22 Predicting Epileptic Seizures—An Update
347
electroencephalogram (EEG), single neuron activities and local field potentials, blood
oxygenation), followed by/or together with heart activities (e.g., electrocardiogram,
heart rate), body movements (e.g., muscle activities) or behavioral symptoms. Given
only fragmentary knowledge about why, when and how the human epileptic brain
transits from apparently normal dynamical regimes to a preseizure and eventually to
a seizure state, time series analysis techniques should be sensitive enough to identify
even subtle modifications in the dynamics of observables recorded during normal
conditions and similarly account for the large inter- and intraindividual variability.
For brain dynamics, a further obstacle arises from its scale-free behavior coexisting
with an oscillatory one [87].
The spatial-temporal onset of a seizure is usually defined on the EEG (so called
electrographic onset as opposed to the clinical onset, which either coincides with or
follows the electrographic onset). One should be aware that a time-series-analysisbased identification of modifications in the dynamics of observables in the order of
a few to a few tens of seconds prior to seizure onset merely reflects a seizure-onset
detection but not an identification of a preseizure state, taking into account the known
uncertainties in reliably and reproducibly defining seizure onset [49, 145, 149, 161].
Time series analysis techniques that are applied to identify a preseizure state
using the EEG or other modalities (for an overview, see [26, 32, 45, 77, 106, 123,
156]) can be divided into three main categories: univariate, bivariate and multivariate
techniques, depending upon whether data from a single system or a single recording
site are analyzed independently or whether data from two or more systems or sites
are analyzed for possible interactions. These approaches can further be subdivided
into the categories linear and nonlinear.
Univariate linear analysis techniques [115] allow one to draw inferences about
preseizure-state-associated alterations of amplitude-, interval-, or period-distributions
along with their statistical moments, of properties of the auto-correlation function
or of power spectral estimates [14, 19, 44, 60, 117]. Univariate nonlinear analysis
techniques allow a more detailed characterization of the dynamics in state space [62],
particularly when used in conjunction with surrogate-based tests for nonlinearity [80,
141]. Quantities such as an effective correlation dimension [86], correlation density [98], entropy-related measures [90, 157], Lyapunov exponents [55], or quantities
based on recurrence quantification analysis [99, 114] allow one to draw inferences
about preseizure-state-associated alterations of the number of degrees of freedom,
the amount of order/disorder, or of the degree of chaoticity or predictability in a single time series. Other univariate nonlinear techniques aim at discriminating between
deterministic and stochastic dynamics [8] or stationary and non-stationary dynamics [22, 128–130] to identify a preseizure state.
Bivariate analysis techniques allow investigating (linear/nonlinear) relationships
between two (linear/nonlinear) systems (e.g., two brain regions or two organs) and
aim at characterizing strength and direction of an interaction [24, 118, 119]. Common bivariate linear approaches comprise estimating the linear correlation coefficient, cross-correlation or cross-spectral functions or (linear) partial coherence [33,
166]. These techniques, however, can mostly provide information about the strength
of an interaction since correlation does not imply causation. As with univariate linear
347
electroencephalogram (EEG), single neuron activities and local field potentials, blood
oxygenation), followed by/or together with heart activities (e.g., electrocardiogram,
heart rate), body movements (e.g., muscle activities) or behavioral symptoms. Given
only fragmentary knowledge about why, when and how the human epileptic brain
transits from apparently normal dynamical regimes to a preseizure and eventually to
a seizure state, time series analysis techniques should be sensitive enough to identify
even subtle modifications in the dynamics of observables recorded during normal
conditions and similarly account for the large inter- and intraindividual variability.
For brain dynamics, a further obstacle arises from its scale-free behavior coexisting
with an oscillatory one [87].
The spatial-temporal onset of a seizure is usually defined on the EEG (so called
electrographic onset as opposed to the clinical onset, which either coincides with or
follows the electrographic onset). One should be aware that a time-series-analysisbased identification of modifications in the dynamics of observables in the order of
a few to a few tens of seconds prior to seizure onset merely reflects a seizure-onset
detection but not an identification of a preseizure state, taking into account the known
uncertainties in reliably and reproducibly defining seizure onset [49, 145, 149, 161].
Time series analysis techniques that are applied to identify a preseizure state
using the EEG or other modalities (for an overview, see [26, 32, 45, 77, 106, 123,
156]) can be divided into three main categories: univariate, bivariate and multivariate
techniques, depending upon whether data from a single system or a single recording
site are analyzed independently or whether data from two or more systems or sites
are analyzed for possible interactions. These approaches can further be subdivided
into the categories linear and nonlinear.
Univariate linear analysis techniques [115] allow one to draw inferences about
preseizure-state-associated alterations of amplitude-, interval-, or period-distributions
along with their statistical moments, of properties of the auto-correlation function
or of power spectral estimates [14, 19, 44, 60, 117]. Univariate nonlinear analysis
techniques allow a more detailed characterization of the dynamics in state space [62],
particularly when used in conjunction with surrogate-based tests for nonlinearity [80,
141]. Quantities such as an effective correlation dimension [86], correlation density [98], entropy-related measures [90, 157], Lyapunov exponents [55], or quantities
based on recurrence quantification analysis [99, 114] allow one to draw inferences
about preseizure-state-associated alterations of the number of degrees of freedom,
the amount of order/disorder, or of the degree of chaoticity or predictability in a single time series. Other univariate nonlinear techniques aim at discriminating between
deterministic and stochastic dynamics [8] or stationary and non-stationary dynamics [22, 128–130] to identify a preseizure state.
Bivariate analysis techniques allow investigating (linear/nonlinear) relationships
between two (linear/nonlinear) systems (e.g., two brain regions or two organs) and
aim at characterizing strength and direction of an interaction [24, 118, 119]. Common bivariate linear approaches comprise estimating the linear correlation coefficient, cross-correlation or cross-spectral functions or (linear) partial coherence [33,
166]. These techniques, however, can mostly provide information about the strength
of an interaction since correlation does not imply causation. As with univariate linear