Seizure Forecasting from Subcutaneous EEG Using Long Short Term Memory Neural Networks: Algorithm Development and Optimization

Seizure forecasting is a great research interest due to its potential in helping patients manage activities or facilitate targeted therapies, specifically with the emergence of new subcutaneous continuous EEG recording systems that have shown promise to be helpful. In work presented here, we used one subject diagnosed with refractory epilepsy with 230 days of monitoring to evaluate seven architectures to design a seizure prediction algorithm using a deep learning RNN classifier. The preliminary results suggest that it is possible to forecast seizures using two-channel chronic subcutaneous EEG recordings. With an average AUC of 0.74947 for architectures found to have better than chance performance. Future work will focus on extending results to additional patients, investigating cross-subject performance, and the importance of the different inputs to the architectures.