Validating the performance of epileptic seizure prediction algorithms using simulated annealing.

There is a growing number of publications claiming that epileptic seizures can be predicted by the analysis of the electroencephalogram (EEG) using different kinds of characterizing measures (e.g. univariate or bivariate, linear or non-linear). Many of these studies suffer from a severe shortcoming, namely the lack of statistical validation. Only rarely results are passed to a statistical test and verified against some null hypothesis in order to quantify their significance. Here we propose a new method to statistically validate the performance of any given algorithm designed to predict epileptic seizures. From time profiles rendered by the application of a measure to the EEG with a moving window technique we first generate a certain number of surrogates. Using the method of simulated annealing we perform a constrained randomization preserving the amplitude distribution as well as essential parts of the autocorrelation function. Then the algorithm, the validity of which is to be tested, is applied to the original time profile as well as to the surrogate time profiles. Depending on the number of surrogates generated, a desired level of significance of the algorithm's performance for the original time profiles can be obtained. We demonstrate our method using a synchronization analysis applied to quasi-continuous EEG recordings of a patient with ten epileptic seizures in five days. Results are compared to those obtained with the recently proposed method of seizure time surrogates.