Epilepsy: A Comprehensive Textbook
2nd Edition

Seizure prediction, anticipation, or forecasting (despite their different meanings, these terms are currently used interchangeably) is a field of great interest in the clinical and basic neuroscience communities. This is not only because of its potential clinical application in warning and therapeutic antiepileptic devices, but also for its promise to increase our understanding of the mechanisms underlying epilepsy and seizure generation. The motivation for research into the predictability of seizures is straightforward. The fact that seizures occur without warning in the majority of cases is one of the most disabling aspects of epilepsy. If it were possible to predict seizures with high sensitivity and specificity, even seconds before their onset, therapeutic possibilities would change dramatically.⁵⁰ One might envision a simple warning system capable of decreasing both the risk of injury and the feeling of helplessness that results from seemingly unpredictable seizures. Side effects from treatment with antiepileptic drugs, such as sedation and clouded thinking, could be reduced by on-demand release of a short-acting drug^48,197 or electrical stimulation^62,175,213 during the pre-ictal state. Paired with other suitable interventions, such as focal cooling⁷⁰ or biofeedback operant conditioning,^56,170,198 such applications could reduce morbidity and mortality, and greatly improve the quality-of-life for people with epilepsy. In addition, identifying a pre-ictal state would greatly contribute to our understanding of the pathophysiologic mechanisms that generate seizures.

In most patients seizures appear to occur unpredictably, with no discernible pattern, while in others, seizures appear to be entrained to biologic rhythms, such as menstrual or sleep–wake cycles. Clustering patterns, where one seizure appears to increase the likelihood of subsequent seizures, also appear common in clinical practice. Despite these observations, analyses of long-term patterns based upon seizure diaries^{4,11,16,74,115,147,204} have yielded inconsistent findings. While some authors conclude that the timing of seizure recurrence is random, others hypothesize that seizures occur in a probabilistic nonlinear fashion. Because of this inconsistency, based upon clinical observations, the transition to a seizure has generally been thought of as an abrupt phenomenon, occurring without warning. Nevertheless, there is physiologic support for the idea that at least certain types of seizures are predictable.

Several seizure-facilitating factors are known. Lennox¹²⁵ defined seizure facilitation as the input of sensory, metabolic, emotional, or other yet unknown factors that “fill up some reservoir until it overflows,” which in turn results in a seizure. State of consciousness, sleep deprivation, being tense, disturbances of electrolytes and acid–base balance, sensory stimulation, and exposure to certain drugs are factors known to potentiate seizures. Apart from the rare exception of sensory-evoked or reflex epilepsies, however, these factors are rather nonspecific and highly variable, since they depend on individual habits, susceptibility, and daily routine.

Clinicians who care for patients with epilepsy have long known that individual patients can identify periods when seizures are more likely to occur, though they can rarely specify an exact time when seizures will happen. Rajna et al.¹⁷⁴ found that the vague sensations that characterize these periods, called “clinical prodromes,” occurred in more than 50% of 562 investigated patients, though the reliability of these reports was not evaluated prospectively. Reported sensations included mood changes, irritability, sleep problems, nausea, and headache. There are also physiologic studies in small numbers of patients, usually collected serendipitously before seizures, that support the existence of a pre-ictal period. Weinand et al.²¹⁶ detected a significant increase in blood flow in the epileptic temporal lobe that started 10 minutes before seizure onset that spread to both temporal lobes 2 minutes before seizure onset. Similarly, Baumgartner et al.¹⁴ demonstrated increased blood flow in the epileptic temporal lobe in two patients, 11 and 12 minutes, respectively, before seizure onset. Using near-infrared spectroscopy in three patients, Adelson et al.² reported an increase in cerebral oxygen availability that began more than 13.5 hours, and was identified as early as 1.5 hours, before documented seizure onset. Pre-ictal changes in other variables, such as R–R interval on the electrocardiogram (ECG)^40,95,159 may also have predictive value, perhaps as epiphenomena related to seizure precursors, in some types of epilepsy. More recently, functional magnetic resonance imaging has demonstrated changes in perfusion prior to seizure onset.⁵⁵

During recent years, a variety of potential ictogenic (seizure-generating) mechanisms have been identified in experimental models of focal epilepsy, including alterations in synaptic and cellular plasticity and changes in the extracellular milieu (see Section II). However, it is still a matter of debate whether these mechanisms can be regarded as specifically ictogenic, apart from their critical role in normal brain function. On the level of neuronal networks, focal seizures are assumed to be initiated by abnormally discharging neurons (so-called bursters^{21,22,34,182,208}; see reference 226 for an overview) that recruit and entrain neighboring neurons into a critical mass. This build-up might be mediated by an increasing synchronization of neuronal activity that is accompanied by a loss of inhibition, or by processes that facilitate seizures by lowering the threshold for excitation or synchronization. In this context the term “critical mass” might be misleading in the sense that it implies an increasing number of neurons that are entrained into an abnormal firing pattern. This mass phenomenon would be easily accessible to conventional electroencephalogram (EEG) analysis, which, to date, has failed to detect it. Rather, the seizure-initiating process might better be visualized as a process in which an increasing number of critical interactions between neurons in a focal region and connected units in an abnormal functional network unfold over time. Indeed, there is now converging evidence from different laboratories that quantitative analyses appears to be capable of characterizing this collective neuronal behavior from the gross EEG, allowing definition of a transitional pre-ictal phase, in a high percentage of cases.

As early as 1975, researchers considered analysis techniques such as pattern recognition, analytic procedures of spectral data,^44,189,214 or autoregressive modeling of EEG data^179,181 for the prediction of seizures. Their findings indicated that EEG changes characteristic for a pre-ictal state may be detectable a few seconds before their actual seizure onset on EEG. None of these mostly linear techniques has been implemented clinically.

Since spikes in the EEG are usually considered the hallmark of an epileptic brain, their possibly altered pre-ictal occurrence was investigated in several studies. Sherwin¹⁸⁸ noted increased correlation of epileptiform activity (spikes) between two adjacent cortical sites in the 15 to 20 minutes prior to EEG onset of focal seizures in a cat model of epilepsy. In 1983, Lange et al.¹⁰⁷ demonstrated a similar correlation of interictal spikes between the side of the seizure focus and the “normal” temporal lobe within the 20 minutes prior to EEG seizure onset in patients with temporal lobe epilepsy. While other authors reported a decrease or even total cessation of spikes before seizures,^63,64,221 re-examination did not confirm this phenomenon in a larger test set.⁹⁴

With the advent of the physical-mathematical theory of nonlinear dynamics (colloquially termed chaos theory) in the early 1980s, new analysis techniques were developed to characterize apparently irregular behavior—a distinctive feature of the EEG—and thus to extract features from the EEG that are not obvious to the human eye (see references 12, 45, 84, 118 for an overview). During the last two decades, these techniques have generated a large body of evidence for the existence of a pre-ictal state. The earliest attempts to use nonlinear time series analysis (see reference 91 for an overview) were started in the 1990s using the “largest Lyapunov exponent” to describe changes in brain dynamics.^78,79,83 The investigators observed transient drops in the temporal evolution of this measure several minutes prior to seizures and proposed that the EEG became progressively less chaotic as seizures approached. The first studies to describe characteristic changes in the EEG shortly before an impending seizure in a larger group of patients used the “correlation dimension” as an estimate for neuronal complexity^{51,53,117,119,121,124} and the “correlation density.”¹⁴¹ These studies were followed by others using measures such as dynamical similarity index,^{108,111,112,113,157} Kolmogorov entropy,^153,210,211 or marginal predictability.^43,126,127 In parallel, other techniques have focused on extracting neurophysiologic features from the EEG associated with epileptiform activity in human and animal physiology, such as bursts of complex epileptiform activity, slowing, chirps, and changes in signal energy.^{60,133,158,225} Other methods focused on defining pre-ictal states include catastrophe theory,^25,26 self-organized criticality,^128,224 recurrent neural networks,¹⁶⁷ and simulated neuronal cell models.¹⁸⁶ Similar to the studies using the largest Lyapunov exponent, all of the above studies showed characteristic changes minutes to hours prior to seizure onset on the EEG, and were interpreted by their authors as defining pre-ictal states of various durations, some lasting hours.

A problem with most of these studies is that the measures used to characterize the EEG are difficult to interpret in terms of their physiologic correlate. Also, since almost all of these measures are univariate (i.e., related to only a single recording site), they fail to reflect any interactions between different regions of the brain. The epileptogenic process, on the other hand, is commonly accepted to be closely associated with changes in neuronal synchronization in a network of components that may be spatially distributed. The analysis of synchronization in the EEG can therefore a priori be regarded as a promising approach for the investigation of the spatiotemporal dynamics of ictogenesis. Based on newly developed physical-mathematical concepts for synchronization (see reference 169 for an overview), some researchers have focused on bivariate or, more generally, multivariate measures over the last 5 to 6 years that permit assessment of synchronous activity from multiple sites.^61,212 These measures include nonlinear interdependence,^9,110 measures for phase synchronization and cross-correlation,^{28,76,109,149,150,152} the difference of the largest Lyapunov exponents of two or more channels,^27,75,80,82 nonlinear causality,²⁹ a classification approach based on a fusion of multiple EEG features from multiple sites.^38,133

Results obtained indicate that seizures are not random events, but rather are related to ongoing dynamical processes that may begin minutes to hours to days beforehand (for an overview, see references 73, 132, 134, 192, 193). The fact that most of the approaches result in different prediction horizons indicates that they may reflect different aspects of ictogenesis, but it is likely that none of these techniques appears to depict the process fully. As many of these studies suggest, seizure precursors may wax and wane in attempts to ignite a clinical event, but the forces both driving and suppressing seizure generation remain hidden. Other concepts abstracted from the above body of work indicate that seizure precursors may begin locally and then expand spatially, and even “entrain” other brain structures before reaching the critical mass required to initiate a clinical seizure. Patterns appear to be patient specific, within a finite range of pattern types, and it appears that different approaches may be required to predict seizures with clinically useful accuracy in different individuals or in different epilepsy syndromes. This may be a function of individual physiology or potentially confounding variables such as electrode placement and the amount and speed of medication taper during inpatient video-EEG monitoring.^132,134

At the beginning of the new millennium there was great enthusiasm for the ability of a variety of analysis methods to define the pre-ictal state. By that time work in the area had also extended to scalp EEG,^{43,71,77,110,173} though the majority of researchers confined their investigations to intracranial EEG recordings. Careful review of the literature at that time, however, revealed considerable contradiction in results from different research groups. Of even more concern was that despite over a decade of excellent work in the field, convincing evidence demonstrating unequivocal seizure prediction in blinded, prospective, randomized clinical trials, with appropriate statistical validation, remained elusive. Central to the problem was the challenge of developing algorithms to detect unknown patterns associated with seizure generation, a process that remains poorly understood. Much of the EEG data analyzed in studies up to that time were highly selected and restricted with regard to seizure type, patient state, signal:noise ratio, duration of recordings, artefacts, etc. In addition, there were no standardized methods or nomenclature for marking continuous EEG data, no accepted methods for assessing algorithm performance, and no agreement on acceptable test data. Even clear definitions of exactly what constitutes seizure onset, seizure prediction, anticipation, and the definition of ictal events either clinically or by EEG were nebulous. For these reasons, beginning with an impromptu meeting at the American Epilepsy Society Meeting in Los Angeles, California, in 2000, the International Seizure Prediction Group (ISPG) was formed to provide an informal
structure for the major groups working in this area to share data and ideas.

The ISPG was established with the specific goal of moving the field of seizure prediction forward from “proof of principle” experiments into validated, well-understood methods that could be applied to both basic science and clinical applications. The first international workshop of this group was held in Bonn, Germany, in 2002, funded by grants from the German Section of the International League Against Epilepsy, the German Section of the International Federation of Clinical Neurophysiology, and the American Epilepsy Society. At the core of the workshop was an assessment of the state of the field at that time by having each major group apply its methods to predict seizures from a shared set of continuous intracranial EEG data.¹²² Findings obtained from applying a large number of analysis techniques are summarized in eight peer-reviewed articles published together in the journal Clinical Neurophysiology.^{38,54,66,81,86,88,114,151} Although substantial efforts were made to provide uniform data in terms of disease type, conditions, and recordings, the results of all these investigations were inconsistent and at times contradictory. Three studies had positive results, predicting seizures for different time horizons; four studies had negative results; and one had both, depending upon which techniques were employed. In none of the investigations, even those with positive findings, could seizures be predicted with any exact timing. Rather, a state of increased seizure likelihood lasting up to several hours was identified. There was agreement that, at present, none of the EEG analysis techniques was sufficient for broad clinical application, and that there were major practical problems to overcome. Nevertheless, much was learned from the exercise, particularly with regard to the need for standardization of analyses, data requirements, performance criteria, and nomenclature. Some of the results were encouraging, while other results illustrated that certain approaches are unlikely to be worthwhile. The current impact of the latter point is stressed by recent controversies about the relevance of nonlinear approaches for the prediction of epileptic seizures^{140,144,145,150,151} and by studies raising doubts about the reproducibility of previously reported claims.^{10,39,67,105,106,123} These contradictory findings emphasize the need for reliable methods for evaluating the performance of seizure prediction techniques.

Over the last three decades two main categories of analysis techniques have been used to extract pre-ictal information from the EEG: Linear and nonlinear techniques. Depending upon whether EEG data from two or more sites are analyzed independently, or for possible interactions, these techniques can further be divided into univariate and multivariate approaches. All techniques permit reduction of large amounts of EEG data to a small number of parameters for downstream processing.

Linear EEG analysis techniques (see reference 135 for an overview) are important contributors to understanding physiologic and pathophysiologic conditions in the brain. Nonparametric linear methods comprise analysis techniques such as evaluation of amplitude, interval or period distributions, and estimation of auto- and cross-correlation functions, as well as analysis in the frequency domain (using the Fast Fourier Transform or other time-frequency transformations) such as power spectral estimates, cross-spectral functions, or linear coherence.²³ Parametric linear methods include, among others, autoregressive (AR) and autoregressive moving average (ARMA) models,^57,87,88 and provide an alternative way to estimate properties of the power spectrum. These main branches are accompanied by pattern recognition methods involving either a mixture of techniques mentioned above or, more recently, taking features extracted from neurosignals and inputting them into a variety of novel classifiers, such as probabilistic artificial neural networks. Since linear methods provide only limited information as to the dynamical aspects of the EEG, it is argued that they cannot fully characterize the complicated, apparently irregular behavior of the complex nonlinear dynamical system brain. In this system, nonlinearity is introduced already on the cellular level, since the dynamical behavior of individual neurons is governed by integration, threshold, and saturation phenomena. There is evidence, that the epileptic process enhances the nonlinear deterministic structure in the EEG.^6,8,24 In order to allow for an improved characterization of complex dynamics, nonlinear analysis techniques have been developed that provide a methodologically different approach to EEG analysis. Within this framework, the dynamical behavior is embedded in a so-called state space. This generally high-dimensional cartesian space is spanned by all state variables (i.e., the number of degrees of freedom) of a system, and the system dynamics generate a trajectory through this space. Properties of the trajectory in state space can then quantitatively be characterized by nonlinear measures (see below). When embedding EEG time series, the number of state variables (i.e., the number of degrees of freedom of the system brain) are unknown. Fortunately, the theorem of Takens²⁰² allows to reconstruct a so-called equivalent state space even from a single time series using the so-called method of delays (time-delay embedding) (see also references 164 and 183). Here the basic assumption is that a single but long enough and accurate measurement of a stationary dynamics is sufficient to capture all the relevant system properties necessary to reconstruct the state space. In terms of EEG analysis, one may assume that an EEG signal reflects the influence of the multiple variables participating in brain dynamics.^49,143 The reconstruction of an m-dimensional state space (here m is the so-called embedding dimension) requires the generation of m time-delayed versions of an EEG time series; that is, each version consists of successive points of the original time series separated by a fixed time delay (τ). A variety of techniques have been proposed that allow one to estimate either m or τ from a measured time series, assuming, however, that the other parameter has been chosen appropriately beforehand. Because of this mutual dependence, the time-delay embedding of an EEG time series in a high-dimensional state space is regarded as a crucial point in nonlinear EEG analysis. An improper state space reconstruction is a common source of errors and can lead to a mischaracterization of the dynamics. In case of multichannel EEG recordings, an alternative embedding scheme would be to use each channel as an axis of the cartesian space. In this case the embedding dimension m is fixed and equals the number of recording channels. Although this spatial embedding is regarded to be the more natural scheme, several assumptions have to be made beforehand that might lead to similar problems as with the time-delay embedding and are matter of debate.^103,172 These problems include the optimal distance between different recording sites, which is usually fixed, among others. It remains to be established whether the combined use of techniques (so-called spatial-temporal embedding¹⁴¹) can be regarded as more appropriate for EEG analysis.

In order to characterize dynamics in state space, a number of univariate and bivariate approaches are available. Quantities such as an effective correlation dimension, correlation density, entropy-related measures, or Lyapunov exponents allow one to draw inferences about the number of degrees of freedom (or complexity), the amount of order/disorder, or the degree of chaoticity or predictability in a single time series. Other univariate measures aim at discriminating between deterministic
and stochastic dynamics⁹² or provide an estimate of the amount of nonstationarity.^{36,176,177,178}

Bivariate measures such as the similarity index,¹¹¹ phase synchronization,^{104,109,152,203} nonlinear interdepen- dency,^9,110,138 and other measures for generalized synchronization allow one to estimate dynamic interactions between two time series; some of these approaches can even provide information about the direction of interdependence.^9,29,180,190 A full description of the physical and mathematical concepts underlying these quantities along with their implementation details for EEG analysis is beyond the scope of this chapter. Instead, we summarize in Table 1 the various measures, concepts, and achievements of univariate and bivariate approaches used for seizure prediction.

Further details can be found in a number of tutorials, conference proceedings, or text books on nonlinear time series analysis.^{1,45,49,52,58,65,69,84,90,91,93,118,134,163,171,187,195}

Despite their great potential for a detection of subtle changes in brain dynamics, both linear and nonlinear analysis techniques must be painstakingly applied, and the results obtained should be interpreted with great care. Many techniques place great demands on the recorded time series with respect to the precision of the data and the absence of noise. Almost all techniques assume the underlying dynamical system from which the recordings were taken to be stationary. None of these requirements, however, can be exactly fulfilled in practice. With respect to nonlinear EEG analysis, numerous studies have identified factors that might alter the absolute value of a measure: Properties of EEG electrodes, the precision of the analog-to-digital converter, amplifier and filter settings, and different recording montages might all strongly influence nonlinear measures. Finally, problems specific to the individual algorithms have to be taken into account. Based on these potential limitations, we feel that it is advisable to avoid claims on the existence of chaotic behavior in the EEG, and instead to use nonlinear measures as tentative indexes of different brain states when analyzing EEG data.

The current state of seizure prediction is that there is strong evidence from several methods for identifiable precursors preceding partial-onset seizures. This conclusion, however, is based on retrospective analyses of mostly intracranial EEG data recorded during presurgical evaluation of patients during evaluation for resective surgery. Up to the time of this writing, no study has been published that demonstrates unequivocal seizure prediction in blinded, prospective, randomized clinical trials. Reasons for this become apparent when considering the major methodologic steps involved in seizure prediction algorithms along with problems posed by respective study design.

Typically, the first step in a retrospective seizure prediction study consists of calculating a certain characterizing measure from the EEG using a moving-window technique. The length of the time window is chosen so that there is a reasonable trade-off between approximate stationarity of the EEG signal and sufficient number of data points to characterize the EEG dynamics. The resulting temporal evolution of the measure (measure profile) is then scanned for prominent characteristics that can be related to the actual seizure times. These features might be
drops or peaks (e.g., quantified as threshold crossings) or any other distinct pattern in the measure profile. In a second step the measure’s capability to distinguish the pre-ictal state from the inter-ictal interval is evaluated with test statistics quantifying the occurrence of these features relative to the seizure times, which results in some kind of performance value. A high performance value is assumed to reflect the existence of a pre-ictal state and the capability of the applied measure to detect it.

The first question that arises from these types of analyses is which of all the characterizing measures presented above is the best, though with their different time horizons and processing methods they constitute different ways of viewing the same process. As some studies have demonstrated, it appears that some combination of measures will probably be required to carry out reliable seizure prediction tailored to individual patients.^38,117 Usually, algorithm performance requires optimizing numerous computational parameter choices and sometimes the choice of an estimation algorithm. Upon close inspection, these choices often rely on a posteriori knowledge, which interposes significant risk of in-sample overtraining. Certainly, what is true for a single measure holds also for a larger number of different measures. The application of a huge variety of measures to the EEG might yield seemingly good results just by chance (particularly on a limited database), if appropriate methods for dealing with statistical issues of multiplicity are not implemented. The resulting explosion in computational degrees of freedom emphasizes the need for control tests and independent validation. This is necessary to avoid the potential for fitting results to the data and to enable independently reproducible results. Many measures have a clear physical meaning since they are derived from mathematical-physical theories that allow one to characterize complex spatiotemporal dynamical systems. However, due to a huge number of influencing factors and constraints associated with analyzing EEG data, strict interpretation of particularly nonlinear measures (e.g., in terms of attractor dimensions, deterministic chaos, or entropy) is rarely justified. Nevertheless, many classical measures can still be used as tentative indexes of different brain states, and more recently developed measures often allow less ambitious but more straightforward interpretation. Ongoing research will clarify whether a meaningful physiologic interpretation in terms of underlying brain dynamics can be achieved.

The next issue concerns the EEG recording. Taking into account that epilepsy is a heterogeneous disorder, what constitutes an optimum spatial and temporal sampling of the ictogenic process? How to select optimum data acquisition parameters such as sampling rate, filter settings, resolution of the analog-to-digital converter, and electrode montage is not straightforward, and may depend on things that are currently unknown, such as the temporal and spatial resolution of physiologic events critical to seizure generation. How many electrodes are necessary? Is a noninvasive recording sufficient or does a good seizure prediction technique require the higher frequency components only available in intracranial recordings? More importantly, what constitutes a good data set for testing sensitivity and specificity (see below) of a seizure prediction algorithm? Many of these more technical issues have not been satisfactorily addressed yet, and there may be a number of reasons for this. Data acquisition parameters like sampling rate and filtering have to follow the demands of the sampling theorem, and have traditionally been based on analog paper-and-pen EEGs and the requirement that they allow for good visual inspection of the EEG. With modern digital equipment, cheaper computer storage, and recent interest in single unit recording, acquisition rates of tens of kilohertz per channel are now possible, but deriving meaning from terabytes of such data is a daunting prospect, particularly when searching for unknown patterns. Traditionally, acquisition parameters are chosen to allow adequate sampling of target waveforms, balanced with the need to keep data storage capacity as small as possible. Almost all EEG data sets that have been studied up to now were recorded with data acquisition parameters set by clinical EEG systems, without specific regard to prediction study requirements. It is well known that these parameters affect virtually all characterizing measures, and it still remains an open issue whether they must be regarded as potentially confounding variables in seizure prediction studies. The placement of EEG electrodes typically follows roughly common protocols, guided by the demands of the presurgical evaluation and limited by the need to protect patients, but there is little standardization from center to center in this regard. Without understanding more about seizure generation in the epileptic network, there is no way to know what might be optimal spatial sampling for seizure prediction studies. Clearly, more sensors would be better, including those placed into deep integrating structures, such as the thalamus, but this type of spatial sampling is currently only possible in animal studies, because of patient safety concerns. Even with the number of channels available from implanted electrodes in humans, the spatial information available for measuring the ictogenic process has been used insufficiently. Most researchers have confined analyses to at most one or several electrodes, relying on a posteriori knowledge as to location and extent of the ictal onset zone. At best, data from sites distant from the ictal-onset zone are included in these studies only for comparison to more “normal” regions. Others have developed optimization schemes that allow one to select certain electrodes out of a large number of electrodes^{27,37,75,76,80,114} relying on a posteriori knowledge as to the dynamics of the ictogenic process. Due to the availability of more powerful computers, it is only recently that EEG data from all electrode sites have entered seizure prediction studies. Interestingly, some studies reported that the site selected as best for prediction was not in close vicinity to the epileptic focus but could be located in remote or even contralateral brain structures.^{38,89,114,150,151,156} This seemingly counterintuitive finding may indicate the importance of brain outside of the ictal-onset zone but within the “epileptic network” in generating clinical seizures (see also references 218 and 219). This is also in accordance with findings showing that the synchronization of specific populations in relation to the epileptic focus may be of crucial importance to determine whether a seizure is likely to occur and to spread.^{114,149,150,151} On the other hand, it may also indicate a rather nonspecific phenomenon whose temporal proximity to seizure onset was just by chance. Obviously, correct site selection would be a practical problem in instances where electrode contacts are limited. At present, it is not clear at all whether optimum recording sites can be identified.

Large amounts of well-documented, continuous, prolonged, multichannel EEG data are very difficult to acquire in a busy clinical environment, and storage of a diverse archive of data is an expensive prospect. Many institutions store short-lasting (typically a few minutes only) inter-ictal and/or peri-ictal epochs for documentation purposes or because of legal requirements for medical records and results. It is therefore not surprising that many past studies lack reference to the interictal state in terms of insufficient control data or baseline epochs. In other words, many studies have focused merely on sensitivity without considering specificity of the applied techniques. Maiwald et al.¹⁴⁰ reviewed 14 seizure prediction studies published between 1998 and 2003 and concluded that in only half of these studies was the performance of the applied seizure prediction technique tested against interictal control data. Proper selection of control data poses another problem, that even during the interictal period the epileptic brain is different from “normal,” and that there may be abnormal dynamical changes that are not necessarily followed by a seizure. Other nonstationary variables such as sleep–wake cycles or different cognitive states must definitely be regarded as potential
confounding variables in seizure prediction studies. In addition, during presurgical monitoring an artificially high seizure frequency (0.15 seizures per hour, or 3.6 seizures per day⁶⁸), seizure clustering, and atypical seizures may occur due to the reduction of anticonvulsive medication. Under normal conditions, patients with pharmacoresistant focal epilepsy have a seizure frequency of about three seizures per month, that is, 0.0042 seizures per hour.¹³ In order to thoroughly evaluate the suitability of possible seizure prediction techniques, reference to continuous, prolonged, multichannel EEG data must be regarded as indispensable.

EEG data accompanying clinical information must be as complete as possible to account for other factors that may modulate seizure generation. Relating mathematical approaches to clinical, video, and neurophysiologic data, a massive undertaking, has begun only recently.^148,156 There are now attempts to identify and mark the broad range of physiologic and pathophysiologic changes that occur in these measures interictally and at seizure onset in a standardized and clear-cut manner. Given imperfect methods for determining exact clinical and EEG onset times for seizures, at present, reference to EEG onset of seizures, as opposed to clinical seizure onset, is preferred.¹²²

The above considerations indicate that seizure prediction studies have to be designed with great care, and with an acute awareness that as we delve deeper into seizure generation we may be confronted with other, yet unforeseen problems. A thorough evaluation of potential predictors and potentially confounding variables requires large, high-quality, meticulously collected and annotated data archives that are well characterized and represent the heterogeneity of patterns and patients found in human epilepsy. Equally important, a review of the literature points to the need for proper statistical validation in order to assess the performance of seizure prediction algorithms. This is a major issue in the field at present, and one that has been neglected by many past studies.

Assessing the predictive performance for a given algorithm usually involves analysis of the so-called receiver operating characteristics (ROC) curve that is based on quantities such as sensitivity and specificity. Sensitivity is defined as the ratio of true positive classifications (or predictions, detections, or warnings) to total number of (true and false) positive classifications. Specificity is defined as the ratio of true negative classifications to total number of negative classifications. The false-positive rate is defined as the average number of false classifications per unit time, and is one accepted measure of specificity. Finally, performance can be defined as the square root of squared sensitivity plus squared specificity. Let us assume that a seizure prediction method requires adjustment of a parameter (e.g., some threshold). With ROC, values of the parameter are continuously varied, and the sensitivity of the classification for these parameter values is plotted against 1 minus the corresponding specificity. The resulting curve is termed the ROC curve, and each point on the curve corresponds to a different parameter value. An ideal classifier has a curve that goes from the bottom left corner to the top left and then from the top left to the top right (two perpendicular lines). An unspecific, random classifier produces a ROC curve that is a diagonal line.

The aforementioned considerations indicate the tight dependence between sensitivity and specificity (a high sensitivity may be achieved at the expense of a poor specificity, and vice versa) and the need to evaluate both quantities together^161,162 in order to achieve a meaningful estimate of the performance of a seizure prediction method. Many past studies have reported on the sensitivity of a given method only, and it appears that the mostly positive results had been achieved by an in-sample overoptimization of parameters and by using insufficient control data. Recently, Winterhalder et al.²²² suggested what they call the seizure prediction characteristic to evaluate seizure prediction methods. The authors proposed to extend the ROC approach (i.e., investigation of sensitivity and false prediction rate) by including additional assessment criteria that are based on clinical and statistical considerations. Since, at present, none of the current prediction methods is able to indicate the exact point in time when a seizure is to occur, the authors suggest considering this uncertainty by referring to the “seizure occurrence period,” defined as the period of time during which a seizure is to be expected. In addition, to render a therapeutic intervention possible, they refer to an important interval as the minimum window of time between an “alarm” raised by the prediction method and the beginning of the seizure occurrence period. They and other investigators refer to this interval as the “seizure prediction horizon.” Due to the interdependence between sensitivity and specificity, and that it may not be possible to avoid false alarms completely, the authors suggest using a maximum false prediction rate as a measure for algorithm performance. This quantity can be derived from the average seizure incidence, which determines sensitivity. Finally, the authors propose to compare the achieved sensitivity values with those obtained from random and periodical prediction methods. Using this framework, the same group shows that the achieved performance values of three previously proposed prediction methods (dynamical similarity index, accumulated energy, and effective correlation dimension) are significantly better than the performance of unspecific methods.¹⁴⁰ They conclude that this finding indicates the existence of specific “predictive” information in pre-ictal epochs and that the investigated methods are sensitive to this information. The resulting seizure prediction characteristics, however, were judged as not yet sufficient for clinical application.

Studies like these are important since they clearly point to deficiencies in current seizure prediction methods. Showing that a method is sensitive enough to detect a pre-ictal phenomenon is only the first step in developing a seizure prediction method. Unfortunately, a high sensitivity does not prove that the method is suitable for clinical applications. Even a highly sophisticated performance estimation, as the one above, is confronted with a number of problems for which there are currently no satisfactory solutions. Many quantities that enter performance evaluation tests require parameter optimization or a choice of parameter ranges. As an example, seizure prediction methods usually require adjustment of some threshold (or reference) level, and a deviation of some characterizing measure from this threshold is assumed to (truly or falsely) indicate a pre-ictal state. The threshold level is usually derived from analyses of control (also called baseline or reference) data that should represent all conditions and states of consciousness from which seizures can emerge (which again calls for large, high-quality, and meticulously annotated data archives). In this context, it should be mentioned that there is now consensus among the major groups in the field that choice of short reference periods, even if done randomly, could exert great influence over results. It is important to note that a properly designed study based upon randomly chosen data epochs could be made statistically sound, if strict statistical criteria are met that dictate a sufficient number of randomly chosen interictal segments. Unfortunately, there would still be concern that all patient states might not be adequately represented. Because of these considerations, recommendations from the First International Collaborative Workshop on Seizure Prediction include that it would be best to avoid the use of short reference periods.¹²²

When calculating a threshold level, the question arises as to whether one should consider it as constant or whether one should account for possible slow drifts in dynamics (e.g., due to changing antiepileptic drug levels during medical tapering,¹²⁰ changes in vigilance states,^54,116,133 or other circadian fluctuations¹⁰¹) and use an adaptive threshold level instead. The latter makes a prospective implementation of an
algorithm for seizure prediction a nontrivial problem. Closely related to appropriately selecting a threshold level is the problem of defining false classifications, which requires a clear-cut delineation of a pre-ictal state from the inter-ictal state. Currently available information indicates that a pre-ictal state might last from minutes to hours to days. An inappropriately selected duration would lead to an increased number of false classifications. More importantly, the model of a pre-ictal state assumed by a prediction algorithm has enormous potential impact upon its performance. For example, the interictal-to-pre-ictal transition is usually assumed to follow some rectangular function, with a sharp boundary between states. If, however, this transition follows another function (e.g., linear or exponential or log-periodic), very early or intermittently occurring precursors would wrongly be classified as false positives. The situation becomes even more complicated when taking into account the spatial distribution of the interictal-to-pre-ictal transition. From their synchronization studies, Mormann et al.¹⁵² concluded that an epileptic seizure might be interpreted as the climax of a process of changes in brain dynamics that starts long before the seizure. In an attempt to relate these dynamic changes to EEG-based physiology, Litt et al.¹³³ suggested a possible evolution of waxing and waning events, such as bursts of complex epileptiform activity, subclinical seizurelike bursts (chirps), and alterations in neuronal activity measured by energy changes that might characterize this process. Unfortunately, we still lack the physiologic or dynamic understanding to delineate a pre-ictal state both in time and space, particularly not from a level that relates to neurophysiology on the cellular and network level. Of interest, a recent study by Kalitzin et al.⁸⁹ indicates that important clues for possible dynamical scenarios that lead to epileptic seizure onsets may be obtained from analyzing the so-called phase demodulation of intracranial EEG recorded interictally during intermittent electrical stimulation.

The above considerations clearly indicate that assessing the performance of a seizure prediction algorithm strongly relies on assumptions of a model of seizure generation, including the spatiotemporal characteristics of the process and, more importantly, on the existence of a pre-ictal state. When reviewing the seizure prediction literature, it is important to note that if a pre-ictal state does not exist (null hypothesis), no information predictive of an impending seizure could be extracted from the EEG, but many algorithms could still perform better than an unspecific, or random, prediction algorithm. This is because any measure tested on recorded data will likely contain fluctuations, generating a nonzero probability of attaining any course within its range of definition. Even though this probability may be very small, the large number of potentially confounding variables discussed above is likely to lead to an increased probability for finding a combination of them, which then leads to a nonzero performance value. The influence of the different degrees of freedom on the statistical significance of the obtained results is difficult, if not impossible, to estimate on a theoretic basis. Hence, it is difficult to decide whether a given performance value indicates the existence of a pre-ictal state or whether it is consistent with the null hypothesis stated above. Although there are statistical methods for addressing these issues, these require careful sampling of distributions of feature values amassed over time, and only with adequate amounts of continuous data allow definition of expected measurement error and requirements for statistically significant results.

In order to address this ambiguity, Andrzejak et al.⁷ proposed an alternative approach to the problem, using the concept of seizure time surrogates. In this method, original seizure onset times are replaced with “surrogate” times randomly chosen from the interictal intervals. Specified properties of the original sequence (such as total number of seizures, distribution of intervals between consecutive seizures, and clustering of seizures) can be imposed as constraints on the surrogate seizure onset times. A seizure prediction algorithm is then applied to the original seizure time sequence and the surrogates. Assuming that a pre-ictal state exists, and that the algorithm is able to detect it, the performance should be higher for the original seizure times than for the surrogate times. Alternatively, Kreuz et al.¹⁰¹ proposed the concept of measure profile surrogates, in which the seizure onset times are kept fixed and instead a constrained randomization of a measure profile is performed using the method of simulated annealing. Using this method, the amplitude distribution of a measure and the autocorrelation function (accounting, for example, for circadian fluctuations) of the measure profile are preserved. As with the seizure time surrogate concept, a seizure prediction algorithm is applied to the original measure profile and the surrogates, and the highest performance is expected for the original measure profile, provided that a pre-ictal state exists and the algorithm is able to detect it. Both methods have conceptual advantages and disadvantages, and the concept of measure profile surrogates has a greater computational burden. Both concepts are based on null hypothesis tests, and the nominal size is determined by the number of surrogates. It is important to keep in mind, however, that the fact that the null hypothesis cannot be rejected does not prove its correctness. Rather, there may be alternative explanations for this result.

Taking into account the above considerations, and using the concept of seizure time surrogates, Mormann et al.,¹⁵¹ within the scope of the First International Seizure Prediction Workshop, investigated the predictability of seizures by retrospectively analyzing multiday intracranial recordings from five patients recorded at different epilepsy centers comprising 51 seizures and a total recording time of 311 hours. They compared the performance (based on ROC statistics) of 30 univariate and bivariate measures, comprising both linear and nonlinear approaches, in terms of their ability to distinguish between the interictal state and the pre-ictal state. Using different evaluation schemes that take into account the majority of potentially confounding variables discussed so far, they were not able to detect a pre-ictal state, if one assumes a consistent effect on the EEG at all electrodes in all seizures of a patient and across patients, without comparison to some adaptive threshold level. If, however, one allows variable effects such as recording sites selected as best for prediction, a prolonged pre-ictal state (up to hours) could be detected using bivariate measures. Their findings suggest that a prospective seizure anticipation system is, in principle, possible and would perform better than random. Whether this is sufficient for a clinical application would need to be decided on an individual basis.

In order to gain deeper insights into the neurobiologic mechanisms underlying pre-ictal dynamical changes, approaches that are based on experimental models of epilepsy are highly desirable. Although there is a considerable bulk of literature on animal models of epilepsy, the phase of transition to the seizure state is not yet fully explored. Early work on the neurophysiology of the interictal-to-ictal transition can be found as early as the late 1960s, in the work of Dichter and Spencer.^41,42 In a penicillin model of epilepsy, these studies demonstrate a progression in the complexity and amplitude of interictal discharges during the interictal-to-ictal transition. While other studies have demonstrated similar changes, it was not until more recently that in vivo studies in acute and chronic animal models reported distinguishable preseizure EEG patterns several minutes prior to seizure onset.^{59,129,165,168} Other investigators²⁰¹ have showed seizure predictability from
stochastic models of temporal interdependence between the ictal and interictal states. Studying four pharmacologic epilepsy models, Widman et al.²²⁰ reported on a reduced dimensional complexity of in vitro hippocampal recordings in the seconds to minutes prior to the manifestation of paroxysmal depolarization shifts in xanthine and penicillin models. The authors went on to demonstrate that there was no preceding loss of complexity in low-magnesium and veratridine models. In contrast, Chiu et al.³² recently analyzed in vitro hippocampal recordings from a low-magnesium model using wavelets and artificial neural networks and claimed prediction of seizurelike events as early as 60 seconds before, and with more than 75% accuracy within a 30-second precision window.

Although these studies indicate that discernible pre-ictal patterns can be detected, at least in some experimental models of epilepsy, they must, at present, be regarded as “proof of principle studies,” since most lack a sound statistical validation, suffering from the same limitations of the human studies discussed above. In addition, a number of the studies used for seizure prediction from animal tissue, both in vivo and in vitro, have focused on acute seizure models whose behavior is not clearly analogous to spontaneous seizures in humans. These studies in animal models of epilepsy have considerable potential to address questions regarding the mechanisms regulating seizure generation; however, fundamental questions about underlying cellular mechanisms still remain unanswered. Earlier in vitro studies of ictogenesis suggested that the pre-ictal transition may reflect gradual changes in network excitability during the interseizure interval. In the 30 to 60 seconds preceding a seizurelike event, induced by high extracellular potassium, the excitability of CA1 pyramidal cells appeared to increase gradually to some threshold level, at which point the next incoming interictal burst precipitates a seizure.²⁰⁹ These pre-ictal changes in neuronal excitability are hypothesized to reflect the slow recruitment of an increasing number of excitatory interconnected neurons into synchronous discharges.¹⁴⁶

In chronic epilepsy, it is known that an increased propensity for hypersynchronization is associated with neuronal loss and synaptic reorganization that can be observed in human mesial temporal lobe epilepsy, and in kindling and kainic acid models.¹⁵ The transition to ictal activity, however, is often attributed to a breakdown of the interneuronal inhibitory control, which is assumed to be the major factor responsible for ictogenesis in human temporal lobe epilepsy and analogous animal models of epilepsy. This assumption is supported by early animal model studies that indicate that the inhibition of neuronal activities peripheral to the epileptic focus can be of crucial importance in determining whether or not a seizure is likely to occur and to spread.¹⁴²

A more recent hypothesis^17,97 suggests that a pre-ictal imbalance between inhibition and excitation might be due to some transient excitatory effect of GABAergic (γ-aminobutyric acid) function, resulting in repetitive high-frequency epileptiform bursts. Supporting this hypothesis, the prolonged repetitive activation of inhibitory neurons has been shown to induce an intracellular accumulation of chloride in principal cells, leading to a transient disinhibition of the local networks and, thus, to a decreased threshold for subsequent epileptiform discharges and ictal events.⁹⁹ The potential ictogenic effect of GABA-mediated depolarization has also been observed in other studies³³ (for a review, see reference 35). Finally, slow changes in extracellular ion concentrations are assumed to contribute to the gradual depression of inhibitory mechanisms. In particular, some seizures are preceded by an increase in the extracellular potassium concentration,¹⁶⁶ and in vitro studies have confirmed a corresponding pre-ictal depolarization of the neuronal membrane potential.⁸⁵

More recently, the cellular basis of pre-ictal changes has been investigated using a hippocampal–entorhinal brain slice preparation exposed to high extracellular potassium⁴⁷ and to low magnesium concentrations.⁹⁸ In this network, predictable alterations in interictal activity were shown to precede the transition to ictal-like activity. These pre-ictal changes were characterized by alterations in the origin and spread of CA3 epileptiform discharges. Spectral analysis of the interictal epileptiform activity preceding the transition to ictal-like activity revealed a prominent increase in the high-frequency range (200 to 400 Hz). In a very active related area of research, microelectrode studies in vivo in the hippocampus of kainate-treated rats and depth EEG recordings from patients with mesial temporal lobe epilepsy have also revealed 200- to 400-Hz oscillations, termed “fast ripples,” which have been hypothesized to identify microscopic regions important to seizure generation.^18,19 These oscillations may uniquely occur in areas that generate spontaneous seizures, and may reflect pathologic hypersynchronous population spikes of bursting pyramidal cells.⁴⁶ In rat models of chronic epilepsy, fast ripples are seen during the latent period between the occurrence of an initial epileptogenic insult and the onset of recurrent spontaneous seizures.²⁰ These oscillations may thus be a possible marker of alterations in interictal activity that precede the transition to ictal activity.^207,208 Further investigations at a cellular level during the preseizure state are necessary to validate these hypotheses in humans.¹⁹⁴

In addition to its scientific appeal, two main motivations for seizure prediction research are (a) its potential to control therapy in antiepileptic devices and (b) to elucidate mechanisms underlying seizure generation. Initial research into antiepileptic devices demonstrated that open-loop stimulation (e.g., paradigms that stimulate in a set “on–off” cycle regardless of underlying brain activity) can reduce seizure frequency and severity.^{72,96,131,139,205} While initial results are promising, this strategy has not yet made many patients seizure free, and none of these devices has achieved sustained clinical use.^72,131,215 One device, targeting the anterior thalamic nucleus, is in active clinical trials, and no definitive performance results are available as of this writing. Open-loop devices are attractive because of their relative simplicity; their analogy to successful cardiac devices, such as early pacemakers; and their modern origin in successful devices to treat movement disorders (e.g., Parkinson disease, tremor, etc.). While this research continues, focusing on different central targets and stimulation paradigms, an interest in more “intelligent” responsive devices, designed using strategies similar to implantable cardiac defibrillators, is under way.

Over more recent years, research into responsive, closed-loop antiepileptic devices, which read and respond to features extracted “on the fly” from implantable sensors and processors, has proceeded actively. This is primarily motivated by an attempt to improve upon results from open-loop devices.¹⁰⁰ While trials of these first-generation devices demonstrate proof of principle well, specifically that clinical seizures can be successfully stopped by electrical stimulation triggered after seizure onset is detected, these devices are not yet ready for widespread clinical deployment. Recent results from an initial safety trial, not designed to test efficacy, suggest that responding to seizure onset can reduce seizures by 50% or more in over 40% of refractory epilepsy patients. This result, however, also indicates that this method might not be the most effective path to making patients seizure free (Worrell et al., unpublished). The current state of these responsive devices suggests a significant potential role for therapy triggered in advance of seizure
onset on the intracranial EEG, perhaps controlled by seizure prediction algorithms.

The case that seizure prediction is a better method for controlling antiepileptic brain stimulation than seizure detection can be made on two fronts, though, unfortunately, there is not yet class one evidence to support this hypothesis. First, there is evidence from a number of seizure prediction studies quoted above^{38,114,150,151} that there is considerable evolution of quantitative features in multiple channels, including changes in synchronization in the epileptic network, prior to the time of detectable seizure onset on the intracranial EEG. Given that abnormal brain activity is often broadly distributed by the time seizures are detected, it seems unlikely that stimulation over a limited region, as is the standard technique with first-generation responsive devices, will be effective in suppressing clinical events. Clinical evidence to support this conjecture is more difficult to come by, as there is little information published in peer-reviewed journals about experience to date with responsive antiepileptic devices. There are anecdotal reports of widespread stimulation, grouping four or more electrode contacts together to form one pole for stimulation of distributed electrodes. To date, only two groups have reported results from responsive stimulation in humans, and were lead by NeuroPace, Inc., and Osorio et al.^100,161 (Worrell et al., unpublished). Because of intense activity in industry to perfect effective antiepileptic devices and bring them to market, current concepts are largely shaped by anecdotal clinical experience, conference abstracts, lectures, posters, and personal communication. With clinical trials currently in progress, it is anticipated that publication in this area should increase substantially over the next 12 to 24 months.

Initial reports indicate that when detection algorithms are tuned to high sensitivity and short detection latencies, so that stimulation is triggered not only by seizures, but also by bursts of epileptiform and rhythmic oscillatory activity often found to occur during periods of increased probability of seizure onset, seizure frequency declined (Pless, unpublished). In addition, under this paradigm, subclinical seizurelike bursts, which appear to increase prior to seizure onset in some patients, have been reported to greatly increase in number but not lead to seizures when repetitively suppressed by responsive stimulation. This result, observed by several investigators, suggests a drive to seize that might be somewhat suppressed by electrical stimulation. An additional explanation might include a critical amount of energy or some other quantity that might need to be dissipated in order to interrupt the process of seizure generation. Again, in the absence of controlled, randomized clinical trials in an adequate number of patients, this information is suggestive, at best, that intervention based on seizure “precursors” may be useful in suppressing clinical events. In responsive stimulation studies by Osorio et al.,¹⁶¹ prediction paradigms were not tested in human stimulation trials.

While discussing the clinical application of seizure prediction algorithms, it is instructive to consider how such an algorithm might function in a responsive antiepileptic device. Since current thinking is that seizure generation is likely a probabilistic phenomenon, most investigators agree that prediction algorithms will, at best, be able to identify periods of increased probability of seizure onset. The most important factor governing application of prediction algorithms for device control is the nature of the therapy being administered. If the treatment has few or no side effects, then high false-positive rates can be tolerated, as administering therapy when seizures are not imminent carries little or no penalty. If the algorithm is able to calculate the probability of an impending seizure at a given time, then the intensity of therapy can be increased in proportion to the likelihood that a seizure is going to occur. This scheme would also make sense if seizures are more difficult to pre-empt as they progress in the generation process. Finally, therapy is administered more intensely and perhaps over a broader geographic area as seizure generation progresses and the probability of onset continues to rise, until a seizure is detected, and maximal therapy is administered. At this point it might be reasonable to issue an alarm to the patient that a seizure is imminent. Should therapy not be immediately effective, it is possible for the antiepileptic device to broadcast an alarm, perhaps over a broadband communications link, or contact some central station to call for help unless the patient deactivates this process, demonstrating that awareness has returned. While this is only one way in which a predictive antiepileptic device might be deployed, it gives some sense of how prediction algorithms might be translated into a practical therapy.

One important aspect of implantable antiepileptic devices is the potential, in some models, to provide data vital to understanding the process of seizure generation. Some current devices have the capability for storing EEG information at different points in the seizure cycle, a capacity that might be possible to expand. This capability provides, for the first time, a way for investigators to look at intracranial EEG data acquired in the steady-state human condition, in the absence of proximate surgery or medication taper. Though the buffering capacity of these devices is likely small, the ability to download information from them over time provides an unprecedented opportunity to observe the natural seizure cycle in human patients. The authors are hopeful that these data will eventually become available for analysis by groups using a variety of different techniques. More confined data downloads at different times in relation to seizures might also be used to reconstruct the process of seizure generation, but as discussed above, relying on fragmented data is a more challenging task than analyzing prolonged, continuous data streams.

The utility of seizure prediction in understanding mechanisms underlying seizure generation also remains largely uncharted territory. While a number of investigators have attempted to predict seizures in a variety of mechanism-driven experiments, this body of work suffers from the same statistical and methodologic limitations as the more clinical studies noted above. Primary knowledge from prediction studies with macroelectrodes that could be extremely useful in understanding ictogenesis are such things as the period of time over which seizures are generated, what components of the epileptic network are necessary for clinical events to occur, how seizure precursors are temporally and spatially distributed in brain, and what effective intervention strategies for modulating this process are. The answers to these questions might guide basic science investigations to specific mechanisms. For example, if the seizure prediction horizon turned out to be on the order of seconds, then mechanisms related to ion channel or neurotransmitter function might be areas of focus for seizure generation experiments. Longer periods of time, perhaps minutes, might point to slower second messenger systems for study. More prolonged periods of seizure generation, perhaps minutes to hours or longer, could point to protein transcription, synaptic plasticity, and other mechanisms that might better fit into this time frame. If the pre-ictal state is found to be a permissive one, waxing and waning between false starts that back to the interictal period and then only occasionally to seizures, this finding might point to some oscillating pathologic process that is gated by another, perhaps unrelated mechanism.

These same types of ideas might also better elucidate the functional structure of the epileptic network. For example, if very high-frequency precursor events are found that spread very rapidly through local cellular regions, mechanisms involving ephaptic conduction and gap junctions might be implicated. If precursor spread is measured in longer intervals, it might be attributed to more typical synaptic conduction. Lateral cortical spread versus dispersion of signals via large fiber tracts are
also questions that could be answered by appropriately crafted seizure prediction studies. These are just a few ways in which seizure prediction algorithms and study might contribute to our basic knowledge of seizure generation, and perhaps to the development of better therapies.

The science of seizure prediction is now approximately 15 years old, and maturing. The overt initial enthusiasm that resulted from finding changes in “chaotic” measures reliably associated with periods prior to seizure onset has now given way to careful introspection. There is now a focus on meticulous data acquisition, annotation, study design, and rigorous statistical design of prediction studies and validation of results. The last 5 years has seen established groups and methods come under increasing scrutiny, as new investigators enter the field and newer techniques are developed to benchmark algorithms against random predictor performance and against the null hypothesis that there is no pre-ictal state. There is also pressure to get an answer and develop methods ready for clinical deployment. This is driven by a keen awareness that patients are likely to benefit considerably through second-generation responsive antiepileptic devices if researchers can field working algorithms for seizure prediction and validate them in blinded, prospective, clinical trials. There is great need for establishing convincing, iron-clad evidence for the existence of a pre-ictal state and an appropriate model for its behavior in human epilepsy. Researchers are acutely aware of the heterogeneity of epilepsy and the fact that one model may not work in all epilepsy syndromes. An improved mathematical modeling of the dynamics underlying the transition to seizures^{130,136,137,199,200,217,218,219} may help to test various hypothesis concerning pre-ictal brain dynamics and its relation to endogenous and exogenous control parameters. There is also a great need for new analysis methods, using multivariate techniques^{5,30,155,184,185,223} to analyze data streams from multiple sensory sites simultaneously, and even to sample at multiple temporal and spatial scales, from single neurons to ensembles, local then global networks, to investigate the physiologic substrate for larger-scale quantitative observations. Along with these observations will come improvements in our ability to collect, store, and analyze longer, more complete, and broader-band data sets. This work is already under way as a collaborative effort through the International Seizure Prediction Group. Following initiation of an international collaborative working group in this area, in Bonn in 2002, the task of establishing an extensive archive of high-quality, broadband, meticulously collected and annotated data archive from humans and animal models of epilepsy is getting under way. New hardware and software platforms¹⁵⁴ will likely provide more computational power, such as through VLSI implementation of algorithms,^31,102,191 while an effort to reduce the complexity of algorithms continues.¹⁹⁶ Finally, more work on developing and understanding spontaneously seizing animal models of epilepsy for use in prediction research will allow us access to deep brain structures and other locations in the epileptic network that cannot be explored in human studies, due to safety concerns. Characterizing and verifying the clinical presentation, reliability, and reproducibly of these models will be vital to interpreting results involving these animals. All of these factors promise increasing progress in the field of seizure prediction, now tempered by experience and a knowledge that this is a complex, long-term problem. Most importantly, we are starting to appreciate exactly what the task is that we are trying to accomplish, including articulating the benchmarks required for knowing when we’ve succeeded in predicting epileptic seizures.

In this chapter we have tried to give a broad overview of the field of seizure prediction and its history, accomplishments, controversies, and potential for future development. In a work of this scope it is inevitable that some contributions may be over- or underemphasized, depending upon the points to be made in the text. Seizure prediction remains an exciting field, with the potential to have significant impact upon the quality of life of our patients with epilepsy, in the form of newer and more effective treatments and in expanding our knowledge of how seizures are generated in the brain. What is clear, after the seminal contributions of a few insightful research groups over 15 years ago, is that the problem is large, and will require the collaborative efforts of a dedicated international group of investigators to solve it. We are hopeful that the pace of discovery in this relatively new field will continue to accelerate, as we begin to deal with our growing appreciation of the complexity of the problem.

Klaus Lehnertz acknowledges support from the Deutsche Forschungsgemeinschaft. Michel Le Van Quyen was funded by the Fondation Française pour la Recherche sur l’Epilepsie (FFRE), La Ligue Française Contre l’Epilepsie (LFCE), La Fondation pour la Recherche Médicale (FRM), and the Institut National de la Santé et de la Recherche Médicale (INSERM). Brian Litt was funded by National Institutes of Health grants (RO1NS041811-01, R01 NS48598-01) and grants from The Klingenstein Foundation, Dana Foundation, Whitaker Foundation, The Epilepsy Project, The Epilepsy Foundation, Jim Jacoby, and the American Epilepsy Society.