Functional Brain Network Based on Improved Ensemble Empirical Mode Decomposition of EEG for Anxiety Analysis and Detection

Neurology Information Science Artificial Intelligence Medicine Psychiatry Biophysics Medical Imaging Life Sciences
Functional Brain Network EEG Anxiety Improved Ensemble Empirical Mode Decomposition Support Vector Machine Phase Lag Index

Brain Functional Network for Anxiety Analysis and Detection Based on Improved Ensemble Empirical Mode Decomposition

Academic Background and Research Objectives

With the increasing stress of modern life, anxiety, a common neurological disorder, has become an urgent issue in global public health. Anxiety not only manifests as mental disorders but also involves abnormal performance in cognitive processes such as attention, memory, and learning. The outbreak of the COVID-19 pandemic has further increased the prevalence of anxiety. According to statistics, the 12-month incidence rate of anxiety is 4.80% for men and 5.20% for women. However, the etiology of anxiety remains unclear, and the probability of spontaneous recovery is low. These complexities and uncertainties make early detection and intervention particularly important. Traditional methods for anxiety detection rely on face-to-face interviews and self-assessments, which are time-consuming, laborious, and influenced by subjective factors from both doctors’ professional experiences and patients’ self-assessments. Therefore, there is a need to explore an objective and accurate method for anxiety analysis and detection.

Among various physiological signals, electroencephalogram (EEG) has garnered increasing attention from researchers due to its high temporal resolution and low cost. Traditional anxiety studies based on isolated EEG electrodes find it difficult to reveal abnormal changes in the brain’s topological structure. Hence, this study proposes a new framework for anxiety detection based on improved ensemble empirical mode decomposition (EEMD) and brain functional networks (BFN).

Source of the Paper and Author Information

This paper was co-authored by Bingtao Zhang, Chonghui Wang, Guanghui Yan, Yun Su, Lei Tao, and Hanshu Cai, from institutions including Lanzhou Jiaotong University, Shaanxi University of Science and Technology, and Northwest Normal University. The paper was published in the “Biomedical Signal Processing and Control” journal on February 5, 2024.

Research Process and Detailed Experimental Design

Research Process

  1. Acquisition and preprocessing of experimental data.
  2. Improvement of the EEMD method and application to EEG signal decomposition.
  3. Construction of BFN based on improved EEMD of EEG signals.
  4. Calculation of inter-group difference BFN.
  5. Exploration of anxiety-related brain regions and potential biomarkers.
  6. Detection of anxiety patients and normal controls (NC).

Improved EEMD Method

EEMD solves the mode mixing problem in EMD’s multi-frequency component decomposition by adding auxiliary random white noise. However, the selection of the amplitude of the random white noise directly affects the performance of EEMD. This study proposes an improved EEMD method based on additional adaptive white noise, which determines the amplitude of white noise at each sampling point adaptively according to the signal-to-noise ratio (SNR) theory, to more accurately simulate independent neuronal signals.

EEG Signal Decomposition

EEG signals are decomposed into 7 intrinsic mode functions (IMFs) using the improved EEMD method. By comprehensively analyzing the distribution of the number of IMFs generated in each EEG time window, the lowest number of IMFs is selected as the threshold to avoid excessive loss of information. Subsequently, the phase lag index (PLI) between each IMF is calculated to construct the binarized BFN.

BFN Construction and Binarization

The network consists of nodes (EEG electrodes) and edges (correlation values between nodes). The connection weights between nodes are calculated based on PLI, and the corresponding BFN matrix is constructed. To reduce computation and ensure the efficiency of BFN, a proportional threshold method is used for binarization of BFN. Based on network density and the number of nodes, the proportional threshold is calculated to be 8.89%.

Exploration of Anxiety-Related Brain Regions

Studies have shown that the functional changes in the brains of anxiety patients are mainly concentrated in the left frontal lobe (LF), left temporal lobe (LT), and left central brain region (LC). By analyzing the differences in typical graph metrics of these brain regions, it is determined whether these metrics can serve as potential biomarkers for anxiety detection.

Performance Evaluation of Anxiety Detection

Support vector machine (SVM) is chosen as the classifier, and a 10-fold cross-validation method is used to evaluate the detection performance of potential biomarkers. The experimental results show that the highest detection accuracy for anxiety is 92.07% when considering the graph metrics of IMF1 and IMF4 comprehensively.

Research Results

Effectiveness Evaluation of Improved EEMD Method

By comparing the effects of constant amplitude and adaptive amplitude white noise on the original signal, it is found that adding adaptive white noise can more comprehensively affect the signal’s extrema, thus more effectively solving the mode mixing problem.

Analysis Results of Anxiety-Related Brain Regions

By comparing the difference matrices of BFN for IMF1 and IMF4, it is found that anxiety patients mainly exhibit changes in functional synchronization in the LF, LT, and LC brain regions. This is consistent with the conclusions of previous studies, further confirming the view that anxiety mainly involves left hemisphere activity.

Potential Biomarkers for Anxiety Detection

A one-way analysis of variance (ANOVA) was conducted on the BFN of IMF1 and IMF4, revealing significant differences in indicators such as LF-CC, LT-eloc, LC-CC, and whole-brain σ between the anxiety group and normal control group, indicating that these indicators can serve as potential biomarkers for anxiety detection.

Performance Evaluation of Anxiety Detection

By comprehensively using the graph metrics of BFN for IMF1 and IMF4, the 10-fold cross-validation results based on the SVM classifier show that the average accuracy of anxiety detection is 87.31%, with a maximum accuracy of 92.07%.

Conclusion

This study, through the improved EEMD method, for the first time, constructed BFN based on decomposed EEG signals and deeply explored the changes in brain functional topological structure of anxiety patients. The results indicate that the BFN of anxiety patients exhibit a certain trend towards randomization and that significant graph metrics can serve as effective biomarkers for anxiety detection. These findings not only provide new ideas for early detection of anxiety but also provide important theoretical basis for studying the mechanisms of neuropsychiatric disorders.

Academic Value and Application Prospects

This study has significant academic value in the following aspects:

  1. The improved EEMD method effectively solves the mode mixing problem in EMD decomposition, improving the accuracy of EEG signal decomposition.
  2. The construction of BFN based on EEG signals using the improved EEMD method provides a new method for anxiety detection.
  3. The study reveals the abnormal topological structure in BFN of anxiety patients, providing a new perspective for understanding the pathological mechanisms of anxiety.
  4. The proposed potential biomarkers for anxiety detection have important application value in practical applications, aiding in the early detection and intervention of anxiety disorders.

By exploring the topological changes in brain functional networks associated with anxiety disorders through the improved EEMD method, this study provides important theoretical foundations and practical evidence for the early diagnosis and treatment of anxiety disorders.