Semisupervised Seizure Prediction in Scalp EEG Using Consistency Regularization

Figure 3

The network architecture used in this study. It includes three convolution blocks, and in each block, a batch normalization layer, a convolution layer with ReLU activation, and a max-pooling layer are built in turn. The first block uses 3D convolution, while the next two adopt 2D convolution. The features of these convolution blocks are flattened and explored by two fully connected layers to generate the final prediction. Both of them have a dropout rate of 0.5.