This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCSII.2020.2992285, IEEE
Transactions on Circuits and Systems II: Express Briefs
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMSłII: EXPRESS BRIEFS, VOL. X, NO. X, X 2020 1

A Novel Spike Detection Algorithm Based on

Multi-channel of BECT EEG Signals
Zimeng Wang, Duanpo Wu, Fang Dong, Jiuwen Cao, Member, IEEE, Tiejia Jiang and Junbiao Liu

Abstract—Benign childhood epilepsy with centro-temporal of the automatic spike detection in academia [5]. Since then,
spikes (BECT) is one of the most common epilepsy syndromes in according to the characteristics of epileptiform spikes, many
childhood which is typically characterized by localized discharges different spike detection algorithms have been researched,
in the central and temporal regions. Traditionally, the recognition
of spikes requires visual assessment of long-term EEG recordings including morphology, signal correlation, sub-band decompo-
which is time consuming and subjective because it depends on sition, and feature engineering [6], [7]. Nishida et al. used
the knowledge and experience of the doctor. Therefore, a novel morphological filter to decompose background EEG and spike
multi-step spike detection algorithm based on average reference [8]. Open-closing and close-opening operation can separate
(AV) channel and bipolar (BP) channel BECT EEG is proposed, positive and negative pulses in EEG signals at the same
including candidate spike detection algorithm, false positive
spike (FPS) elimination, spike feature extraction and random time, but the anti-expandability of open operation and the
forest (RF) classification. The proposed method is evaluated expansibility of close operation will cause statistical bias [9],
using 7 routine EEG recordings. This paper shows that the [10]. Based on this situation, Xu et al. constructed a mor-
sensitivity (Sen), specificity (Spe), selectivity (Sel) and accuracy phological open-closing combination filter, which eliminated
(AC) obtained by the proposed method are 97.4%, 96.5%, 96.6% the statistical bias of amplitude, and effectively suppressed
and 96.9%, respectively. Experimental results show that the
proposed method is capable of detecting BECT spikes efficiently. background EEG while retaining the spike characteristics [11].
This paper adopts a brain partition detection method to re-
duce the computational load of spike classification. Firstly, the
Index Terms—BECT spikes, AV channel, BP channel, candi-
date spike detection, FPS elimination. energy of each average reference (AV) channel is calculated in
the 4 EEG regions divided by brain partition detection method,
and the discharge region of the BECT spike is determined
I. I NTRODUCTION based on the signal energy. Then, the three AV channels in
ENIGN childhood epilepsy with centro-temporal spikes the region are screened for spikes, and the candidate spikes
B (BECT) is the most common form of benign epilepsy in
children [1]. The main population of BECT is school-age chil-
are obtained through an adaptive template matching algorithm.
Next, false positives spikes (FPS) are eliminated and the
dren, also known as benign rolandic epilepsy, which accounts possible spikes are separated into two groups of non-spikes
for 15%-20% of childhood epilepsy [2]. Electroencephalogram and possible spikes. Finally, the proposed spike model is
(EEG) is formed by the sum of postsynaptic potentials that applied to these two groups of spikes for further classification.
occur simultaneously in a large number of neurons during The rest of the paper is organized as follows. In Section
brain activity [3]. The background activity of EEG is the most II, materials and methods are described. The experimental
direct reflection of cerebral cortical discharges [4]. results are given in Section III. Conclusion and discussions
In 1972, Stevens et al. first attempted to extract epileptic are described in Section IV.
spikes from long-term EEG signals, and started the research

This work was partially supported by the Natural Science Foundation of

Zhejiang Province (LGG19F030013 and LGF18F010007), Key R & D II. M ATERIALS A ND M ETHODS
Programs in Zhejiang Province (2020C03038), NSFC-Zhejiang Integration
Joint Fund (U1909209), Special Funds for Information Development in This paper proposes an automatic spike detection algorithm
Shanghai (201801050) and Scientific Research Project of Zhejiang Provincial based on AV channel and BP channel BECT EEG signals. A
Department of Education (Y201942165).(Corresponding authors: Duanpo Wu
and Fang Dong)
detailed flowchart of the proposed system is shown in Fig. 1.
Z. Wang is with School of Communication Engineering, Hangzhou Dianzi
University, Hangzhou 310018, China. e-mail: [email protected]
D. Wu is with School of Communication Engineering, Hangzhou Dianzi
University, Hangzhou 310018, China. e-mail: [email protected]
F. Dong is with College of Information and Electric Engineering, Zhejiang U-
niversity City College, Hangzhou 310015, China. e-mail: [email protected]
J. Cao is with Artificial Intelligence Institute, Hangzhou Dianzi University,
Zhejiang, 310018, China, and with Key Lab for IOT and Information Fusion
Technology of Zhejiang, Hangzhou Dianzi University, Zhejiang, 310018,
China. e-mail: [email protected]
T. Jiang is with the Children’s Hospital, Zhejiang University School of
Medicine, Hangzhou 310052, China. e-mail: [email protected]
J. Liu is with Hangzhou Neuro Science and Technology Co., Ltd., Hangzhou Fig. 1. Flowchart of the automatic spike detection algorithm.
310052, China. e-mail: [email protected]

1549-7747 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on May 10,2020 at 19:41:20 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCSII.2020.2992285, IEEE
Transactions on Circuits and Systems II: Express Briefs
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMSłII: EXPRESS BRIEFS, VOL. X, NO. X, X 2020 2

TABLE I
D URATIONS , L OCATIONS OF SPIKES AND NUMBERS OF SPIKES FOR EACH EEG RECORDING IN TRAINING DATASET AND TEST DATASET

Name Duration of training data Duration of test data Training spikes Test spikes Locations of spikes
S1 10 min 5 s 10 min 5 s 286 257 C3, C4, P3, P4, T3, T5
S2 10 min 6 s 10 min 6 s 454 559 C3, P3, T3, T4, T5, T6
S3 29 min 54 s 29 min 54 s 486 453 C4, P4, F8, T4
S4 10 min 6 s 10 min 6 s 332 386 C3, C4, P3, P4, T4
S5 9 min 58 s 9 min 58 s 567 806 C3, C4, P3, P4, F8, T4
S6 30 min 21 s 30 min 21 s 72 68 C3, C4
S7 40 min 45 s 40 min 45 s 61 63 C4, T4
Sum 141 min 15 s 141 min 15 s 2258 2592 -

A. Subjects and Data Acquisition C. Candidate Detection

EEG signals were recorded by a digital electroencephalo- After identifying the region of spike discharge, candidate
graph in the Children’s Hospital, Zhejiang University School detection of three AV channels in the central and temporal
of Medicine. 21 electrodes including ear-lobes were placed regions of the identified region is performed, which includes
according to the international 10-20 system. The EEG signals 5 steps.
were recorded with a sampling frequency of 1000 Hz. A 5- 1) Universal Template Matching: The universal template
th order IIR Butterworth bandpass filter with the frequency is used for spike pre-selection, and the cross correlation
range of 0.016-70 Hz was used to remove noise and artifacts between a given spike template and all signal frames is
from EEG signals. The electrodes were referenced to a com- calculated by template matching on EEG signals [12]. When
mon referential electrode. 7 recordings were recorded from 7 the cross-correlation and morphological features exceed cross-
patients (S1-S7) which are listed in Table I. Sixteen-channel correlation threshold and feature threshold, this signal frame
EEGs (locations Fp1, Fp2, F3, F4, F7, F8, T3, T4, T5, T6, C3, is considered as a possible candidate peak. The universal
C4, P3, P4, O1 and O2) from the recorded data were reviewed template is set according to Chatrian spike standard [13].
by the neurologists. 2) Temporal Domain Clustering: The candidate peaks ob-
tained in the step 1) are sorted into clusters. The input
of clustering algorithm is composed of time series of each
candidate spikes. Then, spikes are separated into clusters using
B. Brain Partition
standard k-mean algorithm. The temporal domain clustering
This paper finds that the spikes in the BECT EEG signals algorithm can get the spike template of the adaptive template
have specific regions by analyzing and interpreting the EEG matching to adapt to variations in spike morphology.
signals of 7 BECT patients. According to this rule, this paper 3) Cluster Selection: The clusters obtained in step 2) are
divides the human brain into 4 regions (R1, R2, R3 and R4), as checked. If the number of candidate peaks in a cluster is less
shown in Fig. 2. Subsequently, this paper calculates the signal than 5% of the number of candidate peaks, then this cluster
energies of the three AV channels in each region separately, is eliminated, leaving the clusters that meet the conditions.
and judges the region where the spikes are located to reduce 4) Adapted Template Matching: Perform spike detection
the computational load of the spike classification. The signal based on template matching. The detection steps are the same
energies of the 4 regions are compared, and the region with the to step 1), but the templates and thresholds are different. The
highest total signal energy is identified as the spike discharge centroid of the cluster selected in step 3) is used as a new
region. template to extract the spikes of the specific cluster. The cross-
correlation threshold and feature threshold are automatically
adjusted by the morphological parameters of candidate peaks
of different EEG signals.
5) Processing of Detection Results: Finally, the spikes
detected in step 4) are summed to obtain the final detection
result. Since the same spike can be detected by different
centroids, a double-peak spike may appear. Therefore, an
interval threshold is set. If the interval between two adjacent
spikes is less than this threshold, the two spikes are merged
into one.

D. FPS Elimination
Fig. 2. A schematic diagram of brain partition. There is a large number of false spikes in the candidate
spikes, which seriously affects the performance of the random

1549-7747 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on May 10,2020 at 19:41:20 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCSII.2020.2992285, IEEE
Transactions on Circuits and Systems II: Express Briefs
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMSłII: EXPRESS BRIEFS, VOL. X, NO. X, X 2020 3

forest (RF) classification. Therefore, FPS are eliminated by F. RF Classification

determining two BP channels related to the AV channel in this In the proposed method, a RF classifier is selected for BECT
paper. By analyzing the EEG of the AV channel, the spikes spike detection [6]. The RF classifier proposed by Breiman
can be divided into three types. The three discharge types of contains multiple decision trees whose output category is
the AV channel EEG and the two related BP channels are determined by the highest number of votes given by all trees.
shown in Fig. 3. Taking these three AV channels (F3-AV, C3- Through the self-lifting sampling technique, a new training
AV, and P3-AV) as examples, if a spike is only detected on sample set is generated by randomly selecting k samples from
the C3-AV, two related BP channels are F3-C3 and C3-P3; the original training sample set N , and then RF is generated
if a spike is detected on both C3-AV and P3-AV, the two from k individual decision tree classifiers. The morphological
related BP channels are F3-C3 and P3-O1 respectively; if a feature vectors of the spikes are input into the RF classifier,
spike is detected on both F3-AV, C3-AV and P3-AV, then the and finally all possible spikes are divided into BECT spikes
two related BP channels are FP1-F3 and P3-O1. Check all and non-BECT spikes. In this paper, 50% of the total duration
candidate spikes according to this rule and these candidate of each patient EEG recording constitutes the training set, and
spikes that do not meet this rule are eliminated. On the BP the remaining 50% of the EEG constitutes the test set.
channel, the spike is “peak-to-peak”.
G. Spike Detection System Based on the BCI
The proposed method is simple and easy to implement,
which can be applied to various brain-computer interaction (B-
CI) devices. Fig. 5 shows the flowchart of the spike detection
system based on the BCI. The hardware structure of the BCI
system includes electrodes, analog signal processing module
(ASPM), digital signal processing module (DSPM), bluetooth
module and computer spike detection module. ASPM is used
to amplify EEG signals and filter out common mode and high
frequency noise. DSPM controls the A/D converter through
the microcontroller unit (MCU) to complete the conversion
from analog signals to digital signals, then converts the EEG
signals into computer instructions through the ARM processor,
and finally transmits the instructions to the computer proces-
sor through the bluetooth module to realize automatic spike
Fig. 3. Spikes in the AV channel and the corresponding BP channel. detection.

E. Feature Extraction
After FPS elimination, 10 features of each channel are Fig. 5. The block diagram for BCI system.
calculated for subsequent classification. These 10 features are
divided into 4 categories, including duration, amplitude, slope,
and area. Table II lists the calculations and descriptions of III. R ESULTS
these features, and a reference diagram for feature extraction A. Evaluation Metrics
is given by Fig. 4. The 30 features from one AV channel and In order to evaluate the generality of the proposed method,
two related BP channels are extracted as a spike feature vector. confusion matrix (CM) will be calculated to show the differ-
ence between the result given by the proposed method and the
marks of experts as shown in Eq. (1).
[ ]
TN FP
CM = (1)
FN TP
The sensitivity (Sen), specificity (Spe), selectivity (Sel) and
accuracy (AC) are given by Eq. (2)-(5).
TP
Sen = (2)
TP + FN
TN
Spe = (3)
TN + FP
Fig. 4. Reference diagram for feature extraction.
TP
Sel = (4)
TP + FP

1549-7747 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on May 10,2020 at 19:41:20 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCSII.2020.2992285, IEEE
Transactions on Circuits and Systems II: Express Briefs
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMSłII: EXPRESS BRIEFS, VOL. X, NO. X, X 2020 4

TABLE II
F EATURES EXTRACTED FROM ONE EEG CHANNEL

Name Calculation Description

Dur− lef t ACx Duration of left half-wave of spike
Dur− right CBx Duration of right half-wave of spike
Dur− spike ACx + CBx Sum of durations of left and right half-waves of spike
Amp− lef t ACy Amplitude of left half-wave of spike
Amp− right CBy Amplitude of right half-wave of spike
|CBx |
Amp− spike CPy = CBy − |ABx |
ABy Amplitude of spike
Slope− lef t ACy /ACx Slope of left half-wave of spike
Slope− right -CBy /CBx Slope of right half-wave of spike
Slope− spike Slope− lef t - Slope− right Sharpness of spike
∫ xB
Area− spike xA
f (x)dx − S Area of spike (f (x) denotes the EEG data)

obtained by adding the confusion matrices of S1-S7. It can

TP + TN
AC = (5) be seen that the comparison of AC among the patients (the
TP + TN + FP + FN smallest is 93.9% and largest is 98.4%). The main reason is
that some patients have a short recording time but have a lot
B. Testing and Validation Results
of spikes, and some patients have a long recording time but
Fig. 6 illustrates the performance of the system stage by have fewer spikes. For the recordings of S1-S7, the Sen ranges
stage on an original EEG segment. The proposed method first from 91.0% (S3) to 99.3% (S5), the Spe ranges from 88.9%
identifies all candidate peaks on the data segment through (S5) to 97.9% (S6), the Sel ranges from 66.3% (S7) to 99.0%
universal template matching (Fig. 6 (a)), then the centers of (S5). By visual inspection, the amplitudes of the 41 spikes that
the spike clusters obtained by the K-means algorithm are used are not detected in S3 have small changes from the amplitudes
as the templates for adaptive template matching to identify of the background signals.
the candidate spikes (Fig. 6 (b)). However, not all candidate
peaks are real spikes, and these falsely detected spikes are TABLE III
completely removed during the FPS elimination (Fig. 6 (c)). P ERFORMANCE OF THE PROPOSED METHOD
The final result is shown in Fig. 6 (d).
Patients FN Sen(%) Spe(%) Sel(%) AC(%)
S1 6 97.7 96.6 98.1 97.3
S2 3 99.5 92.3 98.6 98.3
S3 41 91.0 97.0 96.9 93.9
S4 7 98.2 96.8 98.4 97.7
S5 6 99.3 88.9 99.0 98.4
S6 4 94.2 97.9 78.3 97.6
S7 1 98.4 95.9 66.3 96.1
Overall 68 97.4 96.5 96.6 96.9

In order to verify the superiority of the proposed method,

Table IV shows the results of spike detection using only
the AV channel EEG. Comparing the results between Table
III and Table IV, it can be found that the performance of
the spike detection model has been greatly improved after
considering BP channel EEG. There are two reasons: 1) The
waveform of electrooculogram (EOG) is similar to spike, and
Fig. 6. Results obtained by the proposed spike detection method. (a) Results BP channel can counteract the EOG artifacts present on AV
after universal template matching. (b) Results after adapted template matching. channel; 2) The spikes of BECT patients usually appear in
(c) Results after FPS elimination. (d) Results after RF classification. (e) Actual
spikes. the area between the electrodes, showing “peak-to-peak” on
BP channel EEG. BP channel EEG can be used to determine
As a result, the performance of the proposed method are the area associated with spike discharge and to eliminate FPS
calculated as shown in Table III. The overall Sen, Spe, Sel and missing spikes through “peak-to-peak”.
and AC are calculated based on the overall confusion matrix The performance of the proposed method is compared with

1549-7747 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on May 10,2020 at 19:41:20 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCSII.2020.2992285, IEEE
Transactions on Circuits and Systems II: Express Briefs
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMSłII: EXPRESS BRIEFS, VOL. X, NO. X, X 2020 5

TABLE IV research, a variety of epileptic spikes, not just BECT spikes,

R ESULTS OF AV CHANNEL EEG will be extracted and detected.
Patients Sen(%) Spe(%) Sel(%) AC(%)
E THICAL S TANDARDS
S1 89.8 97.3 86.3 96.2 This study has been approved by the Children’s Hospital,
S2 82.6 92.3 90.1 87.8 Zhejiang University School of Medicine and registered in
S3 90.3 91.9 69.6 91.6 Chinese Clinical Trail Registry (ChiCTR1900028804). All
S4 85.3 98.4 96.0 94.4 patients gave their informed consent prior to their inclusion
S5 94.3 90.2 97.9 93.6 in the study.
S6 88.4 83.6 30.8 84.0
S7 92.0 89.6 84.3 89.7
R EFERENCES
Overall 89.3 91.2 72.9 90.8
[1] H. Goldberg-Stern et al., “Neuropsychological aspects of benign child-
hood epilepsy with centrotemporal spikes,” Seizure, vol. 19, no. 1, pp.
12–16, 2010.
the recent work available from the literature listed in Table V. [2] S. Völkl-Kernstock et al., “Speech and school performance in children
with benign partial epilepsy with centro-temporal spikes (BCECTS),”
[14] applied a deep learning approach utilizing CNN and a Seizure, vol. 18, no. 5, pp. 320–326, 2009.
combination of RNN, and the overall Sen (69.0%) and AC [3] A. Page et al., “A flexible multichannel EEG feature extractor and
(89.0%) are lower than the proposed method (97.4% and classifier for seizure detection,” IEEE Trans. Circuits Syst. II, Exp. Briefs,
vol. 62, no. 2, pp. 109–113, 2015.
96.9%). Both [15] and [16] used tensor decomposition to [4] D. Gajic et.al, “Classification of EEG signals for detection of epileptic
extract spike features, and its performance is slightly lower seizures based on wavelets and statistical pattern recognition,” Biomed.
than the proposed method. It also can be seen that [18] has the Eng., Appl., Basis Commun., vol. 26, no. 2, pp. 1450021, 2014.
[5] J. R. Stevens et al., “Seizure occurrence and interspike interval: teleme-
highest Sen, which is 99.3%, but the Spe of [18] is only 82.8%. tered electroencephalogram studies,” Arch. Neurol., vol. 26, no. 5, pp.
Different from the proposed method, the data set in [18] 409–419, 1972.
includes the EEG signals of healthy people, but the data set in [6] D. Wu et al, “Automatic epileptic seizures joint detection algorithm based
on improved multi-domain feature of cEEG and spike feature of aEEG,”
this article includes the EEG signals of BECT patients. [19] IEEE Access, vol. 7, pp. 41551–41564, 2019.
has the highest Spe, but its Sen is only 69.3%. The Sen and Spe [7] L. Zou et al., “Removing muscle artifacts from EEG data via underde-
in [17] has small difference which are lower than the proposed termined joint blind source separation: s simulation study,” IEEE Trans.
Circuits Syst. II, Exp. Briefs, vol. 67, no. 1, pp. 187–191, 2020.
method. Clinically, doctors have a higher requirements for Sen [8] S. Nishida et al., “Signal separation of background EEG and spike by
than Spe. Therefore, the proposed method greatly improves the using morphological filter,” Med. Eng. Phys., vol. 21, no. 9, pp. 601–
Spe while ensuring high Sen. 608, 1999.
[9] T. Levi et al., “Development and applications of biomimetic neuronal net-
works toward brainmorphic artificial intelligence,” IEEE Trans. Circuits
TABLE V Syst. II, Exp. Briefs, vol. 65, no. 5, pp. 577–581, 2018.
P ERFORMANCE COMPARISON OF RECENT AUTOMATIC SPIKE DETECTION [10] D. Wu et al., “Epileptic seizure detection system based on multi-domain
METHODS feature and spike feature of EEG,” Int. J. Hum. Robot., vol. 16, no. 4,
pp. 1950016, 2019.
[11] G. Xu et al., “A spike detection method in EEG based on improved
Method Pat. Sen(%) Spe(%) Sel(%) AC(%) morphological filter,” Comput. Biol. Med., vol. 37, no. 11, pp. 1647–
Racz et al. [14] - 69.0 - - 89.0 1652, 2007.
[12] A. Nonclercq et al. “Cluster-based spike detection algorithm adapts to
Dao et al. [15] 17 82.8 96.4 - 93.8 interpatient and intrapatient variation in spike morphology,” J. Neurosci.
Thanh et al.[16] 17 83.0 91.1 - 91.0 Methods, vol. 210, no. 2, pp. 259–265, 2012.
Keshri et al. [17] 30 92.4 95.8 - - [13] G. E. Chatrian et al., “A glossary of terms most commonly used by clin-
ical electroencephalographers,” Electroencephalogr. Clin. Neurophysiol.,
Dao et al. [18] 6 99.3 82.8 - - vol. 37, pp. 538–48, 1974.
Ji et al. [19] 6 69.3 99.9 96.1 - [14] M. Racz et al., “Spike detection and sorting with deep learning,” J.
Salmasi et al. [20] - 90.0 - - - Neural Eng., vol. 17, no. 1, pp. 016038, 2020.
[15] N. T. Anh-Dao et al., “New feature selection method for multi-channel
Proposed method 7 97.4 96.5 96.6 96.9 EEG epileptic spike detection system,” REV J. Science, Comp. Science
and Com. Eng., vol. 35, no. 2, pp.47–59, 2019.
Pat.: Number of patients [16] L. T. Thanh et al., “Multi-channel EEG epileptic spike detection by a
new method of tensor decomposition,” J. Neural Eng., vol. 17, no. 1, pp.
016023, 2020.
[17] M. I. Khalid et al., “Epileptic MEG spikes detection using amplitude
IV. C ONCLUSION AND D ISCUSSIONS thresholding and dynamic time warping,” IEEE Access, vol. 5, pp. 11658-
11667, 2017.
In summary, this paper introduces a novel method for [18] N. T. Anh-Dao et al., “A multistage system for automatic detection of
BECT spike detection as an assistant tool for BECT diagnosis, epileptic spikes,” REV J. Electron. Commun., vol. 8, pp. 1–2, 2018.
especially useful in clinical medicine. To achieve spike detec- [19] Z. Ji et al., “An automatic spike detection system based on elimination
of false positives using the large-area context in the scalp EEG,” IEEE
tion with superior performance, an algorithm combining AV Trans. Biomed. Eng., vol. 58, no. 9, pp. 2478–2488, 2011.
channel and BP channel is proposed to identify BECT spikes. [20] M. Salmasi, U. B´’uttner and S. Glasauer, “Fractal dimension analysis
The method has been successfully implemented and validated for spike detection in low SNR extracellular signals,” J. Neural Eng., vol.
13, no. 3, pp. 036004, 2016.
on real EEG data of BECT patients. It can be shown that
the Sen, Spe, Sel and AC obtained by the proposed method
are 97.4%, 96.5%, 96.6% and 96.9%, respectively. In future

1549-7747 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on May 10,2020 at 19:41:20 UTC from IEEE Xplore. Restrictions apply.