基于长短期记忆网络的脑电相位预测方法研究  

EEG phase prediction method based on long short-term memory network

作　　者：庞紫胭 赵鑫雨 买文姝 赵悦茁 刘志朋[1,2] 殷涛 靳静娜[1,2] PANG Zi-yan;ZHAO Xin-yu;MAI Wen-shu;ZHAO Yue-zhuo;LIU Zhi-peng;YIN Tao;JIN Jing-na(Institute of Biomedical Engineering,Chinese Academy of Medical Sciences&Peking Union Medical College,Tianjin 300192,China;Tianjin Key Laboratory of Neuromodulation and Neurorepair,Tianjin 300192,China;Center for Neuroscience,Chinese Academy of Medical Sciences,Beijing 100730,China)

机构地区：[1]中国医学科学院北京协和医学院生物医学工程研究所,天津300192 [2]天津市神经调控与修复重点实验室,天津300192 [3]中国医学科学院神经科学中心,北京100730

出　　处：《医疗卫生装备》2025年第3期1-8,共8页Chinese Medical Equipment Journal

基　　金：国家重点研发计划项目(2022YFC2402202,2023YFC2412503)。

摘　　要：目的:为了提高经颅磁刺激(transcranial magnetic stimulation,TMS)中脑电相位同步预测的准确性和鲁棒性,提出一种基于长短期记忆网络(long short-term memory network,LSTM)的脑电相位预测方法。方法:首先,构建由输入层、LSTM层、ReLU激活层、全连接层和回归层组成的LSTM,通过输入门、遗忘门和输出门的协同作用捕获脑电信号特征。其次,对30名健康受试者的睁眼静息态脑电数据使用LSTM训练得到预测模型,用于脑电信号和脑电相位预测。最后,对比LSTM方法与传统自回归(autoregressive,AR)方法在总体和个体水平上的相位预测误差以及2种方法对波峰、波谷的预测性能,并采用线性回归模型探究LSTM方法下脑电瞬时幅值、信噪比与相位预测误差的关系。结果:LSTM方法的总相位预测误差为0.04°±5.69°,小于传统AR方法(总相位预测误差为-3.36°±51.13°),且对于每位受试者,LSTM方法的相位预测准确性均优于传统AR方法,差异有统计学意义(P<0.001)。LSTM方法可准确预测约89%的波峰(谷),而传统AR方法仅可准确预测约10%的波峰(谷)。不同于传统AR方法,在LSTM方法下,脑电瞬时幅值和信噪比与相位预测误差均无线性关系(P分别为0.58、0.18)。结论:基于LSTM的脑电相位预测方法具有高准确性和鲁棒性,可为脑电相位同步TMS方式提供一种有效的相位预测方法。Objective To propose a brain electrical phase prediction method based on long short-term memory network(LSTM)to improve the accuracy and robustness of phase synchronization prediction in transcranial magnetic stimulation(TMS).Methods First,an LSTM consisting of an input layer,an LSTM layer,an ReLU activation layer,a fully connected layer and a regression layer was constructed to capture the EEG signal features through the synergistic action of input gates,forgetting gates and output gates.Second,eye-open resting-state EEG data from 30 healthy subjects were trained using the LSTM to obtain a predictive model for EEG signal and EEG phase prediction.Finally,the LSTM method and the traditional autoregressive(AR)method were compared in terms of the phase prediction errors at the overall and individual levels and the prediction performance for peaks and troughs.A regression model was used to explore the relationships between instantaneous EEG amplitude,signal-to-noise ratio and phase prediction error with the LSTM method.Results The LSTM method achieved a total phase prediction error of 0.04°±5.69°,which was lower than that of the traditional AR method(-3.36°±51.13°).For each subject,the LSTM method demonstrated superior phase prediction accuracy compared to the traditional AR method(P<0.001).The accuracy for predicting peaks(troughs)by the LSTM method(about 89%)was higher than that by the traditional AR method(about 10%).Unlike the traditional AR method,the LSTM method didnot result in linear relationships between instantaneous EEG amplitude,signal-to-noise ratio and phase prediction error,with P values being 0.58 and 0.18,respectively.Conclusion The LSTM-based brain electrical phase prediction method shows high accuracy and robustness when used for EEG phase-synchronized TMS.

关 键 词：经颅磁刺激 脑电 脑电相位 长短期记忆网络 自回归 脑电信号 

分 类 号：R318[医药卫生—生物医学工程] TP183[医药卫生—基础医学]