基于CNN算法的稳态体感诱发电位的特征识别

脑-机接口研究可为瘫痪病人的康复带来一种新的治疗方法。已有研究表明对手指或者正中神经施加一定频率的体感刺激,会引发相同频率且具有空间特异性的稳态体感诱发电位。为优化基于稳态体感诱发电位的脑-机接口的性能,通过快速傅里叶变换寻找12个健康被试的个人左手特定共振频率,采用事件相关谱扰动进行时频分析,检测其稳态体感诱发电位信号。基于共振频率对实验诱发的脑电信号进行1 Hz带通滤波,获得特定频带的数据,采用卷积神经网络(CNN)学习算法对其进行分类,并与采用共空间模式和支持向量机的特征提取及特征分类的方法(CSP+SVM)进行比较。所有被试的结果显示:基于共振频率滤波方法,采用CNN学习算法获得的离线分类准确率均高于85%,并且CNN学习算法的分类准确率显著性优于CSP+SVM的分类准确率(91.8%±5.9% vs 77.4%±8.5%,P<0.05)。因此,在基于稳态体感诱发电位的脑机接口的特征识别中,CNN学习算法相比传统使用的机器学习分类算法(如共空间模式+支持向量机)能够显著提升分类准确率,提高脑机接口的整体性能。

Feature Recognition of Steady Somatosensory Evoked Potential Based on Convolutional Neural Network

Brain-computer interface (BCI) system provides a new way of rehabilitation for patients with aphasia. Previous studies have shown that applying a certain frequency of somatosensory stimulation to the finger or median nerve triggers a spatially-specific steady-state somatosensory evoked potential (SSSEP) with the same frequency. In order to enhance the overall performance of a steady-state somatosensory evoked potentials based BCI system, this work aimed at finding the person-specific resonance-like frequencies of left-hand of 12 healthy subjects by using fast fourier transform (FFT), and the significant time and frequency range for signal feature were selected by event related spectral perturbation (ERSP) for detecting steady-state somatosensory evoked potential signals. The experimentally induced EEG signals were filtered by 1 Hz band-pass based on person-specific resonance-like frequencies to obtain the data of the specific frequency band, then the convolutional neural network (CNN) learning algorithm is used to classify the data. The methods of feature extraction and classification using common spatial pattern (CSP) and support vector machine (SVM) was ompared. The results of all the subjects showed that the accuracy of offline classification obtained by CNN learning algorithm was higher than 85% based on the person-specific resonance-like frequencies filtering method, and the classification accuracy of CNN learning algorithm was higher than using of CSP with SVM (91.8%±5.9% vs 77.4%±8.5%,P<0.05). Therefore, compared with the traditional machine learning classification algorithm (such as common spatial pattern with support vector machine), the CNN learning algorithm can significantly improveed the classification accuracy for feature recognition of brain-computer interface based on steady-state somatosensory evoked potential, and improved the overall performance of the brain-computer interface