A forward-looking review of seizure prediction

Abstract

Purpose of review:
Seizure prediction has made important advances over the last decade, with the recent demonstration that prospective seizure prediction is possible, though there remain significant obstacles to broader application. In this review, we will describe insights gained from long-term trials, with the aim of identifying research goals for the next decade.

Recent findings:
Unexpected results from these studies, including strong and highly individual relationships between spikes and seizures, diurnal patterns of seizure activity, and the coexistence of different seizure populations within individual patients exhibiting distinctive dynamics, have caused us to re-evaluate many prior assumptions in seizure prediction studies and suggest alternative strategies that could be employed in the search for algorithms providing greater clinical utility. Advances in analytical approaches, particularly deep-learning techniques, harbour great promise and in combination with less-invasive systems with sufficiently power-efficient computational capacity will bring broader clinical application within reach.

Summary:
We conclude the review with an exercise in wishful thinking, which asks what the ideal seizure prediction dataset would look like and how these data should be manipulated to maximize benefits for patients. The motivation for structuring the review in this way is to create a forward-looking, optimistic critique of the existing methodologies.

Full text

CE: Alpana; WCO/300206; Total nos of Pages: 7;
WCO 300206
C
URRENT
O
PINION
A forward-looking review of seizure prediction
Dean R. Freestone
a
, Philippa J. Karoly
a,b,c
, and Mark J. Cook
a
Purpose of review
Seizure prediction has made important advances over the last decade, with the recent demonstration that
prospective seizure prediction is possible, though there remain significant obstacles to broader application.
In this review, we will describe insights gained from long-term trials, with the aim of identifying research
goals for the next decade.
Recent findings
Unexpected results from these studies, including strong and highly individual relationships between spikes
and seizures, diurnal patterns of seizure activity, and the coexistence of different seizure populations within
individual patients exhibiting distinctive dynamics, have caused us to re-evaluate many prior assumptions in
seizure prediction studies and suggest alternative strategies that could be employed in the search for
algorithms providing greater clinical utility. Advances in analytical approaches, particularly deep-learning
techniques, harbour great promise and in combination with less-invasive systems with sufficiently power-
efficient computational capacity will bring broader clinical application within reach.
Summary
We conclude the review with an exercise in wishful thinking, which asks what the ideal seizure prediction
dataset would look like and how these data should be manipulated to maximize benefits for patients. The
motivation for structuring the review in this way is to create a forward-looking, optimistic critique of the
existing methodologies.
Keywords
devices, machine learning, prediction, seizure dynamics
INTRODUCTION
Epileptic seizure prediction is a challenging problem
that has attracted the attention of a wide range of
disciplines including neuroscience, mathematics,
engineering, science, statistics, machine learning,
and physics. It is a field that has a short but rich
history of speculation, controversy, and achieve-
ments. In this review, we discuss seven lessons that
we have obtained from our long-term datasets.
Epilepsy is a unique disease in that the percentage
of time spent with clinically relevant behavioural
symptoms is very low, yet the impact on a patient’s
life is severe. For example, the drug-refractory patients
in the NeuroVista Corporation, Seattle WA, USA study
(the longest continuous recording undertaken in
humans) experienced clinical seizures on average for
less than 0.05% of the total time in the study (up to
3 years for each of the 15 patients) [1]. Notably, this
cohort was chosen because of their relatively high
seizure rate. Despitethe low frequency ofclinical symp-
toms, uncontrolled seizures have profound effects on
thelivesofpatientsandcaregivers.Thedisablingeffects
of seizures are due to their sometimes-dramatic behav-
ioural manifestations and apparent unpredictability.
More than 99.95% of the time when people with
epilepsy are not experiencing seizures, they should
be able to lead relatively normal lives. However, for
many the uncertainty of seizure occurrence intrudes
into every aspect of day-to-day life. Consequently,
the ability to accurately forecast seizures could be
life-changing for people with epilepsy, permitting a
greater degree of personal freedom and safety.
Despite the large number of new anticonvulsant
medications available over the last 20 years, the
proportion of medically refractory patients has
a
Department of Medicine, The University of Melbourne, St. Vincent’s
Hospital, Fitzroy,
b
Department of Electrical and Electronic Engineering,
The University of Melbourne, Parkville, Victoria, Australia and
c
Center for
NeuroEngineering and Therapeutics, University of Pennsylvania, Phila-
delphia, Pennsylvania, USA
Correspondence to Mark J. Cook, Department of Medicine, The Univer-
sity of Melbourne, St. Vincent’s Hospital, Parkville, VIC 3010, Australia.
E-mail: markcook@unimelb.edu.au

Dean R. Freestone and Philippa J. Karoly contributed equally to the
article and share first authorship.
Curr Opin Neurol 2017, 30:000– 000
DOI:10.1097/WCO.0000000000000429
1350-7540 Copyright ß2017 Wolters Kluwer Health, Inc. All rights reserved. www.co-neurology.com
REVIEW
Copyright © 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.

CE: Alpana; WCO/300206; Total nos of Pages: 7;
WCO 300206
not significantly altered [2]. Seizure prediction could
provide new strategies for the treatment of epilepsy.
Many patients report that their seizures occur at
transition points of the day, such as in the evening
during attempting to going to sleep or in the morn-
ing when waking [3,4]. Others report less well
defined changes in behaviour and mood, or cogni-
tive disturbances prior to seizure onset [5], some-
times hours or days prior to clinical events. These
reports, in conjunction with neuroimaging studies
[6], suggested that there are measurable changes in
brain dynamics prior to seizures [7–9]. The field of
seizure prediction has progressed with the assump-
tion that the changes in brain dynamics are best
tracked by electroencephalography (EEG).
The literature in seizure prediction is dominated
by articles written by electronic engineers or com-
puter scientists with a background in signal process-
ing, who have implemented algorithms to track
complexity or information capacity in the brain
(refer to Mormann et al. [10] and Gadhoumi et al.
[11
&&
] for reviews). Algorithms are typically based
around the presumption that there is a feature in the
EEG that is indicative of imminent seizures [12].
Despite the ingenuity of the algorithm developers,
we can rarely have any certainty if these algorithms
will work for a given patient when applied prospec-
tively. Few of the published works have prospec-
tively applied prediction techniques, and very few
have employed statistically appropriate controls. In
addition, they have typically analysed relatively
small datasets with low seizure numbers. The fact
that the current generation of clinically approved
predictive devices [1] and closed-loop stimulation
systems [13] have used detection features alone
suggests a distinct lack of confidence in the more
sophisticated published algorithms for prediction,
which may permit more efficacious counterstimu-
lation strategies.
Possible factors driving the low uptake of
‘black-box’ predictive algorithms are the necessary
abstraction of EEG features from seizure-inducing
mechanisms, coupled with rare demonstrations of
their reproducibility [10,11
&&
]. These abstract algor-
ithms are convincing when effective; but there is little
opportunity to gain insight into why they fail, and
combined with a low number of test cases (patients
and/or seizures) it is hardly surprising that a great deal
of scepticism exists amongst those outside the field. In
this review, we argue that lack of confidence in
machine-learning algorithms stems from a failure to
properly define the problem of seizure prediction.
Seizure prediction is not a straightforward matter of
classifying data segments into strict binary categories
of interictal and preictal periods, but is more appro-
priately considered as a regression problem, in which
data are continuously filtered through a model of
seizure probability and expressed as a seizure likeli-
hood. Furthermore, with long-term datasets, it has
become clear that there are aspects of epileptic dynam-
ics that are underutilized by the existing pattern recog-
nition paradigms for seizure prediction.
The following sections of this review will
expound on the insights that have been gained from
long-term data collected from devices such as the
NeuroVista [1], NeuroPace Inc, Mountain View CA,
USA [13], and Medtronic, Minneapolis MN, USA
systems [14]. These studies demonstrate that the
field of seizure prediction is on the brink of an
exciting new period, in which hardware for record-
ing the relevant neural signals is now technologic-
ally capable of supporting the elaborate software
used to compute online features.
SEVEN LESSONS IN SEIZURE PREDICTION
Pooling data across patients may not be an
effective strategy
When considering pooled data across patients, we
must recognize that epilepsy is a constellation of
symptoms, not mechanisms [15], with a hetero-
geneous range of causes [16], so it is reasonable to
assume that seizure-generating mechanisms may
differ between patients, especially given evidence
that similar network activity can indeed arise from
disparate network configurations [17]. However,
most studies implicitly assume that different
patients have a common onset mechanism with
‘one-size-fits-all’ algorithms [10,11
&&
].
This uniform seizure forecasting approach has
been an unavoidable consequence of limited data
availability. A typical intracranial EEG recording
occurs over 7–10 days, limited by risk of infection,
hospital costs, and patient compliance. During this
period, it is common for a patient to exhibit rela-
tively few seizures [18], and in most cases the
KEY POINTS
Seizure prediction is possible; however, a new
paradigm is needed.
Long-term recording has provided valuable insights into
seizure dynamics that should inform new
prediction strategies.
The next generation of implantable devices and smart,
wearable technology can support sophisticated
algorithms for seizure prediction, providing great hope
to medically refractory patients.
Seizure disorders
2www.co-neurology.com Volume 30 Number 00 Month 2017
Copyright © 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.

CE: Alpana; WCO/300206; Total nos of Pages: 7;
WCO 300206
number of seizures per patient is not sufficient to
create and validate a patient-specific seizure predic-
tion algorithm, and so data must be pooled to
statistically validate an algorithm. Pooling datasets
with low numbers of seizures leads to issues with
repeatability due to overfitting of algorithms in
heterogeneous groups. Algorithms using pooled
data have demonstrated that seizure prediction is
possible in some patient cohorts [19,20], but this
controversial approach has led to reproducibility
issues with different datasets.
Data that are recorded in the epilepsy
monitoring unit are not representative of
‘normal’ seizures
The time constraints in epilepsy monitoring units
often prompt clinicians to provoke seizures, often
through sleep deprivation patients’ medication
reduction [21]. Seizures that occur in this altered state
may not be representative of a habitual event [18], and
there is now strong evidence that drug tapering will
yield measurable changes in the EEG [22
&
,23]. Implan-
tation of intracranial electrodes can alter seizure
dynamics, with some patients even experiencing
remission of seizures [24]. Furthermore, features
derived from electrocorticography (ECoG) are now
known to be nonstationary for months after electrode
implantation [25], though these signals are typically
used to develop seizure prediction algorithms.
False positives (in the deterministic sense)
may not represent false positives (in the
probabilistic sense)
Many seizure prediction algorithms are based on the
notion that if an EEG signal feature crosses a
threshold, then a seizure is deemed certain to occur
within a forecast horizon. If a clinical seizure does not
occur, then the detection is deemed a false positive.
To quote from a recent review, ‘the performance of an
algorithm should be reported in terms sensitivity and
specificity’ [11
&&
]. Coupled with the correct statistical
assumptions, the sensitivity/specificity performance
evaluation is a powerful method of assessment
[26,27], but rules out the possibility that false detec-
tions are capturing epilepsy-related activity.
For a system that has no internal (corrective)
regulatory mechanism, the assumption that we must
remove all false positives is appropriate; however, the
brain has multiple regulatory mechanisms [28]. As
homeostatic mechanisms may correct abnormal
activity, it is possible to have states of high suscepti-
bility, without events occurring [29,30]. This poses a
problem evaluating seizure prediction algorithms,
given it is not possible to assess whether algorithms
are correctly identifying states of high seizure
susceptibility if no clinical seizure occurs.
A more appropriate method of evaluation is to
move away from a deterministic framework and take
a more probabilistic point of view. In line with this
approach, probabilistic scores such as the Brier score
should be used to evaluate predictors [31,32]. A
constraint is that one must empirically construct a
cumulative density function of the seizure prob-
ability given the available measurements [33], and
many data points (seizures) are required to accu-
rately estimate the density, typically orders of mag-
nitude more than are recorded in typical studies.
An added benefit of the probabilistic approach is
that forecasts can be conditioned on relevant prior
information, including the prior performance of the
algorithm itself [34,35]. For example, in certain
individuals, seizures follow circadian patterns
[36
&&
,37
&
,38,39]. These conditional distributions
could be used to construct a-priori estimates that
can be combined with EEG signal features via a
Bayesian inference approach. In a probabilistic
framework, there is no limit to the number of infor-
mative priors that can be incorporated into a pre-
dictive model. More importantly, the Bayesian
framework also provides a solid foundation for
traditional threshold-based seizure warning sys-
tems. With a probabilistic framework in place, there
are established methods [40] that can be used to
iteratively calibrate the forecast based on perform-
ance metrics and patient-specific dynamics.
Bi/multistability of brain states does not
mean seizure forecasting is impossible
There is evidence that a stable seizure state coexists
with stable normal brain states (wake, sleep, etc.).
Coexistence of multiple stable states is known as bi/
multistability [41–43]. It has been postulated that
seizures may occur because of unpredictable fluctu-
ations in the dynamics of the brain driving the state
from the healthy to the pathological attractor [43].
As the driving force is considered random, one may
sensibly assume that seizures are inherently unpre-
dictable [42]. The alternative scenario involves a
primarily deterministic route to seizures through
variation in underlying parameters [44]. However,
in practice, both scenarios may occur simul-
taneously [45], and so evidence for stochasticity
does not exclude other mechanisms.
We argue that both slow variations of a dynam-
ical variable and ‘noisy’ fluctuations in the brain’s
background activity are necessary conditions for
seizure. Predicting the exact time a seizure may
occur [the deterministic case from the ‘False posi-
tives (in the deterministic sense) may not represent
A forward-lo oking review of seizure prediction Freestone et al.
1350-7540 Copyright ß2017 Wolters Kluwer Health, Inc. All rights reserved. www.co-neurology.com 3
Copyright © 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.

CE: Alpana; WCO/300206; Total nos of Pages: 7;
WCO 300206
false positives (in the probabilistic sense)’ section]
may be impossible in the bistability with noise-
driven transition scenario. On the other hand, esti-
mating seizure susceptibility in a probabilistic sense,
in which the important quantity of interest is the
current (slowly varying) stability landscape, which
can be described by the expected escape time from a
healthy to a pathological attractor [45] is more likely
to be successful.
We can also measure susceptibility [46 – 48], the
energy barrier separating states, or equivalently, to
the expected escape time [45]. It is possible to
measure the proximity of a state transition for a
very general class of systems using universal signa-
tures of resilience [49], and these signatures can be
related to a weaker barrier (escape time) or increased
seizure susceptibility. The effect is known as critical
slowing, as it is characterized by an increasing time
constant as a state transition approaches. In the case
of epilepsy, the transition is a change from a normal
to a seizure state. Critical slowing effects have been
observed in a range of natural systems prior to
catastrophic events, such as population collapse
or natural disasters [50]. There is evidence of critical
slowing before seizures in in-vitro tissue [51] and in-
vivo for animal models [52], by either an increase in
autocorrelation of the EEG (slowing) or a lengthen-
ing of electrical evoked potentials (increased time
constant). More recently, it has been shown that
signatures of a state transition are related to chang-
ing levels of excitability in the brain [22
&
], and for
some patients, signatures of critical slowing are
evident in the lead up to clinical seizures [53].
The relationship of epileptic biomarkers and
seizures is patient-specific
Biomarkers of epilepsy include interictal spikes, inter-
ictal spike-wave discharges, epileptiform bursts and
rhythmic bursts, and high-frequency oscillations
[54,55], which are currently used in diagnosis, surgi-
cal planning, treatment, and have demonstrable
clinical significance [56–58]. There is a wealth of
studies investigating their utility in surgical planning
and more generally how biomarkers relate to epilep-
togenesis and ictogenesis. Taken together, these stud-
ies show variable results [56], which may in part be
due to experimental design [59]. Results showing
individualized patterns of activity of biomarkers
may be inconclusive, or efforts are made to ascribe
individual variability to specific, identifiable factors.
It may be futile to continue the search for simplistic
underlying principles governing the behaviour of
biomarkers at a population level.
Spikes and spike-wave discharges are the most
commonly studied features [60]. Despite poor
agreement between physicians on what constitutes
a spike, several recent studies have shown that spikes
can be reliably detected and occur at far higher
frequency compared with seizures [36
&&
,37
&
]. It has
variously been proposed that spikes prevent seiz-
ures, provoke seizures, or are an epiphenomenon of
epilepsy [61–63]. We now have evidence that these
hypotheses are overly simplistic. There are certain
patterns of circadian variation that may hold on a
population level [30,37
&
,64]; however, individual
variability appears to be at least equally powerful
[36
&&
].
Although epileptic biomarkers have clear
implications for understanding seizures, they can
exhibit highly patient-specific dynamics. It is unfor-
tunate that individual variability and statistical sig-
nificance make such poor bedfellows. However, the
expected variability in prospective cohort studies
can be overcome with long-term monitoring. Con-
sidering the overwhelming evidence, we suggest it is
imperative to move towards long-term studies and
away from a one-size-fits-all approach to analysing
biomarkers of epilepsy.
Seizures cluster and cycle
It is now well established that seizures tend to cluster
in time [65–67]. From a forecasting point of view,
seizure likelihood is increased when a seizure has
recently occurred, pointing to underlying trends in
the data. These trends are described by a phenom-
enon known as long-range dependence (long mem-
ory), which has been shown to characterize seizure
timing [68,69]. Long-range dependence suggests
persistent correlations between interseizure inter-
vals, manifest as clustering over short time scales,
as well as very long seizure-free intervals (longer
than expected based on the baseline seizure fre-
quency) [69]. The key insight here is that high
variability in interseizure intervals is the norm –
on the one hand, complicating patient manage-
ment, yet also suggesting the presence of a non-
random component underlying seizure generation
that can be exploited by predictive algorithms.
In addition to statistical trends, seizures exhibit
strong cyclic patterns at periods from hours up to
months [38,39,70]. The precise causes of periodic
circadian patterns remain largely unknown, though
a variety of endogenous factors may affect neural
dynamics and consequently play a role in increasing
the likelihood of seizures [71]. Longer ultradian
cycles, often considered catamenial patterns, are
also present in men [3,72]. It has long been proposed
that other environmental inputs, relating to sea-
sonal and behavioural changes can alter the chances
of seizures at certain times of the day, week, month,
Seizure disorders
4www.co-neurology.com Volume 30 Number 00 Month 2017
Copyright © 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.

CE: Alpana; WCO/300206; Total nos of Pages: 7;
WCO 300206
or even year [38]. Intriguingly, patterns of other
epileptiform discharges including subclinical seiz-
ures and spike-wave discharges appeared to be well
synchronized with seizure cycles for some patients
but not others [36
&&
].
Individual patients may have multiple types
of seizures with different onset mechanisms,
thereby needing multiple predictor types
Patients who performed poorly in terms of forecast-
ing accuracy in the NeuroVista study had seizures
with durations that followed multimodal distri-
butions, compared with the patients with unimodal
distributions who were the best performers. The link
between seizure duration and differences in presei-
zure EEG features suggests the presence of differing
onset mechanisms for short and long seizures [73].
These same patients were also found to have clus-
tered epileptiform discharges, which showed signifi-
cant differences in the preceding EEG energy
compared with other epileptic seizures [67]. This
finding suggests that different types of seizures
manifest distinct preictal dynamics. Further sup-
porting evidence for a causal link between onset
mechanism and seizure duration is found in an
earlier study, in which different electrical stimu-
lation patterns initiated seizures of two distinct,
yet predictable, classes of duration [74].
Pooling data across patients is a suboptimal due
to the heterogeneity of seizure-onset mechanisms,
and we also argue that development of forecasting
algorithms needs to be seizure-type-specific at the
level of individual patients. For instance, training
data for a prediction algorithm should be separated
into multiple classes based on seizure types, that is
for certain patients (in which there is a distinct
separation between long and short events) seizure
classes are determined by duration. Distinguishing
between seizures and other epileptiform discharges
is also important, the latter confounds predictive
algorithms (as previously mentioned patients with
worse predictive outcomes also had greater rates of
other epileptiform discharges [67]), so distinguish-
ing these from seizures may reduce the false alarm
rate. The ability to classify subclinical activity is also
informative for treatment strategies, such as closed-
loop stimulation systems [13,75].
Obtaining the volume of data necessary to train
an algorithm to distinguish between multiple
classes of seizures is difficult, but it is possible to
create a framework that implicitly accounts for such
differences using a voting scheme to combine and
weigh algorithms [76], which often shows impres-
sive results in EEG signal classification [76,77].
THE IDEAL DATASET FOR A PRACTICAL
SEIZURE PREDICTION DEVICE
The dynamics of epileptic activity span the time
scales of milliseconds to months, and possibly even
years, necessitating long-term continuous monitor-
ing to capture clinically useful data. Safe and reliable
recording via long term intracranial EEG provides
sufficient fidelity for seizure forecasting, and has
been shown to be feasible [1,25]. Seizure likelihood
is influenced by the current brain state, the phase of
the body’s internal rhythms (sleep–wake cycle, time
of the day, day of week, and month of year), and
possibly environmental conditions. Monitoring
exogenous and endogenous factors may also pro-
vide additional useful input for seizure prediction,
and an ideal dataset would include biomarkers such
as heart rate, blood oxygen [78], movement infor-
mation (accelerometer), and information relating to
behaviour status (audio and video). These features
can all be readily recorded with current wearable
technology in conjunction with smartphone devi-
ces. Such data enable more accurate tracking of
clinically relevant events, as well as providing a
measure of other factors that may influence seizure
susceptibility.
Less-invasive systems, such as a subscalp
monitor, would potentially permit a more widely
available system, reducing cost, invasiveness, and
risk, but are yet to be realized [79]. Biomarkers, such
as spike-wave discharges, can be measured from
scalp EEG signals, and when combined with infor-
mative prior information from seizure patterns at
longer time scales such low-impact devices with
simpler algorithms may prove to be the most clin-
ically viable system.
WHAT WOULD WE DO WITH THE IDEAL
DATASET?
Long-term continuous monitoring of the relevant
factors relating to seizure likelihood allows us to
implement two fundamentally important paradigm
shifts. The first is to cast seizure prediction as a
regression problem involving probabilities, rather
than the standard binary classification of interictal
versus preictal data. The second change with an
ideal dataset is to move from standard feature engin-
eering to a machine-learning framework to decode
neural signals.
Decoding electrical activity associated with
seizures from an ECoG dataset is analogous to
an object-recognition task in the field of computer
vision. Recently, machine-learning techniques,
known as deep learning, have revolutionized com-
puter vision [80] and have had a major impact
A forward-lo oking review of seizure prediction Freestone et al.
1350-7540 Copyright ß2017 Wolters Kluwer Health, Inc. All rights reserved. www.co-neurology.com 5
Copyright © 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.

CE: Alpana; WCO/300206; Total nos of Pages: 7;
WCO 300206
in artificial intelligence [81]. These techniques
have the potential to also revolutionize seizure
prediction.
Instead of detecting a particular feature in a
scene, artificial neural networks can detect particu-
lar seizure-related patterns in the brain signals in a
system equivalent to recognizing an object in a
visual scene [82]. A critically important aspect of
deep learning is that it is not necessary to under-
stand the physical rules or generators of data [80], a
feature particularly relevant for the decoding proc-
ess, as the mapping between epileptic states and
corresponding brain signals is not completely
understood. Deep learning overcomes this limita-
tion, allowing us to extract and learn features from
data without explicit knowledge of the structure.
The implementation of deep learning algorithms for
seizure prediction devices has previously been lim-
ited by practical considerations of power consump-
tion. These considerations are no longer restrictive,
and the next generation of implantable seizure
warning devices holds great promise [11
&&
,79].
Recently, the utility of using artificial neural
network to decode EEG data [83] has been demon-
strated. These neural decoding algorithms can be
now implemented on devices such as IBM’s True-
North: IBM Research, Armonk NY, USA neuromor-
phic chip [82,84]. This brain-inspired computer
architecture is very fast and very efficient, enabling
the creation of a small implantable device that has
the power of a super computer, but which consumes
little power. This provides a pathway to creating
real-world fully implantable devices.
CONCLUSION
Long-term continuous data collection has recently
shown us that seizure prediction is possible and has
provided a clear path to developing new clinical
devices. In addition, we have seen that data from
device trials have led to a deeper understanding of
seizure-onset mechanisms, potentially opening new
avenues for intervention. The demonstration that an
individual with stereotype clinical phenomenology
may have multiple populations of seizure, each with
different onset and offset mechanisms, provides
quite a different perspective when approaching
therapy. Moreover, the populations and temporal
patterns of seizure activity vary considerably between
individuals, demanding highly personalized strat-
egies both for predictive algorithms, possibly for
seizure control. Advances in machine learning and
low-power computing systems, combined with less-
invasive systems that capture a variety of additional
biomarkers, will likely bring predictive systems into
wider clinical application.
Acknowledgements
The authors acknowledge our colleagues at Mark Cook’s
lab (St Vincent’s Neuroscience, The University of Mel-
bourne), David Grayden’s lab (Biomedical Engineering,
The University of Melbourne) and our close collaborators
at Brian Litt’s lab (Centre for Neuroengineering and
Therapeutics, University of Pennsylvania), and John
Terry’s lab (Centre for Predictive Modelling in Health-
care, University of Exeter), without whom none of this
research would have been possible.
The authors also acknowledge the contribution of the
NeuroVista Corporation.
Financial support and sponsorship
The research was funded by an Australian National
Health and Medical Research Council Project Grant
(APP1065638). P.J.K. also acknowledges the support
of the Kenneth Myer Foundation and Pro Medicus Ltd.
Conflicts of interest
There are no conflicts of interest.
REFERENCES AND RECOMMENDED
READING
Papers of particular interest, published within the annual period of review, have
been highlighted as:
&of special interest
&& of outstanding interest
1. Cook MJ, O’Brien TJ, Berkovic SF, et al. Prediction of seizure likelihood with a
long-term, implanted seizure advisory system in patients with drug-resistant
epilepsy: a first-in-man study. Lancet Neurol 2013; 12:563 –571.
2. Milton J: Medically intractable epilepsy. In Epilepsy as a dynamic disease.
Edited by: Milton J, Jung P, (Eds). Springer-Verlag Berlin Heidelberg, Ger-
many; 2003:1-14
3. Patry FL. The relation of time of day, sleep, and other factors to the incidence
of epileptic seizures. Am J Psychiatry 1931; 87:789 –813.
4. Zarowski M, Loddenkemper T, Vendrame M, et al. Circadian distribution and
sleep/wake patterns of generalized seizures in children. Epilepsia 2011;
52:1076– 1083.
5. Blum D, Eskola J, Bortz J, Fisher R. Patient awareness of seizures. Neurology
1996; 47:260– 264.
6. Litt B, Esteller R, Echauz J, et al. Epileptic seizures may begin hours in advance
of clinical onset: a report of five patients. Neuron 2001; 30:51 –64.
7. Litt B, Lehnertz K. Seizure prediction and the preseizure period. Curr Opin
Neurol 2002; 15:173– 177.
8. Stead M, Bower M, Brinkmann BH, et al. Microseizures and the spatiotem-
poral scales of human partial epilepsy. Brain 2010; 133:2789 –2797.
9. Badawy R, Macdonell R, Jackson G, Berkovic S. The peri-ictal state: cortical
excitability changes within 24 h of a seizure. Brain 2009; 132:1013 –1021.
10. Mormann F, Andrzejak RG, Elger CE, Lehnertz K. Seizure prediction: the long
and winding road. Brain 2007; 130:314– 333.
11.
&&
Gadhoumi K, Lina J-M, Mormann F, Gotman J. Seizure prediction for ther-
apeutic devices: a review. J Neurosci Methods 2016; 260:270–282.
An excellent overview of the current state of the art in predictive algorithms.
12. Lehnertz K, Mormann F, Kreuz T, et al. Seizure prediction by nonlinear EEG
analysis. IEEE Eng Med Biol Mag 2003; 22:57 –63.
13. Sun FT, Morrell MJ. The RNS system: responsive cortical stimulation for the
treatment of refractory partial epilepsy. Expert Rev Med Devices 2014;
11:563– 572.
14. Stypulkowski PH, Stanslaski SR, Denison TJ, Giftakis JE. Chronic evaluation
of a clinical system for deep brain stimulation and recording of neural network
activity. Stereotact Funct Neurosurg 2013; 91:220 –232.
15. Fisher RS, Boas WvE, Blume W, et al. Epileptic seizures and epilepsy:
definitions proposed by the International League Against Epilepsy (ILAE) and
the International Bureau for Epilepsy (IBE). Epilepsia 2005; 46:470–472.
16. Shorvon SD, Andermann F, Guerrini R. The causes of epilepsy: common and
uncommon causes in adults and children. In: Simon DS, Frederick A, Renzo
G, (eds). Cambridge, UK: Cambridge University Press; 2011.
Seizure disorders
6www.co-neurology.com Volume 30 Number 00 Month 2017
Copyright © 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.

CE: Alpana; WCO/300206; Total nos of Pages: 7;
WCO 300206
17. Prinz AA, Bucher D, Marder E. Similar network activity from disparate circuit
parameters. Nat Neurosci 2004; 7:1345–1352.
18. Ghougassian DF, d’Souza W, Cook MJ, O’Brien TJ. Evaluating the utility of
inpatient video-EEG monitoring. Epilepsia 2004; 45:928 –932.
19. Lehnertz K, Mormann F, Osterhage H, et al. State-of-the-art of seizure
prediction. J Clin Neurophysiol 2007; 24:147 –153.
20. Mormann F, Andrzejak RG. Seizure prediction: making mileage on the long
and winding road. Brain 2016; 139:1625– 1627.
21. Malow BA. Sleep deprivation and epilepsy. Epilepsy Curr 2004; 4:193 –195.
22.
&
Meisel C, Schulze-Bonhage A, Freestone D, et al. Intrinsic excitability measures
track antiepileptic drug action and uncover increasing/decreasing excit ability over
the wake/sleep cycle. Proc Natl Acad Sci U S A 2015; 112:14694– 14699.
Demonstrating that excitability measures related to medications can be tracked
using continuous data recording.
23. Azar NJ, Wang L, Song Y, Abou-Khalil BW. Temporal pattern of oxcarbaze-
pine and phenytoin withdrawal seizures during epilepsy monitoring. Epilepsy
Res 2008; 79:78– 83.
24. Boe
¨x C, Seeck M, Vulliemoz S, et al. Chronic deep brain stimulation in mesial
temporal lobe epilepsy. Seizure 2011; 20:485– 490.
25. Sillay KA, Rutecki P, Cicora K, et al. Long-term measurement of impedance in
chronically implanted depth and subdural electrodes during responsive
neurostimulation in humans. Brain Stimul 2013; 6:718 –726.
26. Andrzejak RG, Chicharro D, Elger CE, Mormann F. Seizure prediction: any
better than chance? Clin Neurophysiol 2009; 120:1465 –1478.
27. Snyder DE, Echauz J, Grimes DB, Litt B. The statistics of a practical seizure
warning system. J Neural Eng 2008; 5:392.
28. Buzsaki G. Rhythms of the brain. Oxford, UK: Oxford University Press; 2006.
29. Badawy R, Freestone D, Lai A, Cook M. Epilepsy: ever-changing states of
cortical excitability. Neuroscience 2012; 222:89 –99.
30. Ly J, Gaggioni G, Chellappa S, et al. Circadian regulation of human cortical
excitability. Nat Commun 2016; 7:11828.
31. Brier GW. Verification of forecasts expressed in terms of probability. Mon
Weather Rev 1950; 78:1– 3.
32. Murphy AH, Winkler RL. A general framework for forecast verification. Mon
Weather Rev 1987; 115:1330– 1338.
33. Bishop CM. Pattern recognition and machine Learning, vol 1. New York:
Springer-Verlag; 2006.
34. Box GE, Tiao GC. Bayesian inference in statistical analysis, vol 40. Hoboken,
NJ USA: John Wiley & Sons Inc; 2011.
35. Coelho C, Pezzulli S, Balmaseda M, et al. Forecast calibration and combina-
tion: a simple Bayesian approach for ENSO. J Clim 2004; 17:1504–1516.
36.
&&
Karoly PJ, Freestone DR, Boston R, et al. Interictal spikes and epileptic seizures:
their relationship and underlying rhythmicity. Brain 2016; 139:1066 –1078.
This publication describes the complex and individual patterns of spikes and
seizure activity.
37.
&
Anderson CT, Tcheng TK, Sun FT, Morrell MJ. Day –night patterns of epilepti-
form activity in 65 patients with long-term ambulatory electrocorticography. J
Clin Neurophysiol 2015; 32:406–412.
Another work demonstrating the patterns of epileptiform changes in relation to the
sleep– wake cycle.
38. Langdon-Down M, Brain WR. Time of day in relation to convulsions in
epilepsy. Lancet 1929; 213:1029– 1032.
39. Griffiths G, Fox JT. Rhythm in epilepsy. Lancet 1938; 232:409 – 416.
40. Wilks DS. Statistical methods in the atmospheric sciences, vol 100. Cam-
bridge MA USA: Academic Press; 2011.
41. Suffczynski P, da Silva FHL, Parra J, et al. Dynamics of epileptic phenomena
determined from statistics of ictal transitions. IEEE Trans Biomed Eng 2006;
53:524– 532.
42. da Silva FHL, Blanes W, Kalitzin SN, et al. Dynamical diseases of brain
systems: different routes to epileptic seizures. IEEE Trans Biomed Eng 2003;
50:540– 548.
43. Suffczynski P, Kalitzin S, Da Silva FL. Dynamics of nonconvulsive epileptic
phenomena modeled by a bistable neuronal network. Neuroscience 2004;
126:467– 484.
44. Jirsa VK, Stacey WC, Quilichini PP, et al. On the nature of seizure dynamics.
Brain 2014; 137:2210– 2230.
45. Benjamin O, Fitzgerald TH, Ashwin P, et al. A phenomenological model of
seizure initiation suggests network structure may explain seizure frequency in
idiopathic generalised epilepsy. J Math Neurosci 2012; 2:1.
46. Freestone DR, Kuhlmann L, Grayden DB, et al. Electrical probing of cortical
excitability in patients with epilepsy. Epilepsy Behav 2011; 22:S110– S118.
47. Kalitzin SN, Velis DN, da Silva FHL. Stimulation-based anticipation and control
of state transitions in the epileptic brain. Epilepsy Behav 2010; 17:310 –323.
48. Kalitzin S, Parra J, Velis DN, Da Silva FL. Enhancement of phase clustering in
the EEG/MEG gamma frequency band anticipates transitions to paroxysmal
epileptiform activity in epileptic patients with known visual sensitivity. IEEE
Trans Biomed Eng 2002; 49:1279– 1286.
49. Dai L, Vorselen D, Korolev KS, Gore J. Generic indicators for loss of resilience
before a tipping point leading to population collapse. Science 2012;
336:1175– 1177.
50. Scheffer M, Bascompte J, Brock WA, et al. Early-warning signals for critical
transitions. Nature 2009; 461:53– 59.
51. Negahbani E, Steyn-Ross DA, Steyn-Ross ML, et al. Noise-induced precur-
sors of state transitions in the stochastic Wilson–Cowan model. J Math
Neurosci 2015; 5:1.
52. Medeiros DC, Oliveira LB, Moura
˜o FAG, et al. Temporal rearrangement of
preictal PTZ induced spike discharges by low frequency electrical stimulation
to the amygdaloid complex. Brain Stimul 2014; 7:170 –178.
53. Karoly PJ, Freestone DR, Woldman W, et al. Epileptic seizures are preceded
by early warning signs of critical transitions. The 38th Annual International
Conference of the IEEE Engineering in Medicine and Biology Society.
Orlando, FA: IEEE; 2016.
54. Schulze-Bonhage A. 4 An Introduction to Epileptiform Activities and Seizure
Patterns Obtained by Scalp and Invasive EEG Recordings. Epilepsy: The
Intersection of Neurosciences, Biology, Mathematics, Engineering, and Phy-
sics. FL USA: CRC Press Boca Raton; 2016; 51.
55. Matsumoto H, Marsan CA. Cortical cellular phenomena in experimental
epilepsy: interictal manifestations. Exp Neurol 1964; 9:286 –304.
56. Engel J, Pitka
¨nen A, Loeb JA, et al. Epilepsy biomarkers. Epilepsia 2013;
54:61– 69.
57. Pitka
¨nen A, Lo
¨scher W, Vezzani A, et al. Advances in the development of
biomarkers for epilepsy. Lancet Neurol 2016; 15:843 –856.
58. Worrell G, Gotman J. High-frequency oscillations and other electrophysiolo-
gical biomarkers of epilepsy: clinical studies. Biomarkers 2011; 5:557 –566.
59. Staba RJ, Stead M, Worrell GA. Electrophysiological biomarkers of epilepsy.
Neurotherapeutics 2014; 11:334– 346.
60. de Curtis M, Avanzini G. Interictal spikes in focal epileptogenesis. Prog
Neurobiol 2001; 63:541– 567.
61. Avoli M, Biagini G, De Curtis M. Do interictal spikes sustain seizures and
epileptogenesis? Epilepsy Curr 2006; 6:203–207.
62. Gotman J. Relationships between interictal spiking and seizures: human and
experimental evidence. Can J Neurol Sci 1991; 18:573 –576.
63. Staley KJ, White A, Dudek FE. Interictal spikes: harbingers or causes of
epilepsy? Neurosci Lett 2011; 497:247–250.
64. Da Silva AM, Aarts J, Binnie C, et al. The circadian distribution of interictal
epileptiform EEG activity. Electroencephalogr Clin Neurophysiol 1984;
58:1– 13.
65. Haut SR. Seizure clustering. Epilepsy Behav 2006; 8:50 –55.
66. Haut SR. Seizure clusters: characteristics and treatment. Curr Opin Neurol
2015; 28:143– 150.
67. Karoly PJ, Nurse E, Freestone D, et al. Bursts of seizures in long-term
recordings of human focal epilepsy. Epilepsia 2016. (in press).
68. Cook MJ, Varsavsky A, Himes D, et al. Long memory processes are revealed in
the dynamics of the epileptic brain. Epilepsy 2014; 5:168.
69. Osorio I, Frei MG, Sornette D, Milton J. Pharmaco-resistant seizures: self-
triggering capacity, scale-free properties and predictability? Eur J Neurosci
2009; 30:1554– 1558.
70. Loddenkemper T, Lockley SW, Kaleyias J, Kothare SV. Chronobiology of
epilepsy: diagnostic and therapeutic implications of chrono-epileptology. J
Clin Neurophysiol 2011; 28:146– 153.
71. Quigg M. Circadian rhythms: interactions with seizures and epilepsy. Epilepsy
Res 2000; 42:43– 55.
72. Taubøll E, Lundervold A, Gjerstad L. Temporal distribution of seizures in
epilepsy. Epilepsy Res 1991; 8:153–165.
73. Cook MJ, Karoly PJ, Freestone DR, et al. Human focal seizures are char-
acterized by populations of fixed duration and interval. Epilepsia 2016;
57:359– 368.
74. Nelson TS, Suhr CL, Lai A, et al. Seizure severity and duration in the cortical
stimulation model of experimental epilepsy in rats: a longitudinal study.
Epilepsy Res 2010; 89:261– 270.
75. Bergey GK, Morrell MJ, Mizrahi EM, et al. Long-term treatment with responsive
brain stimulation in adults with refractory partial seizures. Neurology 2015;
84:810– 817.
76. Brinkmann BH, Wagenaar J, Abbot D, et al. Crowdsourcing reproducible seizure
forecasting in human and canine epilepsy. Brain 2016; 139: 1713–1722.
77. Schalk G, McFarland DJ, Hinterberger T, et al. BCI2000: a general-purpose
brain-computer interface (BCI) system. IEEE Trans Biomed Eng 2004;
51:1034– 1043.
78. Goldenholz DM, Kuhn A, Austermuehle A, et al. Long-term monitoring of cardior-
espiratory patterns in drug-resistant epilepsy. Epilepsia 2017; 58:77 –84.
79. Freestone DR, Karoly PJ, Peterson AD, et al. Seizure prediction: science
fiction or soon to become reality? Curr Neurol Neurosci Rep 2015; 15:1– 9.
80. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015; 521:436 – 444.
81. Gelly S, Silver D. Achieving master level play in 9 9 computer go. AAAI 23rd
Conference on Artificial Intelligence. Chicago, IL; 2008; 1537-1540.
82. Nurse E, Mashford BS, Yepes AJ, Kiral-Kornek I, et al. Decoding EEG and LFP
signals using deep learning: heading TrueNorth. In Proceedings of the ACM
International Conference on Computing Frontiers: ACM: 2016:259 – 266.
83. Nurse ES, Karoly PJ, Grayden DB, Freestone DR. A generalizable brain-
computer interface (bci) using machine learning for feature discovery. PLoS
One 2015; 10:e0131328.
84. Merolla PA, Arthur JV, Alvarez-Icaza R, et al. A million spiking-neuron inte-
grated circuit with a scalable communication network and interface. Science
2014; 345:668– 673.
A forward-lo oking review of seizure prediction Freestone et al.
1350-7540 Copyright ß2017 Wolters Kluwer Health, Inc. All rights reserved. www.co-neurology.com 7
Copyright © 2017 Wolters Kluwer Health, Inc. Unauthorized reproduction of this article is prohibited.