise, lacking objectivity. Electroencephalography (EEG) technology can record changes in brain activity, closely related to human brain activity, and objectively reflect mental states. Additionally, EEG has advantages such as non-invasiveness, speed, and affordability. Thus, researchers worldwide are dedicated to utilizing EEG for depression detection and identifying better biomarkers.

However, existing depression detection methods primarily rely on simple simulations of EEG electrode distributions, neglecting collaborative relationships among brain regions, which limits detection performance. To address this, researchers proposed a novel depression detection model based on time-frequency EEG—Time-Frequency Agent Graph Learning (TFAGL)—aiming to capture the whole-brain collaborative mechanism of depression.

Paper Source

This paper was co-authored by Zihua Xu (IEEE Student Member), C. L. Philip Chen (IEEE Fellow), and Tong Zhang (IEEE Senior Member). They are affiliated with the School of Computer Science and Engineering at South China University of Technology, the Engineering Research Center of the Ministry of Education on Health Intelligent Perception and Parallel Digital-Human, and the Pazhou Laboratory in Guangzhou, China. The paper was published in the IEEE Transactions on Affective Computing journal and is expected to be officially released in 2025.

Research Process

1. Research Objectives

The TFAGL model aims to enhance the accuracy and robustness of depression detection by capturing collaborative mechanisms among brain regions. Specifically, the model simulates global interactions among brain regions by generating agent nodes, thereby constructing a dynamic local-global connectivity graph to capture intra- and inter-region connection patterns.

2. Data Preprocessing

The study utilized three public EEG datasets: MODMA, PRED+CT, and TDBrain. The EEG signals from each dataset were segmented into 2-second non-overlapping fragments, treated as independent samples for processing. Specific dataset information is as follows: - MODMA: Contains EEG data from 29 healthy controls and 24 depressed patients, sampled at 250 Hz. - PRED+CT: Contains EEG data from 76 healthy controls and 46 depressed patients, sampled at 500 Hz. - TDBrain: Contains EEG data from 1167 individuals, including 132 diagnosed with depression, sampled at 500 Hz.

3. Feature Extraction

The study designed temporal feature extractors and frequency domain feature extractors to extract time-series and spectral features from EEG signals, respectively. - Temporal Feature Extractor: An improved causal convolutional layer was adopted to capture potential trend changes in EEG signals. - Frequency Domain Feature Extractor: Fast Fourier Transform (FFT) was used to convert signals into the frequency domain, extracting differential entropy features across five frequency bands.

4. Agent Graph Learning Module (AGL)

The AGL module adaptively generates agent nodes to simulate global interactions among brain regions. Specific steps are as follows: - Agent Node Generation: Based on brain region division, agent node locations and attributes are generated through a self-attention mechanism. - Dynamic Latent Connectivity: A dynamic latent connectivity graph is constructed to capture intra- and inter-region connection patterns. - Position-Based Spatial Connectivity: Spatial connectivity graphs are constructed based on the spatial positions of agent nodes. - Multi-Scale Graph Convolution: Multi-scale graph convolution is applied for interactive learning across different receptive fields.

5. Dual-Domain Feature Aggregation Module (DFA)

The DFA module eliminates redundant features and enhances feature discriminability through cross-domain constraints. Specific steps include: - Intra-Domain Feature Fusion: A dynamic feature informativeness encoding network suppresses irrelevant features while retaining effective ones. - Inter-Domain Feature Fusion: Attention mechanisms are used to fuse temporal and frequency domain features.

Main Results

1. Model Performance

The TFAGL model performed exceptionally well across all three datasets: - MODMA: Accuracy of 94.94%, F1 score of 94.32%. - PRED+CT: Accuracy of 93.94%, F1 score of 91.06%. - TDBrain: Accuracy of 75.20%, F1 score of 74.43%.

2. Significance Testing

Through comparative experiments with existing methods and paired t-tests, the TFAGL model demonstrated significantly superior performance across multiple datasets (p < 0.05).

3. Feature Visualization

Using t-SNE (t-distributed Stochastic Neighbor Embedding) visualization, the TFAGL model progressively learned clear classification boundaries during training, indicating its effective capture of EEG signal features.

Research Conclusions

The TFAGL model successfully captured the whole-brain collaborative mechanism of depression by generating agent nodes and applying multi-scale graph convolution, significantly enhancing the accuracy and robustness of depression detection. The innovations of this model include: 1. Agent Node Generation: By adaptively generating agent nodes to simulate global interactions among brain regions, the model’s generalization ability is enhanced. 2. Multi-Scale Graph Convolution: Through multi-scale graph convolution, the model simultaneously captures local and global features, improving the comprehensiveness of feature extraction. 3. Dual-Domain Feature Aggregation: By fusing temporal and frequency domain features, the model effectively eliminates redundant features and enhances feature discriminability.

Research Value

The TFAGL model not only excels in depression detection tasks but also provides new research ideas for other abnormal EEG pattern detections, such as epilepsy and Parkinson’s disease. In the future, researchers will further explore the collaborative patterns among EEG electrodes, promoting the model’s adaptive application across different individuals.

Research Highlights

1. Whole-Brain Collaborative Mechanism: The TFAGL model is the first to capture the whole-brain collaborative mechanism of depression through agent nodes, filling a gap in existing research.
2. Multi-Scale Graph Convolution: Through multi-scale graph convolution, the model can simultaneously capture local and global features, enhancing the comprehensiveness of feature extraction.
3. Cross-Domain Feature Fusion: By fusing temporal and frequency domain features, the model effectively eliminates redundant features, enhancing feature discriminability.

Other Valuable Information

The study also found that activity in the right prefrontal cortex of depressed patients was significantly higher than in other brain regions, closely related to negative emotional experiences and withdrawal behaviors in depression. This finding provides new evidence for the biological mechanisms of depression.