慢性壓力和焦慮的頻譜特徵反映在大腦微狀態中

Abstract

腦電圖，是探索大腦活動的重要工具之一。將腦電圖從時域轉為頻域，可了解大腦在不同頻率的運作模式及特徵。譜圖與頻譜，為腦電圖在頻域的兩種表現形式。其中譜圖與頻譜不同之處在於其帶有時間資訊，使其可提供大腦在連續時間點下不同頻率的強度變化。若將單一時間點的不同頻率強度視為一種頻譜微狀態，我們就可透過譜圖得到隨著時間變化下的頻譜微狀態變化。在本研究中，我們評估了譜圖是否可作為慢性壓力及焦慮的潛在生物標誌物。如今效率及高壓環境的雙重夾擊下，心理健康備受重視。壓力，對我們的大腦有諸多的影響。除了協助我們應付突發狀況，也對具有壓力或情感激發的事件有著記憶增強的作用，這有助於記住重要信息。雖然如此，長期地曝露在壓力之下，會對我們的身心造成負面影響。此外，焦慮亦對我們的日常作息造成嚴重影響，造成注意力無法集中，導致嚴重的精神健康問題。在本研究中，慢性壓力的研究使用LEMON資料集，先將受試者在靜息狀態下收集的腦電圖資料轉化為譜圖及頻譜，再用於訓練卷積神經網路（CNN），以預測由Perceived Stress Questionnaire（PSQ）測量的症狀。焦慮的研究則使用另一個資料集，【再測】資料集，後續處理流程與慢性壓力的處理流程一樣, 以預測由Self-Rating Anxiety Scale （SAS）測量的症狀。結果顯示在兩個資料集，卷積神經網路在以譜圖及頻譜作為訓練材料皆取得優異表現。除了分類結果上的優異表現，本研究發現，在兩個資料集上，卷積神經網路在譜圖上的表現整體而言優於在頻譜的表現，說明了時間資訊有助於提升模型表現。此外，基於時間資訊的有效性，本研究試圖瞭解譜圖中的時間資訊，即頻譜微狀態在譜圖中扮演的關鍵角色，因此使用另一個也在時間資訊上表現強大的深度學習模型，長短期記憶神經網絡（LSTM）與卷積神經網路作比較。結果顯示卷積神經網路在譜圖上的準確率優於長短期記憶神經網絡在譜圖的表現。這說明了雖然時間資訊有助於提升模型表現，但主要並非由於時間資訊帶來的頻譜微狀態之間的長短期關係（長短期記憶神經網絡），反而是利用卷積神經網路同時掌握各時間點下頻譜微狀態的內部特徵帶來優異表現。總結來說，健康受試者與患者的頻譜微狀態差異可作為有效的生物標誌物，卷積神經網路結合譜圖為適合、有效且強大的分類工具。

Electroencephalogram (EEG) is one of the most important tools for exploring brain activity. Switching EEG from time domain to frequency domain can help us to understand how the brain works and how it behaves at different frequencies. Spectrogram and spectrum are the two manifestations of EEG in frequency domain. The key difference between a spectrogram and a spectrum is that a spectrogram includes time information. This allows it to show how the power of different frequencies in the brain changes over successive time points. If the different frequency intensities at a single time point are considered as a spectral microstate, we can obtain the change in the spectral microstate over time through a spectrogram. In this study, we evaluated spectrogram as potential biomarker of chronic stress and anxiety. Under the dual impact of efficiency and high-pressure environment, mental health has become a major priority. Stress has many effects on our brains. In addition to helping us cope with unexpected situations, it also has a memory-enhancing effect on stressful or emotionally stimulating events, which helps us remember important information. However, long-term exposure to stress can have a negative impact on our physical, mental, and spiritual aspects. In addition, anxiety also has a serious impact on our daily routines, resulting in inability to concentrate and leading to serious mental health problems. In this study, LEMON dataset was used in the study of chronic stress, which first converted the resting state EEG data collected from subjects into spectrogram and spectrum, and then used to train Convolutional Neural Network (CNN), to predict the result of Perceived Stress Questionnaire (PSQ). The anxiety study uses another dataset, test-retest dataset, and the follow-up treatment process is the same as that of chronic stress to predict the result of Self-rating Anxiety Scale (SAS). Results showed that CNN performed well in using spectrogram and spectrum as training materials on both datasets. In addition to the excellent performance of the classification results, this study found that, overall, the model performed better on spectrogram than on spectrum in both datasets, indicating that temporal information helps improve model performance. Furthermore, based on the effectiveness of temporal information, this study attempted to understand the critical role played by temporal information in spectrogram, i.e., the role of spectral microstates in spectrogram. Therefore, we used another deep learning model that also performs strongly with temporal information, Long Short-Term Memory (LSTM) network for comparison with CNN. The results showed that CNN's performance on spectrogram was superior to LSTM's performance on spectrogram. This suggests that although temporal information helps improve model performance, it is not primarily due to the long- and short-term relationships between spectral microstates brought about by temporal information (LSTM). Instead, the superior performance is due to CNN's ability to simultaneously capture the internal information of spectral microstates at each time point. In summary, the differences in spectral microstates between healthy subjects and patients can serve as effective biomarkers, and the combination of CNN and spectrogram is a suitable, effective, and powerful classification tool.