Attention Model based Deep Learning Architecture for EEG Seizure Detection

Abstract

Because voice data is similar to EEG data, they are both temporal and concurrent. In order to explore the general appropriateness of speech recognition model in EEG database, this paper introduces the attention model which has been very good in speech field in recent years, and carries out experiments on TESC data set. The attention model in this paper is based on the traditional HMM model, and Z module is added to introduce the attention mechanism. At the input end of the whole framework, PCANet model is introduced to solve the problem of too high dimension of TESC raw data. The convolution size used in PCANet is the most important group of super parameters. It is optimized by convex optimization alone. Many groups of training have been carried out on the horizontal and vertical dimensions, and the maximum local accuracy has been obtained. In order to improve the recall rate, the penalty of incorrect data classification has been greatly increased. In order to further improve the accuracy, the depth of PCANet model is increased. The input and output dimensions of PCANet model are adjusted to the same size and cascaded in multiple layers. The final accuracy rate was raised to 60%.