Abstracts

Automated seizure detection can substantially improve diagnostics and care of patients with intractable forms of epilepsy. One of our main research topics is the development of automated algorithms based on 3D-accelerometry, to detect motor seizures. Validation of such algorithms requires comparison to a clinical standard. According to current practice, the best available standard is visual interpretation of video registrations by experts. These interpretations are qualitative and depend on the accuracy of the experts during scoring. To get the best possible reference, we investigate the validity of using these interpretations as a standard for evaluating performances of automated detection algorithms., Video is used from nocturnal EEG\video registrations in 7 mentally retarded patients who suffer from severe epilepsy and have minimal 20 seizures per month (2 male, 5 female, mean age 30.3 years [plusmn] 14.2 years). For each patient two data episodes are chosen: one episode containing at least one seizure (according to experts who inspected both video and EEG registrations), and one randomly chosen episode containing a non-seizure movement. Depending on duration and spread of movements over the episodes, one episode has a duration of 2 minutes and the other 3 minutes. Three experts are asked to judge the set of 14 video episodes and mark periods that they consider movement. Based on these scores, periods in the data are defined as [apos]movement events[apos]. The remaining periods are considered as [apos]no movement events[apos]. When raters score the same event, but they score a different on- or offset, we tolerate a timing difference of 3 seconds. These events are then considered as the same event. When one rater scores more events than the other in the same time interval, these events are considered as one event if the timing difference between the events of the more precise rater is within 3 seconds. The individual results of the experts are compared with each other and the interrater agreement is calculated. The measure that is used for the agreement is Cohen[apos]s [italic][kappa][/italic]. In general if [italic][kappa] [gt] [/italic]0.7, the interrater reliability is considered satisfactory. Computations are done with SPSS 14.1., Table 1 lists the values of [italic][kappa][/italic] for each pair of raters., Since[italic] [kappa] [/italic][gt] 0.7 in all cases, we consider the scores to be reliable enough to use as a reference for validation of our algorithm. Nevertheless we must realize this standard is not gold. It is possible that an automated detection algorithm based on accelerometry recordings detects actual movements, not visible in the video, because the moving body part is not within the scope of the camera. Therefore performance measures based on such a standard should be interpreted with care.[table1],