Figure 3.

Architecture of Bayesian convolutional neural network (BCNN) with embedded Bayesian modulator. n is the number of EEG channels, which is 19 for all 30 patients in this study. Unlike a conventional convolutional neural network where each weight is a single value, each weight of a BCNN is a distribution. In our work, we model each weight as a normal distribution with mean value (μ) and standard deviation (σ) are trainable parameters. The BCNN network has three convolutional layers, each is followed by a max-pooling layer (not shown). Extracted features by the convolutional layers are fed to two fully-connected layers. The Bayesian modulator incorporates other relevant data for seizure forecasting into the last fully-connected layer of the network by using Equation (8).