基于迁移学习的癫痫脑电时空预测模型研究OA

To address the deficiencies in cross-subject generalizability and robustness of existing epileptic electroencephalogram(EEG)prediction models,we proposed a transfer learning-based epileptic prediction model with multi-scale spatio-temporal features(TLEP-MST)by integrating signal analysis and deep learning technology.Firstly,the raw data was analyzed through independent com-ponent analysis(ICA)to remove artifacts,and the temporal feature extraction module and wavelet convolutional layer were used to ex-tract the time-frequency information in the EEG signals.Then,the adaptive attention mechanism was applied to perform multi-channel weight assignment of the time-frequency information,and obtain spatial features of the EEG signals.Finally,transfer learning was in-corporated to reduce data distribution discrepancies between source and target domains,enhancing the model generalization perform-ance.The model was experimented on the public dataset CHB-MIT.In the cross-validation,the accuracy rate,specificity and false positive rate of the model was 91.88%,96.49%and 0.0369/h,respectively.In patient-specific experiments,the specificity improved from 67.04%to 85.06%,and the false positive rate decreased from 0.4194/h to 0.3485/h after introducing transfer learning.This re-search is of great value in predicting the EEG of epilepsy across subjects.