Abstracts

Rationale: Worldwide, there are approximately 65 million people with epilepsy, more than Parkinson’s disease, Alzheimer’s disease and Multiple Sclerosis combined. The morbidity and mortality of epilepsy is secondary only to its unpredictability. The seizure prediction problem has, until recently, evaded success from computational approaches utilizing electroencephalograph (EEG) data. The difficulty of the problem arises from the lack of a general and specifiable distinction between pre-ictal and inter-ictal periods of the EEG signal. We aim to develop a prediction model that will reliably identify the pre-ictal state using extra-cranial EEG approximately 11 minutes before an on-coming seizure.Methods: Our approach casts the epilepsy prediction problem as one of anomaly or outlier detection by building a model for baseline EEG activity. This results in a method that capitalizes on the vast amount of EEG data with no seizure activity. To build our baseline model, we construct graphs to model the connectedness of electrodes and use graph mining techniques to extract features during ictal events. Then we identify the subset of our feature set which correlate to baseline EEG activity, using the well-known method of Quadratic Programming Feature Selection. This new set of features is used to construct an auto-regressive model, which can predict how baseline EEG activity would look given the current values of the EEG signal. The patient’s EEG signal is then fed sequentially to our model, which predicts the patient’s baseline activity. By calculating our prediction error and detecting outlying sequences in our prediction error, we can predict ictal events. We utilize two methods for detecting outlying sequences; Chebyshev’s inequality and an information-theoretic approach. We focus on long term recordings with multiple seizures and train our model in an on-line fashion, adapting it for a given individual’s seizure pattern using the patient’s new EEG data. By updating our model constantly, we hope to capture the changes in baseline EEG activity the patient undergoes.Results: Waveforms generated from our new method show a promising distinction between normal EEG activity and anomalies. The figure attached shows the deviation between the actual EEG signal of the patient and our prediction for baseline EEG activity. An abrupt change in prediction error can easily be detected (blue vertical lines) several minutes before the onset of the seizure (red vertical line).Conclusions: Anomaly detection methods for seizure prediction were able to reliably identify the pre-ictal state in this set of patients. We hope that as we augment the model with data from more patients and with more seizures per patient we can begin to understand and improve upon its sensitivity and specificity.