基於機器學習非目標腦波排除

Study on Machine Learning Based Non-target EEG Rejection for Multi-state Brain Computer Interface

蕭玉聰(Yu-Tsung Hsiao)

指導教授 ： 劉益宏

中原大學/工學院/機械工程研究所/碩士(2014年)

https://doi.org/10.6840/cycu201400905

摘要

摘要
腦機介面(Brain Computer Interface, BCI)在近年來的研究與發展已越來越受到重視，其可以讓運動功能嚴重受損的病患(例如中風、漸凍症)利用其腦波(EEG)直接與外界溝通、或控制外界的裝置(例如電動輪椅、復健機器人)。其中一種腦機介面則利用運動想像(Motor Imagery, MI)所設計，藉著想像產生μ節律(8~12Hz)與β 波(13~30Hz)的變化作為腦波特徵，再藉由機器學習訓練一套系統進行辨識，所得結果能使行動不便的人能直接藉由腦波就可以操作外部裝置。
非同步(Asynchronous)腦機介面(Brain Computer Interface, BCI)系統或稱Self-pace BCI系統是一種能應用於現實生活的BCI系統，因此其效能必須具備一定的穩定性。在先前的研究中已提出許多方法來改善非同步BCI效能，而本文亦專注於非同步BCI效能之提升，透過不同的特徵抽取方法以及分類器進行兩類別的運動想像分類，期望能使得非同步BCI系統兼具低偽正率(False Positive Rate, FPR)與高偵測率(True Positive Rate, TPR)之特性，但兩種效能呈現反比趨勢，因此在本文中將題出偽正率控制(False Positive Rate Controlling , FPR Controlling)之比較，以對非同步BCI介面之偵測率與偽正率的平衡進行控制。
特徵抽取方法包含頻帶功率(Band Power, BP)、共同空間模式(Common Spatial Pattern, CSP)與核主成份分析(Kernel Principal Component, KPCA)，而分類器則包含支持向量機器(Support Vector Machine, SVM)、非平衡式支持向量機器(Imbalanced Support Vector Machine, ISVM)、線性鑑別分析(Linear Discriminant Analysis, LDA) 與馬氏鑑別分析(Mahalanobis Discriminant Analysis, MDA)。
本實驗透過兩種資料庫進行兩類別與四類別運動想像分析，實驗結果顯示以CSP特徵抽取方法配合ISVM分類器方法能達幾乎排除所有非目標腦波，在偽正率為0%時偵測率為29.5%且正確率為71.71%。

關鍵字

運動想像 ； 機器學習 ； 目標腦波 ； 分類器 ； 特徵抽取 ； 偵測率 ； 偽正率

並列摘要

Abstract
BCIs (Brain-Computer Interface) have received much more attentions in recent studies and developments, they establish a direct communication between patients suffering for severe motor-disabilities (such as stroke and ALS) and the outside world, or provide control over external devices (such as electric wheelchair and rehabilitation robot) via EEG. One of these BCIs is designed based on Motor Imagery (MI), using the fluctuations of μ-rythem (8~12 Hz) and β-band occured when imagining as features, undergo a machine-learning system for classification, and the output result enables motor-disabled patients to have control over external devices via EEG.
Asynchronous BCI system, also called Self-pace BCI system, is a BCI system that can be applied in real life, therefore its performance requires a certain degree of stablity. In earlier researches many approaches have been suggested to improve the performance of asynchronous BCI system, this thesis also dedicates in improving the performance of asychronous BCI system, seeking through different feature extraction methodologies and classifiers to achieve an asynchronous BCI system with both the characteristics of low False Positive Rate (FPR) and high True Positive Rate (TPR) for a two-class Motor Imagery classification, since these two characteristics show a inverse relation, a comparision of False Positive Rate Controlling (FPR Controlling) was proposed in this thesis gain a balance between the TPR and FPR of the asychronous BCI system.
Feature extration methodology includes Band Power (BP), Common Spatial Pattern (CSP) and Kernel Principal Compnent (KPCA), while classifier includes Support Vector Machine (SVM), Imbalanced Support Vector Machine (ISVM), Linear Discriminant Analysis (LDA) and Mahalanobis Discriminant Analysis (MDA).
In this experiment a two-class classification and a four-class classification was done with two databases, experimental results show that CSP feature extraction method combined with ISVM classifier can nearly reject non-target EEG, obtaining a TPR of 29.5% while FPR is 0% and total correct rate of 71.71%.

並列關鍵字

Motor Imagery ； Target EEG ； Machine Learning ； Classifier ； Feature Extraction ； FPR ； TPR

參考文獻

[1] C. Andrew and G. Pfurtscheller, “On the existence of different alpha band rhythms in the hand area of man,” Neuroscience letters, vol. 222, no. 2, pp. 103-106, Jan 1997.

[2] G. Edlinger, P. Wach, and G. Pfurtscheller. “On the realization of an analytic high-resolution EEG," IEEE Transactions on Biomedical Engineering, vol. 45, no. 6, pp. 736-745, June 1998.

[3] J. Duchen and T. Y. Hohrel, “A EMG generation,” IEEE Transactions on Biomedical Engineering, vol. 47, no. 2, pp. 192-201, Feb 2000.

[4] J. R. Wolpaw, N. Birdaumer and W. J. Heetderks, “Brain-computer interface technology: Review of the first international meetingm,” IEEE transactions on rehabilitation engineering, vol. 8, no. 2, pp. 164-173, June 2000.

[5] M. Murugappan, et al. “Lifting scheme for human emotion recognition using EEG,” in Proceedings - International Symposium on Information Technology, ITSim, 2008.