Zhang Rui, Wang Yinwang, Li Xianpeng, Liu Bo, Zhang Lipeng, Chen Mingming, Hu Yuxia
Annu Int Conf IEEE Eng Med Biol Soc. 2019 Jul;2019:3087-3090. doi: 10.1109/EMBC.2019.8857923.
BCI illiterate subject is defined as the subject who cannot achieve accuracy higher than 70%. BCI illiterate subject cannot produce stronger contralateral ERD/ERS activity, thus most of the frequency band-based algorithms cannot obtain higher accuracy. Deep learning with convolutional neural networks (CNN) has revolutionized in many recent studies to learn features and classify different types of data through end-to-end learning. We designed a CNN to extract motor imagery EEG features and then do classification for BCI illiterate subjects in this work. Results showed that the average classification accuracy increased by 18.4% compared with the CSP+LDA algorithm, and the accuracies obtained by CNN exceed 70% for 9 of 11 subjects particularly. CNN requires only a little prior knowledge, thus the features it extracted are not limited in frequency band, but because the poor interpretability of deep learning, we do not know which kind of feature CNN extracted until now. Our future study will focus on visualizing the extracted features to support our conclusions.
脑机接口文盲受试者被定义为准确率无法高于70%的受试者。脑机接口文盲受试者无法产生更强的对侧事件相关去同步化/事件相关同步化活动，因此大多数基于频带的算法无法获得更高的准确率。在最近的许多研究中，基于卷积神经网络（CNN）的深度学习通过端到端学习来学习特征并对不同类型的数据进行分类，带来了变革。在这项工作中，我们设计了一个CNN来提取运动想象脑电特征，然后对脑机接口文盲受试者进行分类。结果表明，与共空间模式+线性判别分析（CSP+LDA）算法相比，平均分类准确率提高了18.4%，特别是11名受试者中有9名通过CNN获得的准确率超过了70%。CNN只需要很少的先验知识，因此它提取的特征不受频带限制，但由于深度学习的可解释性较差，到目前为止我们还不知道CNN提取了哪种特征。我们未来的研究将集中在对提取的特征进行可视化，以支持我们的结论。
