Forecasting epileptic seizures from electroencephalograms using deep neural networks

The occurence of eplileptic seizures is a problem that makes everyday life very difficult for people who suffer from this disorder. This is particulary true for patients which disorder can not be cured with therapies or drugs. In this work I studied and developed different kind of recurrent neural networks' models (RNNs) that, through analysis of electroencephalography (EEG) signals, aimed to correctly predict the onset of a seizure by detecting the preictal brain state and differentiating it from the interictal state within minutes before the actual epileptic attacks. The two proposed models use different approaches to extract relevant temporal features from the raw EEG signals that are then used for prediction purposes. Many different kind of experiments and evaluation methods were used to asses the capacity of the model to correctly predict the seizures' onsets leading to promising results that can be further enhanceed with more sophisticated models that could extract more relevant temporal features.