Analyzing EGG dataset using graph convolutional neural network and graph neural ODE. Achieved ~5% improvement on brain state identification for epilepsy patients than the previous work. The work is the foundation for implementing the brain graph learning and brain state identification system on chip.
Brain graphs are the representation of brain state in terms of the connectedness between each node (i.e. EEG detector node). Such representation is useful for application such as brain state prediction and identification. Below is an example of an epilepsy patient's brain graph, in chronological order as the patient's brain state changes.
These brain graphs are generated by graph convolutional neural network (GCN), a type of neural network that deal with network topology and connection. In this work, I also tried graph convolutional neural ODE, the neural ODE version of the network.
The accuracy plot is shown below.
It is interesting to see how each brain graph progress, which is different for each neural network. Below shows the brain graph progression for patient 958.
From the hidden state progression, we can see the drastic difference between how GCN and Graph Convolutional Neural ODE process brain state.