基于双向长短时记忆神经网络的步态时空参数脑肌电解码方法

EEG and sEMG Decoding of Gait Spatiotemporal Parameters Based on Bidirectional Long Short-Term Memory Neural Network

针对脑电(EEG)信号对连续步态轨迹解码结果与实际轨迹相关性低的问题,提出一种基于双向长短时记忆(BiLSTM)神经网络的步态参数解码方法。首先,构建基于双向长短时记忆神经网络的步态时空参数解码模型,根据脑肌电信号特性设计解码模型的超参数;其次,同步采集脑电、下肢运动相关肌肉的表面肌电信号(sEMG)和下肢关节运动信号,并对脑电和表面肌电信号的步态相关特征进行分析;然后,以多通道脑电和下肢运动相关表面肌电信号作为解码模型的输入,自动提取脑肌电融合信号中步态相关特征并构建膝踝关节运动轨迹与特征之间的非线性回归模型;最后,以多通道脑电作为解码模型的输入,构建步态相关脑电信号和表面肌电信号之间的非线性回归模型。实验结果表明:所提方法与传统支持向量机方法相比,对踝关节解码轨迹与实测轨迹形状相似性Pearson相关系数提高了0.12;与单独采用脑电、表面肌电信号和脑肌电信号平均绝对值特征融合信号进行解码方法相比,对踝关节解码轨迹与实测轨迹形状相似性Pearson相关系数分别提高了0.81、0.19和0.63。该方法可实现从脑电信号中对部分表面肌电信号波形的解码,解码波形和实测波形的平均Pearson相关系数值接近0.5,证明从脑电信号中可解码出肌肉通道的表面肌电信号波形,为下肢外骨骼主动连续控制的应用提供了新思路。

To solve the problem of low correlation between continuous gait trajectory decoding results and actual trajectory by electroencephalography(EEG)signals,a gait parameter decoding method based on bidirectional long short-term memory(BiLSTM)neural network is proposed.Firstly,a gait spatiotemporal parameter decoding model based on this neural network is constructed,and the hyperparameters of the decoding model are designed according to the characteristics of EEG and surface electromyography(sEMG).Secondly,EEG,lower limb movement-related sEMG and lower limb joint movement signals are collected synchronously,and gait features of EEG and sEMG signals are analyzed.Thirdly,multi-channel EEG and lower limb movement-related sEMG signals are used as input of the decoding model,and gait related features are extracted automatically from EEG and sEMG fusion signals,and the nonlinear regression model between ankle joint motion and gait related features is constructed.Finally,a nonlinear regression model between gait related EEG signals and sEMG signals is constructed with multi-channel EEG as the input of the decoding model.The results show that compared with traditional support vector machines,the Pearson correlation coefficient of shape similarity between decoded trajectory and measured trajectory is improved by 0.12.Compared with the decoding methods using EEG,sEMG and fusion average absolute value of EEG-sEMG,the proposed method improves the Pearson correlation coefficient of shape similarity between decoded trajectory and measured trajectory by 0.81,0.19 and 0.63,respectively.Our decoding method can realize decoding of part of sEMG waveform,the average Pearson correlation coefficient of decoded waveform and measured waveform is close to 0.5.It shows that the sEMG signals can be decoded from EEG signal which provides a new idea for the application of active continuous control of lower extremity exoskeleton.