EEG-Based Classification and Advanced Warning of Epileptic Seizures

Epilepsy is the second most common neurological disease after stroke. Epileptics may suffer hundreds of seizures per day, yet one is enough to put a person in constant fear of the next. The sudden and unexpected onset of seizures has debilitating and sometimes fatal consequences. The development of a real-time seizure prediction and alerting device would greatly improve epileptics’ quality of life. Major challenges for such a device include determining predictive features and discovering the maximum prediction window. Using the novel approach of random forest classification on EEG data, this research investigates the predictive features among the common EEG frequency bands for one patient with partial complex and partial with secondarily generalized seizures. The impact on classifier performance of labeling the transitional brain states is also investigated, using a time-series accuracy graph. Predictive features are found as far as 40 minutes in advance of two seizures, specifically in the beta frequencies of one brain node. The random forest classifier does not perform well, but shows promise for improved performance with minor adjustments in training. The time-series accuracy graphs prove a useful tool for visualization and insight into classifier performance that is lacking in other evaluation methods