IEEE TRANSACTIONS ON CONSUMER ELECTRONICS, VOL. 65, NO.

3, AUGUST 2019 379

eSeiz: An Edge-Device for Accurate Seizure

Detection for Smart Healthcare
Md. Abu Sayeed, Student Member, IEEE, Saraju P. Mohanty , Senior Member, IEEE,
Elias Kougianos , Senior Member, IEEE, and Hitten P. Zaveri , Member, IEEE

Abstract—Epilepsy is one of the most common neurological lation of neurons [1], which is manifested in the form of
disorders affecting a significant portion of the world’s popula- sensory disturbances, convulsions, and may lead to loss of
tion and approximately 2.5 million people in the United States. conscience.
Important biomedical research effort is focused on the devel-
opment of energy efficient devices for the real-time detection There are multiple lines of treatment for epilepsy. Anti-
of seizures. In this paper, we propose an Internet of Medical convulsant drugs are used as a first line to control seizures,
Things (IoMT)-based automated seizure detection system which but a significant portion (more than 30%) of patients remain
will detect a seizure from electroencephalography (EEG) sig- refractory to medication. For patients who are refractory,
nals using a voltage level detector (VLD) and a signal rejection uncontrolled seizures have a potentially devastating effect on
algorithm (SRA). The proposed system analyzes neural signals
continuously and extracts the hyper-synchronous pulses for the the patient’s quality of life. Epilepsy surgery is suitable for
detection of seizure onset. Within a time frame, if the number of some patients with medically refractory seizures, but not if
pulses exceeds a predefined threshold value, a seizure is declared. the patient has multi-focal seizures or if the seizure onset
The SRA reduces false detections, which in turn enhances the area is located in the eloquent cortex [2]. Other possibil-
accuracy of the seizure detector. The design was validated using ities include modification of diet, which may be effective
system-level simulations and consumer electronics proof of con-
cept. The proposed seizure detector reports a sensitivity of 96.9% in some children. Wearable and implantable devices consti-
and specificity of 97.5%. The use of minimal circuitry can lead to tute an important fourth line of treatment, and one whose
reduction of power consumption compared to many contempo- use is growing. Automated, closed loop therapy has shown
rary approaches. The proposed approach can be generalized to efficacy in managing epilepsy. Responsive neural stimulation
other sensor modalities and the use of wearable or implantable (RNS), for example, which is approved by the Food and
solutions, or a combination of the two.
Drug Administration (FDA) for use in the USA, has been
Index Terms—Smart healthcare, IoMT, wearables, epilepsy, shown to reduce the number of seizures experienced by a
seizure, electroencephalography, automated detection, energy patient. Automated seizure detection systems are a growing
efficient systems, low latency systems.
need for the treatment of epilepsy, as early warning can enable
a patient to take protective action, or enable drug delivery or
neurostimulation when necessary [3], [4].
I. I NTRODUCTION There are several ways to diagnose epilepsy by clinical
MART health care is becoming dominant due to the com- examinations. However, the diagnosis can be best performed
S bined pressures of an increasing and aging population, an
increasing demand for excellent care and limited resources.
by electroencephalography (EEG) due to its high temporal
resolution [5], [6]. The EEG can be collected in a number
Smart health achieved through wearables has been focused on of manners, including through the use of a cap, headband,
general wellness, but is now starting to encompass the man- and invasive or subcutaneous sensors. Also possible is the use
agement of acute disorders. One example of such an effort is of other wearable sensors monitoring galvanic skin response
the use of smart health care for the automated real-time con- and heart rate to detect changes within the autonomic ner-
trol of seizures. Epilepsy is a neurological disorder marked by vous system, which is reflective of seizure onset. Manual
spontaneous recurrent seizures. A seizure is the occurrence seizure detection is a tedious and time consuming task, which
of an abnormal hyper-synchronous disturbance of a popu- suggests a need for the development of automated seizure
detection systems which can detect seizures efficiently. In
Manuscript received February 27, 2019; revised April 28, 2019 and May this paper we propose such an automated seizure detection
22, 2019; accepted May 23, 2019. Date of publication May 30, 2019; date of system in the IoMT framework. The proposed seizure detec-
current version July 24, 2019. The work of H. Zaveri was supported in part
by the National Science Foundation and in part by the Swebilius Foundation. tor has reduced design complexity, low sensitivity to noise
(Corresponding author: Saraju P. Mohanty.) and simulation results indicate that it can be implemented
M. A. Sayeed and S. P. Mohanty are with the Department of Computer as a low-power system. Neural activity is monitored continu-
Science and Engineering, University of North Texas, Denton, TX 76201 USA
(e-mail: [email protected]; [email protected]). ously. The input signal is analyzed by a detection circuit and
E. Kougianos is with the Department of Engineering Technology, University potential seizure activity (hyper-synchronization) is detected.
of North Texas, Denton, TX 76201 USA (e-mail: [email protected]). Hyper-synchronous pulses are processed by a different algo-
H. P. Zaveri is with the Department of Neurology, Yale University,
New Haven, CT 06520 USA (e-mail: [email protected]). rithm. A seizure is declared if the pulses exceed a certain
Digital Object Identifier 10.1109/TCE.2019.2920068 threshold. The proposed seizure detector consists of a filter,
1558-4127 c 2019 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: SIDDAGANGA INSTITUTE OF TECHNOLOGY. Downloaded on November 28,2023 at 05:31:25 UTC from IEEE Xplore. Restrictions apply.
 380 IEEE TRANSACTIONS ON CONSUMER ELECTRONICS, VOL. 65, NO. 3, AUGUST 2019

TABLE I
E XISTING C ONSUMER E LECTRONICS W ORKS ON S MART H EALTHCARE

hyper-synchronous pulses within a time frame defines a seizure

onset. The continuous rejection of unwanted signals leads to
a reduction in false detections, which in turn enhances the
performance of the seizure detector. In common with current
IoMT trends, the proposed system provides the advantages of
remote healthcare monitoring and consultation. The patient’s
EEG data are continuously stored and analyzed on the cloud
and a notification is sent to the physician if a seizure occurs.
Physicians can take required actions based on the healthcare
Fig. 1. Seizure Detection Based on EEG data. report from the cloud.
Simulation results of an ideal prototype demonstrate low
power consumption. Combined with the system’s increased
an amplifier, a voltage level detector, and a signal rejection detection accuracy, it is suitable for use as a wearable or
algorithm. An illustration of the proposed system is shown in implantable device for seizure detection.
Fig. 1. The IoT framework associated with the seizure detector
enables recording the patient’s day to day activities and access-
ing healthcare data from anywhere and at anytime, hence the
III. R ELATED P RIOR R ESEARCH
remote connectivity leads to improved treatment.
The remainder of this paper is organized as follows: Table I provides a comparative presentation of relevant
Section II emphasizes the novel contributions of this work. existing research and development in consumer electronics and
Existing relevant research on seizure detection is presented illustrates the contributions of each report to smart healthcare.
in Section III. Section IV provides an architectural overview An IoMT based smart healthcare system has been proposed
of the platform. The proposed seizure detector is discussed in [7] which collects the patient’s medical data including
in Section V. The implementation of the design block is blood pressure and heart rate, and sends them to a physician
presented in Section VI. The simulation results are shown in through the cloud, providing remote healthcare. An architec-
Section VII and conclusions are presented in Section VIII. ture of remote electrocardiogram (ECG) monitoring with a
wireless sensor network has been proposed in [8]–[11] in
which portable sensors transmit data to an ECG server which
II. N OVEL C ONTRIBUTIONS OF THE C URRENT PAPER are then sent to hospitals and physicians for evaluation. An
An accurate and energy-efficient seizure detection system extensive survey has been conducted in [12] which reviews
is proposed in the IoMT framework. The detection of EEG the concepts, applications, current research trends, challenges,
abnormalities is a challenging task because of its high com- opportunities and significance of the IoMT in smart health
plexity. Existing algorithms show promising results for seizure care. U.S. regulators have recently approved the first medical
detection, though accuracy and power consumption remain grade smart watch, a novel consumer electronics product for
critical issues. The novel contributions of this work addresses neurological health, which measures abnormal activity during
these issues as discussed below: an epileptic seizure and sends alerts to a physician for proper
The proposed signal rejection algorithm (SRA) estimates action [13]. While this product has shown promising results for
the seizure onset by accurately eliminating unwanted bursts of the detection of generalized clonic-tonic seizures, significant
pulses and noise. The unwanted signal rejection continues in research remains to be conducted for the detection of partial
a time frame up to the threshold point. A threshold number of seizures. The proposed system advances consumer electronics

Authorized licensed use limited to: SIDDAGANGA INSTITUTE OF TECHNOLOGY. Downloaded on November 28,2023 at 05:31:25 UTC from IEEE Xplore. Restrictions apply.
 SAYEED et al.: eSeiz: EDGE-DEVICE FOR ACCURATE SEIZURE DETECTION FOR SMART HEALTHCARE 381

by bringing seizure detection and control to the smart health

care system.
Several seizure detection algorithms such as wavelet decom-
position [15], phase coherence [16], and signal synchroniza-
tion have been proposed. The implementations of these algo-
rithms are restricted to powerful desktop computers and are
not applicable to wearable or implanted devices. Considerable
research has been focused recently on developing implantable Fig. 2. Proposed eSeiz in the Internet of Medical Things (IoMT).
devices [17]–[19]. An algorithm has been defined for detect-
ing seizures based on the detection of events determined
from time and frequency features [17]. A seizure state is
defined by a threshold voltage associated with EEG abnor-
malities. Detection relies on positive and negative thresholds
but false detection is possible. A method based on sup-
port vector machines (SVM) [20] exhibits excellent detection
accuracy but numerous support vectors are required to dif-
ferentiate between a seizure and normal states. This results
in increased power requirements and high cost. A complex
detector [21] requires an ASIC to meet sensitivity require-
ments. A detection technique based on a CMOS preampli-
fier [19] provides good seizure detection accuracy at the
cost of high power. Due to poor CMOS noise immunity,
noise related issues degrade its performance [22]. In the
current work, to mitigate the noise problem, the detection Fig. 3. Seizure activity in the time domain: (a) intracranial EEG (icEEG)
method proposed is energy efficient and more immune to recording of a seizure, (b) an expanded view (zoom) of the time between
2985-3005 seconds.
noise.
The wavelet transform has the capability of capturing the
non-stationary behavior of EEG signals using time-frequency IV. T HE P ROPOSED S EIZURE D ETECTOR IN AN I NTERNET
localization. A wavelet transform-based seizure detection OF M EDICAL T HINGS P ERSPECTIVE
method [23] has been introduced which provides a sensitivity
Due to the increasing and increasingly aging population,
of 76% and a detection latency of 10 sec. A fast and accu-
traditional healthcare systems are not able to provide the
rate approach [24] of seizure detection has been proposed in
necessary services to everyone. Smart healthcare utilizes lim-
the edge-IoT framework that utilizes a naive Bayes classifier
ited resources in an efficient way to meet needs [12], [30].
for feature classification. The edge-IoT framework offers a
The IoMT in smart healthcare is an integration of universal
reduction in latency compared to the cloud-IoT and enhances
communication and connectivity where all the necessary com-
the detection accuracy for short duration intracranial EEG.
ponents can be connected together [7]. The proposed device
A patient-specific seizure classification method [25] has been
is divided into three components, as shown in Fig. 2.
presented, which uses an SVM as a classifier and achieves a
sensitivity of 96% and a mean detection latency of 4.6 sec.
An alternative approach to EEG is proposed [26], which uses A. Sensor Unit
a single wrist-worn accelerometer device for monitoring and The sensor unit consists of an EEG pre-processing unit and a
detection of convulsive seizures. The use of the accelerom- seizure detector. The input EEG signal is analyzed by the pre-
eter sensor reduces resource and labor requirements of the processing unit. The seizure detector continuously monitors
existing EEG based detection system. In the temporal synchro- the seizure state of the subject. The information relating to the
nization based approach [27], the recurrence pattern of seizure patient’s seizure state is then sent to remote storage through a
and non-seizure behavior is represented by a complex model wireless transfer.
and reports an average latency of 6 sec. In the local mean
decomposition (LMD) based approach [28], the EEG signal B. Transmission and Storage Unit
is decomposed into several product functions (PF) and then
The transmission unit acts as an interface between the sensor
the extracted features are submitted to a classifier for seizure
unit and cloud storage. The main function of the storage unit
detection.
is to store and manage the patient’s data. Cloud storage is
We previously reported an energy efficient seizure detec-
preferred as it enables data to be accessed from anywhere and
tor [29]. The system needs both software and hardware
provides redundancy.
validation with extensive EEG data. In the current extended
article, an accurate and energy efficient seizure detector is
proposed in the IoMT platform. The proposed eSeiz has been C. Access Unit
extensively validated, and uses a signal rejection algorithm for This unit allows health professionals to access data from the
seizure detection. cloud. The information relating to a patient is continuously

Authorized licensed use limited to: SIDDAGANGA INSTITUTE OF TECHNOLOGY. Downloaded on November 28,2023 at 05:31:25 UTC from IEEE Xplore. Restrictions apply.
 382 IEEE TRANSACTIONS ON CONSUMER ELECTRONICS, VOL. 65, NO. 3, AUGUST 2019

Fig. 4. Architecture of the Proposed Seizure Detector.

stored in the cloud. In the case of a seizure, a notification

is sent to the corresponding physician for proper action. The
physician will check the patient’s medication history as well
as documented seizures and adjust the required dosage for
the treatment of epilepsy. The data is also accessible by the
patients [7].

V. T HE P ROPOSED S EIZURE D ETECTOR

The proposed seizure detector (SD) monitors the brain activ-
ity at the seizure onset area. Fig. 3 shows a capture of seizure
onset. The architecture and flowchart of the proposed detector
are shown in Fig. 4 and Fig. 5, respectively. The input EEG
signals are filtered and submitted to an amplification unit. The
amplified signals of desired range are then passed through a
voltage level detector (VLD). The resulting hyper-synchronous
pulses from the VLD are then submitted to the signal rejection
algorithm unit. The SRA unit eliminates unwanted signals and
noise. The elimination of unnecessary signals continues until
the number of hyper-synchronous pulses surpasses the thresh-
old value. Seizure detection is characterized by the following Fig. 5. The Proposed Steps for the Seizure Detection in eSeiz.
equation:

1, seizure, for V(n − i) = 1 · · · and V(n) = 1
VSE (n) =
0, no seizure, otherwise,
(1)

where V(n) is the EEG sample at time sample n and i = Fig. 6. Hyper-synchronous Signal Detection Circuit.
1, 2, 3, . . . , N, with N being the threshold number of samples.
B. Signal Rejection Algorithm (SRA): Detection of Seizure
Onset From Hypersynchronous Signals
A. Hyper-Synchronous Signal Detection Circuit
The hyper-synchronous signals from the VLD are ana-
The proposed circuit in Fig. 6 consists of a band pass fil-
lyzed and spurious pulses are eliminated using the SRA. The
ter, an amplifier, and a voltage level detector (VLD). The
elimination of unwanted signals is performed by:
filter eliminates the unwanted signals and noise associated 
with the scalp-EEG signals and only keeps signals of the 0, V(n − 2) = 0 or V(n − 1) = 0
VSE1 (n) = (3)
desired frequency range. The low amplitude neural signals V(n), otherwise.
need to be amplified prior to analysis. This is achieved by the
Spurious pulses are further eliminated using:
amplification unit. The VLD analyzes the amplified signals

and detects hyper-synchronous pulses. The threshold values 0, V(n − 1) = 0 and V(n) = 1
(Vmax , Vmin ) of the VLD are determined from the analysis of VSE2 (n) = (4)
V(n), otherwise.
known seizures. The detection of the hyper-synchronous signal
is based on the following equation [19], [29]: The seizure onset is defined by [19], [29]:
⎧
 ⎨ Seizure, V(n − 2) = 1 and V(n − 1) = 1,
1, for Vmax > Vmod (n) > Vmin VSE (n) = if V(n) = 1 or 0
Vvld (n) = (2) ⎩
(5)
0, otherwise. 0, otherwise.

Authorized licensed use limited to: SIDDAGANGA INSTITUTE OF TECHNOLOGY. Downloaded on November 28,2023 at 05:31:25 UTC from IEEE Xplore. Restrictions apply.
 SAYEED et al.: eSeiz: EDGE-DEVICE FOR ACCURATE SEIZURE DETECTION FOR SMART HEALTHCARE 383

TABLE II
S IGNAL R EJECTION A LGORITHM (SRA) T ECHNIQUE

obtained from the VLD is input to the SRA unit. In the first
iteration of signal rejection, if a sample is either ‘0’ or ‘1’ and
the previous two samples are ‘0’, the algorithm outputs a 0. In
the next iteration, if the previous sample is ‘0’, the algorithm
outputs a ‘0’. For seizure onset at the (n − k)th iteration, if the
previous sample is ‘1’, the algorithm results a ‘1’. In the n-th
stage, the algorithm defines a seizure onset if the number of
hyper-synchronous pulses, which is denoted by ‘1’, exceeds
the threshold value. The optimal value of threshold, n and k
can be achieved by heuristic analysis of the known seizure
and non-seizure instances. The SRA technique is illustrated
Fig. 7. Design flow of the proposed eSeiz detector. in Table II, where the signals have been analyzed for seizure
onset and non-seizure onset instances.
Power estimation is generally performed using two
Neural signals are continuously monitored and seizure is approaches: pattern-dependent and pattern-independent. The
detected from the hyper-synchronous pulses (Vvld ). Within power consumption of the proposed system was computed
a time frame, the unwanted pulses are eliminated if their using the pattern-independent approach [31]. Different EEG
amplitude is lower than the threshold value. A seizure onset signals of identical size are applied to the design and the
is declared when the SRA completes the n-th iteration. If average of the computed power defines the power dissipation.
the number of hyper-synchronous pulses is greater than the The proposed seizure detector is viewed as a black box and
threshold number, a seizure is declared according to equa- current and voltage values are obtained from current and volt-
tions (3), (4), and (5). age sensors available in the system level simulator libraries.
A hardware-in-the-loop simulation approach was followed for
VI. C ONSUMER E LECTRONICS (CE) P ROOF OF the CE prototyping of the proposed system. A vendor-provided
C ONCEPT OF THE P ROPOSED E S EIZ hardware support package was used in the system level simu-
The EEG signal is initially preprocessed and filtered. lator and the proposed model was run on the actual board.
The low amplitude neural signal is then amplified using an EEG data and seizure information are continuously stored
adjustable gain amplifier. The maximum and minimum volt- on the eSeiz channel in the open IoT cloud. A liquid crys-
ages of the VLD define the hyper-synchronous signal. Using tal display (LCD), which is attached to the board, displays
the proposed SRA, the seizure onset is identified when it information about seizure state. If any seizure occurs, a noti-
occurs. The design flow of the proposed seizure detector is fication is sent to the designated user through the cloud. The
shown in Fig. 7. System-level prototyping of the proposed EEG data and seizure state are sent to the cloud. The proposed
eSeiz was performed as a first step towards CE prototyping. system consists of two channels: the eSeiz channel and the
The system level model of the proposed system is shown in EEG channel. The information on seizure state is stored in
Fig. 8. EEG signals are fed into the system. A band pass fil- the eSeiz channel whereas continuous EEG data are saved in
ter eliminates unwanted noise and extracts all seizure onset the EEG channel. Both patient and medical professionals have
information. The adjustable gain amplifier enables signals to access to the IoT cloud as well as the database using a REST
be amplified to the desired level. Hyper-synchronous signals API [24]. Fig. 9 shows the CE prototype of the proposed eSeiz.
are detected by the VLD. If the voltage is within the range, It should be pointed out that this prototype serves simply as
the function outputs a 1, otherwise it is zero. a proof-of-concept production. It is not optimized and con-
The maximum and minimum voltages of the VLD are deter- sumes substantially more power than a final product would.
mined by heuristic analysis of the amplified signal. For a For these reasons, in this discussion we examined the power
subject with an even number of seizures n, the Vmax and consumption from a simulation point of view since this will
Vmin values are computed from n/2 seizure instances. If n be much closer to an optimized, consumer-ready device.
is odd, the Vmax and Vmin values are obtained from (n − 1)/2
seizure instances. The average optimal values are adjusted by VII. E XPERIMENTAL R ESULTS
trial and error and are then applied to unknown seizure and The continuous and long term EEG recordings are taken
non-seizure instances. The hyper-synchronous signal which is from the CHB-MIT scalp EEG database [25], [32], which

Authorized licensed use limited to: SIDDAGANGA INSTITUTE OF TECHNOLOGY. Downloaded on November 28,2023 at 05:31:25 UTC from IEEE Xplore. Restrictions apply.
 384 IEEE TRANSACTIONS ON CONSUMER ELECTRONICS, VOL. 65, NO. 3, AUGUST 2019

Fig. 8. System-Level Simulator Model of (a) Proposed Seizure Detector. (b) Pattern-Independent Power Measurement Setup.

eliminates unwanted pulses with every iteration. The SRA iter-

ations continue until the number of hyper-synchronous pulses
surpass the threshold number. The unwanted pulses are being
eliminated as SRA completes the n-th iteration. The time
frame (Tf ) is in the range of milliseconds to seconds. For
patient-specific detection, the values of Tf , Vmax , and Vmin can
be varied accordingly. It is also reported that some unwanted
signals with amplitude in the VLD range trigger a false detec-
tion. In order to solve this problem, the statistical energy [24]
in each time frame is calculated for the known seizure and
non-seizure instances and an optimal value is determined by
a heuristic approach, as discussed earlier. The optimal thresh-
old value of energy in each time frame is considered. Even if
an instance is incorrectly detected by the SRA, further analy-
sis using the threshold energy provides correct detection. The
average amplitude of the seizure pattern at onset is between
Fig. 9. CE Prototype of the Proposed eSeiz Device.
150 mV and 450 mV. The frequency range of interest is
between 3 and 29 Hz. The performance of the detector is mea-
sured using sensitivity, specificity, and latency. The sensitivity
consists of EEG recordings from pediatric subjects. This work
and specificity of the classifier were evaluated as follows:
uses common EEG data from the following anonymized sub-
jects: chb01, chb03, chb05, chob08, chb11, chb17, and chb19. True Positive
EEG recordings were obtained at 256 samples per second with Sensitivity = (6)
True Positive + False Negative
16-bit resolution. The EEG electrodes were placed according True Negative
to the International Federation of Clinical Neurophysiology Specificity = (7)
True Negative + False Positive
10-20 placement system. Initially, the EEG signal is passed
through a band pass filter of frequency range between 3 Hz to The latency is defined as the delay between the expert marked
29 Hz, which is then amplified to a certain level. The amplified seizure onset and the seizure onset marked by the seizure
signal is applied to the VLD. The seizure onset information detector.
is extracted using the VLD. The maximum (Vmax ) and min- Figure 10 shows the analysis of the EEG signal and detec-
imum (Vmin ) voltages are computed by heuristic analysis of tion of seizure for epileptic subject 1 (chb01), in which seizure
the known seizures. As a good number of non-seizure sig- starts at 2996 sec. and ends at 3036 sec. The time frame for
nals fall in the VLD category, the VLD produces a number chb01 is selected as 500 ms. Fig. 10(a), 10(b), 10(c), 10(e),
of unwanted pulses. The output of the VLD, namely hyper- and 10(h) represent EEG epochs of different duration. The out-
synchronous pulses (Vvld ), are fed into the SRA. The SRA put of the VLD contains unnecessary and unwanted pulses, as

Authorized licensed use limited to: SIDDAGANGA INSTITUTE OF TECHNOLOGY. Downloaded on November 28,2023 at 05:31:25 UTC from IEEE Xplore. Restrictions apply.
 SAYEED et al.: eSeiz: EDGE-DEVICE FOR ACCURATE SEIZURE DETECTION FOR SMART HEALTHCARE 385

Fig. 10. Transient analysis (a) Input EEG signal of 2800-3200 seconds (b) EEG signal of 2975-3050 seconds (c,e) Zoom 2985-3005 seconds of input signal
(d) Output of VLD at 2985-3005 seconds (f) Output of SRA after 1st iteration (g) SRA output after the 2nd iteration (h) Zoom 2993-3003 seconds of input
signal (i) SRA output after the (n − k)-th iteration (j) SRA output after the n-th iteration.

TABLE III
C OMPARISON TO E XISTING S YSTEMS

depicted in Fig. 10(d). The SRA eliminates the unwanted sig- Fig. 10(i) shows the initiation of seizure detection after the
nals with every iteration. Fig. 10(f) and 10(g) are the output (n − k)-th iteration, which reports a smaller number of pulses.
of the SRA after the first and second iterations, respectively. After the n-th iteration, the SRA eliminates all spurious pulses

Authorized licensed use limited to: SIDDAGANGA INSTITUTE OF TECHNOLOGY. Downloaded on November 28,2023 at 05:31:25 UTC from IEEE Xplore. Restrictions apply.
 386 IEEE TRANSACTIONS ON CONSUMER ELECTRONICS, VOL. 65, NO. 3, AUGUST 2019

with existing seizure detectors is provided in Table III. The

characterization of the seizure detector is shown in Table IV.

VIII. C ONCLUSION AND F UTURE R ESEARCH

We have proposed an SRA-based automated seizure detector
in the IoMT framework. The proposed system was imple-
mented with a system level simulator in an open IoT platform.
The proposed signal rejection algorithm was useful for elim-
inating unwanted signals and extracting hyper-synchronous
activity from the EEG signals for the detection of seizure, thus
enhancing the performance of the seizure detector. Simulation
results also demonstrate that the proposed system can poten-
tially be implemented as a consumer electronics product
consuming low power, which makes it a suitable candidate for
low power wearable applications. In future research, we will
miniaturize the proposed system using the latest integrated cir-
cuit technologies for wearable biomedical applications and test
this solution in an animal model of epilepsy.

R EFERENCES
[1] F. Mormann, R. G. Andrzejak, C. E. Elger, and K. Lehnertz, “Seizure
prediction: The long and winding road,” Brain, vol. 130, no. 2,
pp. 314–333, Feb. 2007.
[2] S. Spencer and L. Huh, “Outcomes of epilepsy surgery in adults and
children,” Lancet Neurol., vol. 7, no. 6, pp. 525–537, 2008.
[3] I. Osorio, H. P. Zaveri, M. G. Frei, and S. Arthurs, Epilepsy: The
Intersection of Neurosciences, Biology, Mathematics, Engineering, and
Physics. Boca Raton, FL, USA: CRC Press, 2011.
[4] B. J. Gluckman and C. Schevon, “Seizure Prediction 6: From mecha-
nisms to engineered interventions for epilepsy,” J. Clin. Neurophysiol.,
Fig. 11. Transient analysis (a) Input EEG signal of 290-310 seconds vol. 32, no. 3, pp. 181–187, Feb. 2015.
(b) Output of VLD at 290-310 seconds (c) Zoom 295-305 seconds of input [5] L. S. Vidyaratne and K. M. Iftekharuddin, “Real-time epileptic seizure
signal (d) SRA output after the (n − k)-th iteration (e) SRA output after the detection using EEG,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 25,
n-th iteration. no. 11, pp. 2146–2156, Nov. 2017.
[6] R. G. Andrzejak, K. Schindler, and C. Rummel, “Nonrandomness,
TABLE IV nonlinear dependence, and nonstationarity of electroencephalographic
C HARACTERIZATION OF THE S EIZURE D ETECTOR recordings from epilepsy patients,” Phys. Rev. E, Stat. Phys. Plasmas
Fluids Relat. Interdiscip. Top., vol. 86, Oct. 2012, Art. no. 046206.
[7] K. Ullah, M. A. Shah, and S. Zhang, “Effective ways to use Internet of
Things in the field of medical and smart health care,” in Proc. Int. Conf.
Intell. Syst. Eng. (ICISE), Islamabad, Pakistan, 2016, pp. 372–379.
[8] N. Dey, A. S. Ashour, F. Shi, S. J. Fong, and R. S. Sherratt, “Developing
residential wireless sensor networks for ECG healthcare monitoring,”
IEEE Trans. Consum. Electron., vol. 63, no. 4, pp. 442–449, Nov. 2017.
[9] S. Ivanov, D. Botvich, and S. Balasubramaniam, “Cooperative wire-
less sensor environments supporting body area networks,” IEEE Trans.
Consum. Electron., vol. 58, no. 2, pp. 284–292, May 2012.
[10] S. Spinsante and E. Gambi, “Remote health monitoring by OSGi tech-
nology and digital TV integration,” IEEE Trans. Consum. Electron.,
vol. 58, no. 4, pp. 1434–1441, Nov. 2012.
[11] S. Raj and K. C. Ray, “A personalized point-of-care platform for real-
time ECG monitoring,” IEEE Trans. Consum. Electron., vol. 64, no. 4,
and declares a seizure. The seizure for subject 11 (chb11) is pp. 452–460, Nov. 2018.
shown in Fig. 11. Fig. 11(a) shows the EEG epoch for 290- [12] P. Sundaravadivel, E. Kougianos, S. P. Mohanty, and M. Ganapathiraju,
“Everything you wanted to know about smart healthcare,” IEEE Consum.
310 sec. The seizure instance originates at 298 sec and ends at Electron. Mag., vol. 8, no. 1, pp. 18–28, Jan. 2018.
320 sec. The outputs of the VLD and SRA (after the (n − k)- [13] E. Dolgin. (Feb. 2018). This Seizure-Detecting Smartwatch Could Save
th iteration) are shown in Fig. 11(b), and 11(d), respectively. Your Life. [Online]. Available: https://spectrum.ieee.org/the-human-
os/biomedical/diagnostics/this-seizuredetecting-smartwatch-could-save-
Finally, the resulting signal after the n-th SRA iteration as well your-life
as the detection of seizure is presented in Fig. 11(e). Overall, [14] S.-Y. Lee, P.-W. Huang, M.-C. Liang, J.-H. Hong, and J.-Y. Chen,
the detector missed one seizure instance for the chosen EEG “Development of an arrhythmia monitoring system and human study,”
dataset. The sensitivity, and specificity of the seizure detector IEEE Tran. Consum. Electron., vol. 64, no. 4, pp. 442–451, Nov. 2018.
[15] A. Shoeb, H. Edwards, J. Connolly, B. Bourgeois, S. T. Treves, and
were 96.9%, and 97.5%, respectively. The average latency of J. Guttag, “Patient-specific seizure onset detection,” Epilepsy Behav.,
the proposed system is found to be 3.6 seconds. A comparison vol. 5, no. 4, pp. 483–498, Aug. 2004.

Authorized licensed use limited to: SIDDAGANGA INSTITUTE OF TECHNOLOGY. Downloaded on November 28,2023 at 05:31:25 UTC from IEEE Xplore. Restrictions apply.
 SAYEED et al.: eSeiz: EDGE-DEVICE FOR ACCURATE SEIZURE DETECTION FOR SMART HEALTHCARE 387

[16] F. Mormann, K. Lehnertz, P. David, and C. E. Elger, “Mean phase coher- Md. Abu Sayeed (S’18) received the bache-
ence as a measure for phase synchronization and its application to the lor’s degree in electrical and electronic engineering
EEG of epilepsy patients,” Phys. D Nonlin. Phenom., vol. 144, nos. 3–4, from the Khulna University of Engineering and
pp. 358–369, Oct. 2000. Technology, Bangladesh, in 2009 and the mas-
[17] S. Raghunathan, S. K. Gupta, M. P. Ward, R. M. Worth, K. Roy, and ter’s degree in electrical engineering from Lamar
P. P. Irazoqui, “The design and hardware implementation of a low-power University, Beaumont, TX, USA, in 2015. He is cur-
real-time seizure detection algorithm,” J. Neural Eng., vol. 6, no. 5, rently pursuing the Ph.D. degree in computer science
Oct. 2009, Art. no. 056005. and engineering with the University of North Texas.
[18] N. C. Bhavaraju, M. G. Frei, and I. Osorio, “Analog seizure detection His research interests include biomedical signal pro-
and performance evaluation,” IEEE Trans. Biomed. Eng., vol. 53, no. 2, cessing for smart healthcare, machine learning, and
pp. 238–245, Feb. 2006. low-power very large-scale integration.
[19] M. T. Salam, M. Sawan, and D. K. Nguyen, “A novel low-power-
implantable epileptic seizure-onset detector,” IEEE Trans. Biomed.
Circuits Syst., vol. 5, no. 6, pp. 568–578, Dec. 2011.
Saraju P. Mohanty (SM’08) received the bach-
[20] N. Verma, A. Shoeb, J. Bohorquez, J. Dawson, J. Guttag, and
elor’s degree (Hons.) in electrical engineering
A. P. Chandrakasan, “A micro-power EEG acquisition SoC with inte-
from the Orissa University of Agriculture and
grated feature extraction processor for a chronic seizure detection
Technology, Bhubaneswar, 1995, the master’s degree
system,” IEEE J. Solid-State Circuits, vol. 45, no. 4, pp. 804–816,
in systems science and automation from the Indian
Apr. 2010.
Institute of Science, Bengaluru, in 1999, and the
[21] K. Patel, C.-P. Chua, S. Faul, and C. J. Bleakley, “Low power real-time
Ph.D. degree in computer science and engineering
seizure detection for ambulatory EEG,” in Proc. Int. Conf. PCTHealth
from the University of South Florida, Tampa, in
Pervasive Health, London, U.K., Apr. 2009, pp. 1–7.
2003.
[22] X. Zhang, H. Jiang, L. Zhang, C. Zhang, Z. Wang, and X. Chen,
He is a Professor with the University of North
“An energy-efficient ASIC for wireless body sensor networks in med-
Texas. His research is in “Smart Electronic Systems”
ical applications,” IEEE Trans. Biomed. Circuits Syst., vol. 4, no. 1,
which has been funded by National Science Foundations, Semiconductor
pp. 11–18, Feb. 2010.
Research Corporation, U.S. Air Force, IUSSTF, and Mission Innovation. He
[23] M. E. Saab and J. Gotman, “A system to detect the onset of epilep-
has authored 300 research articles, 4 books, and invented 4 U.S. patents. His
tic seizures in scalp EEG,” Clin. Neurophysiol., vol. 116, no. 2,
Google Scholar H-index is 31 and i10-index is 104. He was a recipient of
pp. 427–442, 2005.
the six best paper awards, the IEEE-CS-TCVLSI Distinguished Leadership
[24] M. A. Sayeed, S. P. Mohanty, E. Kougianos, V. P. Yanambakha, and
Award in 2018 for services to the IEEE, and to the VLSI research commu-
H. Zaveri, “A robust and fast seizure detector for IoT edge,” in Proc.
nity, and the 2016 PROSE Award for Best Textbook in Physical Sciences and
IEEE Int. Conf. Smart Electron. Syst. (iSES), Hyderabad, India, 2018,
Mathematics category from the Association of American Publishers for his
pp. 156–160.
Mixed-Signal System Design book published by McGraw-Hill in 2015. He
[25] A. Shoeb and J. Guttag, “Application of machine learning to epileptic
has delivered multiple keynote talks at various international conferences. He
seizure detection,” in Proc. Int. Conf. Mach. Learn., Haifa, Israel, 2010,
is the Editor-in-Chief of the IEEE Consumer Electronics Magazine. He served
pp. 975–982.
as the Chair of Technical Committee on VLSI, IEEE Computer Society from
[26] S. Kusmakar, C. K. Karmakar, B. Yan, T. J. O’Brien,
2014 to 2018.
R. Muthuganapathy, and M. Palaniswami, “Automated detection
of convulsive seizures using a wearable accelerometer device,” IEEE
Trans. Biomed. Eng., vol. 66, no. 2, pp. 421–432, Feb. 2019.
[27] M. Fan and C.-A. Chou, “Detecting abnormal pattern of epileptic Elias Kougianos (SM’07) received the B.S.E.E.
seizures via temporal synchronization of EEG signals,” IEEE Trans. degree from the University of Patras, Greece, in
Biomed. Eng., vol. 66, no. 3, pp. 601–608, Mar. 2019. 1985 and the M.S.E.E. degree, the M.S. degree
[28] T. Zhang and W. Chen, “LMD based features for the automatic seizure in physics, and the Ph.D. degree in EE from
detection of EEG signals using SVM,” IEEE Trans. Neural Syst. Rehabil. Lousiana State University in 1987, 1988, and 1997,
Eng., vol. 25, no. 8, pp. 1100–1108, Aug. 2017. respectively.
[29] M. A. Sayeed, S. P. Mohanty, E. Kougianos, and H. Zaveri, “An energy He is a Professor with the Department of
efficient epileptic seizure detector,” in Proc. IEEE Int. Conf. Consum. Engineering Technology, University of North Texas
Electron. (ICCE), Las Vegas, NV, USA, 2018, pp. 1–4. (UNT), Denton, TX, USA. From 1988 to 1997,
[30] M. A. Sayeed, S. P. Mohanty, E. Kougianos, and H. Zaveri, “A fast and he was with Texas Instruments, Inc., Houston
accurate approach for real-time seizure detection in the IoMT,” in Proc. and Dallas, TX, USA. In 1997, he joined Avant!
IEEE Int. Conf. Smart Cities (ISC2), Kansas City, MO, USA, 2018, Corporation (currently, Synopsys), Phoenix, AZ, USA, as a Senior
pp. 1–5. Applications Engineer. In 2001, he joined Cadence Design Systems, Inc.,
[31] U. Albalawi, S. P. Mohanty, and E. Kougianos, “Energy-efficient Dallas, TX, USA, as a Senior Architect in Analog/Mixed-Signal Custom
design of the secure better portable graphics compression architec- IC design. He has been with UNT since 2004. He has authored over 120
ture for trusted image communication in the IoT,” in Proc. IEEE peer-reviewed journal and conference publications. His research interests are
Comput. Soc. Annu. Symp. VLSI (ISVLSI), Pittsburgh, PA, USA, 2016, in the area of analog/mixed-signal/RF IC design and simulation and in the
pp. 302–307. development of VLSI architectures for multimedia applications.
[32] A. L. Goldberger, “PhysioBank, PhysioToolkit, and PhysioNet:
Components of a new research resource for complex physi-
ologic signals,” Circulation, vol. 101, no. 23, pp. e215–e220,
Hitten P. Zaveri (M’10) received the B.S.E.
Jun. 2000.
and M.S.E. degrees in electrical engineering, the
[33] J. Yoo, L. Yan, D. El-Damak, M. A. B. Altaf, A. H. Shoeb,
B.S.E. degree in computer engineering, and the
and A. P. Chandrakasan, “An 8-channel scalable EEG acquisition
M.S. and Ph.D. degrees in bioengineering from the
SoC with patient-specific seizure classification and recording pro-
University of Michigan, Ann Arbor. He was a Post-
cessor,” IEEE J. Solid-State Circuits, vol. 48, no. 1, pp. 214–228,
Doctoral Fellow in epilepsy and neurology at Yale
Jan. 2013.
University, where he is currently the Director of
[34] M. A. B. Altaf, C. Zhang, and J. Yoo, “A 16-channel patient-specific
the Computational Neurophysiology Laboratory. His
seizure onset and termination detection SoC with impedance-adaptive
research interests lie at the intersection of neuro-
transcranial electrical stimulator,” IEEE J. Solid-State Circuits, vol. 50,
science, engineering, and mathematics.
no. 11, pp. 2728–2740, Nov. 2015.

Authorized licensed use limited to: SIDDAGANGA INSTITUTE OF TECHNOLOGY. Downloaded on November 28,2023 at 05:31:25 UTC from IEEE Xplore. Restrictions apply.