Seizure prediction: Quantifying the performance of measures in distinguishing pre-ictal from inter-ictal states

RATIONALE: In the rapidly developing field of seizure prediction more and more interest is directed towards the question of how to quantify the performance of measures applied to the EEG in seperating pre-ictal from inter-ictal states. In this study we compare two different concepts to address this point. Both evaluations are based on the extraction of characteristic features (e.g. positive and negative deviations from a given reference level) derived from time profiles of bivariate measures. While the first approach is aiming at a statistical seperation of the pre-ictal from the inter-ictal states, the second one is an algorithmic approach defining alarms and evaluating their distribution relative to the times of seizure onset in terms of sensitivity and specificity. For the latter approach a new way of weighting sensitivity and specificity to get one overall measure of performance is introduced. METHODS: We analyzed continuous intracranial multichannel EEG recorded from patients suffering from unilateral mesial temporal lobe epilepsy (MTLE). In the first step a number of bivariate measures (e.g. cross correlation) were calculated applying a moving window technique. Secondly, from the resulting time profiles we extracted and parametrized characteristic features (e.g. positive and negative deviations from a given reference level). Using on the one hand a statistical and on the other hand an algorithmic approach the performances of the different measures were evaluated automatically. RESULTS: Within the statistical as well as within the algorithmic approach the different bivariate measures yielded different degrees of performance in distinguishing pre-ictal from inter-ictal states. For the latter approach different ways of weighting sensitivity and specificity to get one overall measure of performance were compared. As a solution to the problem of how to define sensitivity and specificity for continuous long-time recordings we propose the use of the prediction horizon as a common time unit to get a proper normalized measure of performance (similar to the discrete case of diagnostic tests where the natural unit is the single patient). CONCLUSIONS: While the statistical approach is free of parameters and therefore acts as an unbiased criterion for the distinguishability of the two different states, the algorithmic approach offers the possibility to adjust certain parameters. However, as with the parameters of the single measures much care has to be taken to avoid in-sample optimization. Also, for this approach a proper weighting of sensitivity and specificity seems to be of high importance for an unbiased judgement of the performance of any measure.