22 Predicting Epileptic Seizures—An Update
349
taneously within a few minutes [59, 71, 139, 140]. Recent research into the epileptic
network’s local characteristics provides increasing evidence that the epileptic focus—
widely referred to as the seizure-generating or seizure-initiating brain area [94]—is
not even a distinguished part of the network, neither during the seizure-free interval,
nor during the preseizure period, and not even during seizures [31, 46, 47, 132,
159, 165]. Rather, these studies suggest that a rearrangement of the network’s path
structure—possibly triggered by endogenous and/or exogenous factors—and that
results in a formation of bottlenecks [132] which induces the generation of a preseizure state. Earliest indications for such a formation (with lead times up to hours)
can be observed in network vertices which generate and sustain normal, physiological
brain dynamics during the seizure-free interval.
All the aforementioned characteristics can further be used—either independently
or in some combined way—as input to pattern recognition systems, machine learning
algorithms or classifiers (such as artificial neural networks) to identify a preseizure
state [2, 67, 105, 155].
22.3 Performance Evaluation of Seizure Prediction
Algorithms
Predictability of seizures with above-chance performance of prediction algorithms
was claimed by many studies published before and around the turn of the millennium. Follow-up studies indicated, however, that these were mostly over-optimistic
findings obtained by applying highly optimized algorithms to small, selected data
sets and could not be reproduced on unselected, more extended EEG recordings that
are more closely related to the real-life challenge of predicting seizures prospectively
from the continuous EEG [106, 109]. This key fault of historic literature continues to
pervade the field, despite (a) well known minimum requirements to ensure that published prospective and retrospective seizure prediction studies are comprehensible,
comparable and assessable (see guidelines in [77, 106]) as well as (b) availability of
rigorous frameworks to evaluate the performance of seizure prediction techniques.
These frameworks are based on Monte Carlo simulations [5, 7, 72, 109, 146] or on
comparison with analytical results derived from naive (random or periodic) prediction schemes [136, 167, 169]. Awareness of the importance of statistical evaluation
of seizure prediction algorithms—beyond estimating sensitivity and specificity—as
well as adherence to published guidelines is critical for understanding the value of the
results of seizure prediction studies and is indispensable to avoid making too strong
claims that may raise false hopes in professionals and in people with epilepsy [77].
There are similar concerns about the significance of recent seizure-forecasting
contests [28, 76]. Despite availability of large, high-quality databases that provide a pseudo-realistic test bed with continuous, multichannel, multi-day EEG data
recorded during sleep, wakefulness and activities of daily life [56, 76, 160], these
contests solely investigated preselected, small, discontinuous pieces of data. More-
349
taneously within a few minutes [59, 71, 139, 140]. Recent research into the epileptic
network’s local characteristics provides increasing evidence that the epileptic focus—
widely referred to as the seizure-generating or seizure-initiating brain area [94]—is
not even a distinguished part of the network, neither during the seizure-free interval,
nor during the preseizure period, and not even during seizures [31, 46, 47, 132,
159, 165]. Rather, these studies suggest that a rearrangement of the network’s path
structure—possibly triggered by endogenous and/or exogenous factors—and that
results in a formation of bottlenecks [132] which induces the generation of a preseizure state. Earliest indications for such a formation (with lead times up to hours)
can be observed in network vertices which generate and sustain normal, physiological
brain dynamics during the seizure-free interval.
All the aforementioned characteristics can further be used—either independently
or in some combined way—as input to pattern recognition systems, machine learning
algorithms or classifiers (such as artificial neural networks) to identify a preseizure
state [2, 67, 105, 155].
22.3 Performance Evaluation of Seizure Prediction
Algorithms
Predictability of seizures with above-chance performance of prediction algorithms
was claimed by many studies published before and around the turn of the millennium. Follow-up studies indicated, however, that these were mostly over-optimistic
findings obtained by applying highly optimized algorithms to small, selected data
sets and could not be reproduced on unselected, more extended EEG recordings that
are more closely related to the real-life challenge of predicting seizures prospectively
from the continuous EEG [106, 109]. This key fault of historic literature continues to
pervade the field, despite (a) well known minimum requirements to ensure that published prospective and retrospective seizure prediction studies are comprehensible,
comparable and assessable (see guidelines in [77, 106]) as well as (b) availability of
rigorous frameworks to evaluate the performance of seizure prediction techniques.
These frameworks are based on Monte Carlo simulations [5, 7, 72, 109, 146] or on
comparison with analytical results derived from naive (random or periodic) prediction schemes [136, 167, 169]. Awareness of the importance of statistical evaluation
of seizure prediction algorithms—beyond estimating sensitivity and specificity—as
well as adherence to published guidelines is critical for understanding the value of the
results of seizure prediction studies and is indispensable to avoid making too strong
claims that may raise false hopes in professionals and in people with epilepsy [77].
There are similar concerns about the significance of recent seizure-forecasting
contests [28, 76]. Despite availability of large, high-quality databases that provide a pseudo-realistic test bed with continuous, multichannel, multi-day EEG data
recorded during sleep, wakefulness and activities of daily life [56, 76, 160], these
contests solely investigated preselected, small, discontinuous pieces of data. More-