Epilepsy Detection and Monitoring - Quantitative EEG Analysis Methods and Clinical Applications - page 160

In seizure prediction, the effectiveness of seizure prediction techniques tends to
be lower in terms of statistical robustness. This is because the time horizon of these
methods ranges from minutes to hours in advance of an impending seizure and
because the preictal state is not a well-defined state across multiple seizures and
across different patients. Some studies have shown evidence of a preictal period that
could be used to predict the onset of an epileptic seizure with high statistical robust-
ness [13, 49]. However, many of these studies use a posteriori knowledge or do not
use out-of-sample training [14]. This leads to a model that is “overfit” for the data
being used. When this same model is applied to other data, the accuracy of the tech-
nique typically decreases dramatically.
A number of algorithms have been developed solely for seizure detection and not
for seizure prediction. The goal in this case is to identify seizures from EEG signals
offline. Technicians spend many hours going through days of recorded EEG activity in
an effort to identify all seizures that occurred during the recording. A technique that
could automate this screening process would save a great amount of time and money.
Because the purpose is to identify every seizure, any part of the EEG data may be used.
Particularly a causal estimation of algorithmic measures can be used to determine the
time of seizure occurrence. Algorithms designed for this purpose typically have better
statistical performance and can only be used as an offline tool to assist in the identifi-
cation of EEG seizures in long records of EEG data.

6.13 Performance of Seizure Detection/Prediction Schemes

With so many seizure detection and prediction methods available, there needs to be
a way to compare them so that the “best” method can be used. Many statistics that
evaluate how well a method does are available. In seizure detection, the technique is
supposed to discriminate EEG signals in the ictal (seizure) state from EEG signals in
the interictal (nonseizure) state. In seizure prediction, the technique is supposed to
discriminate EEG signals in the preictal (before the seizure) state from EEG signals
in the interictal (nonseizure) state. The classification an algorithm gives to a particu-
lar segment of EEG for either seizure detection or prediction can be placed into one
of four categories:

￿ True positive ( TP ): A technique correctly classifies an ictal segment (preictal
for prediction) of an EEG as being in the ictal state (preictal for prediction).
￿ True negative ( TN ): A technique correctly classifies an interictal segment of an
EEG as being in the interictal state.

False positive ( FP ): A technique incorrectly classifies an interictal segment of
an EEG as being in the ictal state (preictal for prediction).

￿

False negative ( FN ): A technique incorrectly classifies an ictal segment
(preictal for prediction) of an EEG as being in the interictal state.

￿

Next we discuss how these classifications can be used to create metrics for evalu-
ating how well a seizure prediction/detection technique does. In addition, we also
discuss the use of a posteriori information. A posteriori information is used by cer-
tain algorithms to improve their accuracy. However, in most cases, this information