基於機器學習和深度學習應用於預測重度憂鬱症之靜息態功能性磁共振造影訊號

Using rs-fMRI Signals for Prediction of Major Depressive Disorder Based on Machine Learning and Deep Learning

林家軒(Chia-Hsuan Lin)

指導教授 ： 陳中平

共同指導教授 ： 李正達(Cheng-Ta Li)

國立臺灣大學/電機資訊學院/生醫電子與資訊學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202201254

摘要

重度憂鬱症是影響人類的一種常見的世紀重大疾病。而具有一定比率的重度憂鬱症患者在數個抗憂鬱症藥物的治療下沒有好轉就會被定義成頑固型憂鬱症。本研究分析199位受試者的靜息態功能性磁共振造影訊號，包含頑固型憂鬱症患者（無效藥物3種以上）、非頑固型憂鬱症患者（無效藥物2種以下）和健康受試者，我們想要建構一套輔助醫師預測是否為憂鬱症患者並依照嚴重程度預測是否為頑固型憂鬱症患者或是非頑固型憂鬱症患者的系統。
我們從靜息態功能性磁共振造影訊號中提取三種不同面向的特徵，包含區域同質性特徵、圖論式網路特徵和皮爾森相關性特徵，通過結合大量的特徵再選出最為相關且重要的特徵，經由我們的機器學習演算法能夠成功預測並分別在三種模型上（預測是否憂鬱症、預測是否頑固型憂鬱症、複雜情況預測是否頑固型憂鬱症）得到的準確率是84.3%、88.4% 和78.6%。並依照重要特徵能夠找出憂鬱症特徵分布在額葉、顳葉和頂葉，而頑固性嚴重度特徵主要分布在額葉和顳葉。
最後，我們透過深度學習來簡化處理資料步驟，並分別在三種模型上的準確度為87.5%、84.2% 和71.0%。
關鍵字：重度憂鬱症、頑固型憂鬱症、靜息態功能性磁共振造影、機器學習、深度學習

關鍵字

重度憂鬱症 ； 頑固型憂鬱症 ； 靜息態功能性磁共振造影 ； 靜息態功能性磁共振造影 ； 機器學習 ； 深度學習

並列摘要

Major depressive disorder (MDD) is a common and a major disease of the century that affects human beings. Treatment resistant depression (TRD) is described as a fraction of people with MDD who do not cure after being treated with various antidepressant medicines. This study analyzed the resting-state fMRI (rs-fMRI) signals of 199 subjects, including patients with TRD (more than 3 ineffective drugs), Non-TRD (with less than 2 ineffective drugs) and healthy subjects. We want to construct a system that assists physicians in predicting depression and predicting TRD or Non-TRD according to severity.
We extracted three different oriented features from the rs-fMRI signals, including correlation-based feature, Graph-based feature and Between-region connectivity-based features, and then selected the most relevant and important features by combining a large number of features. Our machine learning algorithm can predict the class of subject and the accuracy obtained on the three models (prediction of depression, prediction of resistant severity, and complex case prediction of resistant severity) are 84.3% and 88.4%. and 78.6%. And according to the important features, it can be found that the features of depression are distributed in the frontal lobe, temporal lobe and parietal lobe, and the features of resistant severity are mainly distributed in the frontal and temporal lobe.
Finally, we use deep learning to simplify the data processing step and achieve 87.5%, 84.2% and 71.0% accuracy on the three models, respectively.
Keywords: major depression, refractory depression, resting-state functional magnetic resonance imaging, machine learning, deep learning

並列關鍵字

major depression ； refractory depression ； esting-state functional magnetic resonance imaging ； machine learning ； deep learning

參考文獻

Nnatu, D.I. What is Treatment-Resistant Depression? 2022; Available from: https://www.priorygroup.com/mental-health/depression-treatment/treatment-resistant-depression.

Bergfeld, I.O., et al., Treatment-resistant depression and suicidality. Journal of affective disorders, 2018. 235: p. 362-367.

Kennedy, N. and E. Paykel, Residual symptoms at remission from depression: impact on long-term outcome. Journal of affective disorders, 2004. 80(2-3): p. 135-144.

Fava, M., Diagnosis and definition of treatment-resistant depression. Biological psychiatry, 2003. 53(8): p. 649-659.

Rush, A.J., et al., Star* d. CNS drugs, 2009. 23(8): p. 627-647.