12-24 – DeepEEG: A Deep Representation Learning based Clinical Decision Support System to Predict Anti-Epileptic Drug Response for Paediatric Epilepsy

Background: Approximately 112,000 children in the UK have epilepsy, which disrupt their lives and can be harmful for developing brain. Although we have a number of (Anti-Epileptic Drug) AEDs to stop seizures, these AEDs fail to work in about a third of children. The response of an individual child to a specific AED is unpredictable and is currently a trial-and-error process. Unnecessary trials of AEDs could be avoided in refractory (drug-resistant) children, and delays could be reduced before other therapies or surgeries are explored.
Aim: Currently, there is no test available to identify which children will respond to AED in advance. Use of EEG in diagnosis and classification of epilepsy is well established, however, its value as an AED response predictor has not been studied. We will develop a test that will help clinicians to recognize which children might not respond to AEDs and allow them to consider alternative therapies for these children.
Proposed Methods: This study focuses on idiopathic generalized epilepsy (IGE), which comprises a group of clinical syndromes (childhood absence, juvenile myoclonic epilepsy, and juvenile absence epilepsy) and accounts for 25-35% of all epilepsies. One of the defining features of IGE is a generalized spike wave seen on EEG which is evidence of pathological hypersynchrony.
This research will start with feature selection of these spikes; it is an important and challenging concept in DRL as well as in EEG analysis. It should be noted that when we extracted features from raw EEG channels, the number of features is larger than the original number of channels. We will implement Graph Convolutional Networks (GCN) on multichannel EEGs. GCN follows neuron connectivity approach of the brain to encode epilepsy through spatial proximity. We will the spectral features using GCN for extraction of 10 second windows from multichannel EEGs.