Comparison of three nonlinear seizure prediction methods by means of the seizure prediction characteristic

Epilepsy is characterized by the spontaneous and unforeseeable occurrence of seizures, during which the perception or behavior of patients is disturbed. The predictability of these seizures would render novel therapeutic approaches possible. Several prediction methods have claimed to be able to predict seizures based on EEG recordings minutes in advance. However, the term seizure prediction is not unequivocally defined, different criteria to assess prediction methods exist, and only little attention has been paid to issues of sensitivity and false prediction rate. We introduce an assessment criterion called the seizure prediction characteristic that incorporates the assessment of sensitivity and false prediction rate. Within this framework, three nonlinear seizure prediction methods were evaluated on a large EEG data pool of 21 patients. Altogether, 582 h intracranial EEG data and 88 seizures were examined. With a rate of 1–3.6 false predictions per day, the “dynamical similarity index” achieves a sensitivity between 21 and 42%, which was the best result of the three methods. Sensitivity was between 18 and 31% for the extended, prospective version of the “accumulated energy” and between 13 and 30% for the “effective correlation dimension”. These results still are not sufficient for clinical applications.