Abstracts

During prolonged EEG monitoring of epileptic patients, the continuous recording may be marked where seizures are likely to take place. Several methods of automatic seizure detection exist, but few can operate as an on-line seizure alert system. We propose a seizure detection system that can be used in real-time to alert medical staff to the onset of a patient seizure and hence improve clinical diagnosis. Also, existing methods provide an all-or-nothing approach to detection. We propose a probability-based system that outputs a variable based on the [italic]seizure probability [/italic]of a section of EEG. Final operation features a user-tunable threshold to exploit the trade-off between early detection and an acceptable false detection rate.
[italic]System Training[/italic]: Wavelet decomposition and feature extraction were performed on a total of 298 hours of EEG (52 seizures in 13 patients as well as one 4-hour seizure-free awake recording and one 4-hour seizure-free sleep recording from each patient). Features from three wavelet decomposition scales relevant to seizure activity (3-6 Hz, 6-12 Hz and 12-25 Hz) were stored separately for seizure and non-seizure sections. [italic]A priori[/italic] probabilities were computed based on bin counts from the segmented data distributions. The detection variable for a given 2-second epoch of EEG was formed by summing the [italic]a posteriori[/italic] probabilities from all three scales in one channel and considering the six channels with the highest sums.
[italic]Operation[/italic]: Feature data is computed as it was in the training stage and Bayes[rsquo] formula is applied using the [italic]a priori[/italic] data collected during training. Output is a single variable and operation allows a tunable threshold.
Results based on analysis of training data show a sensitivity of 72.0% (average of individual patient sensitivities) and a false detection rate of 1.25/hr. This average sensitivity is based on the successful detection of 34 out of 52 seizures with a mean detection delay of 7.4 s and a delay of 6 seconds or less in 22 of these. Out of the 18 seizures that were not detected, nine were from two patients and constituted the entire set for those patients. We were able to demonstrate the trade-off between detection delay and false detection rate.
False detections were caused primarily by two phenomena: epileptiform events such as spike and wave complexes and bursts of slow wave activity as well as bursts of alpha activity.
Results are encouraging because the false alarm rate is acceptable and seizures are detected quite early in their development. We expect to be able to improve sensitivity and reduce false alarms, particularly by identifying specifically bursts of alpha activity. The results presented are preliminary, however, in that they describe the behavior of the system with data used in its original development. Further testing using a separate data set will serve as the final evaluation of whether the system will be useful in a clinical setting.
[Supported by: The Canadian Institutes of Health Research (CIHR) under grant MT-10139.]