Neuron, Vol.

30, 51–64, April, 2001, Copyright 2001 by Cell Press

Epileptic Seizures May Begin Clinical Study

Hours in Advance of Clinical Onset:
A Report of Five Patients
Brian Litt,1,7 Rosana Esteller,2,6 Javier Echauz,2,3 available therapy. In most cases, this is focal epilepsy,
Maryann D’Alessandro,2 Rachel Shor,2 in which seizures arise from a region of abnormal brain,
Thomas Henry,5 Page Pennell,5 the epileptic focus, and spread in a stereotyped, individ-
Charles Epstein,5 Roy Bakay,4,8 ualized fashion. 50 years ago, Penfield and colleagues
Marc Dichter,1 and George Vachtsevanos2 found localized abnormal discharges on the EEG prior
1
Department of Neurology and Bioengineering to clinical seizures in some affected individuals (Penfield
3 West Gates and Erickson, 1941). They reported that seizures could
Hospital of the University of Pennsylvania be cured in these patients by removing the brain region
3400 Spruce Street that gave rise to this activity (Penfield and Flanigin, 1950;
Philadelphia, Pennsylvania 19104 Feindel et al., 1952). Since that time, recognizing seizure
2
Georgia Institute of Technology onset on EEG has remained the gold standard for locat-
School of Electrical and Computer Engineering ing epileptic foci in the brain (Sperling, 1986).
777 Atlantic Drive One of the most disabling aspects of epilepsy is that
Atlanta, Georgia 30332 seizures appear to be unpredictable. Investigation to
3
Electrical and Computer Engineering elucidate their underlying mechanisms has traditionally
Department focused on seconds to minutes before clinical seizure
University of Puerto Rico onset. Ralston and Dichter reported preseizure EEG
Mayagüez, Puerto Rico 00681 changes consisting of increased frequency and com-
4
Department of Neurosurgery plexity of epileptiform discharges, commonly called
5
Department of Neurology spikes, which became polyphasic (developed multiple
Emory University School of Medicine peaks), rhythmic, and sustained as seizures approached
1365 Clifton Road, NE (Ralston, 1958; Dichter and Spencer, 1969a, 1969b).
Atlanta, Georgia 30322 They postulated mechanisms initiating rapid changes
6
Dpto. Tecnologı́a Industrial in neuronal function, effected by ion channels, neuro-
Universidad Simón Bolı́var transmitters, and the cell membrane. Despite over 40
Caracas, Edo. Miranda years of investigation into the physiology of epilepsy, it
Venezuela Aptdo. Postal 89000 is still not possible to explain how and over what time
spontaneous clinical seizures emerge from the relatively
normal brain state observed between them.
Summary In recent years, clinical epilepsy monitoring has en-
abled studying human seizures quantitatively, as they
Mechanisms underlying seizure generation are tradi- develop in vivo. This effort has been aided by technologi-
tionally thought to act over seconds to minutes before cal advances such as digital video-EEG monitoring sys-
clinical seizure onset. We analyzed continuous 3- to tems that can process data in real time, and safe, reliable
14-day intracranial EEG recordings from five patients electrodes for implantation directly on or in the brain
with mesial temporal lobe epilepsy obtained during (intracranial electrodes). Most work in quantitative anal-
evaluation for epilepsy surgery. We found localized ysis of human seizures has focused on detecting their
quantitative EEG changes identifying prolonged bursts electrical onset on EEG in the up to 10 s that may pre-
of complex epileptiform discharges that became more cede their first clinical signs (clinical onset) (Gotman,
prevalent 7 hr before seizures and highly localized 1987, 1991; Qu and Gotman, 1997; Osorio et al., 1998).
subclinical seizure-like activity that became more fre- As early as 1970, researchers considered methods for
quent 2 hr prior to seizure onset. Accumulated energy predicting epileptic seizures prior to electrical onset,
increased in the 50 min before seizure onset, com- though none were implemented clinically (Viglione et al.,
pared to baseline. These observations, from a small 1970; Rogowski et al., 1981). Recent research in seizure
number of patients, suggest that epileptic seizures prediction uses the theory of chaotic dynamics to track
may begin as a cascade of electrophysiological events nonlinear parameters, such as the principal Lyapunov
that evolve over hours and that quantitative measures exponent and correlation dimension, over minutes prior
of preseizure electrical activity could possibly be used to EEG onset of temporal lobe seizures (Iasemidis et al.,
to predict seizures far in advance of clinical onset.
1990, 1994; Lehnertz and Elger, 1998; Martinerie et al.,
1998; Le Van Quyen et al., 2001). These studies are
Introduction important because most support the idea that seizures
develop over time, rather than occur as abrupt, isolated
Epilepsy affects 50 million people worldwide, 25% of events. They are somewhat limited, however, because
whom have seizures that cannot be controlled by any most concentrate on relatively few or brief data seg-
7
ments prior to seizures, and the abstract parameters
To whom correspondence should be addressed (e-mail: littb@
they explore are difficult to relate to the neurophysiology
mail.med.upenn.edu).
8
Present Address: Chicago Institute of Neurosurgery and Neurore-
of epilepsy.
search, Department of Neurosurgery, Rush-Presbyterian-St. Luke’s Because many patients report nonspecific, stereo-
Medical Center, 1653 West Congress Parkway, Jelke, Suite 1491, typed symptoms hours in advance of seizures (Rajna et
Chicago, Illinois 60612. al., 1997), we have focused on looking for quantifiable
Neuron
52

Table 1. Patient Characteristics

Sz Ictal Onset Zone(s) on Intracranial Side of Temporal Seizure Risk
Pt Types MRI FDG Activity EEG and Comments Resection and Outcome Factors
1* SP Left hippocampal atrophy Left temporal Left hippocampus (6/6) Left CHI
decrease
CP Incomplete resection due to Class II FS
language localization in ictal
onset zone
2 SP Normal Equal bilateral Left hippocampus (11/13) No surgery None
temporal decrease
CP Right hippocampus (2/13)
Accumulated energy lateralized to
side of onset in all seizures; all
Sz propagated in ⬍ 3 s to contra-
lateral side
3 SP Right frontal venous Bilateral temporal Right hippocampus (9/11) Right None
angioma decrease
CP Left hippocampus (2/11) Class I
Accumulated energy lateralized to
the right in all seizures
4 SP Normal Right temporal Right hippocampus (5/5) Right None
decrease
CP Class II
GTC
5 CP Subependymal heterotopia Left temporal Left hippocampus (6/10) Left FS
adjacent to left lateral decrease
ventricle
Left temporal neocortex (3/10) Class III
Right frontal neocortex (1/10)
Accumulated energy lateralized to
side of onset in all seizures
Class of outcomes refers to “Engel Classification” (Engel, 1987): Class I, free of disabling seizures; Class II, rare disabling seizures; Class III,
worthwhile improvement; Class IV, no worthwhile improvement.
* Earliest EEG changes associated with seizures were localized to one or two adjacent electrode contacts.
Abbreviations: CP, complex partial; GTC, generalized tonic-clonic; SP, simple partial; CHI, closed head injury with loss of consciousness
and/or amnesia; FS, febrile seizures in early childhood.

precursors of seizures far in advance of seizure onset nial electrodes during evaluation for epilepsy surgery.
on the EEG. We analyzed continuous intracranial EEG Subjects were the first five patients meeting these crite-
recordings over many days and used algorithms that ria admitted to the Epilepsy Monitoring Unit after univer-
are sensitive to the spikes, high-frequency rhythmic dis- sity approval of the research protocol. Patients included
charges, and high-amplitude slow waves that are typical four men and one woman. Mean age was 32 years (range
of the EEG in epilepsy. This preliminary investigation 21–50). The etiology of epilepsy was likely closed head
in five patients is a necessary first step in identifying injury in patient #1, although she also had febrile convul-
quantitative parameters worthy of more intensive study sions in early childhood, which have been associated
in a larger, prospective trial. with later development of mesial temporal lobe epilepsy.
We chose to study patients with mesial temporal lobe Patient #5 also had febrile convulsions, but no other risk
epilepsy, in whom seizures arose from the amygdala, factors for epilepsy. The other three patients had no
hippocampus, or adjacent cortex in one or both tempo- identified risk factors.
ral lobes because: (1) mesial temporal lobe epilepsy Intracranial monitoring was required in these patients
frequently cannot be controlled by antiepileptic drugs, because brain imaging, scalp EEG, and other noninva-
(2) it is common and is among the best understood of sive studies did not definitively identify the site(s) from
focal epilepsies, (3) it is common practice to record from which seizures originated. Evaluations included: (1) re-
intracranial electrodes in these patients during evalua- cording seizures with scalp EEG electrodes, (2) high-
tion for epilepsy surgery, (4) the likelihood of electrodes resolution brain magnetic resonance imaging (MRI) to
being placed in or very close to the region generating look for focal lesions, (3) brain positron emission tomog-
the earliest precursors of seizures is high in these cases, raphy scanning (PET, using FDG, or [18F]2-fluoro-2-
and (5) electrical seizure onsets are usually well localized deoxyglucose) to identify regions of decreased brain
on intracranial EEG in mesial temporal lobe epilepsy, metabolism usually associated with epileptic foci, and
have high signal to noise ratio, and offer the best avail- (4) neuropsychological testing to identify focal impair-
able data for studying seizure generation in humans. ment of cognitive function corresponding to brain re-
gions that give rise to seizures. A clinical summary of
Clinical Background the study patients, including brain imaging results and
We analyzed data from five patients with mesial tempo- clinical outcome, is provided in Table 1.
ral lobe epilepsy who required monitoring with intracra- Location of seizure precursors was validated by clini-
Precursors of Temporal Lobe Seizures
53

Figure 1. Long-Term Energy, Patient #1

Intracranial EEG electrode placement in pa-
tient #1, imaged on MRI scan. Subtemporal
strip electrodes were placed under the tem-
poral lobes bilaterally through burr holes
through the temporal bones. Bilateral tempo-
ral depth electrodes were placed stereotaxi-
cally through burr holes in the occipital re-
gions bilaterally.

cal outcome after epilepsy surgery in four of the five Each patient had one six-contact “depth” electrode
study patients (Table 1). Patient #2 chose not to have placed stereotaxically in each temporal lobe, using coor-
epilepsy surgery despite a majority of seizures arising dinates generated from a high-resolution MRI scan of
from the right hippocampus. This patient had normal the brain referenced to a coordinate frame fixed to the
brain imaging and seizures that spread from one side head of the patient for surgery. These electrodes were
to the other within several seconds, predicting a poorer placed through a burr hole in the occipital region of the
chance of significant reduction in seizures after epilepsy skull and advanced forward so that the most anterior
surgery. contact lay in the amygdala and more posterior contacts
were located along the head and body of the hippocam-
pus. Additional electrodes were placed directly on the
Table 2. Time between Seizures
surface of the brain (subdural electrodes) over and under
Mean Time to Standard the temporal lobes bilaterally through a single burr hole
Patient Next Seizure (Hr) Deviation (Hr) Range (Hr)
made in the temporal bone on each side of the skull.
1 12.7 10.2 0.9–22.8 Three patients (#3, #4, and #5) also had subdural elec-
2 16.3 30.9 0.7–102.3 trodes placed on the surface of the frontal lobes bilater-
3 19.2 12.7 1.7–35.1
ally. Figure 1 demonstrates placement of intracranial
4 27.5 35.2 2.3–103.7
5 9.8 15.1 0.6–35.6
depth and subdural temporal electrodes in one study
patient.
Spacing of seizures over time for each patient, listed as time to next All EEG recordings were reviewed visually by four
seizure, with standard deviation and range of values.
epileptologists certified by the American Board of Clin-
Neuron
54

Figure 2. Long-Term Energy Burst, Raw EEG

Prolonged energy calculated from two elec-
trodes in a bipolar montage placed proximate
to the ictal onset zone prior to a single seizure
in patient #1. Similar prolonged bursts of en-
ergy were seen prior to other seizures in pa-
tients #1, #3, and #4, but not during baselines.

ical Neurophysiology in EEG (BL, TH, PP, CE), for contacts. Unequivocal EEG onsets of seizures were ste-
purposes of clinical decision-making. The first author reotyped for each patient, suggesting a reproducible,
visually reviewed the EEG for each seizure before quan- individualized pattern of initiation and spread.
titative analyses of these data and marked the following Measures of energy in the EEG signal formed the core
times: (1) the earliest EEG change associated with sei- of quantitative methods used to analyze seizure-related
zures: this point in time was found by identifying un- activity (see Experimental Procedures). These were cho-
equivocal seizure activity on the EEG and then moving sen because: (1) their calculation is simple and fast,
backward in time to the point at which the first clear, consisting only of squaring the voltage at each point
sustained change from the patient’s EEG baseline prior measured in the digital EEG and summing this up over
to the seizure was detected; and (2) the unequivocal a particular period of time; and (2) signal energy, as we
EEG onset of seizures: this marked the time at which defined it, is sensitive to EEG waveforms associated
an EEG pattern typically associated with seizures (Ris- with seizures and epileptic foci in the brain, such as
inger et al., 1989) first became unquestionably clear and epileptiform spikes, rhythmic sharply contoured wave-
could be identified independent of knowing that a sei- forms, and high-amplitude slow (0.5–4 Hz) activity seen
zure followed. Intracranial EEG, simultaneous video, and after bursts of epileptiform activity.
medical records were reviewed in their entirety for each In the first experiment, signal energy in the EEG was
patient. Important clinical activity, sleep-wake cycles, computed over the entire hospital stay in the three pa-
times, and doses of medication administered during the tients, #1, #3, and #4, who had highly localized earliest
monitoring period were recorded and analyzed. EEG changes prior to seizure onset. Mean time between
seizures was 12.7 hr for patient #1 (stdev ⫽ 10.2 hr,
Results range 0.9–22.8 hr), 19.2 hr for patient #3 (stdev ⫽ 12.7
hr, range 1.7–35.1), and 27.5 hr for patient #4 (stdev ⫽
In two patients (#1 and #4), all seizures began in one 35.2 hr, range 2.3–103.7 hr). This information is listed
temporal lobe and earliest EEG changes were very focal, for all patients in Table 2. Periodic elevations in signal
confined to one or two adjacent electrode contacts. This energy appeared to occur in the epileptic focus intermit-
implies that these contacts were placed in or very close tently throughout the hospital stay and increased in
to the epileptic focus in these patients. Seizures arose number beginning approximately 7 hr prior to seizures.
independently from both temporal lobes in the re- Figure 2 shows a representative long-term energy plot
maining three patients (#2, #3, #5). Earliest EEG changes beginning 8 hr prior to a single seizure in patient #1. The
were poorly localized in patients #2 and #5, indicating figure demonstrates several prolonged bursts of energy,
either less favorable electrode placement or that sei- some lasting more than 30 min, as seizures approach.
zures arose from a more diffuse area in these individuals. Bursts were defined as sustained elevations of signal
In patient #3, earliest EEG changes were localized after energy of greater than two standard deviations above
ⵑ0.5 s demonstrating a loss of normal EEG background the interburst baseline for periods of 5 min or more in
in the anterior hippocampal electrodes bilaterally. In all a running average of 5 min duration. Bursts occurred in
five patients, unequivocal onset of seizure activity on the epileptic focus region in patients #1 and #4, though
EEG was localized to one or two adjacent electrode substantially lower amplitude synchronous activity in
Precursors of Temporal Lobe Seizures
55

Figure 3. Number of Long-Term Energy Bursts versus Time to Seizure

Segment of 28.5 min of unprocessed EEG recorded from an electrode in the seizure onset zone in patient 1 (left hippocampus in this patient)
including a period of time constituting a long-term energy burst lasting almost 25 min. Each horizontal row contains 100 s of continuous EEG.
Bursting activity begins ⵑ130 s into the record (marked by arrow, second row) and continues until the third row from the bottom (also marked).
A small box is drawn around 20 s of EEG in the sixth row of EEG (9 min into the figure), which is enlarged at the bottom of the figure. This
segment demonstrates typical high-amplitude epileptiform discharges and slowing that comprise long-term energy bursts.

the opposite mesial and subtemporal electrodes were relationship between long-term energy bursts and time
often present. In patient #3, long-term energy bursts of onset of electrographic seizures. Bursts appeared to
often appeared with approximately equal amplitude bi- be concentrated between 7 hr and 30 min prior to seizure
laterally, perhaps related to the independent bitemporal onset. Conclusions could not be drawn about the inci-
unequivocal EEG seizure onsets seen in this patient, but dence of long-term energy bursts more than 12 hr prior
were more frequently seen in the right temporal region. to seizure onset, due to data inclusion criteria (see Ex-
Increases in long-term energy were not detected at con- perimental Procedures). Long-term energy was elevated
trol electrode sites in the frontal and posterior temporal after each seizure and remained increased for an aver-
lobes. Visual inspection of long-term energy bursts re- age of 4–6 hr. This agrees with the findings of Franaszcuk
vealed increasingly complex spikes accompanied by and Bergey, who found increased synchrony in the re-
high-amplitude low-frequency activity (0.5–4 Hz) and gion of the epileptic focus after seizures to the end of
loss of normal background activity in the epileptic focus their 2 hr analysis period (Franaszcuk and Bergey, 1999).
channel during these periods. Figure 3 displays raw EEG For this reason, long-term energy bursts could only be
in a single channel in the epileptic focus comprising a evaluated before the first seizure when seizures oc-
typical long-term energy burst that occurred 4.5 hr prior curred less than 6 hr apart. This suggests that mecha-
to seizure onset and lasted ⵑ25 min. The figure includes nisms giving rise to clustered seizures are likely different
a “close-up” of typical epileptiform discharges and slow from those leading to isolated seizures. These findings
activity seen during these bursts for patient 1. Prolonged suggest that seizure generation may begin far earlier
bursts of signal energy were more common while pa- than 1 hr prior to clinical seizures.
tients were asleep, as is typical of epileptiform dis- In a second observation, patient #1 demonstrated ste-
charges, but were not attributable to EEG patterns asso- reotyped 6–12 s runs of very focal, rhythmic, high-fre-
ciated with sleep. Figure 4 demonstrates the temporal quency activity in the left anterior hippocampus (the
Neuron
56

reported previously. Of note, two of the patient’s six

seizures, which were excluded from the accumulated
energy experiment (below) because the EEG was briefly
obscured by artifact in the 3 hr leading up to seizure
onset, were included in the above analysis.
In a third experiment, “accumulated energy”—a run-
ning sum of the square of the EEG voltage sampled at
200 Hz—was calculated from EEG data beginning 50 min
prior to unequivocal seizure onset on EEG and ending 10
min after onset. Accumulated energy was computed at
the electrode site recording the earliest or maximal (if
several contacts were involved simultaneously) un-
equivocal EEG onset for each seizure. This was com-
pared to two controls: (1) baseline controls: randomly
chosen, 1 hr EEG segments recorded from the seizure
focus, but separated from the beginning or end of any
seizure by at least 3 hr of uninterrupted, artifact-free
EEG, and (2) contralateral controls: data acquired during
epochs leading up to seizures but from analogous elec-
trodes in the temporal lobe opposite the seizure focus.
Seizures and baseline data were compared only within
the same state of consciousness (e.g., awake versus
asleep), as baseline energy was found to be higher dur-
ing sleep than wakefulness, due to normal EEG wave-
Figure 4. Number of Long-Term Energy (LTE) Bursts versus Time
forms occurring during sleep (Table 3).
to Seizure (Ten Seizures, Three Patients)
80 1 hr EEG epochs were analyzed: 30 leading to
Time of occurrence of long-term energy bursts in relation to seizure
seizures, and 50 baselines. Figure 7 displays all baseline
onset is displayed for patients #1, #3, and #4. There was no clear
relationship between burst amplitude or duration and time to seizure and preseizure epochs recorded from the seizure focus
onset. Epochs of data not meeting inclusion criteria (e.g., at least in all five patients, normalized to include both sleep
12 contiguous hr without interruptions of 5 min or more by artifact and awake records for all subjects. For each patient,
or gaps in the data) were not analyzed, due to the lack of ability to accumulated energy during baseline periods was very
determine if these data epochs led to or followed subclinical sei- similar (dashed red lines) and accumulated energy in
zures. Because of these strict inclusion criteria, and frequent sei-
preseizure recordings (solid blue lines) clearly rose
zures and clusters induced by medication taper, little information
is available about the incidence of long-term energy bursts at times
above these baselines within 50 min prior to seizure
greater than 12 hr prior to electrical seizure onset. These findings onset. Table 3 displays “preseizure declaration times”
support the hypothesis that increasing incidence of long-term en- calculated for each patient. This was defined as the time
ergy bursts appear to correlate in time with approaching electro- prior to seizure onset at which each accumulated energy
graphic seizures beginning ⵑ7 hr before seizure onset. curve for a particular patient crossed the level of the
highest baseline accumulated energy for that patient.
Using this simple threshold function, data epochs were
focus in this patient) just prior to seizures. Each of these
correctly classified as either preseizure or baseline with
runs of activity began abruptly at 20–25 Hz then gradu-
89% accuracy, a sensitivity of 90%, and specificity of
ally increased in amplitude while slowing to 6–10 Hz, 88%. The choice of optimal threshold by using this
sometimes spreading to adjacent electrodes before ei- method was verified by experimentally generating a re-
ther leading to seizure onset or ending in higher ampli- ceiver-operating characteristic curve (Figure 8), plotting
tude, irregular 1–2 Hz activity. Typical examples of these sensitivity versus one minus specificity for a wide range
discharges are displayed in Figure 5, prior to the pa- of threshold values. Mean time from declaration of im-
tient’s first seizure. These discharges were consistent pending seizure to unequivocal EEG onset of seizures
with brief, highly localized seizures on EEG, but did not by this simple algorithm was 18.5 min (⫾13.4 min). Close
cause any clinical symptoms. Inspection of the patient’s examination of Figure 7 demonstrates that accumulated
entire 4 day intracranial EEG recording demonstrated energy actually rose above baselines much earlier than
that these electrical events clustered around clinical sei- this time in many cases. This suggests that it should
zures and were extremely rare at other times. Figure 6 be possible to devise classification algorithms, perhaps
demonstrates the rate of occurrence of these brief, very more sophisticated than a simple linear threshold, capa-
focal “subclinical seizures” in relation to unequivocal ble of identifying preseizure epochs even farther in ad-
electrical onset of the patient’s six typical seizures. The vance of electrical seizure onset than the above results.
majority of these 37 electrical events occurred within 2 These findings complement reports of “seizure anticipa-
hr of seizure onset. Discrete, self-limited, subclinical tion” using “nonlinear analysis” 2–6 min prior to seizure
seizures in intracranial EEG recordings are reported to onset in 17 of 19 patients by Martinerie et al. (1998), Le
have important value in identifying the location of the Van Quyen et al. (2001), and in 20 seizures processed
epileptic focus in patients with mesial temporal lobe by Lehnertz and Elger (1998).
epilepsy (Sperling and O’Connor, 1990; Schiller et al., Baseline signal energy was increased following clus-
1998; Schuh et al., 2000). Their timing, predictive value, ters of two or more seizures within a 6 hr period. When
and potential role in seizure generation have not been this was taken into account, four EEG epochs following
Precursors of Temporal Lobe Seizures
57

Figure 5. Rate of Subclinical Seizures versus Time to Electrical Onset of Symptomatic Seizures
Focal subclinical seizures prior to one seizure of patient #1 recorded from left hippocampal depth electrode contacts LT2 and LT3 referenced
to a remote electrode. These electrical events began with stereotyped fast activity and were of variable length. The last subclinical seizure,
#4, leads to a clinical seizure with UEO marked by the arrow. Subclinical seizures did not occur within 6 hr after electrographic seizures
except when seizures clustered together over short periods of time. There was no evidence that these discharges were associated with the
postseizure period.

clusters initially mislabeled as “preseizure” were cor- cal to generating seizures in patient #3 was likely located
rectly reclassified as baselines, improving overall accu- in the right hemisphere, perhaps near pathways allowing
racy to 94%. There were no clear EEG changes responsi- rapid conduction or diversion of activity to the left hippo-
ble for the two remaining false positive declarations; campus under some circumstances. This hypothesis is
however, these could have occurred during prolonged supported by the patient’s seizure-free outcome after
long-term bursts of energy between seizures (see surgical removal of the right temporal lobe.
above), or asymptomatic seizures hidden by gaps in Visual inspection of preseizure EEG recordings did
recorded data found more than 3 hr prior to these base- not demonstrate obvious findings that corresponded to
line segments. There were no specific characteristics changes in accumulated energy. The actual number of
or obvious systematic errors to account for misclassifi- epileptiform discharges (spikes) during the 50 min be-
cation of three false negative epochs. The short predic- fore seizures did not differ statistically from baseline
tion horizon for patient #4 sets him apart from the other periods, in agreement with Lange et al. (1983) and Katz
study patients, as does the unusually short time delay et al. (1991). The frequency of these discharges during
between electrical and clinical seizure onset in this pa- longer periods prior to seizure onset and their correlation
tient. These findings are of unclear significance. with impending seizures was not analyzed. Changes
In patients with seizure onsets confined to one tempo- in accumulated energy were not clearly confined to a
ral lobe (#1, #4), and in two of the patients with seizures particular frequency band in the EEG.
arising independently from both temporal lobes (#2, #5), Further analyses were performed to determine if high-
increased accumulated energy was found only on the amplitude epileptiform activity during long-term energy
same side as unequivocal EEG onset for each seizure. bursts or focal subclinical seizures accounted for the
In patient #3, of the seven seizures that met inclusion rise in accumulated energy prior to clinical seizures.
criteria, unequivocal EEG onset was seen in the right Long-term energy bursts occurred prior to 8 of 30 sei-
temporal lobe for five events and on the left side in the zures during the 50 min before unequivocal electro-
other two. Pre-seizure changes in accumulated energy graphic seizure onset in which the accumulated energy
predicted onset on the right side for all seizures. These was calculated. Given that accumulated energy was ele-
findings suggest that the epileptic focus, or region criti- vated above threshold prior to seizure onset in 27 of
Neuron
58

an average difference of ⬍3.5% in accumulated energy

was found between these curves at any given point.
This difference was not felt to be clinically significant and
did not significantly alter preseizure declaration times.
These findings suggest that the buildup in accumulated
energy in the 50 min prior to seizures cannot be ac-
counted for solely by changes in long-term energy or
the presence of subclinical seizures.

Discussion

These results are important for several reasons. First,

they demonstrate that there are changing forms and
amounts of abnormal electrical activity in the epileptic
focus that increase over long periods of time as seizures
approach. Second, they give hints to the mechanisms
underlying seizure generation and their time course.
Third, they provide important insight into the experimen-
tal design, data collection, and processing required for
seizure prediction trials. Finally, these results may be
exploitable in an implantable, therapeutic device. It is
important to point out that these results do not consti-
tute genuine seizure prediction, which would require a
Figure 6. Focal Subclinical Seizures Prior to Clinical Seizure #1 prospective analysis of data when time to seizure is
Rate of occurrence of subclinical seizures versus time to electrical unknown. This study is currently under way. Rather, they
seizure onset for patient #1: the rate of occurrence of subclinical provide evidence for a series of reproducible events
seizures relative to time of unequivocal electrographic onset (UEO) comprising seizure generation and a means for measur-
of seizures is displayed. Rates are normalized to the number of
ing them that may make it possible to reliably predict
events per hour. The graph demonstrates that beginning ⵑ2 hr prior
to UEO, there is marked increase in the occurrence of subclinical seizures.
seizures. This increase rises dramatically beginning in the period
50–10 min prior to UEO (A) and more in the period 10–1 min prior
to UEO (B). The rate of occurrence rises most dramatically within 1
Mechanisms
min (C) prior to electrical seizure onset. Subclinical seizures were Our findings suggest that seizure generation may be a
rarely more than 3 hr removed from seizures. multistage process that evolves over hours. Intermittent
bursts of energy in the epileptic focus, when influenced
by factors that promote seizures, may become more
these 30 epochs, it was felt that long-term energy bursts common and then give rise to brief, localized subclinical
alone could not account for preseizure changes in accu- seizures. As these rhythmic discharges occur more fre-
mulated energy. To examine the contribution of focal quently, adjacent tissue is recruited, increasing accumu-
subclinical seizures to accumulated energy, energy gen- lated energy and heralding oncoming seizures. This
erated by this activity was subtracted from the accumu- agrees with reports of increasing coherence in chaotic
lated energy calculation each time one of these events parameters near the seizure focus during the preseizure
was detected, substituting the energy from the adjacent period (Iasemidis et al., 1996, 1997).
previous EEG segment during these time periods. There Numerous mechanisms might account for these find-
was no significant difference in the accumulated energy ings, likely generated by hypersynchronous synaptic as
tracings before and after removal of energy due to sub- well as intrinsic voltage-dependent currents with longer
clinical seizures, when comparing plots of these curves time constants (Dichter and Ayala, 1987). Internal fac-
on the same graph. When the two curves were compared tors, such as sleep deprivation and hormone levels may
quantitatively for each preseizure period in this patient, modulate bursts of localized epileptiform activity. Neu-

Table 3. Accumulated Energy: Number of Records, Classification, and Prediction

Seizures Baselines
Patient Mean Preseizure Stdev of Pred False False
Number Declaration Time (Min) Time (Min) Negatives Positives Asleep Awake Asleep Awake
1 23.3 15.25 1 3 1 3 6 6
2 20.78 15.77 1 2 2 2 6 4
3 17.29 10.15 0 0 1 6 5 4
4 1.3 1.83 1 0 2 3 4 5
5 21.89 11.61 0 1 7 3 4 6
All Patients 18.49 13.42 3 6 13 17 25 25
Mean prediction horizon, standard deviation, and number of false positive and false negative predictions are displayed for each patient, in
addition to the number of seizure and baseline epochs and their distribution between wakefulness and sleep.
Precursors of Temporal Lobe Seizures
59

Figure 7. Normalized Accumulated Energy for Five Study Patients

Accumulated energy trajectories for all five patients, including 30 preseizure and 50 baseline epochs. Data are normalized within patients to
include both sleep and awake epochs. Baseline accumulated energy trajectories (dotted red lines) cluster together at the bottom of the figure.
Preseizure trajectories (blue lines) deviate from baselines within 50 min prior to seizures, though a number of the preseizure epochs demonstrate
deviation from baseline trajectories already in progress at 50 min prior to seizure onset, suggesting earlier preseizure change. Time ⫽ 0 marks
unequivocal electrographic seizure onsets, marked by the first author, according to standard clinical criteria.

ronal excitability could be enhanced by slow accumula- and the corresponding processes could be modified
tion of extracellular K⫹ (Dichter et al., 1972), alkaliniza- prior to seizure onset, leading to specific, targeted new
tion (Tang et al., 1990), or changes in intracellular therapies. Our findings, for the first time, suggest the
osmolarity (Schwartzkroin et al., 1998), which could ef- need to evaluate these processes in broader time
fect slow potentiation of excitatory synapses via local windows.
electrical coupling or neuromodulatory substances.
Changes in the redox state of the focus related to altered The Next Experimental Steps
perfusion, O2 extraction or neuronal activity has been Because collection of human data is ethically limited to
shown to significantly affect NMDA receptor activity. clinical necessity, we are employing animal models of
Increases in energy could also trigger transcriptional human temporal lobe epilepsy in an attempt to under-
changes in neurons within the epileptic focus. The re- stand the preseizure cascade. Some models, such as
sulting gene products may initiate a cascade of intracel- systemically delivered pilocarpine and sustained electri-
lular changes leading to seizure onset (Liang and Jones, cal status epilepticus, cause widespread neuronal in-
1997). This idea is appealing, because communication jury. Others, such as intrahippocampal injection of kainic
between cell surface membrane and the nucleus may acid, are more focal. All are problematic because the
require minutes and sometimes hours to occur. In addi- frequency and severity of associated seizures is difficult
tion, long-term potentiation is known to influence gene to control. Genetically engineered animal models of epi-
transcription, providing one potential mechanism for lepsy (i.e., knockouts) are less variable, though most are
electrical activity at the synapse to affect protein expres- more similar to human primary generalized epilepsy than
sion (Bailey et al., 1996). These events could be analo- to seizures of focal origin (Noebels et al., 1990; Noebels
gous to changes that occur in cortical neurons when and Rutecki, 1990; Burgess and Noebels, 1999; Noebels,
memories are stored. 1999). Transgenic mouse epilepsy models may be trans-
Although speculative, these hypotheses are testable, ferable to primates (Kafri et al., 2000; Wang et al., 2000),
Neuron
60

Figure 8. Receiver-Operating Characteristic

Curve
The empirical receiver-operating characteris-
tic curve (ROC) graph compares performance
of seizure prediction sensitivity and specific-
ity based upon accumulated energy as a
function of raising and lowering the prediction
threshold (see Figure 7). Each point on the
curve corresponds to a different threshold.
The vertical axis of this graph represents the
sensitivity of prediction based upon accumu-
lated energy and is plotted against 1-specific-
ity of the prediction algorithm at that specific
threshold value. In a practical sense, points at
the bottom left corner of the curve represent
performance for thresholds at which there are
no false alarms, and none of the seizures can
be predicted. Positions at the top right corner
represent thresholds at which all seizures are
predicted, but all baselines are false alarms
as well. An ideal detector 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). If a classifier pro-
duces an ROC curve that is a diagonal line,
this implies that the system randomly classi-
fies data segments as preseizure or baseline
with error probabilities that sum to one. The
threshold chosen for the accumulated energy
seizure prediction algorithm described above
(circled point) is demonstrated to be optimal
for this system, given its place at the “corner”
of the ROC curve.

offering potential benefits for preclinical testing of algo- Though some of the variability in findings between
rithms and implanted devices. our five study patients is likely due to electrode place-
ment and clinical factors, each subject was found to
Clinical Research: Experimental Issues have a unique preseizure electrophysiology. Our experi-
All tapes were reviewed in their entirety to screen out ence suggests that individual variability will likely fall
artifacts and gaps in EEG data, which necessitated ex- within a small number of pre-ictal patterns, analogous
cluding some epochs from analysis (see Experimental to seizure onset patterns recorded from depth elec-
Procedures). In retrospect, strict review and inclusion trodes in temporal lobe epilepsy (Spencer et al., 1992).
criteria were justified, given our findings that postseizure This implies that prediction algorithms will likely need
elevations in EEG signal energy lasted up to 6 hr and to be individually optimized, though utilizing a limited
that a large percentage of apparent “false positive” pre- number of adjustable parameters.
cursors occurred in anticipation of previously unrecog- All of the findings described in this study refer to
nized subclinical seizures. These findings reinforce the temporal and not extratemporal epilepsy, a discussion
need for long, continuous, and complete data sets that of which is beyond the scope of this paper.
are as free as possible from artifacts and recording gaps
for this type of research. An Implantable Diagnostic-Therapeutic Device
All study patients underwent rapid tapering of their Understanding mechanisms of seizure generation will
antiepileptic drugs to precipitate seizures, which is ultimately produce the best therapies, however, an im-
standard during presurgical evaluation, and facilitates plantable antiseizure device is a natural interim goal.
seizure clustering. Haut et al. have demonstrated that Figure 9 displays a schematic of such a device, based
clustered seizures demonstrate different EEG charac- upon algorithms we have developed to predict seizures
teristics than single spontaneous seizures in the same and trigger therapy as they approach. The device contin-
patient (Haut et al., 1997). Studies of seizure precursors uously monitors features extracted from the intracranial
in the absence of medication withdrawal will be impor- EEG and computes the probability of an approaching
tant to validate prediction algorithms for application out- seizure, while continuously checking for electrical sei-
side of the monitoring setting. zure onset. A wavelet neural network is trained, and its
Accumulated energy measurements were only useful performance individually optimized, based upon multi-
in separating preseizure from baseline epochs when ple features extracted from continuous data containing
state of consciousness was taken into account. Other both seizures and long seizure-free and preseizure inter-
investigators have dealt with this issue by examining vals. The patient or physician sets the probability thresh-
seizures only during wakefulness. Future trials will need olds and rate of rise that will trigger a warning or inter-
to account for EEG changes during all states of vention.
awareness. Electrical stimulation and local drug infusion are two
Precursors of Temporal Lobe Seizures
61

Figure 9. A Seizure Prediction Device

A schematic of an implanted device for seizure prediction and intervention. EEG is recorded from electrodes implanted in the area of the
seizure focus. Instantaneous features, such as signal energy, and historical features, such as the time of the last subclinical seizure, are
recorded. A neural network compares the behavior of these parameters to known behavior leading up to seizures for the given subject and
continuously outputs the probability that a seizure will occur over four time horizons: 1 day, 1 hr, 10 min, and 1 min. A warning of impending
seizure or intervention in the form of electrical stimulation or infusion of drugs in the area of the seizure focus are triggered when the probability
of an oncoming seizure, or the increase in probability over time, exceeds a predetermined threshold value.

main therapies being investigated in conjunction with arrhythmias. Implantable cardiac defibrillators process
these devices. As we envision this, when seizures ap- one channel of data, detect arrhythmias after their onset,
proach with low probability, lower amplitude stimuli that and deliver graded shocks to “reset” the heart’s electri-
are less likely to disrupt normal brain function might be cal system. Side effects such as pain and alteration or
applied. Algorithms might include standard or “chaos loss of consciousness are accepted, since the goal of
control” pacing (Glanz, 1994; Lopes da Silva et al., 1994; therapy is to preserve life. Epilepsy devices will likely
Heffernan, 1996), low-level depolarization or hyperpolar- require processing multichannel data in parallel, with
ization of the focus region. As the probability of seizure the goal of preventing clinical symptoms, such as loss
onset increases, stimulation intensity, duration, and area of awareness. Delaying therapy until electrical seizures
are increased, perhaps extending to subcortical struc- are detected may be too late to prevent clinical symp-
tures, which may modulate seizure generation. Care toms, given the rapidity with which seizures may spread.
must be taken to avoid stimuli that might cause tissue Finally, epilepsy devices will likely require optimization
injury or promote kindling of seizures in otherwise for each patient, and perhaps periodic retraining, given
healthy tissue (Lothman and Williamson, 1994). Of inter- the potential for seizures and electrical stimulation to
est, peripheral nerve stimulation and applying magnetic alter network dynamics in the brain over time.
fields to brain also hold therapeutic promise for treating
seizures (Fanselow et al., 2000; Gluckman et al., 2001).
Infusing small amounts of antiepileptic drugs locally Conclusion
in the epileptic focus may also inhibit seizure generation. Our results suggest that the changes in cellular and
Candidate drugs for infusion include benzodiazepines, network function that lead to epileptic seizures likely
lidocaine, standard, or newer antiepileptic agents. This develop over hours, providing exciting clues to potential
approach has been the subject of considerable interest mechanisms underlying seizure generation. Understand-
for the past several years (Eder et al., 1997; Ward and ing these mechanisms should produce new, more effec-
Rise, 1998; Stein et al., 2000). tive and specific therapies for epilepsy. Our results also
suggest that seizures in mesial temporal lobe epilepsy
Analogy to Cardiology are predictable within time horizons that make it possi-
Intelligent devices to treat seizures may prove more ble to prevent them, opening an immediate opportunity
difficult to implement than those used to treat cardiac to develop intelligent, implantable therapeutic devices.
Neuron
62

As seizure precursors become better understood, it may energy at the given sample rate of 200 Hz. A sliding rectangular
become possible to trace the earliest precursors of sei- window was used to process the data.
Average energy values were computed by averaging every 250
zures back to their site(s) of origin, perhaps enabling
points, equivalent to 1.25 s of the instantaneous energy at the given
ablation of critical regions of abnormal function by mini- sample rate of 200 Hz. Results were computed in a moving average,
mally invasive surgery or other techniques to replace overlapping consecutive calculations by 0.8 s (160 points) to give
large-scale brain resections that are now standard. Most a higher resolution. This was accomplished by shifting the computa-
important, these results provide new neurophysiologic tion window in Equation 2 0.45 s (90 points) each time:
evidence that may bring researchers closer to under- 1
Ek ⫽ ⌺ x2[n], (2)
standing how epilepsy begins and perhaps how to pre- N
vent it.
where Ek represents the kth value of the average energy over a
window of 250 points, N is the total number of instantaneous energy
Experimental Procedures values averaged, in this case N ⫽ 250.
A second level of averaging was performed to calculate the accu-
Continuous intracranial EEG and video were collected using a digi- mulated energy. Ten consecutive values of the average energy are
tal, 64-channel, 12 bit Nicolet BMS-5000 epilepsy monitoring system added, divided by 10, and then added to the running tally of accumu-
and were stored on videotape. Referentially recorded EEG was lated energy. This process is equivalent to integrating the absolute
downloaded from tape and archived to CD-ROM for processing. value of the EEG voltage as it is recorded on the intracranial EEG.
EEGs were digitized at 200 Hz and recorded after filtering through An overlap of 50%, or five points, is used in this measure, again
a bandpass of 0.1–100 Hz. Data were processed using a combina- to improve resolution in time. Equation 3 summarizes how each
tion of MATLAB programs and custom-written C⫹⫹ code for more accumulated energy value is obtained:
prolonged, computationally intensive tasks on a network of four 410
1
MHz IBM personal computers running Windows NT and equipped AEm ⫽ ⌺ Ek ⫹ AEm⫺1, (3)
10
with parallel processors. Hospital stays varied from 3–14 days in
the five study patients, yielding roughly 50 gigabytes of raw data where AEm is the mth value of the accumulated energy, AEm⫺1 is the
available for processing. Bipolar electrode montages were used to mth ⫺ 1, or previous value, of the accumulated energy.
eliminate common mode artifact. A digital 60 Hz notch filter was
employed to eliminate line noise. Videotapes were watched in their Accuracy Calculation
entirety for clinically important events, and approximation of periods Classification accuracy was calculated as the percentage of total
of sleep and wakefulness. Finer resolution of sleep-wake cycles number of correct predictions with respect to total number of re-
was achieved by examining the EEG. cords analyzed (Echauz et al., 1998):
Strict inclusion criteria were used in the study of long-term energy
to ensure that the relationship between long-term energy bursts
and seizures could be properly explored. Seizures that occurred
冤
A ⫽ 100 1 ⫺
1 N
N i⫽1 冥
兺 兩yi ⫺ ŷi兩 , (4)

with less than 6 hr between them were excluded from analysis, where A is the percentage of accurately classified data epochs,
except for the first seizure in each cluster. In addition, all processed where yi, is the true classification of the ith record (1 ⫽ preseizure,
data segments had to be preceded and followed by at least 6 hr of 0 ⫽ baseline), ŷi is the ith system’s classification of the same event
uninterrupted, contiguous EEG, containing no artifacts or gaps of (1 ⫽ preseizure, 0 ⫽ baseline), and N is the total number of records
5 min duration or longer. These strict criteria resulted in elimination classified.
of approximately 1/3 of baseline and preseizure data from the analy-
sis in the three patients evaluated in this portion of the study. Data Handling and Processing
In the study of accumulated energy, all preseizure epochs meeting The large computational burden of these experiments required
criteria for inclusion were used. Acceptable baseline epochs were adopting a system of procedures to deal with the more than 100
divided into hour-long segments, assigned numbers, and then ran- CDs, each containing 480 MB of digital data, processed for the
domly selected for analysis. To be accepted, preseizure and base- above experiments. A custom toolbox of programs, written in MATLAB
line epochs had to be preceded by at least 3 hr of uninterrupted and C⫹⫹ was developed, with an emphasis on flexibility, automa-
data, excluding brief artifacts lasting less than 10 s. This excluded tion of feature extraction from raw data, and tabulation of results.
14 preseizure epochs from processing. A 15th seizure that arose Two major groups of programs were developed: generic programs
outside of the temporal lobes (patient #5) was not included in the that were run on all data sets, including feature extraction, window
analysis. Not all excluded data epochs were rejected due to artifact. length adjustment, and data shifting routines; and patient-specific
Some occurred near the beginning or end of a tape where previous programs used to generate a set of individual features for each data
or trailing data were lost, interrupted, or obscured. Additionally, epoch and to compile results for each trial. Batch programs were
only one baseline epoch was allowed from a single 3 hr period, to generally run overnight on the network, and results were automati-
eliminate biasing results. Baseline epochs had to be recorded at cally saved into patient-specific work folders labeled by patient
least 80% during sleep or wakefulness. Target numbers for base- number. Of great importance, descriptive naming conventions were
lines were arbitrarily set at six each collected during sleep and established for each feature generated from each patient’s data,
wakefulness for each patient in this study. Table 3 demonstrates all for each data epoch, for different states of consciousness, and for
usable, randomly chosen baseline segments available for pro- different data classes (preseizure, baseline, control channels, etc.).
cessing, using the stringent selection criteria listed above. These naming conventions were vital to handling the enormous
amount of data processed for this study. Other issues, such as
Calculation of Signal Energy and Accumulated Energy organization and control of access to the CD library of raw data,
Signal energy was calculated in a sliding window of 1.25 s to assure organization of files on computer hard disk drives, and assembling
signal stationarity, chosen empirically, based upon the work of Qin descriptive logs of clinical and quantitative data for each patient
(1995). The instantaneous energy was obtained by squaring each were labor intensive, but essential to maintaining easy access to
value of the intracranial EEG sequence as follows: raw and preprocessed data, calculated features, and overall results.

Ei[n] ⫽ x2[n], (1) Acknowledgments

where x[n] is the nth value of the intracranial EEG signal in the focus This work is supported by funding from the Whitaker Foundation,
channel within the ith time window, indexed by time. Average energy Epilepsy Foundation, American Epilepsy Society, University of
values were computed by averaging every 250 points (indexed by Pennsylvania Research Foundation, Jim Jacoby, and the National
n in Equation 1 above), equivalent to 1.25 s of the instantaneous Institutes of Health grant #MH-62298RO1. Drs. Litt, Esteller, Echauz,
Precursors of Temporal Lobe Seizures
63

and Vachtsevanos are founders of a small company, IntelliMedix, dependencies in the occurrences of epileptic seizures: a nonlinear
which supports this research. The authors express their apprecia- approach. Epilepsy Res. 17, 81–94.
tion to Steven Cranstoun, Peter Crino, Francisco Gonzalez-Scarano, Iasemidis, L., Pappas, K., Gilmore, R., Roper, S., and Sackellares, J.
Steven Scherer, Dasakumar Navaratnam, Leif Finkel, and David (1996). Preictal entrainment of a critical cortical mass is a necessary
Raizen for reviewing this manuscript. condition for seizure occurrence. Epilepsia 37, 90.
Iasemidis, L., Gilmore, R., Roper, S., and Sackellares, J. (1997).
Received February 1, 2000; revised February 8, 2001.
Dynamical interaction of the epileptogenic focus with extrafocal
sites in temporal lobe epilepsy (TLE). Ann. Neurol. 42, 429.
References
Kafri, T., van Praag, H., Gage, F.H., and Verma, I.M. (2000). Lentiviral
Bailey, C.H., Bartsch, D., and Kandel, E.R. (1996). Toward a molecu- vectors: regulated gene expression. Mol. Ther. 1, 516–521.
lar definition of long-term memory storage. Proc. Natl. Acad. Sci. Katz, A., Marks, D., McCarthy, G., and Spencer, S. (1991). Does
USA 93, 13445–13452. interictal spiking change prior to seizures? Electroencephalog. Clin.
Burgess, D.L., and Noebels, J.L. (1999). Single gene defects in mice: Neurophysiol. 79, 153–156.
the role of voltage-dependent calcium channels in absence models. Lange, H.H., Lieb, J., Jr., Engel, J., and Crandall, P. (1983). Temporo-
Epilepsy Res. 36, 111–122. spatial patterns of pre-ictal spike activity in human temporal lobe
Dichter, M., and Spencer, W.A. (1969a). Penicillin-induced interictal epilepsy. Electroencephalog. Clin. Neurophysiol. 56, 543–555.
discharges from the cat hippocampus. I. Characteristics and topo- Le Van Quyen, M., Martinerie, J., Navarro, V., Boon, P., D’Havé, M.,
graphical features. J. Neurophysiol. 32, 649–662. Adam, C., Renault, B., Varela, F., and Baulac, M. (2001). Anticipation
Dichter, M., and Spencer, W.A. (1969b). Penicillin-induced interictal of epileptic seizures from standard EEG recordings. The Lancet 357,
discharges from the cat hippocampus. II. Mechanisms underlying 183.
origin and restriction. J. Neurophysiol. 32, 663–687. Lehnertz, K., and Elger, C. (1998). Can epileptic seizures be pre-
Dichter, M.A., Herman, C.J., and Selzer, M. (1972). Silent cells during dicted? Evidence from nonlinear time series analysis of brain electri-
interictal discharges and seizures in hippocampal penicillin foci. cal activity. Phys. Rev. Lett. 80, 5019–5022.
Evidence for the role of extracellular K⫹ in the transition from the Liang, F., and Jones, E.G. (1997). Zif268 and Fos-like immunoreactiv-
interictal state to seizures. Brain Res. 48, 173–183. ity in tetanus toxin-induced epilepsy: reciprocal changes in the epi-
Dichter, M.A., and Ayala, G.F. (1987). Cellular mechanisms of epi- leptic focus and the surround. Brain Res. 778, 281–292.
lepsy: a status report. Science 237, 157–164. Lopes da Silva, F.H., Pijn, J.P., and Wadman, W.J. (1994). Dynamics
Echauz, J., Kim, S., Ramani, V., and Vachtsevanos, G. (1998). Neuro- of local neuronal networks: control parameters and state bifurca-
fuzzy approaches to decision making: a comparative study with an tions in epileptogenesis. Prog. Brain Res. 102, 359–370.
application to check authorization. Journal of Intelligent and Fuzzy Lothman, E.W., and Williamson, J.M. (1994). Closely spaced recur-
Systems 6, 259–278. rent hippocampal seizures elicit two types of heightened epilepto-
Eder, H.G., Stein, A., and Fisher, R.S. (1997). Interictal and ictal genesis: a rapidly developing, transient kindling and a slowly devel-
activity in the rat cobalt/pilocarpine model of epilepsy decreased oping, enduring kindling. Brain Res. 649, 71–84.
by local perfusion of diazepam. Epilepsy Res. 29, 17–24. Martinerie, J., Adam, C., Quyen, M.L.V., Baulac, M., Clemenceau,
Engel, J., Jr. (1987). Outcome with respect to seizures. In Surgical S., Renault, B., and Varela, F. (1998). Epileptic seizures can be antici-
Treatment of the Epilepsies, J. Engel, Jr., ed. (New York: Raven pated by non-linear analysis. Nat. Med. 4, 1173–1176.
Press), pp. 553–571. Noebels, J.L. (1999). Single-gene models of epilepsy. Adv. Neurol.
Fanselow, E., Reid, A., and Nicolelis, M. (2000). Reduction of pentyl- 79, 227–238.
enetetrazole-induced seizure activity in awake rats by seizure-trig- Noebels, J.L., and Rutecki, P.A. (1990). Altered hippocampal net-
gered trigeminal nerve stimulation. J. Neurosci. 20, 8160–8168. work excitability in the hypernoradrenergic mutant mouse tottering.
Feindel, W., Penfield, W., and Jasper, H. (1952). Localization of Brain Res. 524, 225–230.
epileptic discharge in temporal lobe automatism. Trans. Am. Neurol. Noebels, J.L., Qiao, X., Bronson, R.T., Spencer, C., and Davisson,
Assoc. 77, 14–17. M.T. (1990). Stargazer: a new neurological mutant on chromosome
Franaszcuk, P., and Bergey, G. (1999). An autoregressive method 15 in the mouse with prolonged cortical seizures. Epilepsy Res. 7,
for the measurement of synchronization of interictal and ictal EEG 129–135.
signals. Biol. Cybernet. 81, 3–9. Osorio, I., Frei, M., and Wilkinson, S. (1998). Real-time automated
Glanz, J. (1994). Do chaos-control techniques offer hope for epi- detection and quantitative analysis of seizures and short-term pre-
lepsy? Science 265, 1174. diction of clinical onset. Epilepsia 39, 615–627.
Gluckman, B., Nguyen, H., Weinstein, S., and Schiff, S. (2001). Adap- Penfield, W., and Erickson, T.C. (1941). Epilepsy and cerebral local-
tive electric field control of epileptic seizures. J. Neurosci. 21, ization (Springfield, IL: Charles C. Thomas).
590–600. Penfield, W., and Flanigin, H. (1950). Surgical therapy of temporal
Gotman, J. (1987). Computer analysis of the EEG in epilepsy. In lobe seizures. Arch. Neurol. Psychiatr. 64, 491–500.
Methods of Analysis of Brain Electrical and Magnetic Signals: EEG Qin, D. (1995). A comparison of techniques for the prediction of
Handbook, A. Gevins and A. Rémond, eds. (Amsterdam: Elsevier), epileptic seizures. In Eighth IEEE Symposium on Computer-Based
pp. 171–204. Medical Systems (Lubbock, TX: IEEE Computer Society Press), pp.
Gotman, J. (1991). Relationships between interictal spiking and sei- 151–157.
zures: human and experimental evidence. Canadian J. Neurol. Sci. Qu, H., and Gotman, J. (1997). A patient-specific algorithm for the
18, 573–576. detection of seizure onset in long-term EEG monitoring: possible
Haut, S.R., Legatt, A.D., O’Dell, C., Moshe, S.L., and Shinnar, S. use as a warning device. IEEE Transactions of Biomedical Engi-
(1997). Seizure lateralization during EEG monitoring in patients with neering 44, 115–122.
bilateral foci: the cluster effect. Epilepsia 38, 937–940. Rajna, P., Clemens, B., Csibri, E., Dobos, E., Geregely, A., Gottschal,
Heffernan, M.S. (1996). Comparative effects of microcurrent stimula- M., Gyo Horvath, A., Horvath, F, Mezofi, L., et al. (1997). Hungarian
tion on EEG spectrum and correlation dimension. Integr. Physiol. multicentre epidemiologic study of the warning and initial symptoms
Behav. Sci. 31, 202–209. (prodrome, aura) of epileptic seizures. Seizure 6, 361–368.
Iasemidis, L.D., Sackellares, J.C., Zaveri, H.P., and Williams, W.J. Ralston, B. (1958). The mechanism of transition of interictal spiking
(1990). Phase space topography and the Lyapunov exponent of foci into ictal seizure discharge. Electroencephal. Clin. Neurophys-
electrocorticograms in partial seizures. Brain Topogr. 2, 187–201. iol. 10, 217–232.
Iasemidis, L., Olson, L., Savit, R., and Sackellares, J. (1994). Time Risinger, M., Engel, J.J., VanNess, P., Henry, T., and Crandall, P.
Neuron
64

(1989). Ictal localization of temporal seizures with scalp-sphenoidal

recordings. Neurology 39, 1288–1293.
Rogowski, Z., Gath, I., and Bental, E. (1981). On the prediction of
epileptic seizures. Biol. Cybernet. 42, 9–15.
Schiller, Y., Cascino, G., Busacker, N., and Sharbrough, F. (1998).
Characterization and comparison of local onset and remote propa-
gated electrographic seizures recorded with intracranial electrodes.
Epilepsia 39, 380–388.
Schuh, L.A., Henry, T.R., Ross, D.A., Smith, B.J., Elisevich, K., and
Drury, I. (2000). Ictal spiking patterns recorded from temporal depth
electrodes predict good outcome after anterior temporal lobectomy.
Epilepsia 41, 316–319.
Schwartzkroin, P.A., Baraban, S.C., and Hochman, D.W. (1998). Os-
molarity, ionic flux, and changes in brain excitability. Epilepsy Res
32, 275–285.
Spencer, S., Guimaraes, P., Katz, K., Kim, J., and Spencer, D. (1992).
Morphological patterns of seizures recorded intracranially. Epilepsia
33, 537–545.
Sperling, M. (1986). Electrophysiology of the ictal-interictal transition
in humans. In Mechanisms of Epileptogenesis: The Transition to
Seizure, M.A. Dichter, ed. (New York, NY: Plenum Press), pp. 17–38.
Sperling, M., and O’Connor, M. (1990). Auras and subclinical sei-
zures: characteristics and prognostic significance. Ann. Neurol. 28,
320–328.
Stein, A.G., Eder, H.G., Blum, D.E., Drachev, A., and Fisher, R.S.
(2000). An automated drug delivery system for focal epilepsy. Epi-
lepsy Res. 39, 103–114.
Tang, C.M., Dichter, M., and Morad, M. (1990). Modulation of the
N-methyl-D-aspartate channel by extracellular H⫹. Proc. Natl. Acad.
Sci. USA 87, 6445–6449.
Viglione, S., Ordon, V., and Risch, F. (1970). A methodology for
detecting ongoing changes in the EEG prior to clinical seizures (West
Huntington Beach, CA: McDonnel Douglas Astronautics Co.).
Wang, L., Nichols, T.C., Read, M.S., Bellinger, D.A., and Verma,
I.M. (2000). Sustained expression of therapeutic level of factor IX in
hemophilia B dogs by AAV-mediated gene therapy in liver. Mol.
Ther. 1, 154–158.
Ward, S., and Rise, M. (1998). Techniques for treating epilepsy by
brain stimulation and drug infusion (US Patent #5713923).