micromachines

Article
Reliability Analysis of an Epileptic Seizure Detector
Powered by an Energy Harvester
Sunhee Kim 1, *, Suna Ju 2 and Chang-Hyeon Ji 2
1 Department of System Semiconductor Engineering, Sangmyung University, Cheonan-si 31066, Korea
2 Department of Electronic and Electrical Engineering, Ewha Womans University, Seoul 03760, Korea;
suna5290@gmail.com (S.J.); cji@ewha.ac.kr (C.-H.J.)
* Correspondence: happyshkim@smu.ac.kr; Tel.: +82-41-550-5357




Received: 8 December 2019; Accepted: 27 December 2019; Published: 30 December 2019


Abstract: Due to a limited lifetime of a battery, energy harvesters have been studied as alternative
energy sources for implantable biomedical devices such as an implantable stimulator for epileptic
seizure suppression. However, energy harvesters have weakness in providing stable power. We
designed a neural recording circuit powered solely by a piezoelectric energy harvester, and applied
its output to a seizure detector to analyze the reliability of the recorded signal. Performance of
the seizure detector was evaluated. We found that the average time differences between with and
without voltage variances were about 0.05 s under regular vibrations and about 0.07 s under irregular
vibrations, respectively. The ratio of average true positive alarm period varied within about 0.02%
under regular vibrations and 0.029% under irregular vibrations, respectively. The ratio of average
false positive alarm period varied within about 0.004% under regular vibrations and 0.014% under
irregular vibrations, respectively. This paper presents a reliability analysis of an epileptic seizure
detector with a neural signal recording circuit powered by a piezoelectric energy harvester. The
results showed that a supply voltage variance within ±10% could be acceptable for reliable operation
of a seizure detector.

Keywords: energy harvesting; epilepsy; medical signal detection; performance estimation; and
power supplies

1. Introduction
Epilepsy is the fourth most common neurological disorder and affects approximately 65 million
people around the world [1–4]. It is characterized by unpredictable and recurrent seizures [4,5].
Medication can control the majority of epileptic seizures and is almost the first therapy [4–6]. The
25%–30% of patients whose seizures cannot be controlled with anti-epileptic drugs may need other
treatments such as seizure-trigger-area-removing surgery and neuro stimulation [4–6]. A neuro
stimulation device for epileptic seizure is usually implanted within the skull under the scalp [5] and
consists of one or two electrodes, a neural signal recording circuit, a seizure detector, and a stimulator.
The implantable device has usually used a battery as its power source. The battery has to be
replaced surgically after the battery is dead. This can lead to economic burden and psychological
distress as well as physical pain [7]. Therefore, energy harvesting devices have attracted interest
in the implant device field as alternative power sources because they can convert ambient energy
into electrical energy [8–12]. Human-body motion based-harvesters, such as kinetic-energy or
vibration-energy harvesters, are more useful [7]. Piezoelectric energy harvesting technique is one
of the mechanical-to-electrical converting types. It has relatively high conversion efficiency, simple
configuration, and high power density [13,14]. However, vibrations occurred at a natural environment,
unlike an industrial environment, are distributed in low-frequency bands and vary unpredictably

Micromachines 2020, 11, 45; doi:10.3390/mi11010045 www.mdpi.com/journal/micromachines

Micromachines 2019, 10, x 2 of 12
Micromachines 2020, 11, 45 2 of 12

unpredictably from time to time [15]. Therefore, energy harvesters including piezoelectric energy
harvesters
from time to have
timebeen
[15].examined
Therefore,toenergy
find proper conditions
harvesters to be
including used as stable
piezoelectric energyor tolerable
harvesters power
have
sources. We studied
been examined to findconditions under which
proper conditions to beaused
targeted piezoelectric
as stable or tolerableenergy
powerharvester
sources.canWe support
studied
stable power
conditions to a neural
under which asignal recording
targeted circuit.energy
piezoelectric We analyzed thecan
harvester performance of anpower
support stable epileptic
to aseizure
neural
detector that processes
signal recording circuit.data
We from the neural
analyzed recording circuit
the performance under the
of an epileptic studied
seizure powerthat
detector conditions.
processes
data from the neural recording circuit under the studied power conditions.
2. Materials and Methods
2. Materials and Methods
We have estimated performance of a seizure detector when the seizure detector was connected
We haverecording
to a neural estimatedcircuit
performance
powered of abyseizure detector when
a piezoelectric energythe harvester.
seizure detector
We couldwas connected
not test theto
a neural recording
harvester and circuitscircuit powered
in vivo. by a piezoelectric
We tested and analyzed energy
power harvester. We could by
supply generated notthe
testharvester
the harvester
and
and circuits
circuits in vivo.
powered from Wethe
tested and analyzed
generated power supply
power supply, generated
particularly in theby thethat
case harvester and circuits
the power supply
powered from
fluctuated. thewe
Then, generated
made theirpower supply,circuit
equivalent particularly
modelsinandthe simulated
case that the power
seizure supply
data with fluctuated.
the circuit
Then, we
models tomade their
estimate equivalent circuit
performance. Detailed models and simulated
description seizure data with the circuit models to
is as follows.
estimate performance. Detailed description is as follows.
2.1. Harvester and Circuits
2.1. Harvester and Circuits
We have previously reported a frequency-up-converting impact-based piezoelectric energy
We have
harvester [16].previously reported
Figure 1 shows a adiagram
frequency-up-converting impact-based piezoelectric
of the used frequency-up-converting energy
impact-based
harvester [16]. Figure 1 shows a diagram of the used frequency-up-converting
piezoelectric energy harvester. It consisted of a 26 mm × 8 mm × 8 mm cuboid, a spherical ball with a impact-based
piezoelectric
radius of 2.5 energy harvester.
mm inside It consisted
the cuboid a 26 mm × 8 fiber-based
and aofpiezoceramic mm × 8 mmmacro cuboid,fiber
a spherical
compositeball (MFC)
with a
radius One
beam. of 2.5end
mmofinside the cuboid
the beam andat
was fixed a piezoceramic
the cuboid and fiber-based
the other macro fiber composite
end supported a proof(MFC) beam.
mass like a
One end of the beam was fixed at the cuboid and the other end supported a
cantilever. Two electrical ports worked as a reference signal and a generated voltage signal,proof mass like a cantilever.
Two electrical ports worked as a reference signal and a generated voltage signal, respectively.
respectively.

Figure
Figure 1.
1. Diagram
Diagramof
ofthe
theused
usedfrequency-up-converting
frequency-up-converting impact-based
impact-based piezoelectric
piezoelectric energy
energy harvester.
harvester.

When thethe harvester

harvestershook,
shook,the
theball
ball was
was bumped
bumped intointo
the the cuboid
cuboid because
because the ballthecould
ball could
move move
freely.
freely. This impact
This impact enabledenabled the flexible
the flexible MFC beam MFC tobeam to be changed
be changed in shape in andshape and the
the MFC MFC converted
converted this force
this
into force into an electrical
an electrical charge. Thecharge. The generated
generated voltage
voltage signal signal
had higherhadfrequency
higher frequency
than thethan
forcethe force
applied
applied to the harvester.
to the harvester. It couldthe
It could increase increase
energythe energy by
utilization utilization
converting by low-frequency
converting low-frequency
mechanical
mechanical vibrations
vibrations such as humansuchbody
as human body movements
movements into high-frequency
into high-frequency electrical signals.
electrical signals.
We composed power transfer circuits, which extracted power from the harvester and supplied
stable power to load circuits, by using an inductor (LM M),),aafull-bridge
full-bridgerectifier,
rectifier, aa storage
storage capacitor
capacitor (CSS),),
and a dual power supplier as shown in Figure 2. The inductor, LM M,, was
was used
used as a matching element
because
because the piezoelectric harvester could generate voltage maximally at inductive loads [17]. The
rectifier converted
convertedsparsely
sparselygenerated
generated voltage
voltage signals
signals to DC tovoltage,
DC voltage,
which which
was stored wasatstored at the
the capacitor.
capacitor. The dual
The dual power power generated
supplier supplier generated
a positivea power
positivesupply
powervoltage
supply voltage
(VDD) and (VDD) and a negative
a negative power
power
supplysupply
voltagevoltage (VSS)
(VSS) for loadfor load circuits.
circuits.
Neural signal recording circuits for an epileptic seizure detector consisted of an amplifier, a low
pass filter (LPF), a high pass filter (HPF), and an analog-to-digital converter(ADC) as shown in Figure 2.
An instrumentation amplifier was used at the first stage as an input buffer amplifier. The LPF and the
Micromachines 2019, 10, x 3 of 12

Micromachines 2020, 11, 45 3 of 12

HPF selected desired frequency components and removed DC offset. The ADC sampled processed
signals and transferred
Micromachines 2019, 10, x them to a host. 3 of 12

Figure 2. Power transfer circuits and neural signal recording circuits.

Neural signal recording circuits for an epileptic seizure detector consisted of an amplifier, a low
pass filter (LPF), a high pass filter (HPF), and an analog-to-digital converter(ADC) as shown in Figure
2. An instrumentation amplifier was used at the first stage as an input buffer amplifier. The LPF and
the HPF selected desired frequency components and removed DC offset. The ADC sampled
processed signals and Figure
2.2.Power
Powerthem
transferred
Figure transfer circuits
a host.and neural
tocircuits
transfer neural signal
signalrecording
recordingcircuits.
circuits.

2.2. TestNeural
2.2. Test of
of the
the Harvester
Harvester and
and the
signal recording Circuits
the circuits
Circuitsfor an epileptic seizure detector consisted of an amplifier, a low
pass filter (LPF), a high pass filter (HPF), and an analog-to-digital converter(ADC) as shown in Figure
SB120 Schottky
Schottky barrier
barrier rectifiers
rectifiers (Vishay General
General Semiconductor,
Semiconductor, Malvern,Malvern, PA, PA, USA) were used
2. SB120
An instrumentation amplifier was(Vishayused at the first stage as an input buffer amplifier. USA)
The LPFwereand
used
for the
forthe full-bridge
the HPF
full-bridge rectifier.
rectifier. AD8500 CMOS operational amplifiers (Analog Devices, Norwood, MA,
selected desiredAD8500frequency CMOS operational
components andamplifiers
removed (Analog
DC offset. Devices,
The ADCNorwood,
sampledMA,
USA)
USA) were used
were used for the dual power supplier, the instrumentation amplifier, and the LPF. The matching
processed signalsforand the dual power
transferred themsupplier,
to a host. the instrumentation amplifier, and the LPF. The
inductor
matchingwas 27 mH.
inductor wasAn 8-bit An
27 mH. ADC ADS7868
8-bit ADC ADS7868(Burr-Brown, Tucson,Tucson,
(Burr-Brown, AZ, USA) AZ,wasUSA) used. A field
was used. A
programmable
2.2.programmable
field gate
Test of the Harvesterarray (FPGA)
and the(FPGA)
gate array Starter including an Altera XC3S200 (Libertron,
CircuitsStarter including an Altera XC3S200 (Libertron, Seoul, Korea), Seoul, Korea), a
universal
a universal asynchronous receiver/transmitter (UART) connector was used to transfer the sampled data
SB120asynchronous
Schottky barrier receiver/transmitter
rectifiers (Vishay General(UART) connector wasMalvern,
Semiconductor, used toPA, transfer
USA) thewere sampled
used
to a
data host. A test system was built by using a LDS V406 M4 shaker (Brüel and Kjær, Narum, Denmark),
fortothea full-bridge
host. A testrectifier.
systemAD8500was built CMOSby using a LDSamplifiers
operational V406 M4(Analogshaker Devices,
(Brüel and Kjær, Narum,
Norwood, MA,
aDenmark),
33220A waveforma 33220A generator
waveform (Agilent,
generatorSanta(Agilent,
Clara, CA, USA),
Santa a R300PLUS
Clara, CA, USA), amplifier
a (Inter-M,
R300PLUS Yangju,
amplifier
USA) were used for the dual power supplier, the instrumentation amplifier, and the LPF. The
Korea),
(Inter-M, a SDS6062
Yangju, digital oscilloscope
Korea), a SDS6062 (OWON,
digital Zhangzhou,
oscilloscope China),
(OWON, and an Agilent
Zhangzhou, E3631A an
China), DCAgilent
power
matching inductor was 27 mH. An 8-bit ADC ADS7868 (Burr-Brown, Tucson, AZ, USA)andwas used. A
supply
E3631A (Agilent,
DC power Santa
supplyClara, CA, USA).
(Agilent, Santa Clara, CA, USA).
field programmable gate array (FPGA) Starter including an Altera XC3S200 (Libertron, Seoul, Korea),
Figure
Figure 33 shows
a universal shows measured
measured
asynchronous open-circuit
open-circuit output
receiver/transmitter output voltage
voltage
(UART) signals
signalswas
connector of
of the
the
usedharvester
harvester when
when
to transfer the
the shaker
shaker
the sampled
excited
data to
excited the harvester
thea harvester
host. A test at acceleration
system was built
at acceleration of 3 g
of 3 gbyandand
usingfrequency
a LDS V406
frequency of 20
of 20 M4 Hz
Hz shakerin the same direction
(Brüeldirection
in the same and Kjær, as shown
asNarum,
shown in
in Figure
Denmark),
Figure 1. As shown
a 33220A
1. As shown in Figure
waveform
in Figure 3a, the
generator
3a, the harvester
harvester(Agilent,generated voltage
Santa voltage
generated Clara, CA, signals
USA),
signals twice
a R300PLUS
twice in one period
amplifier
in one period of
of the
the applied
(Inter-M,
applied low-frequency
Yangju,
low-frequency Korea), avibration
SDS6062
vibration because
digital
because the the freely
oscilloscope
freely movable
(OWON,
movable ball
Zhangzhou,
ball impactedimpacted at both
China),
at both and
the anthe bottom
Agilent
bottom and
and
top end in one cycle. Because expansion and contraction coefficients of the MFC as well as the
top
E3631A endDC in one
power cycle.
supply Because
(Agilent, expansion
Santa and
Clara, contraction
CA, USA). coefficients of the MFC as well as the
collision
collision forces
Figure
forces 3at both
shows
at both sides
measuredwerewere
sides different,
open-circuit two different
output
different, two waveform
voltage signals
different signals
of
waveform were
the harvester generated
signals when
werethe alternately
shaker
generated
as shown the
excited
alternately inasFigure
shown3b,c.
harvester The generated
inatFigure
acceleration of 3 signal
3b,c. The oscillated
ggenerated
and frequency
signalofaround
20 Hz infrequency
oscillated the same of
around 8.5 KHz,
direction
frequency attenuated
as of
shown in
8.5 KHz,
almost
Figure
attenuatedexponentially
1. almost
As shown after
exponentiallypositive-
in Figure 3a, negative
the
after harvester
positive- peaks, and returned
generated
negative voltage
peaks, andto zero. twicetoinzero.
signals
returned one period of the
applied low-frequency vibration because the freely movable ball impacted at both the bottom and
top end in one cycle. Because expansion and contraction coefficients of the MFC as well as the
collision forces at both sides were different, two different waveform signals were generated
alternately as shown in Figure 3b,c. The generated signal oscillated around frequency of 8.5 KHz,
attenuated almost exponentially after positive- negative peaks, and returned to zero.

(a) (b) (c)

Figure 3. Measured
Measured open-circuit
open-circuit voltage
voltage at regular force: 20
20 Hz
Hz excitation
excitation with
with 33 g acceleration,
acceleration, (a)
150ms,
during 150 ms,(b)
(b) zoomed
zoomed bottom-collision
bottom-collision voltage
voltage during
during 1.5and
1.5 ms, ms,(c)and (c) zoomed
zoomed up-collision
up-collision voltage
voltage1.5
during during
ms. 1.5 ms.
(a) (b) (c)
To analyze
ToFigure 3. the
analyze effects
the
Measured of the
effects ofgenerated VDD
thevoltage
open-circuit generated onVDD
theforce:
at regular neural signal
on 20the recording
Hz neural withcircuits,
signal
excitation we experimented
3recording circuits,
g acceleration, (a) we
as follows.
experimentedFigure
during 150 4 shows
as follows.
ms, measured
Figure
(b) zoomed signals
4 shows of the
measured
bottom-collision circuits when
signals
voltage during of1.5 the
the harvester
circuits
ms, shook
and (c)when
zoomed at 20 Hz with
theup-collision
harvester 3g
shook
and an voltage
input signal
duringof 1.5the
ms.instrumentation amplifier was a 50 Hz sinusoidal voltage. The capacitance

To analyze the effects of the generated VDD on the neural signal recording circuits, we
experimented as follows. Figure 4 shows measured signals of the circuits when the harvester shook
Micromachines 2020, 11, 45 4 of 12
Micromachines 2019, 10, x 4 of 12

values ofHz
at 20 Cswith
in Figure
3 g and4a–d were
an input 1 mF
signal of and 0.33 uF, respectively.
the instrumentation amplifierThe
wasVDD ofsinusoidal
a 50 Hz the circuits increased
voltage.
whenThe ancapacitance
impact occurred,
values of Cs in Figure 4a–d were 1 mF and 0.33 uF, respectively. The VDD ofCs
and then decreased until the next event. When the capacitance thewas 1
mF, acircuits
ripple increased
of the VDD whenwasan±1.2%
impact(Figure 4b).and
occurred, Whenthenthe capacitance
decreased until Cs
thewas
next0.33 uF, When
event. a ripple
theof the
VDDcapacitance
was ±10%Cs was 1 mF,
(Figure 4d), awhich
ripple we
of the
setVDD
as thewas ±1.2%
worst (Figure
case. 4b). When the
Accordingly, the capacitance Cs was
output voltage signals
of the0.33 uF, a ripple of the
instrumentation VDD wasthe
amplifier, ±10%
LPF,(Figure
and the4d),HPF,
which we the
and set as the worst
input voltagecase. Accordingly,
signal the had
of the ADC
output voltage
distortions signals
and spikes, of the instrumentation
especially when the VDD amplifier, the LPF, and the HPF, and the input voltage
rose sharply.
signal of the ADC had distortions and spikes, especially when the VDD rose sharply.

(a) (b)

(c) (d)
Figure
Figure 4. Measured
4. Measured data
data signals.
signals. (a)(a)Output
Outputsignals
signals of
of the
the amplifier
amplifier and
andthe
thelow
lowpass
passfilter (LPF)
filter at at Cs
(LPF)
Cs = 1 mF, (b) the positive supply voltage (VDD), the output signal of the high pass filter (HPF), and
= 1 mF, (b) the positive supply voltage (VDD), the output signal of the high pass filter (HPF), and the
the input signal of the analog-to-digital converter(ADC) at Cs = 1 mF, (c) output signals of the
input signal of the analog-to-digital converter(ADC) at Cs = 1 mF, (c) output signals of the amplifier
amplifier and the LPF at Cs = 0.33 uF, and (d) the positive supply voltage VDD, the output signal of
and the LPF at Cs = 0.33 uF, and (d) the positive supply voltage VDD, the output signal of the HPF, and
the HPF, and the input signal of the ADC at Cs = 0.33 uF.
the input signal of the ADC at Cs = 0.33 uF.
We tested the ADC separately by using a commercial power supply instead of the harvester.
We tested
Figure the ADC
5b shows separately
the sampled byanusing
data by a commercial
8-bit ADC power
when the supply supply
voltages of instead
the ADCof the harvester.
varied with
Figure 5b shows
slopes as showntheinsampled data
Figure 5a. Theby an 8-bitofADC
frequency when voltages
the supply the supply
wasvoltages
20 Hz (a 1ofmsthe ADC
rising varied
time and with
slopes as ms
a 49 shown in time).
falling FigureThe5a.magnitudes
The frequency
were of(2 ±the
α)supply
V, which voltages
α was 0,was 200.04,
0.02, Hz (a 1 ms
0.06, andrising
0.08, time
and arespectively.
49 ms fallingThetime).
input signals of the ADC were
The magnitudes were fixed at 1V,V,which
(2 ± α) which α
waswasexpected
0, 0.02,as0.04,
their0.06,
sampled
and 0.08,
data wereThe
respectively. 127input
at a 2signals
V supply voltage.
of the ADC When the supply
were fixed at 1 V,voltage decreased,
which was expected input
as analog signals data
their sampled
werebecome
127 at alarger relatively
2 V supply and viceWhen
voltage. versa.the
Therefore,
supplythe sampled
voltage data, withinput
decreased, an average
analogvalue of 127,
signals become
were changed on the contrary to the supply voltages. In addition, even when the supply voltage was
larger relatively and vice versa. Therefore, the sampled data, with an average value of 127, were
fixed at 2 V by a commercial power supply, the sampled ADC data had variations due to various
changed on the contrary to the supply voltages. In addition, even when the supply voltage was fixed
noises.
at 2 V by a commercial power supply, the sampled ADC data had variations due to various noises.
Figure 6 shows standard deviations of the sampled ADC data according to supply voltages and
input signals of the ADC. The magnitudes of the supply voltages were (2 ± α) V, which α was 0, 0.02,
0.04, 0.06, 0.08, and 0.10, respectively. The amplitudes of the input signals were from 0.1 V to 1.9 V.
When the variation of the supply voltage increased, the standard deviation of the sampled ADC data
also increased. In addition, when the magnitude of the input signals increased, the standard deviation
also increased.
In summary, the supply voltage generated by the harvester had some ripple and caused signal
distortions and spikes. In addition, when processed signals was sampled at the ADC, they were
modified according to their magnitude and the supply voltage magnitude at that time.
Micromachines 2020, 11, 45 5 of 12
Micromachines 2019, 10, x 5 of 12

(a) (b)
Figure 5. The sampled ADC data when an ADC was powered by a ramp signal. (a) Supply voltage
signals with a (2 ± α) V ramp shape, and (b) the sampled ADC data.

Figure 6 shows standard deviations of the sampled ADC data according to supply voltages and
input signals of the ADC. The magnitudes of the supply voltages were (2 ± α) V, which α was 0, 0.02,
0.04, 0.06, 0.08, and 0.10, (a)
respectively. The amplitudes of the input signals(b) were from 0.1 V to 1.9 V.
When the variation of the supply voltage increased, the standard deviation of the sampled ADC data
Figure 5.
Figure 5. The
also increased. The sampled ADC
sampled
In addition, ADC
whendata
data when
thewhen an ADC
an ADCofwas
magnitude was powered
thepowered by aa ramp
by
input signals ramp signal.the
signal.
increased, (a) standard
(a) Supply voltage
Supply voltage
deviation
signals with a (2 ± α) V ramp shape, and (b) the sampled ADC data.
signals with a (2 ± α) V ramp shape, and (b) the sampled ADC data.
also increased.

Figure 6 shows standard deviations of the sampled ADC data according to supply voltages and
input signals of the ADC. The magnitudes of the supply voltages were (2 ± α) V, which α was 0, 0.02,
0.04, 0.06, 0.08, and 0.10, respectively. The amplitudes of the input signals were from 0.1 V to 1.9 V.
When the variation of the supply voltage increased, the standard deviation of the sampled ADC data
also increased. In addition, when the magnitude of the input signals increased, the standard deviation
also increased.

Figure 6. The
Figure 6. The standard
standard deviation
deviation of
of the
the sampled
sampled ADC
ADC data
data according
according to
to VDDs
VDDs and input voltage
and input voltage
signals of the ADC in the range of 0.1 V to 1.9 V.
signals of the ADC in the range of 0.1 V to 1.9 V.

2.3. Seizure Detection

In summary, theAlgorithm
supply voltage generated by the harvester had some ripple and caused signal
distortions and spikes. In addition,
We used a generic Osorio when processed
Frei algorithm signals
(GOFA) [18] amongwas sampled
known at the
seizure ADC, algorithms.
detection they were
modified according to their magnitude and the supply voltage magnitude at that time.
The GOFA is based on recorded data from intracranial electro- encephalography (icEEG) to extract the
characteristic features of seizures: Energy and spectral analysis. The GOFA enables seizure detection
Figure Detection
2.3. Seizure 6. The standard deviation of the sampled ADC data according to VDDs and input voltage
Algorithm
by adjusting detection parameters according to energy and spectral elements. The GOFA and its
signals of the ADC in the range of 0.1 V to 1.9 V.
modified
We usedalgorithms usually
a generic consistFrei
Osorio of the following(GOFA)
algorithm procedures[18] [18–21].
among known seizure detection
algorithms. The GOFA is based on recorded data from
In summary, the supply voltage generated by the harvester had intracranial electro-
someencephalography
ripple and caused (icEEG)
signal
2.3.1.
to Filtering
extract the characteristic features of seizures: Energy and spectral analysis. The GOFA enables
distortions and spikes. In addition, when processed signals was sampled at the ADC, they were
seizure
modified detection
A level-3
according by to
DAUB4 adjusting
their detectionfinite
wavelet-based
magnitude parameters
and theimpulseaccording
supply voltage to
response energy
(FIR)
magnitude and
filter spectral
atextracts
that elements. The
seizure-related
time.
GOFA
frequency andbandits modified
componentsalgorithms usually
from icEEG thatconsist of the
recorded at following
240 Hz. If procedures
the recorded[18–21].
data are denoted as
k |kSeizure
2.3.
{x = 1, 2, .Detection
. . }, the filtered
Algorithmdata are given by
2.3.1. Filtering
We used a generic Osorio Frei algorithm Xp−1 (GOFA) [18] among known seizure detection
A level-3 DAUB4 wavelet-based
algorithms. The GOFA is based on recorded data from y
finite
k = impulse b jresponse
xk− j (FIR) filterencephalography (1)
extracts seizure-related
j=0 intracranial electro- (icEEG)
frequency
to extract the bandcharacteristic
components features
from icEEG that recorded
of seizures: Energy at 240
andHz. If the recorded
spectral analysis. data are denoted
The GOFA as
enables
where
{x |k = {b
1, 0 ,
2, b ,
…},
1 · · · b
the } and
filtered
p−1 p are
data the
are coefficients
given by and the order of the FIR filter,
seizure detection by adjusting detection parameters according to energy and spectral elements. The
k respectively.
GOFA and its modified algorithms usually consist of the following procedures [18–21].
2.3.2. Calculating Foreground Sequences 𝑦 = 𝑏 𝑥 (1)
2.3.1.The
Filtering
filtered data yk are squared and passed through a median filter. The median filter processes a
where 0, b1, ⋯ bp−1} and p are the coefficients and the order of the FIR filter, respectively.
moving{bwindow with the most recent 2 s of data and can separate short bursts such as certain artifacts
A level-3 DAUB4 wavelet-based finite impulse response (FIR) filter extracts seizure-related
or single spike. The resulting sequence is termed “foreground {FGk }” and given by
frequency band components from icEEG that recorded at 240 Hz. If the recorded data are denoted as
{xk|k = 1, 2, …}, the filtered data are given by n 2 2 o
FGk = median yk , yk−1 , . . . , y2k−OFG+1 (2)
𝑦 = 𝑏 𝑥 (1)
where OFG = 480 is the order of the median filter for the foreground sequence.

where {b0, b1, ⋯ bp−1} and p are the coefficients and the order of the FIR filter, respectively.
2.3.2. Calculating Foreground Sequences
The filtered data yk are squared and passed through a median filter. The median filter processes
a moving window with the most recent 2 s of data and can separate short bursts such as certain
artifacts or single spike. The resulting sequence is termed “foreground {FGk}” and given by
Micromachines 2020, 11, 45 6 of 12
𝐹𝐺 = median{𝑦 , 𝑦 ,…,𝑦 } (2)
where OFG = 480 is
2.3.3. Calculating the order ofSequences
Background the median filter for the foreground sequence.
To search the changes in the foreground, the foreground sequence is compared with a reference,
2.3.3. Calculating Background Sequences
which is called background {BGk }. The foreground sequence is sampled every 0.5 s and passed through
To search
another median the changes in the foreground, the foreground sequence is compared with a reference,
filter.
which is calledbackground {BG n k}. The foreground sequence o is sampled every 0.5 s and passed
 (1 − λ)median FG , FG , . . . , FG
through another  median filter. k k−s k−(OBG−1)s + (λ)BGk−1 , if k = ns
BGk =  (3)
 BGk−1 , i f n(s − 1)k ns
1 − 𝜆 median 𝐹𝐺 , 𝐹𝐺 , … , 𝐹𝐺 + 𝜆 𝐵𝐺 , if 𝑘 = 𝑛𝑠

𝐵𝐺 = (3)
𝐵𝐺 , 𝑖𝑓 𝑛 𝑠 − 1 𝑘 𝑛𝑠
where n = 0, 1, 2, · · · s = 120, λ = 0.999807 (forgetting factor) and OBG = 240 (the order of the median
filter for
where n =the 1, 2, ⋯ s = 120,
0, background sequence).
λ = 0.999807 (forgetting factor) and OBG = 240 (the order of the median
filter for the background sequence).
2.3.4. Decision
2.3.4.To
Decision
decide whether seizure occurs or not, the dimensionless ratio, rk , is calculated as
To decide whether seizure occurs or not, the dimensionless
FGk ratio, rk, is calculated as
rk = . (4)
𝑟 = BGk. (4)

Then,Then, two parameters,

two parameters, threshold,
threshold, Thon,Th on ,duration,
and and duration, Donused.
Don, are , are used.
WhenWhen rk remains
rk remains at a given
at a given Thon
Thabove
or on or above for a given
for a given Donsignal
Don, the , the signal is decided
is decided as a seizure.
as a seizure.
2.4. Simulation
2.4. Simulation
2.4.1. Supply Voltage Conditions
2.4.1. Supply Voltage Conditions
We have previously reported equivalent circuit models of the frequency-up-converting
We have previously
impact-based piezoelectricreported
energyequivalent
harvestercircuit models
and its of the
validity frequency-up-converting
[17,22]. We simulated to find impact-
the
based piezoelectric energy harvester and its validity [17,22]. We simulated to find the lowest
lowest available frequency of external vibrations to support stable power supply voltages to the circuits available
frequency
by using theof equivalent
external vibrations to support stable power supply voltages to the circuits by using the
circuit model.
equivalent
Figurecircuit
7 showsmodel.
simulation results at a regular vibration: 13.5 Hz excitation with 3.5 g acceleration.
As shown in Figure 7a,simulation
Figure 7 shows the harvester results at a voltage
generated regular signals
vibration: 13.5 Hz
regularly excitation
twice with 74
in one period, 3.5ms.
g
acceleration. As shown in Figure 7a, the harvester generated voltage signals regularly
Peak values were 17.6 V and 8 V, respectively. Figure 7b shows a VDD and a VSS of the neural recording twice in one
period,
circuits.74
Asms. Peakinvalues
shown Figurewere 17.6
7c, the V and
VDD was8stable
V, respectively. Figure
at an average 7b shows
of 1.7953 V witha VDD andofa ±0.05%.
a ripple VSS of
the neural recording circuits. As shown in Figure 7c, the VDD was stable at an average
In other words, when the harvester shook at 13.5 Hz excitation with 3.5 g acceleration, charged and of 1.7953 V
with a ripple
discharged of ±0.05%.
energy In other
at Cs were wellwords,
balancedwhen the harvester shook at 13.5 Hz excitation with 3.5 g
at 1.795V.
acceleration, charged and discharged energy at Cs were well balanced at 1.795V.

(a) (b) (c)

Figure
Figure 7.
7. Simulation
Simulation results
results in
in case
case of
of aa regular
regular vibration,
vibration, (a)
(a) harvester
harvester voltage
voltage outputs,
outputs, (b)
(b) aa VDD
VDD
and
and aa negative
negative power
power supply
supply voltage
voltage (VSS)
(VSS) for
for the
the neural
neural recording
recording circuits,
circuits, and
and (c)
(c) zoomed-in
zoomed-in VDDVDD
from
from (b).
(b).

Figure 8 shows a voltage signal generated by the harvester and a VDD in case of irregular
vibrations. The reference vibration period was 74 ms (13.5 Hz) and its variance was ±40%. In other
words, the period was in the range of 44.4 ms (22.5 Hz) to 103.6 ms. (9.7 Hz). We assumed that the
average acceleration was 3.0 g and the variance of the generated voltage was also ±40%. As a result,
the VDD was changed within ±0.006% of 1.795 V.
 Figure 8 shows a voltage signal generated by the harvester and a VDD in case of irregular
Figure 8 shows a voltage signal generated by the harvester and a VDD in case of irregular
vibrations. The reference vibration period was 74 ms (13.5 Hz) and its variance was ±40%. In other
vibrations. The reference vibration period was 74 ms (13.5 Hz) and its variance was ±40%. In other
words, the period was in the range of 44.4 ms (22.5 Hz) to 103.6 ms. (9.7 Hz). We assumed that the
words, the period was in the range of 44.4 ms (22.5 Hz) to 103.6 ms. (9.7 Hz). We assumed that the
average acceleration was 3.0 g and the variance of the generated voltage was also ±40%. As a result,
average acceleration
Micromachines 2020, 11, 45was 3.0 g and the variance of the generated voltage was also ±40%. As a result,
7 of 12
the VDD was changed within ±0.006% of 1.795 V.
the VDD was changed within ±0.006% of 1.795 V.

(a) (b)
(a) (b)
Figure 8. Simulations results in case of irregular vibrations: (a) Reference vibration is 13.5 Hz
Figure 8. Simulations
Simulationsresults
resultsin in case
case of of irregular vibrations: (a) Reference vibration is excitation
13.5 Hz
Figure 8.
excitation with 3.0 g acceleration, a irregular
generatedvibrations:
voltage by(a) the
Reference vibration
harvester, is 13.5of
(b) VDD Hz the neural
excitation
with with 3.0 g
3.0 g acceleration,acceleration, a generated voltage by the harvester, (b) VDD of the
a generated voltage by the harvester, (b) VDD of the neural recording circuits.neural
recording circuits.
recording circuits.
In
In comparison
comparison withwith the the generated
generated voltage
voltage at at a
a regular
regular vibration
vibration asas shown
shown in in Figure
Figure 7a,7a, the
the
In comparison
generated voltages with
at an the generated
irregular voltage
vibration as shownat ainregular
Figure vibration
8a had as shown
many different inmagnitudes
Figure 7a, andthe
generated voltages at an irregular vibration as shown in Figure 8a had many different magnitudes
generated
intervals. voltages
When an at an irregular
interval betweenvibration aswas
collisions shown in Figure
increased, the 8a had many
reduction different
of the magnitudes
and intervals. When an interval between collisions was increased, the reduction ofsupply voltage
the supply was
voltage
and intervals.
increased. WhenWhenthean interval
interval wasbetween collisions
decreased, was increased,
the supply voltage theoverall.
rose reduction of the supply
Therefore, the VDDvoltage
at an
was increased. When the interval was decreased, the supply voltage rose overall. Therefore, the VDD
was increased.
irregular When
vibration hadthewider
interval was decreased,
variation than at a the supply
regular voltage rose overall. Therefore, the VDD
vibration.
at an irregular vibration had wider variation than at a regular vibration.
at an irregular vibration had wider variation than at a regular vibration.
2.4.2. Simulation Method
2.4.2. Simulation Method
2.4.2. WeSimulation Method
used 15,300-s segments of seizure data and 1510-min segments of non-seizure data from 9
We used 15,300-s segments of seizure data and 1510-min segments of non-seizure data from 9
epilepsy patients. Each
We used 15,300-s segmentsseizure segment haddata
of seizure one and
seizure event, segments
1510-min the duration of which ranged
of non-seizure data from
from30 9
epilepsy patients. Each seizure segment had one seizure event, the duration of which ranged from 30
to 120 s. patients.
epilepsy Figure 9 shows one recorded
Each seizure segmentseizure
had onedata and itsevent,
seizure spectrogram. The seizures
the duration of whichstart at about
ranged from120 30
to 120 s. Figure 9 shows one recorded seizure data and its spectrogram. The seizures start at about
s and
to 120 have a high-amplitude
s. Figure (>300 uV)seizure
9 shows one recorded and high-frequency oscillation (>20
data and its spectrogram. TheHz).
seizures start at about
120 s and have a high-amplitude (>300 uV) and high-frequency oscillation (>20 Hz).
120 s and have a high-amplitude (>300 uV) and high-frequency oscillation (>20 Hz).

Figure 9.
Figure 9. One
One recorded
recorded ictal
ictal EEG
EEG data and its
data and its spectrogram.
spectrogram.
Figure 9. One recorded ictal EEG data and its spectrogram.
modified recorded
We modified recorded data
data on
on the
the basis
basis of
of the
the simulation
simulation results.
results. Figure 10 shows an example
We modified recorded data on the basis of the simulation results. Figure 10 shows an example
explaining the effects of supply
supply voltage
voltage variations
variations by using
using aa sine
sine wave
wave (Figure
(Figure 10a).
10a). A signal in
explaining the effects of supply voltage variations by using a sine wave (Figure 10a). A signal in
irregular supply
Figure 10b reflected distortions due to irregular supply voltage
voltage variations.
variations. An envelope of the signal
Figure 10b reflected distortions due to irregular supply voltage variations. An envelope of the signal
the form
took the form ofofan
anirregular
irregularramp.
ramp.Figure
Figure10c
10cshows
showsa amodified
modifiedsignal
signal
byby adding
adding noise
noise to to
thethe data
data in
took the form of an irregular ramp. Figure 10c shows a modified signal by adding noise to the data
in Figure 10b. Figure 11 shows another example by using a 300-s seizure segment (Figure
Figure 10b. Figure 11 shows another example by using a 300-s seizure segment (Figure 11a). Figure 11b 11a). Figure
in Figure 10b. Figure 11 shows another example by using a 300-s seizure segment (Figure 11a). Figure
11b shows
shows a modified
a modified signal,signal,
whichwhich distortions
distortions and noiseandwere
noiseapplied
were applied to.11c
to. Figure Figure
shows 11c
theshows the
enlarged
11b shows a modified signal, which distortions and noise were applied to. Figure 11c shows the
enlarged
signal signalcomparison.
for easier for easier Thecomparison. The modified
modified signal signal
had irregular had Significant
spikes. irregular differences
spikes. Significant
between
enlarged signal for easier comparison. The modified signal had irregular spikes. Significant
differences
the raw signalbetween
and the themodified
raw signal andoccurred
signal the modified
mostlysignal
when occurred mostlyofwhen
an amplitude a rawansignal
amplitude of a
was large.
differences between the raw signal and the modified signal occurred mostly when an amplitude of a
raw signal was large.
raw signal was large.
Micromachines 2020, 11, 45 8 of 12
Micromachines
Micromachines2019,
2019,10,
10,xx 88 of
of 12
12

(a)
(a) (b)
(b) (c)
(c)
Figure
Figure10.
Figure 10.Example
10. Exampleexplaining
Example explainingthe
explaining theeffects
the effectsof
effects ofsupply
of supplyvoltage
supply voltagevariations
voltage variationsby
variations byusing
by usingaaasine
using sinewave.
sine wave.(a)
wave. (a)A
(a) AAraw
raw
raw
sine
sine signal,
signal, (b)
(b) a
a signal
signal in
in which
which distortions
distortions caused
caused by
by supply
supply voltage
voltage variations
variations were
were reflected,
reflected,
sine signal, (b) a signal in which distortions caused by supply voltage variations were reflected, and (c) a and
and
(c)
(c) aamodified
modified
modified signal
signalsignal by
by adding
by adding adding noise
noisenoise to
to the the
thesignal
tosignal b. b.
signal b.

Figure
Figure11.
Figure 11.Example
11. Exampleexplaining
Example explaining
explaining the
the effects
theeffects of supply
ofof
effects supply voltage
voltage
supply variations
variations
voltage by
byusing
variations using
by aaseizure
using seizure segment.
segment.
a seizure (a)
(a)
segment.
A
A raw
raw seizure
seizure segment,
segment, (b)
(b) data
data in
in which
which distortions
distortions caused
caused by
by supply
supply voltage
voltage variations
variations
(a) A raw seizure segment, (b) data in which distortions caused by supply voltage variations and noise and
and noise
noise
were
werereflected,
were reflected,and
reflected, and(c)
and (c)zoomed-in
(c) zoomed-in data
zoomed-in dataof
data of(a,b).
of (a,b).
(a,b).

2.4.3.Performance
2.4.3.
2.4.3. PerformanceMeasures
Performance Measures
Measures
To analyze
To
To analyzedetection
analyze detectionperformance,
detection performance,
performance, wewe
we considered
considered
considered three
three
three terms:
terms:
terms: Detection
Detection
Detection time,time,
time, specificity,
specificity,
specificity, and
and
and sensitivity.
sensitivity.
sensitivity.Specificity Specificity
Specificityand and
andsensitivitysensitivity
sensitivitywere were
weredefined
definedas defined
asdescribedas
describedby described
byEquations by
Equations(5) Equations
(5)and
and(6), (5) and
(6),respectively
respectively(6),
respectively [23].
[23].
[23].
TN
specificity = 𝑇𝑁
𝑇𝑁 (5)
specificity =
specificity = TN + FP (5)
(5)
𝑇𝑁
𝑇𝑁 ++𝐹𝑃𝐹𝑃
TP
sensitivity = 𝑇𝑃
𝑇𝑃 (6)
sensitivity
sensitivity == TP + FN (6)
(6)
𝑇𝑃
𝑇𝑃 ++𝐹𝑁𝐹𝑁
where TN and FP represented the true negatives and false positives, respectively, and TP and FN
where
where TN
TN and
and FP represented the true negatives and
and false
false positives, respectively, and TP
TP and
and FN
represented theFPtruerepresented
positives andthefalse
truenegatives,
negatives respectively.positives, respectively,
We calculated FP by the andpercentage FNof
represented
represented the
the true
true positives
positives and
and false
false negatives,
negatives, respectively.
respectively. We
We calculated
calculated FP
FP by
by the
the percentage
percentage
time spent under false positive alarms (%FP_time) [23] instead of the absolute number of false positives
of
of time spent under false positive
positive alarms (%FP_time) [23] instead of
of the absolute number of
of false
of time spent
the false underrate.
positive falseSimilarly, alarms (%FP_time)
we calculated TP by[23]
theinstead
percentage the absolute
of time spent number
under the false
true
positives
positives of
of the
the false
false positive
positive rate.
rate. Similarly,
Similarly, we
we calculated
calculated TP
TP by
by the
the percentage
percentage of
of time
time spent
spent under
under
positive alarms (%TP_time) (Appendix A Figure A1). As a result, we measured the performance in
the
the true
true positive alarms (%TP_time) (Appendix Figure
FigureA1).
A1).AsAsaaresult,
result,we
wemeasured
measuredthe theperformance
terms ofpositive
detection alarms
time,(%TP_time)
%FP_time, and (Appendix
%TP_time. performance
in
in terms
terms ofof detection
detection time,time, %FP_time,
%FP_time, and and %TP_time.
%TP_time.
Micromachines 2020, 11, 45 9 of 12

Micromachines 2019, 10, x 9 of 12

Micromachines 2019, 10, x 9 of 12
3. Results
3. Results
3. Results
Figure 12a shows average detection time values over ±0.2 V regular variations of a 2 V supply
Figure 12a shows average detection time values over ±0.2 V regular variations of a 2 V supply
Figure
voltage 12a shows
as shown average
in Figure 5a.detection
When the time valuesof
variance over
the ±0.2 V regular
supply voltagevariations
was zero,ofaaseizure
2 V supply
was
voltage as shown in Figure 5a. When the variance of the supply voltage was zero, a seizure was
voltage as shown in Figure 5a. When the variance of the supply voltage was zero, a seizure
detected at 130.895 s. The seizure detection times were varied from 130.855 s to 130.920 s according was
detected at 130.895 s. The seizure detection times were varied from 130.855 s to 130.920 s according
detected at 130.895
to the voltage s. TheAs
variances. seizure detection
a result, times time
the average weredifference
varied from 130.855
between s toand
with 130.920 s according
without regular
to the voltage variances. As a result, the average time difference between with and without regular
to the voltage
voltage variances.
variances As a0.05
was about result,
s. the average time difference between with and without regular
voltage variances was about 0.05 s.
voltage variances was about 0.05 s.

(a) (b)
(a) (b)
Figure 12.
12. (a)
(a) Average
Average detection time
time and (b)
(b) average %TP_time
%TP_time and %FP_time
%FP_time according to
to regular
Figure
Figure 12. (a) Average detection
detection time and
and (b) average
average %TP_time and
and %FP_time according
according to regular
regular
variations
variations of the supply voltage.
variations of
of the
the supply
supply voltage.
voltage.

Figure 12b
Figure 12bshows
showsaverage
average %TP_time
%TP_time andand average
average %FP_time,
%FP_time, respectively,
respectively, when the when the voltage
supply supply
Figure 12b shows average %TP_time and average %FP_time, respectively, when the supply
voltageregularly
varied varied regularly
in the rangein of
the±10%.
rangeWhen
of ±10%. Whenvoltage
the supply the supply voltage variance,
was without was without variance,
%TP_time was
voltage varied regularly in the range of ±10%. When the supply voltage was without variance,
%TP_time was 18.401% and %FP_time was 0.1258%. These results indicated the
18.401% and %FP_time was 0.1258%. These results indicated the true positive alarms were raised fortrue positive alarms
%TP_time was 18.401% and %FP_time was 0.1258%. These results indicated the true positive alarms
were
an raisedoffor
average 55.2an average
s over of seizure
a 300-s 55.2 s over a 300-s
segment seizure
and the falsesegment and the
positive alarms false positive
occurred alarms
for an average
were raised for an average of 55.2 s over a 300-s seizure segment and the false positive alarms
occurred
of for an
0.75 s over average
a 10-min of 0.75 s over
inter-seizure a 10-min
data. %TP_time inter-seizure
and %FP_time data.varied
%TP_time
withinand %FP_time
about varied
0.02% (0.06 s)
occurred for an average of 0.75 s over a 10-min inter-seizure data. %TP_time and %FP_time varied
within about 0.02% (0.06 s) and about 0.004% (0.024 s), respectively, according to the
and about 0.004% (0.024 s), respectively, according to the regular supply voltage variances from 0% to regular supply
within about 0.02% (0.06 s) and about 0.004% (0.024 s), respectively, according to the regular supply
voltage variances from 0% to ±10%.
±10%.
voltage variances from 0% to ±10%.
Figure 13a
Figure 13a shows
showsaverage
averagedetection
detectiontime
timevalues
valueswhen
when thethe harvester
harvester shaken
shaken irregularly
irregularly so that
so that an
Figure 13a shows average detection time values when the harvester shaken irregularly so that
an average voltage was 2 V and maximum voltage variances were in the range of
average voltage was 2 V and maximum voltage variances were in the range of 0 V to 0.2 V at 0.02 V 0 V to 0.2 V at 0.02
an average voltage was 2 V and maximum voltage variances were in the range of 0 V to 0.2 V at 0.02
V intervals.
intervals. Seizures
Seizures weredetected
were detectedatatananaverage
averageofof130.860
130.860s swithout
withoutsupply
supplyvoltage
voltagevariation.
variation. The
V intervals. Seizures were detected at an average of 130.860 s without supply voltage variation. The
seizure detection times varied between 130.857 s and 130.921 s in the maximum
seizure detection times varied between 130.857 s and 130.921 s in the maximum voltage variation voltage variation
range
seizure detection times varied between 130.857 s and 130.921 s in the maximum voltage variation
range
of ±0.2 of
V. ±0.2 V. Consequentially,
Consequentially, the the average
average detection detection
time was time was changed
changed within within
0.07 s when0.07the
s when the
harvester
range of ±0.2 V. Consequentially, the average detection time was changed within 0.07 s when the
harvester
shaken shaken irregularly.
irregularly.
harvester shaken irregularly.

(a) (b)
(a) (b)
Figure 13. (a) Average detection time and (b) average %TP_time and %FP_time according to irregular
Figure 13.
13. (a) Average
(a)the
Average detection time and (b) average %TP_time and %FP_time according to irregular
Figure
variations of supplydetection
voltage. time and (b) average %TP_time and %FP_time according to irregular
variations of
variations of the
the supply
supply voltage.
voltage.

Figure 13b shows average %TP_time and average %FP_time, respectively, when the supply
Figure 13b
13bshows
showsaverage
average %TP_time
%TP_time andand average
average %FP_time,
%FP_time, respectively,
respectively, when
when the the voltage
supply supply
voltage was varied irregularly in the range of ±10%. When the supply voltage was without irregular
voltage
was wasirregularly
varied varied irregularly in the
in the range of range
±10%.of ±10%.
When When
the thevoltage
supply supplywas
voltage was irregular
without without irregular
variance,
variance, %TP_time was 18.404% and %FP_time was 0.132%. %TP_time and %FP_time varied within
variance, %TP_time
%TP_time was 18.404%was 18.404%
and andwas
%FP_time %FP_time
0.132%.was 0.132%.and
%TP_time %TP_time
%FP_timeand %FP_time
varied withinvaried within
about 0.029%
about 0.029% (0.087 s) and about 0.014% (0.084 s), respectively, according to the irregular supply
about 0.029% (0.087 s) and about 0.014% (0.084 s), respectively, according to the irregular supply
voltage variances from 0% to ±10%.
voltage variances from 0% to ±10%.
Micromachines 2020, 11, 45 10 of 12

(0.087 s) and about 0.014% (0.084 s), respectively, according to the irregular supply voltage variances
from 0% to ±10%.

4. Discussion
We found that the average time differences between with and without voltage variances were
about 0.05 s under regular vibrations and about 0.07 s under irregular vibrations, respectively. They
were considered to be tolerable when considering that Osorio et al. [18] showed the mean delay
from the electrographic onset to automated detection was varied from −0.34 s to 3.9 s according to
detection parameters under an ideal voltage source. In addition, %TP_time and %FP_time were nearly
impervious to ±10% regular and irregular variations of the supply voltage.
The variance of the supply voltage could cause non-linear distortions and spikes with the same
frequency as the variance of the supply voltage. Most of the seizure detection algorithms, including
the GOFA, could distinguish between seizures and spikes. The GOFA filtered out spikes primarily
using a median filter and then decided on a seizure when a state transition from non-seizure to seizure
occurred for at least a threshold, Don [18]. Furthermore, they have usually concentrated on spectral
analysis and entropy. Therefore, non-linear distortions and spikes caused by the variance of the supply
voltage within ±10% affected performance a little.
The proposed frequency-up-converting impacted-based piezoelectric energy harvester may still
be heavy to be implantable. But as research continues, it is expected to shrink to a reasonable size in
the near future. In addition, we were aware that we did not test the harvester and circuits together
in vivo. However, we used each model that reflected its behavior characteristics under implanted
conditions. Therefore, our models and study results could help the design of energy harvesters, power
transfer circuits, implantable hardware devices, and algorithms for implantable devices.

5. Conclusions
We estimated the performance of a seizure detection algorithm with the neural signal recording
circuits powered by the harvester. We tested and analyzed the piezoelectric energy harvester and its
load circuits: The power conversion circuits and the neural recording circuit. To operate the neural
signal recording circuits under the lower supply voltage variation of ±10%, the frequency-up-converting
impact-based piezoelectric energy harvester has to shake regularly at a rate of at least 13.5 Hz with
3.5 g acceleration, or irregularly at a frequency range of 9.7 to 22.5 Hz with a maximum acceleration
of 4 g. By using equivalent circuits for the harvester and circuits, we simulated them under the
operating conditions. According to simulation results, seizure and non-seizure data were modified as
they were processed under supply voltage variance ranging from 0% to ±10%. The seizure detection
algorithm GOFA was used to estimate performance variation about these data. The onset detection
time, %TP_time, and %FP_time were changed to less than 0.1s, 0.03% and 0.02%, respectively. The
results showed that a supply voltage variance within ±10% could be acceptable to a seizure detection
algorithm. Therefore, it shows that energy harvesters have the potential to become a reliable power
source for implantable devices and reduce replacement of implantable devices due to dead batteries.

Author Contributions: Conceptualization, S.K.; methodology, S.K., S.J., and C.-H.J.; validation, S.K.; formal
analysis, S.K.; investigation, S.J., and C.-H.J.; resources, S.J., and C.-H.J.; writing, S.K; project administration, S.K.
All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Acknowledgments: Authors thank Hyang Woon Lee for providing icEEG recording data.
Conflicts of Interest: The authors declare no conflict of interest.

Appendix A
Figure A1 shows the concept of %TP_time. One seizure segment was tested by two different
detection algorithms. Detection results were displayed. If TP was to be measured according to
Micromachines 2020, 11, 45 11 of 12

the absolute number of true positive alarms, result 2 would appear to have the better performance.
However, if TP was to be measured in terms of %TP_time, result 1 and result 2 had 50% and 31.25%
%TP_time, respectively. Result 1 was therefore considered to have the better performance than
result 2 [24].
Micromachines 2019, 10, x 11 of 12

Figure A1. Examples

Examples of
of different
different detection results during a seizure.

References
1. Young, C.P.;Liang,
Young, C.P.; Liang,S.F.;
S.F.;Chang,
Chang,D.W.;
D.W.; Liao,
Liao, Y.C.;
Y.C.; Shaw,
Shaw, F.Z.;
F.Z.; Hsieh,
Hsieh, C.H.C.H. A portable
A portable wireless
wireless online
online closed-
closed-loop
loop seizure controller in freely moving rats. IEEE Trans. Instrum. Meas. 2011;
seizure controller in freely moving rats. IEEE Trans. Instrum. Meas. 2011, 60, 513–521. [CrossRef] 60, 513–521.
2.
2. Carey, J. Brain
Carey, J. Brain Facts:Facts:
A Primer A on Primer
the Brain onand the Brain System.
Nervous and Nervous System.http://www.brainfacts.
Available online: Available online:
http://www.brainfacts.org/book
org/book (accessed on 10 December 2019). (accessed on 10 December 2019).
3.
3. Bhavaraju,
Bhavaraju, N.; N.;Frei,
Frei,M.;M.; Osorio,
Osorio, I. Analog
I. Analog Seizure
Seizure Detection
Detection and Performance
and Performance Evaluation.
Evaluation. IEEEBiomed.
IEEE Trans. Trans.
Biomed. Eng. 2006, 53, 238–245.
Eng. 2006, 53, 238–245. [CrossRef] [PubMed]
4. What is Epilepsy? Available online: http://www.epilepsy.com/learn/epilepsy-101/what-epilepsy (accessed
4. What is Epilepsy? Available online: http://www.epilepsy.com/learn/epilepsy-101/what-epilepsy (accessed
on 10 December 2019).
on 10 December 2019).
5. Epilepsy Health Center. Available online: http://www.webmd.com/epilepsy/ (accessed on 10 December
5. Epilepsy Health Center. Available online: http://www.webmd.com/epilepsy/ (accessed on 10 December
2019).
6. 2019). J.; Pedley, T.A.; Aicardi, J. Epilepsy: A Comprehensive Textbook, 2nd ed.; Lippincott Williams & Wilkins:
Engel,
6. Engel, J.; Pedley,
Philadelphia, PA,T.A.;
USA, 2008. J. Epilepsy: A Comprehensive Textbook, 2nd ed.; Lippincott Williams & Wilkins:
Aicardi,
7. Philadelphia, PA, USA, 2008.
Cadei, A.; Dionisi, A.; Sardini, E.; Serpelloni, M. Kinetic and thermal energy harvesters for implantable
7. Cadei, A.;
medical Dionisi,
devices andA.; Sardini, E.;
biomedical Serpelloni,sensors.
autonomous M. KineticMeas.and
Sci.thermal
Technol.energy harvesters
2014, 25, 012003. for implantable
8. medical
Yang, Y.;devices andLiu,
Wei, X.-J.; biomedical autonomous
J. Suitability sensors. Meas.
of a thermoelectric powerSci. Technol.for
generator 2014, 25, 012003.
implantable [CrossRef]
medical electronic
8. Yang, Y.; J.Wei,
devices. Phys.X.-J.; Liu, J.Phys.
D: Appl. Suitability of a5790–5800.
2007, 40, thermoelectric power generator for implantable medical electronic
9. Almouahed,
devices. J. Phys.S.; Gouriou,
D: Appl. Phys.M.; Hamitouche, C.; Stindel,
2007, 40, 5790–5800. E.; Roux, C. The Use of Piezoceramics As Electrical
[CrossRef]
9. Energy
Almouahed,Harvesters WithinM.;
S.; Gouriou, Instrumented
Hamitouche,Knee ImplantE.;
C.; Stindel, During
Roux, Walking.
C. The Use IEEE/ASME Trans. As
of Piezoceramics Mechatronics
Electrical
2011,
Energy 16,Harvesters
799–807. Within Instrumented Knee Implant During Walking. IEEE/ASME Trans. Mechatronics
10. Luciano, V.; Sardini,
2011, 16, 799–807. E.; Serpelloni, M.; Baronio, G. Analysis of an electromechanical generator implanted
[CrossRef]
10. in a human total knee prosthesis.
Luciano, V.; Sardini, E.; Serpelloni, M.;InBaronio,
Proceedings of theof2012
G. Analysis IEEE Sensors Applications
an electromechanical Symposium
generator implanted in a
Proceedings; Institute of Electrical and Electronics Engineers (IEEE), Brescia,
human total knee prosthesis. In Proceedings of the 2012 IEEE Sensors Applications Symposium Proceedings, Italy, 7–9 February2012; pp.
1–5.
Brescia, Italy, 7–9 February 2012.
11. Nasiri, A.; Zabalawi, S.A.; Jeutter, D.C. A Linear Permanent Magnet Generator for Powering Implanted
11. Nasiri, A.; Zabalawi, S.A.; Jeutter, D.C. A Linear Permanent Magnet Generator for Powering Implanted
Electronic Devices. IEEE Trans. Power Electron. 2011, 26, 192–199.
Electronic Devices. IEEE Trans. Power Electron. 2011, 26, 192–199. [CrossRef]
12. Tashiro, R.; Kabei, N.; Katayama, K.; Tsuboi, E.; Tsuchiya, K. Development of an electrostatic generator for
12. Tashiro, R.; Kabei, N.; Katayama, K.; Tsuboi, E.; Tsuchiya, K. Development of an electrostatic generator for a
a cardiac pacemaker that harnesses the ventricular wall motion. J. Artif. Organs 2002, 5, 239–245.
13. cardiac pacemaker
Erturk, that harnesses
A. Electromechanical the ventricular
Modeling wall motion.
of Piezoelectric Energy J. Artif. Organs Ph.D.
Harvesters. 2002, 5, 239–245.
Thesis, [CrossRef]
Virginia Tech,
13. Erturk, A. Electromechanical
Blacksburg, VA, USA, 2009. Modeling of Piezoelectric Energy Harvesters. Ph.D. Thesis, Virginia Tech,
14. Blacksburg,
Shen, VA, USA,Energy
D. Piezoelectric 2009. Harvesting Devices for Low Frequency Vibration Applications. Ph.D. Thesis,
14. Shen, D. University,
Auburn Piezoelectric Energy AL,
Auburn, Harvesting
USA, 2009. Devices for Low Frequency Vibration Applications. Ph.D. Thesis,
15. Auburn K.;
Ashraf, University, Auburn,OAL,
Khir, M.H.M.; USA, J.;
Dennis, 2009.
Baharudin, Z. A wideband, frequency up-converting bounded
15. Ashraf, K.;
vibration Khir, harvester
energy M.H.M.; for O Dennis, J.; Baharudin,
a low-frequency Z. A wideband,
environment. Smart Mater. frequency up-converting
Struct. 2013, 22, 25018.bounded
16. Ju, S.; Chae,
vibration energyS.H.; Choi, Y.;
harvester forJun, S.; Park, S.M.;
a low-frequency Lee, S.J.; Lee,
environment. SmartH.W.; Ji, Struct.
Mater. C.H. Impact Based
2013, 22, Piezoelectric
25018. [CrossRef]
16. Vibration
Ju, S.; Chae, Energy Harvester
S.H.; Choi, Y.; Jun,Using Spherical
S.; Park, S.M.; Lee,Metal
S.J.; Ball
Lee, and
H.W.;MFC. In Impact
Ji, C.H. Proceedings
Basedof the 7th Asia-Pacific
Piezoelectric Vibration
Conference
Energy Harvesteron Transducers
Using Sphericaland Micro/Nano
Metal Ball andTechnologies, Daegu, Korea,
MFC. In Proceedings of the297th
June–2 July 2014.
Asia-Pacific Conference on
17. Kim, S.H.; Ju,
Transducers andS.;Micro/Nano
Ji, C.H.; Lee, S.J. Equivalent
Technologies, Daegu, circuit
Korea, model of anJuly
29 June–2 impact-based
2014. piezoelectric energy
harvester. J. Phys. Conf. Ser. 2014, 557, 012094.
18. Osorio, I.; Frei, M.G.; Wilkinson, S.B. Real-time automated detection and quantitative analysis of seizures
and short-term prediction of clinical onset. Epilepsia 1998, 39, 615–627.
19. Ravindran, S.; Cole, R. Low complexity algorithms for heart rate and epileptic seizure detection. In
Proceedings of the 2009 2nd International Symposium on Applied Sciences in Biomedical and
Communication Technologies, Bratislava, Slovakia, 24–27 November 2009; pp. 1–5.
20. Sridhara, S.R.; DiRenzo, M.; Lingam, S.; Lee, S.-J.; Blazquez, R.; Maxey, J.; Ghanem, S.; Lee, Y.-H.; Abdallah,
Micromachines 2020, 11, 45 12 of 12

17. Kim, S.H.; Ju, S.; Ji, C.H.; Lee, S.J. Equivalent circuit model of an impact-based piezoelectric energy harvester.
J. Phys. Conf. Ser. 2014, 557, 012094. [CrossRef]
18. Osorio, I.; Frei, M.G.; Wilkinson, S.B. Real-time automated detection and quantitative analysis of seizures
and short-term prediction of clinical onset. Epilepsia 1998, 39, 615–627. [CrossRef] [PubMed]
19. Ravindran, S.; Cole, R. Low complexity algorithms for heart rate and epileptic seizure detection. In
Proceedings of the 2009 2nd International Symposium on Applied Sciences in Biomedical and Communication
Technologies, Bratislava, Slovakia, 24–27 November 2009; pp. 1–5.
20. Sridhara, S.R.; DiRenzo, M.; Lingam, S.; Lee, S.-J.; Blazquez, R.; Maxey, J.; Ghanem, S.; Lee, Y.-H.; Abdallah, R.;
Singh, P.; et al. Microwatt Embedded Processor Platform for Medical System-on-Chip Applications. IEEE J.
Solid-State Circuits 2011, 46, 721–730. [CrossRef]
21. Liu, X.; Hao, H.; Yang, L.; Li, L.; Zhang, J.; Yang, A.; Ma, Y. Epileptic seizure detection with the local field
potential of anterior thalamic of rats aiming at real time application. In Proceedings of the 2011 Annual
International Conference of the IEEE Engineering in Medicine and Biology Society, Boston, MA, USA, 30
August–3 September 2011; Volume 2011, pp. 6781–6784.
22. Raghunathan, S.; Gupta, S.K.; Markandeya, H.S.; Roy, K.; Irazoqui, P.P. A hardware-algorithm co-design
approach to optimize seizure detection algorithms for implantable applications. J. Neurosci. Methods 2010,
193, 106–117. [CrossRef] [PubMed]
23. Kim, S.; Ju, S.; Ji, C.-H.; Lee, S. Realistic Circuit Model of an Impact-Based Piezoelectric Energy Harvester. J.
Semicond. Technol. Sci. 2015, 15, 463–469. [CrossRef]
24. Kim, S.; Choi, Y.S.; Choi, K.; Lee, J.; Lee, B.-U.; Lee, H.W.; Lee, S. Performance Estimation of an Implantable
Epileptic Seizure Detector with a Low-power On-chip Oscillator. J. Biomed. Eng. Res. 2015, 36, 169–176.
[CrossRef]

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).