350
K. Lehnertz
over, it is to be expected that an approach that inputs a large (>100) number of
(mostly correlated) characteristics into various machine-learning algorithms but that
does not reasonably control the type-I error (due to multiple testing) will indeed
perform better than some random predictor. Even the testing against a random predictor is a debatable point, given that the probability of occurrence of seizures and
other epilepsy-related pathophysiologic phenomena is influenced and modulated by
various cycles (hormonal, sleeps-wake, circadian, multidien) [17, 50, 51, 63, 66,
72, 78, 81, 137, 143, 154]. Monte-Carlo-based approaches that preserve such periodicities [72] are better suited for such cases.
22.4 Devices for Seizure Prediction
Following studies that demonstrated the feasibility of hardware implementations
of seizure prediction algorithms [79, 122], between 2010 and 2012 an Australian
group enrolled 15 participants in the first and to date only prospective clinical trial
of a fully functioning, ambulatory seizure prediction system [37, 39]. The system
consisted of 16 intracranial electrodes that were directly in contact with one brain
hemisphere and were connected by subdermal wires to a subdermal telemetry unit
implanted in the chest, which wirelessly transmitted the EEG data to a hand-held
unit. The system enabled data processing in real time and issued warnings in the
form of colored lights indicating an impending seizure, intermediate, or safe states.
In a few subjects with epilepsy, recordings were performed over a period of up to
three years. For nine subjects above-chance warnings could be issued if subjectspecific seizure prediction algorithms were employed that were trained on a great
amount of data. Together with many previous studies, this trial demonstrated that
not all people with epilepsy have seizures that can be predicted and that not all of
a subject’s seizures are predictable. Future work will be needed to define epilepsy
phenotypes or endophenotypes that are associated with predictability [85]. This study
demonstrated that long-term recordings are possible, that people with epilepsy are
willing to volunteer for such studies and that research ethics committees see the
benefits of these trials. Although this scientifically successful study opened a door
for more long-term prospective trials, the device was never commercialized due to a
lack of investment.
Currently, other clinical trials explore the feasibility of personalized forecasts
from a mobile seizure diary app [64] and of wearables to better understand epilepsy
at a large scale [29]. There are also developments towards sub-scalp minimally invasive EEG devices [162] that—if feasible clinically—could represent an interesting
alternative to invasive recording techniques that are not without risk [57].
K. Lehnertz
over, it is to be expected that an approach that inputs a large (>100) number of
(mostly correlated) characteristics into various machine-learning algorithms but that
does not reasonably control the type-I error (due to multiple testing) will indeed
perform better than some random predictor. Even the testing against a random predictor is a debatable point, given that the probability of occurrence of seizures and
other epilepsy-related pathophysiologic phenomena is influenced and modulated by
various cycles (hormonal, sleeps-wake, circadian, multidien) [17, 50, 51, 63, 66,
72, 78, 81, 137, 143, 154]. Monte-Carlo-based approaches that preserve such periodicities [72] are better suited for such cases.
22.4 Devices for Seizure Prediction
Following studies that demonstrated the feasibility of hardware implementations
of seizure prediction algorithms [79, 122], between 2010 and 2012 an Australian
group enrolled 15 participants in the first and to date only prospective clinical trial
of a fully functioning, ambulatory seizure prediction system [37, 39]. The system
consisted of 16 intracranial electrodes that were directly in contact with one brain
hemisphere and were connected by subdermal wires to a subdermal telemetry unit
implanted in the chest, which wirelessly transmitted the EEG data to a hand-held
unit. The system enabled data processing in real time and issued warnings in the
form of colored lights indicating an impending seizure, intermediate, or safe states.
In a few subjects with epilepsy, recordings were performed over a period of up to
three years. For nine subjects above-chance warnings could be issued if subjectspecific seizure prediction algorithms were employed that were trained on a great
amount of data. Together with many previous studies, this trial demonstrated that
not all people with epilepsy have seizures that can be predicted and that not all of
a subject’s seizures are predictable. Future work will be needed to define epilepsy
phenotypes or endophenotypes that are associated with predictability [85]. This study
demonstrated that long-term recordings are possible, that people with epilepsy are
willing to volunteer for such studies and that research ethics committees see the
benefits of these trials. Although this scientifically successful study opened a door
for more long-term prospective trials, the device was never commercialized due to a
lack of investment.
Currently, other clinical trials explore the feasibility of personalized forecasts
from a mobile seizure diary app [64] and of wearables to better understand epilepsy
at a large scale [29]. There are also developments towards sub-scalp minimally invasive EEG devices [162] that—if feasible clinically—could represent an interesting
alternative to invasive recording techniques that are not without risk [57].