Predicting Seizures in Intracranial EEG Recordings

If seizure forecasting systems could reliably identify periods of increased probability of seizure occurrence, patients who suffer from epilepsy would be able to avoid dangerous activities and lead more normal lives. The goal of this project is to differentiate between the preictal and interictal states by analyzing intracranial EEG recordings. Data for each hour are organized into six ten-minute time sequences. Logistic Regression and support vector machines (SVM) are applied to each time sequence to calculate the average test false negative rate. The idea of combining data from various time sequences is also experimented with to examine if false negative rates can be reduced. The results show that SVM provides better prediction results for patients, and that combining training examples from different time sequences does not help improve prediction results. Since the pathogenesis of epilepsy may vary across different species, applying the same training models to both dogs and patients may be problematic. Keywords—seizure, classification, logistic regression, support vector machines (SVM)