Content uploaded by Chao Li
Author content
All content in this area was uploaded by Chao Li on Mar 06, 2024
Content may be subject to copyright.
Content uploaded by Bowen Yang
Author content
All content in this area was uploaded by Bowen Yang on Jan 25, 2024
Content may be subject to copyright.
Localizing seizure onset zone by a cortico-cortical evoked potentials-
based machine learning approach in focal epilepsy
Bowen Yang
a
, Baotian Zhao
a
, Chao Li
b
, Jiajie Mo
a
, Zhihao Guo
a
, Zilin Li
a
, Yuan Yao
a
, Xiuliang Fan
a
,
Du Cai
a
, Lin Sang
c
, Zhong Zheng
c
, Dongmei Gao
a
, Xuemin Zhao
d
, Xiu Wang
a,e
, Chao Zhang
a,e
,
Wenhan Hu
a,e
, Xiaoqiu Shao
f
, Jianguo Zhang
a,e
, Kai Zhang
a,e,
⇑
a
Department of Neurosurgery, Beijing Tiantan Hospital, Capital Medical University, Beijing, China
b
Department of Radiology, Third Affiliated Hospital of Sun Yat-sen University, Sun Yat-sen University, Guangzhou, China
c
Department of Neurosurgery, Beijing Fengtai Hospital, Beijing, China
d
Department of Neurophysiology, Beijing Neurosurgical Institute, Capital Medical University, Beijing, China
e
Stereotactic and Functional Neurosurgery Laboratory, Beijing Neurosurgical Institute, Capital Medical University, Beijing, China
f
Department of Neurology, Beijing Tiantan Hospital, Capital Medical University, Beijing, China
highlights
A novel automated approach was proposed for localizing seizure onset zone (SOZ) by cortico-cortical evoked potentials (CCEPs).
Cortical hyperexcitability at the stim-SOZ/rec-SOZ sites was increased in patients with hippocampus sclerosis compared to those with focal cortical
dysplasia IIa.
The patients with complex partial seizures have more distinctive characteristics of CCEP than those with generalized tonic-clonic seizures in areas with
different epileptogenicity.
article info
Article history:
Accepted 19 December 2023
Available online
Keywords:
Seizure onset zone
Intracranial electroencephalography
Cortico-cortical evoked potentials
Machine learning
Presurgical evaluation
abstract
Objective: We aimed to develop a new approach for identifying the localization of the seizure onset zone
(SOZ) based on corticocortical evoked potentials (CCEPs) and to compare the connectivity patterns in
patients with different clinical phenotypes.
Methods: Fifty patients who underwent stereoelectroencephalography and CCEP procedures were
included. Logistic regression was used in the model, and six CCEP metrics were input as features: root
mean square of the first peak (N1RMS) and second peak (N2RMS), peak latency, onset latency, width
duration, and area.
Results: The area under the curve (AUC) for localizing the SOZ ranged from 0.88 to 0.93. The N1RMS val-
ues in the hippocampus sclerosis (HS) group were greater than that of the focal cortical dysplasia (FCD)
IIa group (p< 0.001), independent of the distance between the recorded and stimulated sites. The sensi-
tivity of localization was higher in the seizure-free group than in the non-seizure-free group (p= 0.036).
Conclusions: This new method can be used to predict the SOZ localization in various focal epilepsy phe-
notypes.
Significance: This study proposed a machine-learning approach for localizing the SOZ. Moreover, we
examined how clinical phenotypes impact large-scale abnormality of the epileptogenic networks.
Ó2024 International Federation of Clinical Neurophysiology. Published by Elsevier B.V. All rights
reserved.
1. Introduction
Approximately 30% of patients with epilepsy have persistent
seizures despite taking antiepileptic drugs, known as refractory
epilepsy (Kwan and Brodie, 2000). Surgery is a potentially curative
option for some of these patients, requiring the removal of the
https://doi.org/10.1016/j.clinph.2023.12.135
1388-2457/Ó2024 International Federation of Clinical Neurophysiology. Published by Elsevier B.V. All rights reserved.
⇑
Corresponding author at: Department of Neurosurgery, Beijing Tiantan Hospi-
tal, Capital Medical University, No.119 South 4th Ring West Road, Fengtai District,
Beijing, China. Fax: +86 10 59976713.
E-mail address: zhangkai62035@163.com (K. Zhang).
Clinical Neurophysiology 158 (2024) 103–113
Contents lists available at ScienceDirect
Clinical Neurophysiology
journal homepage: www.elsevier.com/locate/clinph
seizure onset zone (SOZ) to achieve a seizure-free (SF) status.
Delineating the SOZ using only noninvasive electroencephalog-
raphy (EEG) is challenging, and intracranial EEG is essential for
identifying the SOZ responsible for seizure generation (Sugano
et al., 2021). However, conventional stereoelectroencephalography
(SEEG) monitoring has some limitations such as the risk of intracra-
nial infection due to the long waiting time for a spontaneous seizure
and the possibility of failure to capture habitual seizures. Therefore,
there is an urgent need for an SEEG-based approach to localize the
SOZ that does not depend on spontaneous seizures.
In addition to ictal signals (Enatsu et al., 2012; Iwasaki et al.,
2010; Mouthaan et al., 2016), increasing attention has been paid
to interictal EEG biomarkers (such as high-frequency oscillations
(Jacobs et al., 2010)) for localizing the SOZ. Furthermore, SOZs
derived from both ictal and interictal intracranial EEG have been
shown to coexist in regions with prominent corticocortical evoked
potentials (CCEPs) (Matsumoto et al., 2017; Mouthaan et al., 2016;
Yaffe et al., 2015), suggesting the potential value of CCEPs in repre-
senting the epileptogenicity of brain regions. In addition, the con-
nectivity patterns of CCEPs are positively correlated with
epileptogenic networks, as revealed by ictal electrical propagation
(Enatsu et al., 2012; Iwasaki et al., 2010), as well as structural neu-
roimaging data (Crocker et al., 2021; Donos et al., 2016). Notably,
CCEPs can potentially probe the epileptogenicity of large-scale
brain networks and aid in delineating seizure onset zones.
Some recent studies have explored the value of CCEPs in delin-
eating the epileptogenic network (Guo et al., 2020; Li et al., 2020;
Mouthaan et al., 2016), with varying performances using different
stimulation parameters (Hays et al., 2023; Kamali et al., 2020; van
Blooijs et al., 2018) and input features of the machine learning
models (Davis et al., 2018). Additionally, deep learning, transfer
function models, and other new approaches have been used to
explore the localization value of CCEPs in the SOZ (Kamali et al.,
2020; Smith et al., 2022; van Blooijs et al., 2018). However, these
studies have some limitations. Their limited sample size hindered
the comparison of localization performance between patients with
various clinical phenotypes. Moreover, most studies have been
conducted at the group level, limiting further clinical applications
at the patient-specific level.
In this study, (1) we established a machine learning model
based on multiple CCEP metrics to identify the SOZ at the channel
level. In addition to combining all six metrics, we separately used
each of the six CCEP metrics in the localization models and com-
pared the localizing performance to investigate the impact of CCEP
metrics. (2) Patients were grouped according to pathology, surgical
outcome, and seizure type, and then the localization model perfor-
mance and connectivity patterns of different groups were com-
pared. This session examined how clinical phenotypes impact
large-scale abnormality of the epileptogenic networks.
2. Materials and methods
2.1. Demographics
In this retrospective study, patients with refractory focal epi-
lepsy who underwent epilepsy surgery between January 2015
and December 2021 at Beijing Tiantan and Beijing Fengtai Hospi-
tals were included if they: (1) had unifocal epilepsy confirmed by
SEEG, (2) achieved worthwhile improvement (Engel I-III)
(Durnford et al., 2011) after surgery to ensure the accurate localiza-
tion of the SOZ during presurgical evaluation, and (3) underwent
CCEP procedures. Patients with seizures induced during the CCEPs
data acquisition were excluded. The hospital’s institutional review
board approved the study, and informed consent was obtained
from all participants.
2.2. Neuroimaging acquisition
Magnetic resonance imaging (MRI) was performed using a 3 T
Siemens Prisma scanner. In the preoperative session, patients
underwent a 3D T1 magnetization prepared rapid gradient echo
sequence (MPRAGE). After electrode implantation, computed
tomography (CT) was performed to identify the electrode contact
positions. Parameters used for the 3D T1 scan included: 176 sagit-
tal slices with a slice thickness of 1.0 mm, 256 256 in-plane
matrix, repetition time of 1560 ms, time to echo of 1.70 ms, inver-
sion time of 778 ms, and field-of-view of 240 240 mm
2
.
2.3. Electrical stimulation
While the subject was awake and resting, a single pulse electri-
cal stimulation (SPES) of current (6 mA, biphasic, 300
l
s per phase)
was applied between two adjacent intracranial electrode channels
at 0.5 Hz. Each stimulation trial consisted of at least 40 pulses.
Electrophysiological signals were recorded using a Nihon Kohden
system at a sampling rate of 1000 Hz.
2.4. Analysis of evoked responses
To remove stimulus artifacts, the SEEG signal samples from the
interval of 2 ms before the stimulation to 10 ms after the stimula-
tion pulse were discarded. The remaining gaps were interpolated
using autoregressive modeling and implemented through the
MATLAB fillgaps function. The time series were filtered between
1 Hz and 300 Hz using a Butterworth filter and then re-
referenced using a bipolar montage. In addition, bad trials were
rejected if the overall amplitude, denoted by the root mean square
(RMS) value, was more than three times the interquartile range
(IQR) below the first quartile (Q1) or above the third quartile
(Q3). If the channel distance was shorter than 10 mm for the
reported and stimulated channels, the CCEP responses may be con-
taminated by the stimulation. Thus, those CCEP metrics were
discarded.
Six metrics were used to evaluate the response induced by
SPES: root mean square of the first peak (N1RMS), root mean
square of the second peak (N2RMS), peak latency, onset latency,
width duration, and area (Fig. 1A). The preprocessing steps for cal-
culating CCEP metrics included two parts: (1) the first peak (N1)
and the second peak (N2) were detected automatically using the
findpeaks function in Matlab during the time range from stimula-
tion event to 200 ms after the stimulation, with a threshold for
local maxima as 5 (the normalized value). To determine if there
was a significant response, we first normalized the evoked
potential to the baseline defined as 50 to 5 ms. In detail, the
‘‘peak-start” and ‘‘peak-end” were defined by the findpeaks func-
tion automatically in Matlab. The function will tell us when the
peak started to leave the baseline and returned to the baseline;
(2) the RMS values of N1, onset latency, peak latency, width dura-
tion, and area were calculated based on the detected N1, and the
RMS values of N2 was calculated based on the detected N2.
Six CCEP metrics were calculated: (1)The root mean square of
N1 and (2) N2 were calculated using the rms function based on
the peak amplitude. (3) Onset latency was measured as the period
from the stimulus to the start point of the N1. (4) Peak latency was
measured as the period from the stimulus to the point with the
maximal absolute value of the first peak. (5) Width duration was
measured from when the peak left the baseline to when the peak
returned to the baseline. (6) The area was determined fully based
on the first peak, measuring the area under the response curve.
Considering that the peak’s direction (upward or downward) is
obscure for a bipolar montage, there was a situation when not
the whole area was above the baseline, then there might be 1
B. Yang, B. Zhao, C. Li et al. Clinical Neurophysiology 158 (2024) 103–113
104
baseline crossing within the N1. In that case, the area would be
defined as a summation of the first peak’s positive and negative
parts. In summary, the CCEP metrics reflected the connectivity
properties between two brain regions from different perspectives.
2.5. Anatomical localization of electrode contacts
The iELVis toolbox (Groppe et al., 2017) was used to visualize
the electrode contacts. Co-registration was conducted using post-
operative CT and preoperative T1 images, extracting the electrode
contact positions displayed on the postoperative CT as coordinates
and transforming them into Montreal Neurological Institute (MNI)
space. The brain regions were segmented according to the auto-
mated anatomical labeling (AAL) atlas (Tzourio-Mazoyer et al.,
2002) to reflect connectivity properties, assigning each channel
to a region based on MNI space coordinates.
Specifically, the presurgical assessment involves a comprehen-
sive consideration of the patient’s medical history, EEG, MRI, Posi-
tron Emission Tomography, and Computed Tomography (PET-CT).
Then, all patients underwent SEEG implantation surgery, and ulti-
mately, the determination of which bipolar channels are within
the SOZ is based on the onset patterns identified in the SEEG record-
ings during seizure onset (Perucca et al., 2014). If both channels of
the bipolar channels exhibit seizure onset patterns indicative of the
SOZ, then the bipolar channels are defined as being within the SOZ.
Anatomical information was used to compare connectivity
between the two groups region by region. The CCEP metrics
belonging to specific brain regions were collected. Statistical anal-
yses were performed for specific regions. In region-by-region com-
parisons, the marginal distribution patterns of the CCEPs were
analyzed separately in each category after the data were grouped
into four categories: stimulating(stim)-SOZ/recording(rec)-SOZ,
stim-SOZ/rec-nSOZ, stim-nSOZ/rec-SOZ, and stim-nSOZ/rec-nSOZ.
The four classifications are based on whether the stimulated and
recorded channels were in SOZ (Fig. 1B & 1D). To compare the
CCEP metrics between patients with different clinical phenotypes,
the metric recorded and stimulated between two of three areas
were analyzed: hippocampus, anterior cingulate gyrus (ACC), and
orbitofrontal gyrus (OFC). The reason the three regions were
selected is that they are involved in the temporal lobe epilepsy
and limbic network (Chou et al., 2020; Feng et al., 2023; Vogt,
2019).
2.6. Machine learning model for SOZ localization
A novel model for automatically localizing SOZ was constructed
following the steps below. (1) Feature selection, six CCEP metrics of
each bipolar channel were inputted (Fig. 1C), including N1RMS,
N2RMS, peak latency, onset latency, width duration, and area; (2)
the missing values were checked using the nan function in the
NumPy tool, and then replaced by the median of the specific col-
umn; (3) the unbalanced sample sizes of SOZ and non-SOZ in the
training set were balanced by upsampling using the Ran-
domOverSampler function in the Scikit-learn tool; (4) the dataset
was split into five folds according to the cross-validation approach,
with four folds used as the training set and one fold used as the
testing set for each iteration; (5) z-score normalization was per-
formed for the training set and testing set; (6) to avoid multi-
collinearity, correlation coefficient between the feature pairs was
calculated using the corrcoef function in the NumPy tool, and then
one of the features with coefficiency greater than 0.8 was removed;
(6) the cross-validation strategy with the least absolute shrinkage
and selection operator (Lasso) was applied using the LassoCV func-
tion; (7) for hyperparameter optimization, the random parameter
search strategy was applied using the RandomizedSearchCV func-
tion; (8) the best parameters were used for model fitting; (9) the
predictive probability and predictive label of each bipolar channel
were calculated; and (10) evaluation of model performance was
conducted using the model_evaluator function in eslearn tool, in
which the sensitivity, specificity, and area under the curve were
calculated. Regarding the machine learning algorithm, logistic
regression was applied in the localization analysis in patients with
various clinical phenotypes (Figs. 2–5). In addition, logistic regres-
sion, stochastic gradient descent (SGD), and support vector
machine for classification (SVC) were used to further explore the
impact of the algorithms on model performance (Figure S2).
Fig. 1. Schematic outline of the study. (A) Six CCEP metrics were used in the SOZ localization model. (B) The CCEPs were recorded from each contact when stimulating a pair
of bipolar channels. (C) For each channel, the six CCEP metrics are used as input for each bipolar channel. (D) Four categories of connection are defined based on the
epileptogenicity of stimulated and recorded sites. (E) The overlap between predicted SOZ and actual SOZ reflects the performance of the SOZ localization model, represented
by the AUC of ROC (F). CCEP, corticocortical evoked potential, AUC, area under the curve; ROC, receiver operation characteristic; N1, first peak; N2, second peak; RMS, root
mean square; Conn, connectivity; SOZ, seizure onset zone; Chan, channels.
B. Yang, B. Zhao, C. Li et al. Clinical Neurophysiology 158 (2024) 103–113
105
2.7. Evaluation of the model in patients with different clinical
phenotypes
To verify the robustness of the model, we divided patients into
different groups based on their clinical phenotypes. We compared
the localizing performance and connectivity patterns. Patients
were grouped into (1) focal cortical dysplasia (FCD) IIa and hip-
pocampal sclerosis (HS) groups according to pathology, (2) SF
(Engel Ia) and non-seizure-free (NSF) (Engel Ib-III) groups accord-
ing to surgical outcome, and (3) complex partial seizure (CPS)
and generalized tonic-clonic seizure (GTCS) groups according to
seizures types (patients in GTCS group also had CPSs as habitual
seizures).
2.8. Statistics
Data are presented as the mean ± SD for those with a normal
distribution and median (range) for those with a skewed distribu-
tion. Continuous variables with a normal distribution and homo-
geneity of variance were analyzed for intergroup comparisons
using independent t-tests. Continuous variables with skewed dis-
tributions were analyzed using the Mann–Whitney U test. To com-
pare the connectivity patterns of the CCEP between groups with
various clinical phenotypes, the percentile bootstrap on trimmed
means method was used (B = 10,000 bootstraps) (Shahabi et al.,
2021). To quantify the predictive model’s performance, sensitivity
and specificity at the group and patient-specific levels and the area
under the receiver operating characteristic (ROC) curve was deter-
mined (Fig. 1E & 1F). Sensitivity was calculated as true positive
(TP)/(TP + false negative (FN)), and specificity was calculated as
true negative (TN)/(TN + false positive (TP)). A true positive
referred to a channel defined as SOZ by visual inspection, which
was also identified as positive by the localization model. All statis-
tical analyses were conducted using SPSS 23.0, R, Python 3.8, and
MATLAB software.
3. Results
3.1. Demographics of patients
Fifty patients were included in this study. The detailed clinical
characteristics of the patients are presented in Table 1. All bipolar
channels were electrically stimulated, and the responses were
recorded in the other channels. The total number of channels in
the 50 patients was 6804, including 2053 channels in the gray mat-
ter, whereas only those within gray matter were included in the
SOZ localization model (Figure S1). Among the 2053 bipolar chan-
nels within the gray matter, there were 63 channels with NaN val-
ues obtained from the semi-automatic analysis protocols of CCEP
metrics. For each patient, the total number of bipolar channels
was 107.34 ± 23.77; the number of bipolar channels located within
the gray matter was 34.90 ± 8.11; and the channels located within
SOZ were 8.18 ± 7.55, respectively.
3.2. Demonstrative cases with different localizing performance
3.2.1. A case with superior localizing performance
Illustrative Case 1 (ID = 27) was a 21-year-old man with a 2-
year history of refractory seizures that manifested as left facial pain
and the inability to speak, followed by hypermotor seizure (‘‘bicy-
cling” movement). The transmantle sign in the right frontal oper-
culum is shown in the FLAIR image (Fig. 2A). The SEEG recording
showed low-voltage fast activity originating from the posterior
short insular gyrus and anterior long insular gyrus (Fig. 2B, 2C,
2D, Table S2). The patient underwent tailored resection of the cor-
responding area and achieved seizure-free status at 39 months
postoperatively.
3.2.2. A case with moderate localizing performance
Illustrative Case 2 (ID = 28) was a 21-year-old woman with a
10-year history of refractory seizures manifesting as dialeptic sei-
zures or ictal vocalization, followed by oral and manual automa-
tism. The gray-white matter junction blurring and
hypometabolism were detected in the right fusiform gyrus
(Fig. 2E), from which low-voltage fast activities originated
(Fig. 2F, 2G, 2H, Table S3). The patient underwent a tailored resec-
tion of the corresponding area and achieved a seizure-free status
40 months postoperatively.
3.3. Comparison of SOZ localization performance and connectivity
patterns represented by CCEP in patients with various clinical
phenotypes
A logistic regression-based model was constructed to compare
the SOZ localization performance between the groups based on
the six metrics of CCEPs.
3.3.1. Pathological types
Electrode coverage in the FCD IIa group (number of individu-
als = 15) was concentrated in the frontal and temporal lobes,
whereas that in the HS group (number of individuals = 14) was
concentrated in the orbitofrontal gyrus, insula, and temporal lobes
(Fig. 3A, 3B). Concerning SOZ localizing performance in the FCD IIa
group, the area under the curve (AUC) of the ROC curve was 0.93
(number of channels = 548), with accuracy, sensitivity, and speci-
ficity of 0.87, 0.82, and 0.89, respectively (Fig. 3C). The HS group
(number of channels = 491) had an AUC of 0.88, with accuracy, sen-
sitivity, and specificity of 0.83, 0.76, and 0.85, respectively
(Fig. 3D).
A comparison between FCD IIa and HS for the N1RMS values
recorded in the stim-SOZ/rec-SOZ regions is illustrated in Fig. 3E.
The marginal distribution in the y-axis shows that the HS group
(10.65 ± 0.84) had greater N1RMS values than the FCD IIa group
(5.65 ± 0.27) (p< 0.001, percentile bootstrap on trimmed means
[B = 10
4
bootstraps, with 20% trimmed on each side], number of
stim/rec channel pairs = 440 for FCD IIa, and n= 130 for HS). Com-
paring the distances of contact from the stimulated channels,
revealed no significant difference between the HS (19.56 ± 0.33)
and FCD IIa groups (18.67 ± 0.53) (p= 0.214, [B=10
4
bootstraps,
with 20% trimmed on each side], number of stim/rec channel
pairs = 440 for FCD IIa, and n= 130 for HS).
The matrix containing the N1RMS values reflected the connec-
tivity properties in stim-nSOZ/rec-nSOZ regions (Fig. 3F, 3G,
Table S1). To obtain a sufficient sample size, only AAL regions with
more than 10 channels in each group were analyzed when N1RMS
was compared between the two pathological groups (Fig. 3H). The
connection strength revealed by N1RMS was higher in the HS
group than in the FCD IIa group, which were illustrated between
two of three brain regions including the left hippocampus (Hipp),
orbital frontal gyrus (OFC), and anterior cingulate cortex (ACC)
(p= 0.001 for OFC-ACC, Mann–Whitney U test, number of stim/
rec channel pairs = 57 for FCD IIa and n= 17 for HS; p< 0.001 for
Hipp-OFC, Mann–Whitney U test, number of stim/rec channel
pairs = 35 for FCD IIa and n= 26 for HS) (Fig. 3I,3J,3K).
3.3.2. Surgical outcomes
Regarding the SOZ localizing performance in the SF group
(number of patients = 28, number of channels = 986), the AUC of
the ROC curve was 0.92, with an accuracy, sensitivity, and speci-
ficity of 0.87, 0.85, and 0.88, respectively (Fig. 4C). In contrast,
B. Yang, B. Zhao, C. Li et al. Clinical Neurophysiology 158 (2024) 103–113
106
the NSF group (number of patients = 22, number of channels = 750)
had an AUC of 0.92, and accuracy, sensitivity, and specificity of
0.86, 0.83, and 0.86, respectively (Fig. 4D). In the y-axis (Fig. 4E),
the marginal distribution shows that SF and NSF groups do not
have significant differences in terms of N1RMS values (p= 0.194,
percentile bootstrap on trimmed means [B = 10
4
bootstraps, with
20% trimmed on each side], number of stim/rec channel pairs = 766
for SF, and n= 251 for NSF). Furthermore, the connectivity proper-
ties of stim-nSOZ/rec-nSOZ, as reflected by N1RMS are illustrated
in Fig. 4F, 4G, and 4H. Specifically, the connection strength
revealed by N1RMS between two of the three brain regions (left
Hipp, OFC, and ACC) was higher in the NSF group than in the SF
group (p= 0.006 for Hipp-ACC, Mann–Whitney U test, number of
stim/rec channel pairs = 13 for SF, and n= 53 for NSF; p< 0.001
for OFC-ACC, Mann–Whitney U test, number of stim/rec channel
pairs = 47 for SF, and n= 40 for NSF; p< 0.001 for Hipp-OFC
Mann–Whitney U test, number of stim/rec channel pairs = 29 for
SF, and n= 50 for NSF) (Fig. 4I, 4 J, and 4 K).
3.3.3. Seizure onset types
The predictive models showed comparable localization perfor-
mance in the GTCS (AUC = 0.90, number of patients = 24, number
of channels = 1089) (Fig. 5A, 5C) and CPS groups (AUC = 0.93, number
of patients = 26, number of channels = 1211) (Fig. 5B, 5D). The half-
violin chart revealed that the N1RMS of stim-SOZ/rec-SOZ in the
GTCS group (4.37 ± 3.77) was lower than that in the CPS group (6.9
0 ± 5.95) (p< 0.001, Mann–Whitney U test, number of stim/rec chan-
nel pairs = 534 for GTCS and number of stim/rec channel pairs = 480
for CPS) (Fig. 5F). The N1RMS of stim-nSOZ/rec-SOZ were higher in
GTCS than in CPS (p< 0.001, Mann–Whitney U test, number of
stim/rec channel pairs = 5468, and number of stim-rec channel
pairs = 5237). The N1RMS of stim-nSOZ/rec-nSOZ was also higher
in GTCS than in CPS (p< 0.001, Mann–Whitney U test, both stim-
nSOZ/rec-SOZ, number of stim/rec channel pairs = 58639, and num-
ber of stim-rec channel pairs = 57178) (Fig. 5H, 5I). In summary, the
difference in connectivity strength between the SOZ and nSOZ was
more pronounced in the CPS group than in the GTCS group.
Fig. 2. Illustrative cases with superior and moderate predictive performance at the patient-specific level. Case 1 (ID = 27) (A-D) and Case 2 (ID = 28) (E-H) are illustrated,
respectively. The presurgical evaluation data include preoperative MRI, PET/MRI fusion, and postoperative CT and SEEG (A, B for Case 1; E, F for Case 2). (C, G) Bipolar channels
within gray matter. The blue dot represents channels that were predicted as SOZ, and the red dot represents channels that were predicted as non-SOZ. (D, H) the predicted
probability. The y-axis represents the predicted probability of bipolar channels located within SOZ, while the x-axis represents channels located within gray matter. MRI,
magnetic resonance imaging; PET, positron emission tomography; CT, computed tomography; SEEG, stereoelectroencephalography; PS, posterior short insular gyrus; AL,
anterior long insular gyrus; PL, posterior long insular gyrus; PCC, posterior cingulate cortex; SMA, supplementary motor area; STS, superior temporal sulcus; TP, true positive;
FN, false negative; FT, false positive; TN, true negative.
B. Yang, B. Zhao, C. Li et al. Clinical Neurophysiology 158 (2024) 103–113
107
3.4. Correlation between localizing performance at the patient-specific
level and surgical outcomes
Regarding localization accuracy, the SF also performed better
than the NSF group at the patient-specific level. Specifically, the
sensitivity of the SF group (0.88 ± 0.19) was higher than that of
the NSF group (0.72 ± 0.32) (p= 0.001, Mann–Whitney U test,
n= 28 for SF and n= 22 for NSF) in the logistic regression model
(Fig. S2A), and the sensitivity of the SF group (0.87 ± 0.19) was also
higher than that of the NSF group (0.72 ± 0.32) (p= 0.028) in the
support vector machine model (Fig. S2C).
3.5. Exploration of localization models based on multiple algorithms
and CCEP metrics
Three machine learning algorithms (logistic regression, SGD
Classifier, SVC) and six CCEP metrics were used separately to train
the models. Feature selection revealed that regardless of the algo-
rithm used, the models with only N1RMS demonstrated the best
performance among the six metrics (Figure S3)(AUC = 0.93 for
all three algorithms).
4. Discussion
Evidence from intracranial electrophysiology suggests that the
SOZ and nSOZ have distinct distribution patterns of local responses
and connectivity properties (Guo et al., 2020; Hays et al., 2023;
Lagarde et al., 2018; Mouthaan et al., 2016; van Blooijs et al.,
2018). Some studies have demonstrated the potential value of CCEPs
in identifying the SOZ (Hays et al., 2023; Kamali et al., 2020; Smith
et al., 2022). However, the limitations include small sample size,
inconsistent intracranial electrophysiology approaches (SEEG and
electrocorticography), and some patients not receiving surgical
treatment to validate whether the SOZ localization was precise. To
Fig. 3. The SOZ localizing performance for patients with FCD IIa/HS and the comparison of CCEP probability distributions. (A, B) The locations of channels within gray matter.
(C, D) the ROCs represent the localizing performance of the logistic regression model based on six CCEP metrics and are calculated separately for each group. (E) Each point in
the scatter plot represents a unique recording channel from a stimulation site. The box at the bottom right displayed the trimmed mean and confidence interval for each
group and the p-value for their comparison. Marginal distribution depicts the increased connectivity for HS in comparison to FCD IIa. (F, G) The matrix containing N1RMS
values reflects the connectivity properties in stim-nSOZ/rec-nSOZ regions, where each row corresponds to a stimulated region. According to the AAL atlas, each column
corresponds to a responsive area. The brain regions are re-ordered based on the sides of the hemisphere (Table S1). (H) Statistical values for comparing the two pathological
groups, where * indicates that the comparison of N1RMS between two groups in the specific region is statistically significant (p< 0.05); (I, J, K) a comparison of N1RMS in the
specific regions between of the two pathological groups. FCD, focal cortical dysplasia; HS, hippocampal sclerosis; ROC, receiver operating characteristic; CCEP, cortical-cortical
evoked potential; N1RMS, root mean square of the first peak; stim, stimulating; rec, recording; Hipp, hippocampus; ACC, anterior cingulate cortex; OFC, orbital frontal cortex;
FCD, focal cortical dysplasia; HS, hippocampus sclerosis.
B. Yang, B. Zhao, C. Li et al. Clinical Neurophysiology 158 (2024) 103–113
108
address these issues, this study aimed to improve the performance of
SOZ localization methods in presurgical evaluation and to test their
performance in patients with refractory epilepsy with different clin-
ical phenotypes. The results confirmed the robustness of the model
in various types of refractory focal epilepsy.
4.1. Methodological aspects
In terms of machine learning model algorithms, this study
explored three different algorithms for localizing SOZ using CCEP
metrics. The SVC (or support vector machine) has been widely used
in the detection of epileptic foci in MRI (Mo et al., 2019), delin-
eation of the epileptic zone by the fingerprint of SEEG signals (Li
et al., 2020), and detection of epileptic seizures in long-term EEG
(Raghu et al., 2019). In addition to the commonly used SVC, logistic
regression and SGD were also used to explore the utilization and
limitations of machine learning models in SOZ localization.
Although the three algorithms achieved comparable localization
performance, their underlying mechanisms and applications dif-
fered. SVC uses kernel-based optimization to transform the input
Fig. 4. The SOZ localizing performance channels for patients in SF and NSF groups and the comparison of CCEP distributions. (A, B) The localization of channels within gray
matter. (C and D) The ROCs show the predictive performance of the logistic regression model based on six CCEP metrics. The ROCs are separately calculated based on each
group. (E) Each point in the scatter plot represents a unique recording channel from a stimulation site. The box at the bottom right displays the trimmed mean and confidence
interval for each group and the p-value for their comparison. Marginal distribution depicts the increased connectivity for HS in comparison to FCD IIa. (F, G) The matrix
containing N1RMS values reflected the connectivity properties in stim-nSOZ/rec-nSOZ regions, where each row corresponds to a stimulated region. According to the AAL
atlas, each column corresponds to a responsive area. The brain regions are re-ordered based on the sides of the hemisphere (Table S1). (H) statistical values for comparing the
two pathological groups, where * indicates that the comparison of N1RMS between two groups in the specific region is statistically significant (p< 0.05). (I, J, K) a comparison
of N1RMS in the specific regions between the two pathological groups. SOZ, seizure onset zone; ROC, receiver operating characteristic; CCEP, cortical-cortical evoked
potential; FCD, focal cortical dysplasia; HS, hippocampus sclerosis; N1RMS, root mean square of the first peak; stim, stimulating; rec, recording; Hipp, hippocampus; ACC,
anterior cingulate cortex; OFC, orbital frontal cortex; SF, seizure-free; NSF, non-seizure-free.
B. Yang, B. Zhao, C. Li et al. Clinical Neurophysiology 158 (2024) 103–113
109
data into complex data (unraveled), thereby identifying more com-
plex boundaries between classes. The primary advantages of the
SVC method are that it is effective in high-dimensional spaces
and does not require hyperparameter tuning. However, this wors-
ens with an increase in the number of samples. In contrast, using a
gradient descent optimization technique, the SGD classifier gener-
ally scales better for massive data, where the iteration process
determines the optimum coefficients. SGD only needs to compute
small mini-batches in each iteration, thereby eliminating the most
redundant computations. In this study, the SVC, SGD, and logistic
regression models achieved comparable performances in localizing
the SOZ, illustrating the robustness of these localization models in
presurgical evaluation. However, when the sample size increases
under practical conditions, the SGD model may be applied because
of its better performance and faster computation for massive data.
In terms of the cross-validation method used in the model, in each
iteration, the testing pool scores belong to completely different
individuals compared to the data selected for training. As a result,
our models had the potential to be used in real clinical situations
due to their generalizability when applying the models to a com-
pletely new dataset of subjects.
Previous studies have shown that CCEP can be associated with
areas of seizure propagation when stimulating the SOZ (Enatsu
et al., 2012; Mouthaan et al., 2016). In addition, accentuated CCEP
amplitudes were higher in SOZ than in non-SOZ (Iwasaki et al.,
2010). Based on these findings, the CCEP metrics may reflect
epileptogenicity. To delineate epilepsy networks using CCEPs, a
series of metrics have been used, including those in the time
domain (N1RMS, N2RMS) (Guo et al., 2020; Veit et al., 2021) and
those extracted from frequency transformation (Mouthaan et al.,
2016) and those based on graph theory (Hays et al., 2021). Further-
more, some studies have attempted to localize the SOZ based on
CCEP metrics (Smith et al., 2022), and this study shows a noticeable
improvement in localization performance. A significant contribu-
tion to this improvement may be the multiple input features of
the prediction model. Additional perspectives on the morphologi-
cal features of CCEP that could more precisely represent epilepto-
genicity were included in our model. We used more CCEP
metrics because the underlying mechanism of CCEPs is not fully
understood (Di Giacomo et al., 2019; Kunieda et al., 2015;
Levesque et al., 2018; Matsumoto et al., 2017; Mouthaan et al.,
2016). The N1 of a CCEP may represent the excitation of pyramidal
cells, whereas the N2 may be associated with long-lasting inhibi-
tion (Keller et al., 2014).
In addition to training the model using all six CCEP metrics,
each metric was used as an independent input to train the SOZ
localization model. Using the AUC as the criterion, the model
trained by N1RMS alone had the best performance compared to
models trained using the other CCEP metrics. Furthermore, the fea-
ture weights suggested that N1RMS contributed more to localizing
the SOZ than the other metrics. Consistent with our findings, pre-
vious studies have indicated that N1RMS was a more robust and
sensitive metric (Prime et al., 2020). Consequently, N1RMS was
selected to analyze the marginal distribution of CCEPs in group
comparisons with various clinical phenotypes. In terms of response
latency, Valentín et al. suggested that only delayed responses
(>100 ms) could be used to localize SOZ (Valentín et al. 2005). In
contrast, Mouthaan et al. demonstrated that early responses
(<100 ms) were also associated with SOZ localization and the sei-
zure propagation area (Mouthaan et al., 2016). Based on these con-
tradictory findings, we included our model’s early and delayed
responses to predict SOZ.
4.2. Relation to findings in the literature
Although whole-brain connectivity patterns of focal epilepsy
have been illustrated through neuroimaging or EEG (Englot et al.,
2016), intracranial evidence for connectivity properties revealed
by CCEPs is limited, especially the association between pathologi-
cal types and connectivity (Shahabi et al., 2021). This study pro-
vides novel insights into epileptogenic network distribution in
patients with different clinical phenotypes through CCEPs.
Regarding neurophysiological mechanisms of how CCEPs repre-
sent the connectivity between brain areas, we discussed this issue
from two perspectives: the stimulation-induced neural activity
changes and pathways for the propagation of cortical stimulation.
First, the stimulation into the cortex can affect local pyramidal cells
in three possible ways (Keller et al., 2014): (1) generating an action
potential in the neurons in cortical layers 2, 3, 5, and 6 through
direct depolarization of the superficial dendritic trees of pyramidal
cells; (2) leading to an indirect decrease in pyramidal cell firing
through the depolarization of layer 2/3 inhibitory interneurons
Fig. 5. The localizing performance in patients of GTCS/CPS groups and comparison of CCEP distributions. (A, B) The locations of channels within gray matter. (C, D) the ROCs
show the predictive performance of the logistic regression model based on six CCEP features and are calculated separately based on each group. (E, F, G, H, I) N1RMS are
depicted. The SF and NSF group comparisons are performed for the four regions of various epileptogenicity, illustrated together (E) or separately (F-I). ROC, receiver operating
characteristic; CCEP, cortical-cortical evoked potential; N1RMS, root mean square of the first peak; SF, seizure-free; NSF, non-seizure-free stim, stimulating; rec, recording;
GTCS, generalized tonic-clonic seizure; CPS, complex partial seizure.
B. Yang, B. Zhao, C. Li et al. Clinical Neurophysiology 158 (2024) 103–113
110
Table 1
Summary of clinical data.
ID Sex Age
(years)
Duration of
epilepsy
(years)
Seizure type
(GTCS = 1,
CPS = 2)
MRI Pathology Locations of SOZ Numbers of bipolar
Channels(All/Gray
matter/SOZ)
Surgery
Outcome
(Engel)
Postoperateive
follow-up
(months)
1 F 24 12 2 R HS FCD IIIa R Hipp 83/30/4 I 77
2 M 30 25 1 Negative FCD I R OFC,insula,ACC 134/52/15 I 70
3 M 25 19 2 Negative FCD IIIa R Amy, R Hipp 76/32/7 I 69
4 F 22 12 2 Negative FCD IIa L Amy, Hipp, OFC 122/32/9 I 68
5 M 29 20 2 R HS FCD IIIa R Hipp, ParaHipp 125/46/8 I 67
6 F 34 25 2 Negative FCD I L Insula, Rolandic
operculum, Heschl’s
gyrus
81/42/7 III 67
7 M 22 20 1 Negative FCD IIIa R Hipp 86/25/2 I 66
8 M 27 13 1 Negative Scar R inferior parietal
lobule, Rolandic
operculum
72/29/8 II 65
9 M 27 16 1 Negative Nonspecific L inferior parietal
lobule
136/40/3 I 65
10 M 31 22 2 Negative FCD IIa L Sup Frontal, SMA 174/51/6 I 65
11 M 37 3 1 Negative Scar R OFC 86/27/4 III 65
12 M 19 6 2 R HS FCD IIIa R Amy, Hipp,
ParaHipp
80/32/7 I 65
13 M 18 3 2 L HS HS L Hipp 99/23/4 II 52
14 M 34 12 1 L HS HS R Occipital 85/27/5 I 52
15 M 16 7 1 Negative FCD IIa R insula 103/40/5 I 53
16 F 20 3 2 R HS FCD IIb R Hipp, ParaHipp 109/23/5 I 55
17 F 30 9 2 Negative FCD IIb L SMA, MCC 114/29/7 I 48
18 M 35 3 1 Negative HS R Amy, Hipp 129/40/4 II 47
19 F 23 10 1 Negative FCD IIa R Rolandic
opeculum, insula,
supramarginal
132/39/16 I 46
20 M 27 4 1 Negative FCD I? L OFC, insula 73/14/3 III 44
21 F 24 13 1 Negative FCD I? R insula, Rolandic
operculum
107/46/11 III 44
22 M 30 6 1 Negative FCD IIa R Amy, Hipp,
ParaHipp
97/34/10 II 43
23 F 22 18 2 Negative Nonspecific L Amy, Hipp,
ParaHipp
113/32/10 I 42
24 M 26 8 2 R Choroidal Fissure Cyst GG R Hipp 96/31/5 I 42
25 M 47 19 2 Negative FCD IIa R insula, OFC 101/27/5 I 42
26 M 19 6 2 R Temporal HS R Hipp 99/34/6 I 41
27 M 21 2 2 Negative FCD IIb R Rolandic
operculum
109/45/2 I 39
28 F 21 10 1 R basal temporal-occipital FCD IIa R basal temporal-
occipital
74/25/2 I 40
29 F 21 6 2 R Occipital, Scar Scar L Occipital 95/30/7 I 37
30 M 27 24 2 L HS HS L Temporal 107/33/6 II 37
31 F 32 20 2 R ACC, blurring of the
cortical gray and white
matter boundary
FCD IIb L ACC, SMA, Sup
Frontal
104/41/14 I 36
32 F 18 16 2 L HS HS L temporal, insula 76/27/6 II 36
33 M 34 14 2 Negative FCD IIa L Hipp, Post middle
temporal gyrus
93/29/12 I 37
34 F 19 11 2 L Occpital Gliosis L Occpital 93/30/7 III 40
35 M 47 40 1 R HS HS R Hipp, ParaHipp 106/34/2 II 35
36 M 20 12 1 L HS HS? L Hipp, ParaHipp 161/37/9 I 33
37 M 24 3 1 L periventricular
heterotopia
heterotopia L gray matter
heterotopia
90/31/13 I 32
38 M 37 8 1 Negative Nonspecific L insula 143/37/4 I 39
39 F 30 8 1 Negative FCD IIa L Sup frontal gyrus 104/33/10 I 39
40 F 25 1 2 Negative HS R Hipp, ParaHipp,
fusiform
94/37/12 I 39
41 M 30 11 1 Negative FCD IIa R SMA, Sup Frontal 85/37/17 I 38
42 M 32 15 2 Negative HS R Hipp 129/38/6 I 24
43 F 22 10 1 Negative HS R Hipp, ParaHipp 135/47/9 I 26
44 F 23 11 1 R Occipital, Scar Scar R Occipial 140/34/4 I 24
45 M 18 11 1 R sulcus centralis insulae,
transmentle
FCD IIb R sulcus centralis
insulae
114/30/3 I 25
46 F 44 16 2 Negative FCD IIa L OFC 140/55/9 I 25
47 F 23 16 1 Negative FCD IIa L pre-SMA 89/36/53 I 24
48 F 25 19 1 Negative FCD IIa R insula 123/44/11 I 24
49 F 35 22 2 Negative FCD IIa L OFC 137/37/2 I 30
50 M 32 26 2 R Post Inf frontal gyrus,
transmentle
FCD IIa R Pos Inf frontal
gyrus
114/41/13 II 29
Hipp, hippocampus; HS, hippocampal sclerosis; FCD, focal cortical dysplasia; ACC, anterior cingulate cortex; OFC, orbital frontal cortex; SMA, supplementary motor area;
ParaHipp, para-hippocampus gyrus; Sup, superior; Inf, inferior, Post, posterior; SOZ, seizure onset zone.
B. Yang, B. Zhao, C. Li et al. Clinical Neurophysiology 158 (2024) 103–113
111
that synapse near the soma on adjacent pyramidal cells; (3) gener-
ating action potentials propagating orthodromically and
antidromically by depolarizing long-range axons traversing the
region of stimulation. Second, for pathways of propagation of cor-
tical stimulation, the generation of CCEP primarily involves activat-
ing middle and deep pyramidal cells, followed by the propagation
of signals down their axons to regions connected with synapses
through cortico-cortical and cortico-subcortical projections
(Steriade and Amzica, 1996). Based on the propagation pathways
of neural activity during cortico-cortical evoked potential, our find-
ings about epileptogenic network distribution provided insights
into the brain network characteristics in patients with different
clinical phenotypes of refractory epilepsy.
This study not only proposed an approach for localizing the SOZ
by CCEPs but also compared the cortical excitability between HS
and FCD IIa through the probability distribution of CCEPs. Cortical
excitability at the stim-SOZ/rec-SOZ sites was increased in patients
with HS compared to those with FCD IIa. Simultaneously, the dis-
tance between the stimulated and recorded channels was not sig-
nificantly different between the two groups, confirming that
hyperexcitability was independent of electrode implantation.
Using CCEPs, a previous study showed that patients with FCD I
had a significantly greater degree of stimulation-evoked responses
than those with FCD II (Shahabi et al., 2021). However, the
excitability of HS was not compared with that of FCD II. Our find-
ings provide evidence that fills this gap. This study only analyzed
FCD IIa rather than other types of FCD, because of the insufficient
sample size; there were only five patients with FCD IIb and four
with FCD I. The hippocampus is involved in the limbic system
and may be closely connected to other structures in the limbic sys-
tem, as suggested by electrophysiological (Piper et al., 2022) and
neuroimaging studies (Urquia-Osorio et al., 2022). However, corti-
cal excitability is influenced by several factors, including seizure
duration and control (Pawley et al., 2017). From the perspective
of cortical architecture, the SOZs of patients with HS and FCD IIa
are different types of cerebral cortices (allocortex versus neocor-
tex) (Lopes da Silva et al., 1990). Regarding the spatial distribution,
(Shahabi et al., 2021) illustrated a more restrictive area of
hyperexcitability for FCD II than FCD I. Our study did not demon-
strate a significant difference between the FCD IIa and HS in terms
of spatial distance, confirming that the difference in the excitability
between the HS and FCD IIa groups was not caused by spatial
distribution.
In the SOZ-SOZ regions, the N1RMS values of the SF and NSF
patients did not differ significantly, nor did the distances between
the stimulated and recorded sites. Notably, the SOZ localization
was correct in all patients in this cohort because even patients with
NSF underwent presurgical iEEG monitoring with typical ictal pat-
terns. This was also confirmed by the postoperative worthwhile
improvement. Consistently, at the channel level, the comparative
localization accuracies between SF and NSF patients were reason-
able. In contrast, at the patient level, localization accuracies were
better in the SF groups than in the NSF group. To further explore
the potential reason for the patient-level differences, the connec-
tivity in the nSOZ-nSOZ regions between the three specific struc-
tures (Hipp, ACC, and OFC) was compared between the SF and
NSF. The results showed that the connectivity was stronger in
the NSF than in the SF in the nSOZ-nSOZ regions. As mentioned
above, in the SOZ-SOZ regions, the connectivities were comparable
for NSF and SF. According to the subtraction rule, by comparing the
SOZ and nSOZ, the NSF had more indistinguishable CCEP metrics
than the SF in determining which channel was in the SOZ. This
finding suggests that seizure recurrence may be partially due to
excessive connections between distant regions with weak epilep-
togenicity (nSOZ-nSOZ). Consistently, the literature on large-scale
brain networks indicates that the more complex the epileptogenic
network, the worse the surgical outcome may be (Royer et al.,
2022).
To demonstrate the connectivity patterns in the GTCS and CPS,
the N1RMS was compared separately for four categories of regions
based on epileptogenicity: SOZ-SOZ, SOZ-nSOZ, nSOZ-SOZ, and
nSOZ-nSOZ. The CPS network was more restricted, combined with
a higher degree of cortical excitability than the GTCS network.
Notably, the connectivity patterns of the GTCS and CPS revealed
by CCEPs were consistent with a more widespread epileptic net-
work for GTCS than for the CPS (Englot et al., 2016). The CPS had
greater cortical excitability in the SOZ-SOZ regions than the GTCS
and lower cortical excitability in the nSOZ-nSOZ. Thus, The differ-
ence in N1RMS between SOZ-SOZ and nSOZ-nSOZ was more appar-
ent in CPS than in GTCS. In other words, the patients with CPS had
more distinctive characteristics of CCEP than patients with GTCS in
areas with different epileptogenicity. In summary, all comparisons
of CCEP patterns between different clinical phenotypes provided
new insights into the neurophysiological substance of the epilepto-
genic network.
4.3. Limitations
This study has the following limitations: (1) although the sam-
ple size of the channel was sufficient to establish the SOZ localiza-
tion model, not all regions according to the AAL atlas were covered
when comparing the CCEPs in different groups; (2) the distance
between the stimulating and recording channels was not included
in the prediction models, as incorporating this information as input
features contradicts the purpose of the predictive model; (3) the
CCEP protocol in our center only uses a constant stimulating
parameter, preventing the application of our model to changeable
stimulating parameters. Future research is needed to evaluate
CCEPs under various stimulation parameters in the SOZ localiza-
tion model, providing more insights into presurgical evaluation
and electrical stimulation therapy.
5. Conclusions
We developed a new approach for localizing the SOZ by using
multiple CCEP metrics. The proposed system had some improve-
ments in contrast to the limitations of the traditional method of
localizing the SOZ, including a long monitoring time, increased risk
of infection due to long-term monitoring, and the possibility of no
spontaneous onset. In addition, the proposed localization model
exhibited a stable performance for various clinical phenotypes.
Acknowledgments
This work was supported by the National Natural Science Foun-
dation of China (82071457, 82271495, 82201603, 82201600) and
Capital’s Funds for Health Improvement and Research (2022-1-
1071, 2020-2-1076).
Conflict of interest
None of the authors have potential conflicts of interest to be
disclosed.
Appendix A. Supplementary data
Supplementary data to this article can be found online at
https://doi.org/10.1016/j.clinph.2023.12.135.
B. Yang, B. Zhao, C. Li et al. Clinical Neurophysiology 158 (2024) 103–113
112
References
Chou CC, Lee CC, Lin CF, Chen YH, Peng SJ, Hsiao FJ, et al. Cingulate gyrus
epilepsy: semiology, invasive EEG, and surgical approaches. Neurosurg Focus
2020;48(4) E8.
Crocker B, Ostrowski L, Williams ZM, Dougherty DD, Eskandar EN, Widge AS, et al.
Local and distant responses to single pulse electrical stimulation reflect
different forms of connectivity. Neuroimage 2021;237 118094.
Davis TS, Rolston JD, Bollo RJ, House PA. Delayed high-frequency suppression after
automated single-pulse electrical stimulation identifies the seizure onset zone
in patients with refractory epilepsy. Clin Neurophysiol 2018;129(11):2466–74.
Di Giacomo R, Uribe-San-Martin R, Mai R, Francione S, Nobili L, Sartori I, et al.
Stereo-EEG ictal/interictal patterns and underlying pathologies. Seizure
2019;72:54–60.
Donos C, Maliia MD, Mindruta I, Popa I, Ene M, Balanescu B, et al. A connectomics
approach combining structural and effective connectivity assessed by
intracranial electrical stimulation. Neuroimage 2016;132:344–58.
Durnford AJ, Rodgers W, Kirkham FJ, Mullee MA, Whitney A, Prevett M, et al. Very
good inter-rater reliability of Engel and ILAE epilepsy surgery outcome
classifications in a series of 76 patients. Seizure 2011;20(10):809–12.
Enatsu R, Jin K, Elwan S, Kubota Y, Piao Z, O’Connor T, et al. Correlations between
ictal propagation and response to electrical cortical stimulation: a cortico-
cortical evoked potential study. Epilepsy Res 2012;101(1–2):76–87.
Englot DJ, Konrad PE, Morgan VL. Regional and global connectivity disturbances in
focal epilepsy, related neurocognitive sequelae, and potential mechanistic
underpinnings. Epilepsia 2016;57(10):1546–57.
Feng T, Yang Y, Wei P, Wang C, Fan X, Wang K, et al. The role of the orbitofrontal
cortex and insula for prognosis of mesial temporal lobe epilepsy. Epilepsy Behav
2023;138 109003.
Groppe DM, Bickel S, Dykstra AR, Wang X, Megevand P, Mercier MR, et al. iELVis: An
open source MATLAB toolbox for localizing and visualizing human intracranial
electrode data. J Neurosci Methods 2017;281:40–8.
Guo ZH, Zhao BT, Toprani S, Hu WH, Zhang C, Wang X, et al. Epileptogenic network
of focal epilepsies mapped with cortico-cortical evoked potentials. Clin
Neurophysiol 2020;131(11):2657–66.
Hays MA, Coogan C, Crone NE, Kang JY. Graph theoretical analysis of evoked
potentials shows network influence of epileptogenic mesial temporal region.
Hum Brain Mapp 2021;42(13):4173–86.
Hays MA, Smith RJ, Wang Y, Coogan C, Sarma SV, Crone NE, et al. Cortico-cortical
evoked potentials in response to varying stimulation intensity improves seizure
localization. Clin Neurophysiol 2023;145:119–28.
Iwasaki M, Enatsu R, Matsumoto R, Novak E, Thankappen B, Piao Z, et al.
Accentuated cortico-cortical evoked potentials in neocortical epilepsy in areas
of ictal onset. Epileptic Disord 2010;12(4):292–302.
Jacobs J, Zijlmans M, Zelmann R, Olivier A, Hall J, Gotman J, et al. Value of electrical
stimulation and high frequency oscillations (80–500 Hz) in identifying
epileptogenic areas during intracranial EEG recordings. Epilepsia 2010;51
(4):573–82.
Kamali G, Smith RJ, Hays M, Coogan C, Crone NE, Kang JY, et al. Transfer function
models for the localization of seizure onset zone from cortico-cortical evoked
potentials. Front Neurol 2020;11 579961.
Keller CJ, Honey CJ, Megevand P, Entz L, Ulbert I, Mehta AD. Mapping human brain
networks with cortico-cortical evoked potentials. Philos Trans R Soc Lond B Biol
Sci 2014;369(1653) 20130528.
Kunieda T, Yamao Y, Kikuchi T, Matsumoto R. New approach for exploring cerebral
functional connectivity: Review of cortico-cortical evoked potential. Neurol
Med Chir (Tokyo) 2015;55(5):374–82.
Kwan P, Brodie MJ. Early identification of refractory epilepsy. N Engl J Med
2000;342(5):314–9.
Lagarde S, Roehri N, Lambert I, Trebuchon A, McGonigal A, Carron R, et al. Interictal
stereotactic-EEG functional connectivity in refractory focal epilepsies. Brain
2018;141(10):2966–80.
Levesque M, Shiri Z, Chen LY, Avoli M. High-frequency oscillations and mesial
temporal lobe epilepsy. Neurosci Lett 2018;667:66–74.
Li J, Grinenko O, Mosher JC, Gonzalez-Martinez J, Leahy RM, Chauvel P. Learning to
define an electrical biomarker of the epileptogenic zone. Hum Brain Mapp
2020;41(2):429–41.
Lopes da Silva F, Witter MP, Boeijinga PH, Lohman AJPR. Anatomic organization and
physiology of the limbic cortex. Physiol Rev 1990;70(2):453–511.
Matsumoto R, Kunieda T, Nair D. Single pulse electrical stimulation to probe
functional and pathological connectivity in epilepsy. Seizure 2017;44:27–36.
Mo J, Liu Z, Sun K, Ma Y, Hu W, Zhang C, et al. Automated detection of hippocampal
sclerosis using clinically empirical and radiomics features. Epilepsia 2019;60
(12):2519–29.
Mouthaan BE, van ’t Klooster MA, Keizer D, Hebbink GJ, Leijten FSS, Ferrier CH, et al.
Single Pulse Electrical Stimulation to identify epileptogenic cortex: Clinical
information obtained from early evoked responses. Clin Neurophysiol 2016;127
(2):1088–98.
Pawley AD, Chowdhury FA, Tangwiriyasakul C, Ceronie B, Elwes RD, Nashef L, et al.
Cortical excitability correlates with seizure control and epilepsy duration in
chronic epilepsy. Ann Clin Transl Neurol 2017;4(2):87–97.
Piper RJ, Richardson RM, Worrell G, Carmichael DW, Baldeweg T, Litt B, et al.
Towards network-guided neuromodulation for epilepsy. Brain 2022;145
(10):3347–62.
Prime D, Woolfe M, Rowlands D, O’Keefe S, Dionisio S. Comparing connectivity
metrics in cortico-cortical evoked potentials using synthetic cortical response
patterns. J Neurosci Methods 2020;334 108559.
Raghu S, Sriraam N, Vasudeva Rao S, Hegde AS, Kubben PL. Automated detection of
epileptic seizures using successive decomposition index and support vector
machine classifier in long-term EEG. Neural Comput & Applic 2019;32
(13):8965–84.
Royer J, Bernhardt BC, Lariviere S, Gleichgerrcht E, Vorderwulbecke BJ, Vulliemoz S,
et al. Epilepsy and brain network hubs. Epilepsia 2022;63(3):537–50.
Shahabi H, Taylor K, Hirfanoglu T, Koneru S, Bingaman W, Kobayashi K, et al.
Effective connectivity differs between focal cortical dysplasia types I and II.
Epilepsia 2021;62(11):2753–65.
Smith RJ, Hays MA, Kamali G, Coogan C, Crone NE, Kang JY, et al. Stimulating native
seizures with neural resonance: a new approach to localize the seizure onset
zone. Brain 2022;145(11):3886–900.
Steriade M, Amzica F. Intracortical and corticothalamic coherency of fast
spontaneous oscillations. Proc Natl Acad Sci U S A 1996;93(6):2533–8.
Sugano H, Iimura Y, Suzuki H, Tamrakar S, Mitsuhashi T, Higo T, et al. Can
intraoperative electrocorticography be used to minimize the extent of resection
in patients with temporal lobe epilepsy associated with hippocampal sclerosis?
J Neurosurg 2021:1–8.
Tzourio-Mazoyer N, Landeau B, Papathanassiou D, Crivello F, Etard O, Delcroix N,
et al. Automated anatomical labeling of activations in SPM using a macroscopic
anatomical parcellation of the MNI MRI single-subject brain. Neuroimage
2002;15(1):273–89.
Urquia-Osorio H, Pimentel-Silva LR, Rezende TJR, Almendares-Bonilla E, Yasuda CL,
Concha L, et al. Superficial and deep white matter diffusion abnormalities in
focal epilepsies. Epilepsia 2022;63(9):2312–24.
Valentín A, Alarcón G Fau - García-Seoane JJ, García-Seoane Jj Fau - Lacruz ME,
Lacruz Me Fau - Nayak SD, Nayak Sd Fau - Honavar M, Honavar M Fau - Selway
RP, et al. Single-pulse electrical stimulation identifies epileptogenic frontal
cortex in the human brain. 2005(1526-632X (Electronic)).
van Blooijs D, Leijten FSS, van Rijen PC, Meijer HGE, Huiskamp GJM. Evoked
directional network characteristics of epileptogenic tissue derived from single
pulse electrical stimulation. Hum Brain Mapp 2018;39(11):4611–22.
Veit MJ, Kucyi A, Hu W, Zhang C, Zhao B, Guo Z, et al. Temporal order of signal
propagation within and across intrinsic brain networks. Proc Natl Acad Sci U S A
2021;118(48).
Vogt BA. Cingulate cortex in the three limbic subsystems. Handb Clin Neurol
2019;166:39–51.
Yaffe RB, Borger P, Megevand P, Groppe DM, Kramer MA, Chu CJ, et al. Physiology of
functional and effective networks in epilepsy. Clin Neurophysiol 2015;126
(2):227–36.
B. Yang, B. Zhao, C. Li et al. Clinical Neurophysiology 158 (2024) 103–113
113
