Prediction of epileptic seizures
Epilepsy is a severe disease characterized by temporary changes in the bioelectrical functioning of the brain. These changes cause abnormal neuronal synchronization and seizures which affect awareness, movement or sensation. An early in time prediction would considerably increase the quality of life of those patients who cannot be successfully treated by common therapeutic strategies.
The long-term goal of the project is to develop algorithms that are able to predict epileptic seizures with high sensitivity and specificity. These prediction algorithms could be utilized in a "brain defibrillator", in analogy to cardiac defibrillators. A prediction of seizures at an early stage could trigger an intervention to suppress the upcoming seizure by for instance electrical stimulation. Alternatively, a seizure warning device could be invented that enables behavioral adjustments.
To this aim, electroencephalogram (EEG) data recorded from invasive and scalp electrodes are analyzed in an interdisciplinary project between
- Epilepsy Center, University Hospital Freiburg
- Bernstein Center for Computational Neuroscience (BCCN), Freiburg
- Freiburg Center for Data Analysis and Modeling (FDM).
In the Epilepsy Center epilepsy patients undergo a presurgical monitoring with invasive and scalp electrodes. The EEG data are examined visually by epileptologists in order to localize the epileptic focus for surgery. The contiguous records over several days contain about 60 channels and are sampled with at least 256 Hz.
Furthermore, a high quality data pool comprising EEG data is publicly available. For 21 patients, about five seizures including preictal data and at least 24 hours of seizure-free data are available, yielding a few tens Gigabytes of EEG data.
Register here to access the database.
Assessment of seizure predictability
Seizure Prediction Characteristic
Basic requirements for genuine seizure prediction are the existence of distinct periods in the EEG. Between the prediction and the earliest occurrence of the seizure, there has to be a time period, called the intervention time (IT), necessary for an intervention. Moreover, a time interval is necessary during which the predicted seizure has to occur, the seizure occurrence period (SOP), restricting the period in which intervention is necessary. (s. figure)
Temporal properties of a prediction achieved by these two time intervals have to be related to the number of false alarms, because too many false predictions would lead to impairment due to possible side-effects of interventions or loss of patients' acceptance of seizure warning devices.
Motivated by these requirements for a successful therapeutic intervention the seizure prediction characteristic (SPC) was introduced as a methodology to assess individual seizure prediction performance. It yields a functional relationship depending on these prediction parameters.
The seizure prediction characteristic enables not only an efficient way to obtain information about the sensitivity of seizure prediction methods but also a comparison between different prediction methods.
Besides calculating sensitivity and specificity as well as the temporal aspects of a prediction, an assessment of seizure prediction methods has to include a statistical validation. A predictor should at least perform better than random. We have introduced an approach based on comparing seizure prediction characteristic values with critical sensitivity values obtained for an analytic significance level. The significance level is calculated on basis of a random prediction following a Poisson process.
Sensitivity and synchronisation indices R und Smin for three patients as a function of the IT for SOP=10 min and FPRmax=0.15 FP/h. The maximum, statistically significant sensitivity values for the three different classes of combinations between the focal contacts (foc/foc), between the focal and extra-focal contacts (foc/ext) and between the extra-focal contacts (ext/ext) are represented.