應用圖卷積類神經網路於腦波特徵學習之研究

Research on Applying Graph Convolution Neural Networks to Electroencephalography Feature Learning

邱益鴻(Yi-Hung Chiu)

指導教授 ： 歐陽振森

義守大學/智慧科技學院/資訊工程學系/碩士(2021年)

摘要

注意力不足過動症(Attention Deficit Hyperactivity Disorder; ADHD)是一種神經發展類型疾病之一，常會影響患者的學習與生活。現今以量表評估為主的診斷方法易受主觀性影響而產生偏差，因此如何準確且客觀地診斷注意力不足過動症已成為臨床醫學重要問題。本研究提出一種基於時頻與空頻特徵融合圖卷積類神經網路分類模型，用以學習具鑑別注意力不足過動症患者與健康者的腦波特徵，進而預測輸入腦波為注意力不足過動症或是健康者類別。為驗證本研究所提出分類模型的優越性，除了比較多種用以取得各電極頻道相應關係拓樸圖的montage maps之外，亦與其他三個卷積與遞迴混合型類神經網路進行比較。實驗結果顯示相較於其他方法，本研究所提分類模型的具有較佳的整體測試表現。在基於資料的分類表現指標方面，達到敏感度78.30%、特異度79.00%、準確率78.66%與特徵曲線下面積85.86%；在基於人的分類表現指標方面，達到敏感度83.60%、特異度82.52%、準確率82.17%與特徵曲線下面積88.22%。因此，本研究所提出的基於時頻與空頻特徵融合圖卷積類神經網路分類模型可以提供準確且客觀的注意力不足過動症診斷參考。

關鍵字

注意力不足過動症 ； 腦波 ； 深度學習 ； 特徵學習 ； 圖卷積類神經網路

並列摘要

Attention deficit hyperactivity disorder (ADHD) is a kind of neurodevelopmental disorder and usually affects patients’ learning and life. The current diagnosis approach based on scale evaluation is easily affected by subjectivity, resulting in the bias. Therefore, how to obtain precise and objective diagnosis of ADHD has been an important problem in clinical medicine. In this study, we propose a graph convolution neural network classification model based on the fusion of temporal features and spatial features. The proposed classification model is employed to lean electroencephalogram (EEG) features for discriminating ADHD patients from healthy people and further predict the inputted EEG as ADHD or healthy. To verify the superiority of the proposed classification model, in addition to making a comparison between five montage maps used to obtain relative relationships among electrode channels, the comparison between the proposed model and the other three models is also made. Experimental results have shown that compared with other models, our proposed model possess the better overall test performance. For the data-based classification performance indices, our proposed model achieves sensitivity 78.30%, specificity 79.00%, accuracy 78.66%, and AUC 85.86%. For the person-based classification performance indices, our proposed model achieves sensitivity 83.60%, specificity 82.52%, accuracy 82.17%, and AUC 88.22%. Therefore, our proposed graph convolution neural network classification model can provide a precise and objective reference for the diagnosis of ADHD.

並列關鍵字

attention deficit hyperactivity disorder ； electroencephalogram ； deep learning ； feature learning ； graph convolution neural networks

參考文獻

[1] American Psychiatric Association, Diagnostic and Statistical Manual of Mental Disorders, Five Edition: DSM-V, Arlington: American Psychiatric Publishing, 2013.

[2] N. A.Md Norani, W.Mansor, and L. Y.Khuan, “A review of signal processing in brain computer interface system,” in Proc. 2010 IEEE EMBS Conf. Biomed, pp. 443–449, 2010.

[3] M. R.Nuwer, J.Buchhalter, and K. M.Shepard, “Quantitative EEG in attention-deficit/hyperactivity disorder,” Neurol. Clin. Pract., vol. 6, no. 6, pp. 543–548, 2016.

[4] P.Bashivan, I.Rish, M.Yeasin, and N.Codella, “Learning representations from EEG with deep recurrent-convolutional neural networks,” in Proc. International Conference on Learning Representations, pp. 1–15, 2016.

[5] L. J.-S.Moon, Seong-Eun, and Jang Soobeom, “Convolutional neural network approach for EEG-based emotion recognition using brain connectivity and its spatial information seong-Eun moon soobeom jang jong-seok lee republic of korea,” in Proc. 2018 IEEE Int. Conf. Acoust. Speech Signal Process, pp. 2556–2560, 2018.