A Novel HFO-based Method for Unsupervised Localization of the
Seizure Onset Zone in Drug-Resistant Epilepsy
Paige M. Murphy*, Adam J. von Paternos*, and Sabato Santaniello, Member, IEEE
Abstract— High frequency oscillations (HFOs) are potential
biomarkers of epileptic areas. In patients with drug-resistant
epilepsy, HFO rates tend to be higher in the seizure onset zone
(SOZ) than in other brain regions and the resection of HFO-
generating areas positively correlates with seizure-free surgery
outcome. Nonetheless, the development of robust unsupervised
HFO-based tools for SOZ localization remains challenging.
Current approaches predict the SOZ by processing small
samples of intracranial EEG (iEEG) data and applying patient-
specific thresholds on the HFO rate. The HFO rate, though,
varies largely over time with the patient’s conditions (e.g., sleep
versus wakefulness) and across patients. We propose a novel
localization method for SOZ that uses a time-varying, HFO-
based index to estimate the epileptic susceptibility of the iEEG
channels. The method is insensitive to the average HFO rate
across channels (which is both patient- and condition-specific),
tracks the channel susceptibility over time, and predicts the
SOZ based on the temporal evolution of the HFO rate. Tested
on a preliminary dataset of continuous multi-day multichannel
interictal iEEG recordings from two epileptic patients (117±
97.6 h/per patient, mean ± S.D.), the reported SOZ prediction
had an average 0.70±0.18 accuracy and 0.67±0.07 area under
the ROC curve (mean ± S.D.) across patients.
I. INTRODUCTION
High frequency oscillations (HFOs, 80-500 Hz) are short
(i.e., typically 150 ms or less) events detectable in high-pass
filtered intracranial EEG (iEEG) recordings [1, 2]. HFOs are
currently investigated as potential biomarkers of epileptic
areas. Epileptiform HFOs may have distinctive shapes when
compared with normal physiologic HFOs [3], occur more
often during slow-wave sleep than in wakefulness [4, 5], and
the rate of epileptiform HFOs tend to be higher in the seizure
onset zone (SOZ) that in non-SOZ regions, both during
interictal stages and at the onset of seizures [6-10]. Moreover,
the surgical removal of high-rate HFO regions significantly
correlates with seizure-free surgery outcomes while poor
surgery outcomes have been reported when the resected
regions overlap only in part with the HFO-generating regions
[11-17]. Overall, these results indicate that the occurrence of
epileptiform HFOs can be a sign of seizure susceptibility and
higher HFO rates may indicate a greater risk of seizures [7].
Despite the interest in using HFOs as a predictor of the
SOZ, the development of unsupervised HFO-based methods
for SOZ localization remains challenging. Studies [4, 6-9, 12-
*Equal contribution to the research. Research supported in part by the
US National Science Foundation through grant NSF ECCS-1518672.
P. M. Murphy and A. J. von Paternos are with the Biomedical
Engineering Department, University of Connecticut, Storrs, CT 06269 USA
(e-mail: {paige.murphy, adam.von_paternos}@uconn.edu).
S. Santaniello is with the Biomedical Engineering Department and the
CT Institute for Brain and Cognitive Sciences, University of Connecticut,
Storrs, CT 06269 (corresponding author; phone: +1 860-486-4701; fax: +1
860-486-2500; e-mail: sabato.santaniello@uconn.edu).
978-1-5090-2809-2/17/$31.00 ©2017 IEEE 1054
16] used short intervals of iEEG recordings (i.e., less than 15
minutes of data per intervals) chosen arbitrarily, with often
only one interval per channel. Furthermore, the SOZ and
non-SOZ channels included in these studies were only a
subset of the implanted iEEG recording channels and were
chosen non-contiguous in space to better capture differences
between SOZ and non-SOZ. Finally, in [4, 6-9, 14-16], HFOs
were manually labelled by trained clinicians through visual
inspection of the iEEG time series.
Recent studies [18, 19], instead, have considered a more
clinically-relevant scenario and proposed HFO-rate-based
criteria to automatically localize the SOZ. First, each of these
solutions has included a clinically-validated algorithm for the
unsupervised detection of the HFOs. Then, for each patient,
they have included iEEG data collected across the entire set
of recording channels, thus extending their analysis to critical
regions that are located in between the clearly separable SOZ
and non-SOZ. However, the proposed solutions were tested
either on small, non-consecutive samples of iEEG recordings
[19] or on averaged HFO rates calculated over several hours
of recordings [18], thus combining interictal and peri-ictal
data. Furthermore, each of these solutions used a threshold-
based policy to identify the SOZ, i.e., a channel is part of the
SOZ if the HFO rate in that channel is higher than a chosen
threshold. Since the rate of HFOs may vary largely across
patients [12] and – in the same patient – over time [5], this
means that the method must be calibrated to the specific
patient and iEEG dataset.
In this preliminary study, we propose a novel HFO-based
method for automatic localization of the SOZ and we tested it
on multi-day multichannel iEEG data from two patients in
the iEEG Portal database [20]. Differently from current
solutions, the method uses a series of HFO rates estimated
over consecutive 10-min-long windows to build a time-
varying normalized vector index. The vector assigns a value
between 0 and 1 to each iEEG channel based on the level of
epileptogenic susceptibility in that channel and allows to
automatically identify the candidate channels for the SOZ.
We tested the algorithm on continuous interictal recordings
spanning several days per patient (i.e., periods of sleep and
wakefulness are included) and we performed a receiver
operating characteristic (ROC) analysis to determine the
optimal parameters across the population of patients involved
in the study. Data from ictal periods was initially excluded
given the low frequency of seizure events compared to the
interictal data and the specific nature of ictal HFOs [21, 22].
We found that our solution can provide robust predictions of
the SOZ with 20 minutes of consecutive iEEG recordings,
regardless of when recordings are collected (i.e., wakefulness
or sleep), and that the accuracy of the method remains
statistically significant over time (above the 95th percentile)
with average value 53.0±4.2% across patients (mean ± S.D.).
 TABLE I
EXPERIMENTAL DATASET
Subject ID Study-004-2 Study-023
Age (y)/Sex 27 / F 16 / M
PS/GTC CPS
Type of Seizures
Engel Class IV I
No. iEEG channels 56 88
Seizure Focus RT LO
Duration of iEEG
186 48
recordings (h)
Legend: F= female; M= male; PS= partial seizure; GTC= generalized tonic-
clonic seizure; CPS= complex partial seizure; RT= right temporal lobe;
LO= left occipital lobe.
II. METHODS
A. Dataset and HFO Detection
Continuous multichannel iEEG recordings from two
patients in the iEEG Portal database (https://www.ieeg.org/)
[20] were included in this study and processed in common-
reference mode. See Table I for details. For each patient, the
channels covering the SOZ were identified as described in
[23] by board-certified epileptologists along with the clinical
seizures. For each patient, iEEG recordings collected at least
2 hours after the last clinical seizure were included in this
study (interictal data) and sampled at 500 Hz. The decision of
excluding recordings collected before the first seizure event
or between seizures was made to guarantee that the analysis
is conducted on consecutive interictal data, thus representing
a scenario where seizures are rare events and the SOZ must
be localized with no prior information about the occurrence
of clinical seizures. Data from ictal periods was excluded as
well given the low frequency of seizure events compared to
the interictal data and the specific nature of ictal HFOs [21,
22]. Inclusion of ictal data in our analysis, however, did not
significantly affect results reported in section III.
HFO were detected on each iEEG channel independently
by using the unsupervised algorithm proposed by Staba et al.
[24] with artifact removal, where the artifacts were detected
as in [18]. Briefly, each iEEG time series was divided in
consecutive, non-overlapping 10-min-long intervals and, for
each interval, the signal was high-pass filtered at 100 Hz
(fourth order Butterworth filter). HFOs were then detected as
oscillatory events in the filtered signal that last 100ms or less
and contain at least 6 peaks, each peak having amplitude of at
least 10 µV (absolute value) and being at least 3 times the
standard deviation of the mean rectified signal. The algorithm
was also run on the common average reference and HFOs
detected on this signal were considered artifact and removed
from the analysis.
B. Index of Epileptic Susceptibility and ROC Analysis
Denoted with N the number of iEEG channels recorded in
a generic 10-min-long interval W, we estimated the HFO rate
fj of channel j, j=1, 2, 3, …, N, as the average number of
detected HFOs per minute in the channel and we ranked all
the channels based on the rate values, i.e., the higher the rate
the lower the rank position of the channel. Let us denote with
R = [r1 r2 r3 … rN] the vector of rank positions, where rj is the
rank of channel j, j=1, 2, 3, …, N, and 1 ≤ rj ≤ N. For each
patient, the rank vector R was recalculated for consecutive,
1055
Figure 1. A-B) Average HFO rate (mean±S.D.) for each iEEG channel
estimated across consecutive 10-min-long windows in Study-004-2 (A) and
Study-023 (B). C-D) Box-plot of the components of Sk corresponding to the
SOZ and non-SOZ regions across consecutive windows in Study-004-2 (C)
and Study-023 (D). For each box-plot, the median (red line), the 25th and
75th percentiles (blue lines), the 98.5th percentile (whiskers), and outliers
(red asterisks) are reported.
non-overlapping time windows Wk, k=1, 2, 3,.., i.e., R = Rk,
and the vector index
Sk = exp(1– Rk) ∙ exp(1– Rk–1) (1)
was introduced. Intuitively, vector Sk aims to characterize the
level of epileptic susceptibility [25] in each channel in the
window Wk. The components of Sk are uniquely related to the
iEEG channels (i.e., the j-th component corresponds to
channel j, j=1, 2, …, N), they all range between 0 and 1, and
low values correspond to discharge patters with occasional
HFOs (i.e., low susceptibility) while high values correspond
to sequences of HFO rates that are consistently above the
average trend across channels (i.e., high susceptibility). In
(1), the epileptic susceptibility of a channel is associated with
the occurrence of HFOs in the previous 20 minutes (i.e., the
current window and the previous window) and the relative
rate of HFOs compared to the remaining iEEG channels. The
number of past windows included in (1) was chosen by
minimizing the fluctuations of the susceptibility index on a
training set including 3 h/channel of interictal activity (i.e.,
18 windows) per patient.
To assess the predictive power of index Sk, we performed
a receiver operating characteristic (ROC) curve analysis [26]
for each window Wk across both patients. We also repeated
the ROC curve analysis over the entire set of windows Wk for
each patient, thus determining the average performance over
multiple days. Regions that were either surgically removed or
otherwise marked as belonging to the SOZ in the clinical
notes were labelled as “true positives” and all the remaining
regions were “true negatives”. For each Sk, channels were
marked as part of the SOZ if the correspondent components
in Sk were above a threshold α, and the ROC curve was built
by varying α in [0,1]. For each ROC curve, the area under the
curve (AUC) and the values of sensitivity, specificity, and
accuracy at cutoff point (i.e., the point maximizing the
Youden index [27]) were stored.
The significance of the AUC values was evaluated using
the nonparametric permutation test described in [28]. Briefly,
the components of each vector Sk, k=1, 2, …, were randomly
Figure 2. A-B) ROC curves for the vector index Sk (black lines) and the
vector of average HFO rates Fk (blue lines), respectively, estimated on the
entire dataset for Study-004-2 (A) and Study-023 (B). Curves are reported
along with the 99% confidence intervals. In A-B), cutoff points are marked
with red circles and estimated AUC values are reported in the legend.
sampled without replacement (permutation) and the ROC
analysis was repeated for each permutation. The procedure
was repeated 1,000 times per vector to generate an ensemble
of simulated permuted AUC values and, for each ensemble,
the mean permuted AUC value and 95% confidence intervals
were computed. Finally, a vector Sk was deemed indicative of
the SOZ (P-value, P<0.05) if the correspondent AUC value
was above the upper 95% bound of the mean permuted AUC
value. The threshold α* that maximized the average AUC
value across all SOZ-indicative vectors Sk and patients was
then selected.
III. RESULTS
Two subjects (ID: Study-004-2 and Study-023) were
considered. Both subjects underwent surgery to remove the
SOZ and both experienced a substantial amelioration of the
overall clinical conditions, even though only patient Study-
023 was reported seizure-free after surgery (Engel class I
[29], Table I). In Study-004-2, the annotated SOZ included
channels 11-12, 19, 40-41, 45-48, and 53-56, which were
located on the right temporal lobe. In Study-023, the
annotated SOZ included channels 7, 38, 45, 64, 66-68, 73-75,
and 77, which covered the left occipital lobe.
A total of 9,608 and 2,138 consecutive, non-overlapping
10-min-long intervals were processed for HFO detection in
Study-004-2 and Study-023, respectively. Fig. 1A-B reports
the average HFO rate in each channel and patient (average
across the intervals). Consistently with previous studies [12,
14, 18], the channels with the highest HFO rates overlapped
with the SOZ and the average rate per channel was largely
dependent on the patient (Study-004-2: 0.15±0.11 HFOs/min;
Study-023: 0.22±0.08 HFOs/min; mean ± S.D.). However,
several SOZ and non-SOZ channels had comparable HFO
rates (Fig. 1A-B), which can be explained by observing that
HFOs may consistently occur in several brain regions outside
the epileptic network [30]. The similarity between HFO rates
in several SOZ and non-SOZ channels, though, limited the
ability to identify the SOZ based on the sample probability
distribution function of the HFO rates. To further support this
statement, we considered the vector Fk = [f1,k f2,k … fN,k] of
average HFO rates in the window Wk, for k=1,2,.., and we
performed the ROC curve analyses outlined in section II.B
for all vectors Fk. In both patients, the predictive power of
vector Fk across the entire dataset was slightly above chance
(Fig. 2A-B), with low values of sensitivity and specificity at
cutoff point (Fig. 2A-B, red circles). In both patients, instead,
the components of the vector index Sk were, on average,
1056
Figure 3. A-D) ROC curve analysis performed on consecutive 10-min-long
windows in Study-004-2 (A,C) and Study-023 (B,D), respectively, for the
vector index Sk (A,B) and the vector of average HFO rates Fk (C,D). In A-
D), black lines denote the sequences of estimated AUC values (one value
per window), shaded red lines and strips denote the mean permuted AUC
values and the simulation-derived 95% confidence intervals, respectively.
significantly higher for the SOZ channels than the non-SOZ
channels (Fig. 1C-D; two-sample t-test, P-value P<0.05).
Moreover, the average values of Sk for SOZ and non-SOZ
channels were similar in both patients, with small prevalence
in Study-023 (SOZ: 0.033±0.02 for Study-004-2 and 0.045±
0.03 for Study-023; non-SOZ: 0.0046±0.004 for Study-004-2
and 0.0013±0.001 for Study-023; mean ± S.D.), which may
indicate the robustness of the index across patients having
different patterns of HFO activity. Finally, the ROC curve
analysis performed over the entire dataset (Fig. 2A-B) shows
that, in both patients, the index Sk outperformed the rate
vector Fk in terms of AUC value, sensitivity, and specificity
(Z-test, P-value P<0.01).
Furthermore, we investigated the temporal evolution of Sk
and the stability of the predictions based on this index over
consecutive intervals. Fig. 3A-B report the results of the
ROC analysis for Study-004-2 and Study-023, respectively,
over consecutive intervals. In both patients, the AUC value
was consistently above 0.6 and high AUC values (i.e., above
0.7) were reported in 69.5±21.1% of the intervals (Study-
004-2: 84.4%; Study-023: 54.6%). More importantly, the
index-based prediction of the SOZ was significantly above
chance (i.e., the AUC was above the 95% upper bound of the
permuted AUC, which corresponds to the chance level) in
97.7% of intervals in Study-004-2 and all the intervals in
Study-023, which suggests that the predictive value of the
index was stable in time and comparable across patients. To
test whether index Sk outperforms the average HFO rate, and
we repeated the ROC curve analysis outlined in section II.B
for vectors Fk. Results are reported in Fig. 3C-D and indicate
that, in our specific set of patients, the average HFO rates
were poorly predictive of the SOZ, which is consistent with
the observation that several SOZ and non-SOZ channels had
comparable HFO rates. For the ROC curves used in Fig. 3A-
B, the average values of sensitivity, specificity, and accuracy
at cutoff point across the available intervals were 0.77±0.12,
0.84±0.29, and 0.82±0.13, respectively (Study-004-2), and
0.91±0.17, 0.52±0.26, and 0.57±0.11 respectively (Study-
023), mean ± S.D., with similar trends across patients.
 Finally, we chose the threshold α* that maximizes the
average AUC value across the entire dataset (i.e., all intervals
and patients) and we assessed the performance of such
threshold on each patient individually. We found that the
average prediction accuracy remained high in both patients
(Study-004-2: 0.80±0.12; Study-023: 0.85±0.1; mean ± S.D.)
along with the sensitivity (Study-004-2: 0.77±0.12; Study-
023: 0.45±0.17) and specificity (Study-004-2: 0.81±0.29;
Study-023: 0.91±0.26), thus suggesting that the proposed
index may need minimal customization to the patient.
IV. CONCLUSIONS
In this study, we propose a novel method to automatically
localize the SOZ in patients with multichannel iEEG. We use
a time-varying HFO-ranking-based index to measure the
epileptogenic susceptibility of each iEEG channel and we
choose candidate SOZ channels based on the value of the
index over a 20-min-long interval. Tested on a preliminary
dataset of two patients, the method was robust against
fluctuations in the HFO rate over time, showed consistent
results across patients, and provided predictions of the SOZ
that were significantly more accurate than a traditional HFO-
count-based approach.
REFERENCES
[1] J. D. Jirsch, E. Urrestarazu, P. LeVan, A. Olivier, F. Dubeau, and J.
Gotman, "High-frequency oscillations during human focal seizures,"
Brain, vol. 129, pp. 1593-608, Jun 2006.
[2] J. Jacobs, R. Staba, E. Asano, H. Otsubo, J. Y. Wu, M. Zijlmans, I.
Mohamed, P. Kahane, F. Dubeau, V. Navarro, and J. Gotman, "High-
frequency oscillations (HFOs) in clinical epilepsy," Prog Neurobiol,
vol. 98, pp. 302-15, Sep 2012.
[3] A. Matsumoto, B. H. Brinkmann, S. Matthew Stead, J. Matsumoto, M.
T. Kucewicz, W. R. Marsh, F. Meyer, and G. Worrell, "Pathological
and physiological high-frequency oscillations in focal human
epilepsy," J Neurophysiol, vol. 110, pp. 1958-64, Oct 2013.
[4] A. P. Bagshaw, J. Jacobs, P. LeVan, F. Dubeau, and J. Gotman,
"Effect of sleep stage on interictal high-frequency oscillations
recorded from depth macroelectrodes in patients with focal epilepsy,"
Epilepsia, vol. 50, pp. 617-28, Apr 2009.
[5] M. Dumpelmann, J. Jacobs, and A. Schulze-Bonhage, "Temporal and
spatial characteristics of high frequency oscillations as a new
biomarker in epilepsy," Epilepsia, vol. 56, pp. 197-206, Feb 2015.
[6] E. Urrestarazu, R. Chander, F. Dubeau, and J. Gotman, "Interictal
high-frequency oscillations (100-500 Hz) in the intracerebral EEG of
epileptic patients," Brain, vol. 130, pp. 2354-66, Sep 2007.
[7] J. Jacobs, R. Zelmann, J. Jirsch, R. Chander, C. E. Dubeau, and J.
Gotman, "High frequency oscillations (80-500 Hz) in the preictal
period in patients with focal seizures," Epilepsia, vol. 50, pp. 1780-92,
Jul 2009.
[8] J. A. Blanco, M. Stead, A. Krieger, W. Stacey, D. Maus, E. Marsh, J.
Viventi, K. H. Lee, R. Marsh, B. Litt, and G. A. Worrell, "Data
mining neocortical high-frequency oscillations in epilepsy and
controls," Brain, vol. 134, pp. 2948-59, Oct 2011.
[9] L. Andrade-Valenca, F. Mari, J. Jacobs, M. Zijlmans, A. Olivier, J.
Gotman, and F. Dubeau, "Interictal high frequency oscillations
(HFOs) in patients with focal epilepsy and normal MRI," Clin
Neurophysiol, vol. 123, pp. 100-5, Jan 2012.
[10] U. Malinowska, G. K. Bergey, J. Harezlak, and C. C. Jouny,
"Identification of seizure onset zone and preictal state based on
characteristics of high frequency oscillations," Clin Neurophysiol, vol.
126, pp. 1505-13, Aug 2015.
[11] T. Akiyama, B. McCoy, C. Y. Go, A. Ochi, I. M. Elliott, M. Akiyama,
E. J. Donner, S. K. Weiss, O. C. Snead, 3rd, J. T. Rutka, J. M. Drake,
and H. Otsubo, "Focal resection of fast ripples on extraoperative
intracranial EEG improves seizure outcome in pediatric epilepsy,"
Epilepsia, vol. 52, pp. 1802-11, Oct 2011.
1057
[12] J. R. Cho, D. L. Koo, E. Y. Joo, D. W. Seo, S. C. Hong, P. Jiruska,
and S. B. Hong, "Resection of individually identified high-rate high-
frequency oscillations region is associated with favorable outcome in
neocortical epilepsy," Epilepsia, vol. 55, pp. 1872-83, Nov 2014.
[13] H. Fujiwara, H. M. Greiner, K. H. Lee, K. D. Holland-Bouley, J. H.
Seo, T. Arthur, F. T. Mangano, J. L. Leach, and D. F. Rose,
"Resection of ictal high-frequency oscillations leads to favorable
surgical outcome in pediatric epilepsy," Epilepsia, vol. 53, pp. 1607-
17, Sep 2012.
[14] J. Jacobs, M. Zijlmans, R. Zelmann, C. E. Chatillon, J. Hall, A.
Olivier, F. Dubeau, and J. Gotman, "High-frequency
electroencephalographic oscillations correlate with outcome of
epilepsy surgery," Ann Neurol, vol. 67, pp. 209-20, Feb 2010.
[15] P. N. Modur, S. Zhang, and T. W. Vitaz, "Ictal high-frequency
oscillations in neocortical epilepsy: implications for seizure
localization and surgical resection," Epilepsia, vol. 52, pp. 1792-801,
Oct 2011.
[16] K. Kerber, M. Dumpelmann, B. Schelter, P. Le Van, R.
Korinthenberg, A. Schulze-Bonhage, and J. Jacobs, "Differentiation of
specific ripple patterns helps to identify epileptogenic areas for
surgical procedures," Clin Neurophysiol, vol. 125, pp. 1339-45, Jul
2014.
[17] C. Haegelen, P. Perucca, C. E. Chatillon, L. Andrade-Valenca, R.
Zelmann, J. Jacobs, D. L. Collins, F. Dubeau, A. Olivier, and J.
Gotman, "High-frequency oscillations, extent of surgical resection,
and surgical outcome in drug-resistant focal epilepsy," Epilepsia, vol.
54, pp. 848-57, May 2013.
[18] S. V. Gliske, Z. T. Irwin, K. A. Davis, K. Sahaya, C. Chestek, and W.
C. Stacey, "Universal automated high frequency oscillation detector
for real-time, long term EEG," Clin Neurophysiol, vol. 127, pp. 1057-
66, Feb 2016.
[19] S. Liu, Z. Sha, A. Sencer, A. Aydoseli, N. Bebek, A. Abosch, T.
Henry, C. Gurses, and N. F. Ince, "Exploring the time-frequency
content of high frequency oscillations for automated identification of
seizure onset zone in epilepsy," J Neural Eng, vol. 13, p. 026026, Apr
2016.
[20] A. A. Azarion, J. Wu, A. Pearce, V. T. Krish, J. Wagenaar, W. Chen,
Y. Zheng, H. Wang, T. H. Lucas, B. Litt, J. C. Gee, and K. A. Davis,
"An open-source automated platform for three-dimensional
visualization of subdural electrodes using CT-MRI coregistration,"
Epilepsia, vol. 55, pp. 2028-37, Dec 2014.
[21] S. A. Weiss, G. P. Banks, G. M. McKhann, Jr., R. R. Goodman, R. G.
Emerson, A. J. Trevelyan, and C. A. Schevon, "Ictal high frequency
oscillations distinguish two types of seizure territories in humans,"
Brain, vol. 136, pp. 3796-808, Dec 2013.
[22] P. Perucca, F. Dubeau, and J. Gotman, "Intracranial
electroencephalographic seizure-onset patterns: effect of underlying
pathology," Brain, vol. 137, pp. 183-96, Jan 2014.
[23] F. Rosenow and H. Luders, "Presurgical evaluation of epilepsy,"
Brain, vol. 124, pp. 1683-700, Sep 2001.
[24] R. J. Staba, C. L. Wilson, A. Bragin, I. Fried, and J. Engel, Jr.,
"Quantitative analysis of high-frequency oscillations (80-500 Hz)
recorded in human epileptic hippocampus and entorhinal cortex," J
Neurophysiol, vol. 88, pp. 1743-52, Oct 2002.
[25] F. Bartolomei, P. Chauvel, and F. Wendling, "Epileptogenicity of
brain structures in human temporal lobe epilepsy: a quantified study
from intracerebral EEG," Brain, vol. 131, pp. 1818-30, Jul 2008.
[26] T. Hastie, R. Tibshirani, and J. H. Friedman, The elements of
statistical learning : data mining, inference, and prediction, 2nd ed.
New York, NY: Springer, 2009.
[27] W. J. Youden, "Index for rating diagnostic tests," Cancer, vol. 3, pp.
32-5, Jan 1950.
[28] S. P. Burns, S. Santaniello, R. B. Yaffe, C. C. Jouny, N. E. Crone, G.
K. Bergey, W. S. Anderson, and S. V. Sarma, "Network dynamics of
the brain and influence of the epileptic seizure onset zone," Proc Natl
Acad Sci U S A, vol. 111, pp. E5321-30, Dec 09 2014.
[29] J. Engel, Surgical treatment of the epilepsies, 2nd ed. New York:
Raven Press, 1993.
[30] R. Alkawadri, N. Gaspard, Goncharova, II, D. D. Spencer, J. L.
Gerrard, H. Zaveri, R. B. Duckrow, H. Blumenfeld, and L. J. Hirsch,
"The spatial and signal characteristics of physiologic high frequency
oscillations," Epilepsia, vol. 55, pp. 1986-95, Dec 2014.