基於機器學習應用操弄後的腦波訊號來預測憂鬱症嚴重性頑固性及自殺風險

Abstract

重度憂鬱症是一種反覆發作且嚴重影響生活質量的精神疾病，需要早期且有效的介入與治療。若未能及早緩解症狀，往往會變成慢性病。重度憂鬱症的症狀可能包括長時間的悲傷、極度疲勞、食慾和睡眠模式的改變、注意力困難、自我價值感的減低，甚至反覆出現有關死亡或自殺的念頭。部分患者患有輕度至中度嚴重的憂鬱症，但有些人可能有嚴重的憂鬱症狀（即多種症狀或在憂鬱評分量表中得分較高）。一些病患對抗憂鬱症藥物的治療反應不佳，可能需要個別化或其他進一步的治療方式，例如：重複經顱磁刺激（rTMS）。早期識別高度難治性（即對不同類型的抗憂鬱藥物反應不佳）的病患能夠避免浪費無效治療的時間並且及時決定下一步的抗憂鬱藥物治療（如rTMS）。此外，重度憂鬱症中最極端的症狀為自殺意念，甚至是企圖自殺。及早識別病患症狀並及時介入對於應對自殺風險至關重要。因此，能夠在臨床治療之前預測重度憂鬱症患者的憂鬱嚴重程度、高度難治性和自殺風險，並提供個性化、精準和有效的治療，將是未來的重要技術發展。 本研究使用了209名重度憂鬱症患者的靜息和操弄的臨床腦電圖數據，訓練了多個機器學習算法，以預測憂鬱嚴重程度、高度難治性和自殺風險。我們的特徵集包括來自前額葉區域7個通道的電極（FP1、FP2、F7、F3、FZ、F4、F8），5個子頻帶（Alpha、Beta、Delta、Theta、Gamma）和6個線性和非線性特徵（LLE、DFA、ApEn、KFD、HFD、Welch）。本研究所訓練出的機器學習模型能夠顯著預測憂鬱嚴重程度、高度難治性和自殺風險（校正後的p<0.05），而最佳三個模型的準確率分別為82.8%、86.7%和97.6%。並依照模型重要特徵能夠發現，用於分類憂鬱嚴重程度的特徵位於FP2通道的Alpha頻帶。用於分類高度難治性的特徵主要分布在FP1通道的Theta頻帶和F7通道的Beta頻帶。用於分類自殺風險的特徵分佈在FP1、F3和FZ通道的Beta和Gamma頻帶。 本研究亦設計一個方便且易於操作的使用者介面，顯示了憂鬱嚴重程度、高度難治性和自殺風險的預測結果，以及每個模型在進行這些預測時所考慮的三個最重要特徵。該介面有助於患者更好地理解醫護人員所想要傳達的信息。

Major depressive disorder (MDD) is a recurrent and highly disabled mental illness that warrants early and effective interventions. Without early symptomatic remission, MDD tends to be chronic. Symptoms of MDD may include prolonged periods of sadness, fatigue, changes in appetite and sleep patterns, difficulty concentrating, feelings of worthlessness, and even recurrent thoughts of death or suicide. Some patients have mild to moderate severity of depression, but some may have severe depression of severity (i.e., multiple symptoms or a higher rating in depression rating scale). A significant proportion of MDD individuals may not respond well to antidepressant treatments, requiring personalized or further treatments such as repetitive transcranial magnetic stimulation (rTMS). Early identifying patients with high levels of refractoriness (i.e., poor responses to different kinds of antidepressants) can avoid a waste of time on ineffective treatment and enable the timely decision for next-line antidepressant treatment (e.g., rTMS). Moreover, the most extreme symptoms of MDD are suicidal ideations and even attempts. Early recognition of symptoms and prompt intervention are vital in addressing this risk. Hence, predicting the severity of depression, high refractoriness of treatments, and suicide risk of MDD patients prior to clinical treatment, and delivering personalized, precise, and effective treatments, will be a crucial technological advancement in the future. The present study used resting and modulated electroencephalography (EEG) data of 209 patients with MDD to train several machine learning algorithms to predict the severity of depression, high refractoriness, and suicide risk. Our feature set included 7 channels from the frontal region (FP1, FP2, F7, F3, FZ, F4, F8), 5 sub-bands (Alpha, Beta, Delta, Theta, Gamma), and 6 linear and nonlinear features (LLE, DFA, ApEn, KFD, HFD, Welch). The machine learning model in the study significantly predicted depression severity, high refractoriness, and suicide risk (corrected p<0.05), while the accuracy obtained on the best three models were 82.8%, 86.7%, and 97.6%. The feature for classifying depression severity lies in alpha band of FP2 channel. The feature for classifying high refractoriness is predominantly found in theta band of FP1 channel and beta band of F7 channel. The features for classifying suicide risk are distributed across beta and gamma band of FP1, F3, and FZ channel. The study also developed a user-friendly interface for predicting the three dimensions of severity in MDD. The interface displays the predicting results for severity of depression, high refractoriness, and suicide risk, along with the top three most important features considered by each model in making these predictions. It also provides visualizations of EEG signals and explanations of the meaning behind the extracted features. As an aid for clinicians to identify levels of severity in MDD, the interface demonstrated objective data and meaningful information when explaining the results to patients. Such features facilitate patients' understanding of the information conveyed by their healthcare providers.