Poster paper: Predicting seizures from electroencephalography recordings: A knowledge transfer strategy

Epilepsy, a brain disorder afflicts nearly 1% of the world's population, is characterized by the occurrence of spontaneous seizures. For most epilepsy patients, the drugs are either not effective or produce severe side-effects. Seizure forecasting systems have the potential to help patients with epilepsy lead more normal lives. Recently multi-center clinical studies showed evidence of premonitory symptoms in 6.2% of 500 patients with epilepsy, and some interviews of epilepsy patients also found that a certain amount of patients felt "auras". All these are promising signs suggesting that seizure might be predictable. In this paper, we will study the application of deep learning techniques for seizure prediction with EEG signals. Deep learning methods have been shown to be very effective on exploring the latent structures from continuous signals and they have achieved state-of-the-art performance on speech analysis. One potential requirement for deep learning algorithms to work is a huge training set, which could be difficult for a specific medical problem. Therefore we specifically investigated a transfer learning strategy: we performed the major seizure prediction task on the data from American Epilepsy Society Seizure Prediction Challenge1, and we adopted another 6 publicly available EEG datasets2, which are not directly related to seizure prediction, as auxiliary information to pre-train the deep neural network for getting a good initial point. Our results show that with those auxiliary information, the prediction performance can be boosted. This observation is validated with different predictive models, which opens another gate for effective integration and utilization of medical data resources.