以休息狀態腦磁波進行情感性疾病之分類

Abstract

近年來，受情感性疾病 (Mood Disorder) 所苦的病患日益增加，此類疾患嚴重擾亂病人的情緒，進而對日常生活層面造成不良影響，而其中又屬躁鬱症 (Bipolar Disorder) 以及重鬱症 (Major Depressive Disorder) 最廣為所知。情緒性疾病已漸漸成為現代人的主要疾病之一，關於此類疾病的各方面研究也在近數十年內蓬勃發展，其中，患者腦部結構與功能的異常被認為是情感性疾病的重要病因之一。

關於情感性疾病在腦部異常的研究，主要分為腦結構影像與腦波訊號兩方面。然而現今對於患者腦波的研究仍顯不足，最主要的困難之一在於如何自腦波訊號中擷取具有鑑別力的訊號特徵。在本篇論文當中，我們和台北榮民總醫院合作，取得情緒性疾病患者在休息狀態的腦磁波 (Magnetoencephalography) 訊號量測資料。受試者包含二十六位躁鬱症患者、二十二位重鬱症患者以及二十五位做為對照組的健康受試者。在本篇研究中我們分析研究這三個群組的腦磁波訊號，提出具有鑑別力的訊號特徵並且對此三群組加以分類。

在本篇論文中我們使用三種類型的特徵擷取方法，其一是從功率頻譜密度(Power Spectrum Density)中所擷取的特徵，其次為時序訊號上的複雜度，包含Lempel-Ziv Complexity以及Sample Entropy，最後再總合前兩類型特徵以取得左右半腦非對稱性的特徵。針對所擷取的特徵，我們使用統計學中的T檢定(t-test)以及線性判別分析(Linear Discriminant Analysis)的方法，挑出有鑑別力的訊號特徵並藉以將特徵空間的維度降至合理的範圍。在本篇論文中我們對所擷取的特徵做了詳細的分析與探討，此外並使用支援向量機(Support Vector Machine)作為分類器。最後，在任兩群組以及三個群組的分類中得到良好的分類正確率，證明用於本篇論文中的訊號特徵對於情感性疾病具有一定程度的鑑別能力。

Recently, more and more people are suffering from mood disorders such as Bipolar Disorder (BD) and Major Depressive Disorder (MDD). These mood disorders have become one of the major illnesses of modern people. Therefore, researchers are attempting to study these disorders in different areas, including the abnormality of brain structure and brain signals.

However, studies about the abnormality of brain signals are still insufficient and inconsistent. One of the main difficulties is to obtain significant features for further analysis. In this work, we studied three groups of resting Magnetoencephalographic signal data collected by Taipei Veterans General Hospital, including 26 patients with BD, 22 patients with MDD, and 25 normal controls. We then proposed a procedure to classify the three study groups from each others.

In this work, we studied features obtained from power spectrum density, Lempel-Ziv complexity, sample entropy, multi-scale entropy, and hemispheric asymmetry. After the feature extraction, t-test and Linear Discriminant Analysis were applied as feature selection and also to reduce the features to a reasonable number. We provided methodical analysis of the selected features. Furthermore, we applied Support Vector Machine to classify the three groups. The results showed an almost 100\% accuracy in the classification, verifying the significance of our features.