A Neural Network Study of Predicting Seizures in Epilepsy

Background Epilepsy is a group of chronic neurological disorders characterized by recurrent seizures [1]. A seizure is a transient abnormal condition caused by a sudden hyper-synchronous neuronal activity in the brain [1] . One of the most problematic aspects of epilepsy is the apparent unpredictability of the seizures, although patients spend just a marginal part of their lives having seizures [2]. A methodology for seizure prediction from raw EEG signals is essential for the development of a new generation of warning/medication devices for epilepsy. The potential impact on patients’ life quality of the development of such a device would be enormous, especially for those patients affected by currently medical intractable epilepsy. Methods The seizure prediction problem was formulated as a classification task. Relying on the hypothesis that epileptic activity is preceded by specific brain dynamics, seizures can be anticipated recognizing these dynamics by means of a classifier able to distinguish them from the normal brain activity. The proposed seizure prediction procedure consists of four sub-systems: The Raw Data Elaboration Manager is responsible of the raw EEG windowing and features extraction process, the Classifier takes in input the extracted features and decides if the current window must be considered as preictal or not, the Queue stores the most recent classification results, the Queue Manager checks the queue status at regular intervals and raises alarms when a security threshold is exceeded. Results The proposed methodology allowed to forecast 85 over a total number of 87 seizures from 21 patients affected by medically intractable epilepsy with anticipation of several minutes. Conclusions The seizure prediction methodology proposed in this work achieved very good results on the FSEEG Database, also with an average anticipation time suitable for clinical intervention. In order to seriously consider the development of an actual medication device, these findings must be confirmed on an extended dataset. However the results achieved in this work are encouraging. Regardless the performance of the described system, we believe that the real value of this work is the developing of a general seizure prediction methodology. Given its flexible structure, the proposed system can be easily used in future studies in order to test and verify new hypothesis.