22 Predicting Epileptic Seizures—An Update
351
22.5 Open Questions and Outlook
The field of seizure prediction has made remarkable progress over the last three
decades. Proceeding from preliminary descriptions of preseizure phenomena, the
formulation of guidelines, standardizations, build-up of large databases, prediction
contests, and developments of novel analysis concepts together with rigorous statistical tests for performance evaluation resulted in seizure prediction algorithms suitable
for clinical applications. These innovations led to the first in-man proof-of-concept
study of an implantable seizure prediction system. It can be expected that further
miniaturization of electronics, development of connected devices and advances in
engineering of neural systems will continue in the coming years; seizure prediction
is thus not unthinkable anymore [77, 150]. The field has also impacted on other
advances such as single neuron recording techniques in humans [34], brain control
techniques based on neuromodulation [4], and the paradigm shift from the concept
of an epileptic focus to a network theory of seizure generation [77]. Despite the many
achievements, there are important issues that would need to be addressed in the near
future.
Most of the time series analysis techniques presented in Sect. 22.2 allow an interpretation of findings in terms of a physiologic and/or pathophysiologic correlate.
They are thus important to elucidate mechanisms underlying seizure dynamics and
its interactions with normal physiology as well as to derive generic models. Findings achieved so far suggest that physiological understanding of the preseizure state
must be improved to determine whether there are universal mechanisms that lead to
the variety of observed preseizure states. A better understanding of the underlying
mechanisms will also improve understanding of seizure generation, will allow better
translation of the information collected into methods for the detection, prediction,
and control of seizures, and eventually will allow a better understanding of what
a seizure is. Current artificial-intelligence-based approaches, though commendable
for their pragmatism, do not readily reveal which physiological aspects underlie the
predictive characteristics of EEG or other modalities and have resulted in a limited
understanding of the mechanistic underpinnings of the preseizure state. A meaningful combining of data from different modalities requires identification and matching
of the relevant though vastly different spatial and temporal scales.
There are similar challenges for computational modeling approaches [97, 163].
The information gained from multi-scale and multi-modal studies of epilepsy results
in increasingly sophisticated modeling approaches that are used to gain insights into
possible mechanisms of seizure generation and of controlling seizures. Models are
important for identifying changes in network or other control parameters or of inputs
(environmental and/or endogenous) that might not always be evident or accessible
to an observer but are associated with or cause the initiation of seizures. Moreover,
models allow to test in silico hypotheses concerning preseizure brain dynamics and
their relation to endogenous and exogenous parameters as well as the effectiveness of
control strategies [96]. Nevertheless, we are still lacking appropriate tools to bridge
351
22.5 Open Questions and Outlook
The field of seizure prediction has made remarkable progress over the last three
decades. Proceeding from preliminary descriptions of preseizure phenomena, the
formulation of guidelines, standardizations, build-up of large databases, prediction
contests, and developments of novel analysis concepts together with rigorous statistical tests for performance evaluation resulted in seizure prediction algorithms suitable
for clinical applications. These innovations led to the first in-man proof-of-concept
study of an implantable seizure prediction system. It can be expected that further
miniaturization of electronics, development of connected devices and advances in
engineering of neural systems will continue in the coming years; seizure prediction
is thus not unthinkable anymore [77, 150]. The field has also impacted on other
advances such as single neuron recording techniques in humans [34], brain control
techniques based on neuromodulation [4], and the paradigm shift from the concept
of an epileptic focus to a network theory of seizure generation [77]. Despite the many
achievements, there are important issues that would need to be addressed in the near
future.
Most of the time series analysis techniques presented in Sect. 22.2 allow an interpretation of findings in terms of a physiologic and/or pathophysiologic correlate.
They are thus important to elucidate mechanisms underlying seizure dynamics and
its interactions with normal physiology as well as to derive generic models. Findings achieved so far suggest that physiological understanding of the preseizure state
must be improved to determine whether there are universal mechanisms that lead to
the variety of observed preseizure states. A better understanding of the underlying
mechanisms will also improve understanding of seizure generation, will allow better
translation of the information collected into methods for the detection, prediction,
and control of seizures, and eventually will allow a better understanding of what
a seizure is. Current artificial-intelligence-based approaches, though commendable
for their pragmatism, do not readily reveal which physiological aspects underlie the
predictive characteristics of EEG or other modalities and have resulted in a limited
understanding of the mechanistic underpinnings of the preseizure state. A meaningful combining of data from different modalities requires identification and matching
of the relevant though vastly different spatial and temporal scales.
There are similar challenges for computational modeling approaches [97, 163].
The information gained from multi-scale and multi-modal studies of epilepsy results
in increasingly sophisticated modeling approaches that are used to gain insights into
possible mechanisms of seizure generation and of controlling seizures. Models are
important for identifying changes in network or other control parameters or of inputs
(environmental and/or endogenous) that might not always be evident or accessible
to an observer but are associated with or cause the initiation of seizures. Moreover,
models allow to test in silico hypotheses concerning preseizure brain dynamics and
their relation to endogenous and exogenous parameters as well as the effectiveness of
control strategies [96]. Nevertheless, we are still lacking appropriate tools to bridge