Epilepsy Detection and Monitoring - Quantitative EEG Analysis Methods and Clinical Applications - page 162

dictions or detections divided by the length of time of the recorded data (FP/ T ). It
gives an estimate of the number of times that the algorithm under consideration
would produce a false prediction or detection in a unit time (usually an hour).
Morman et al. [50] also point out that the prediction horizon is important when
considering the specificity rate of prediction algorithms. The prediction horizon is
the amount of time before the seizure for which the given algorithm is trying to pre-
dict it. The reason is false positives are more costly as the prediction horizon
increases. A false positive for an algorithm with a larger prediction horizon causes
the patient to spend more time expecting a seizure that will not occur. This is in
opposition to an algorithm with a smaller prediction horizon. Less time is spent
expecting a seizure that will not occur when a false positive is given. To correct this,
they suggest using a technique that reports the portion of time from the interictal
period during which a patient is not in the state of falsely awaiting a seizure [50].
Another issue that should be considered when assessing a particular seizure
detection/prediction technique is whether or not a posteriori information is used by
the technique in question. A posteriori information is information that can be used
to improve an algorithm's accuracy, but is specific to the dataset (EEG signal) at
hand. When the algorithm is applied to other datasets where this information is not
known, the accuracy of the algorithm can drop dramatically. In-sample optimiza-
tion is one example of a posteriori information used in some algorithms [14, 50].
With in-sample optimization, the same EEG signal that is used to test the given tech-
nique is also used to train the technique. When training a given algorithm, certain
parameters are adjusted in order to come up with a general method that can distin-
guish two classes. When training the technique, the algorithm is optimized to clas-
sify the training data. Therefore, when the same data that is used to test a technique
is used to train the technique, the technique is optimized (“overfit”) for the testing
data. Although this produces promising results as far as accuracy, these results are
not representative of what would be produced when the algorithm is applied to
nontraining, that is, out-of-sample, data.
Another piece of a posteriori information that is used in some algorithms is opti-
mal channel selection. When testing, other algorithms are given the channel of the
EEG that produces the best results. It has been shown that out of the available EEG
channels, not every channel provides information that can be used to predict or
detect a seizure [48, 50]. Other channels provide information that would produce
false positives. So when an optimal channel is provided to a given algorithm, the
results produced from this technique again will be biased. Therefore, the algorithm
does not usually generalize well to the online case when the optimal channel is not
known.

6.14 Closed-Loop Seizure Prevention Systems

The majority of patients with epilepsy are treated with chronic medication that
attempts to balance cortical inhibition and excitation to prevent a seizure from
occurring. However, anticonvulsant drugs only control seizures for about
two-thirds of patients with epilepsy [51]. Electrical stimulation is an alternative
treatment that has been used [52]. In most cases, open-loop simulation is used. This