RESEARCH ARTICLE
Time se ies segmen a ion o ecogni ion o
epilep i o m pa e ns eco ded ia
mic oelec ode a ays in i o
Gab iel Galeo e-ChecaID
1
, Gab iella PanuccioID
2,4‡
, Angel Canal-AlonsoID
3‡
,
Be nabe Lina es-Ba ancoID
1‡
*, Te esa Se ano-Go a edona
1‡
1Ins i u o de Mic oelec o
´nica de Se illa (IMSE-CNM), Consejo Supe io de In es igaciones Cien ı
´ icas
(CSIC) and Uni e sidad de Se illa, Se illa, Spain, 2Enhanced Regene a i e Medicine Lab, Ins i u o I aliano
di Tecnologia, Genoa, I aly, 3Ins i u o de In es igacio
´n Biome
´dica de Salamanca-IBSAL, Consejo Supe io
de In es igaciones Cien ı
´ icas and Uni e si y o Salamanca, Salamanca, Spain, 4Is i u o I aliano di
Tecnologia, Genoa, I aly
‡ GP, ACA, BLB and TSG con ibu ed equally o his wo k in supe ision and ideas.
*[email p o ec ed]
Abs ac
Epilepsy is a p e alen neu ological diso de ha a ec s app oxima ely 1% o he global
popula ion. App oxima ely 30-40% o pa ien s espond poo ly o an iepilep ic medica ions,
leading o a signi ican nega i e impac on hei quali y o li e. Closed-loop deep b ain s imu-
la ion (DBS) is a p omising ea men o indi iduals who do no espond o medical he apy.
To achie e e ec i e seizu e con ol, algo i hms play an impo an ole in iden i ying ele an
elec og aphic bioma ke s om local ield po en ials (LFPs) o de e mine he op imal s imu-
la ion iming. In his ega d, he de ec ion and classi ica ion o e en s om ongoing b ain
ac i i y, while achie ing low powe consump ion h ough compu a ionally inexpensi e imple-
men a ions, ep esen s a majo challenge in he ield. To add ess his challenge, we he e
p esen wo algo i hms, he Zdensi yRODE and he AMPDE, o iden i ying ele an e en s
om LFPs by u ilizing ime se ies segmen a ion (TSS), which in ol es ex ac ing di e en
le els o in o ma ion om he LFP and ele an e en s om i . The algo i hms we e alida ed
alida ed agains epilep i o m ac i i y induced by 4-aminopy idine in mouse hippocampus-
co ex (CTX) slices and eco ded ia mic oelec ode a ay, as a case s udy. The Zdensi yR-
ODE algo i hm showcased a p ecision and ecall o 93% o ic al e en de ec ion and 42%
p ecision o in e ic al e en de ec ion, while he AMPDE algo i hm a ained a p ecision o
96% and ecall o 90% o ic al e en de ec ion and 54% p ecision o in e ic al e en de ec-
ion. While ini ially ained speci ically o de ec ing ic al ac i i y, hese algo i hms can be
ine- uned o imp o ed in e ic al de ec ion, aiming a seizu e p edic ion. Ou esul s sugges
ha hese algo i hms can e ec i ely cap u e epilep i o m ac i i y, suppo ing seizu e de ec-
ion and, possibly, seizu e p edic ion and con ol. This opens he oppo uni y o design new
algo i hms based on his app oach o closed-loop s imula ion de ices using mo e elabo a e
decisions and mo e accu a e clinical guidelines.
PLOS ONE
PLOS ONE | h ps://doi.o g/10.1371/jou nal.pone.0309550 Janua y 24, 2025 1 / 20
a1111111111
a1111111111
a1111111111
a1111111111
a1111111111
OPEN ACCESS
Ci a ion: Galeo e-Checa G, Panuccio G, Canal-
Alonso A, Lina es-Ba anco B, Se ano-
Go a edona T (2025) Time se ies segmen a ion
o ecogni ion o epilep i o m pa e ns eco ded
ia mic oelec ode a ays in i o. PLoS ONE 20(1):
e0309550. h ps://doi.o g/10.1371/jou nal.
pone.0309550
Edi o : Gennady S. Cymbalyuk, Geo gia S a e
Uni e si y, UNITED STATES OF AMERICA
Recei ed: Decembe 3, 2023
Accep ed: Augus 9, 2024
Published: Janua y 24, 2025
Copy igh : ©2025 Galeo e-Checa e al. This is an
open access a icle dis ibu ed unde he e ms o
he C ea i e Commons A ibu ion License, which
pe mi s un es ic ed use, dis ibu ion, and
ep oduc ion in any medium, p o ided he o iginal
au ho and sou ce a e c edi ed.
Da a A ailabili y S a emen : The algo i hm
implemen a ions and expe imen al esul s ha
suppo he indings o his s udy a e openly
a ailable in Zenodo eposi o y (h ps://doi.o g/10.
5281/zenodo.10154332).
Funding: This wo k was pa ially suppo ed by he
Eu opean Union (h ps:// esea ch-and-inno a ion.
ec.eu opa.eu/index_en) h ough g an s 824164
(HERMES, wi h P incipal In es iga o s TSG and
In oduc ion
Epilepsy is a p e alen neu ological diso de ha a ec s app oxima ely 1% o he global popu-
la ion [1]. The s anda d ea men o epilepsy elies on medical he apy; howe e , 30-40% o
he pa ien s espond poo ly, i a all, o an i-seizu e medica ions, alling in o D ug-Resis an
Epilepsy (DRE) pa ien s. Su gical emo al o he epilep ic ocus is he cu en gold s anda d
o hose DRE pa ien s. Howe e , he success o abla i e su ge y highly depends on he accu-
a e iden i ica ion o he seizu e ocus. B ain s imula ion has eme ged as a p omising and less
adical al e na i e o abla i e neu osu ge y. In ecen yea s, he e has been a apid g ow h o
b ain implan able de ices, d i en by he ecen con ibu ions in wi eless powe ansmission
echniques [2–4], lexible elec onics o implan able de ices [3,5], and de ice design minia-
u iza ion [6–8]. Among he di e se de ices unde p e-clinical and clinical esea ch, Neu o-
Pace (Moun ain View, CA), which is he i s FDA-app o ed esponsi e neu os imula ion
sys em o ocal epilepsy, o e s a median seizu e educ ion ange o 50-70% ac oss di e en
s udies [9]. Gi en he p omising esul s in seizu e supp ession by esponsi e b ain s imula ion
and he la es ad ances in b ain implan able de ices, he g ow h o hese de ices o ea ing
epilepsy is expec ed o accele a e in he nex yea s. Howe e , impo an aspec s such as long-
e m iabili y, biocompa ibili y, powe ha es ing, and he need o mo e e icien and op i-
mized algo i hms make clinical esul s s ill imp o able.
To ope a e e ec i ely, b ain s imula ion de ices should de ec ele an elec og aphic bio-
ma ke s o de e mine he op imal s imula ion iming and pa e n o supp ess, o ideally p e-
en , he seizu e. Typically, local ield po en ial (LFP) eco dings a e used o his pu pose.
Va ious bioma ke s o in e es , such as high- equency oscilla ions (HFOs) [10–12], can be
de ec ed using embedded algo i hms in he implan able de ice, allowing low la ency eal- ime
seizu e supp ession. In essence, he iden i ica ion o hose di e en bioma ke s in he LFP
would help o c ea e decision models o closed-loop b ain s imula ion. Howe e , he iden i i-
ca ion o elec og aphic bioma ke s emains challenging due o hei di e se cha ac e is ics in
e ms o mo phology, ime/ equency ea u es, and hei co ela ion wi h seizu e e en s [12].
To add ess his challenge, a ious algo i hms ha e been p oposed, such as p incipal compo-
nen analysis [13], linea suppo ec o machine [14], app oxima e en opy [15], K-nea es
neighbo and naï e Bayes classi ica ion [14], ea u e ex ac ion using suppo ec o machines
[16], o complex mul i- ea u e-based decision models [17]. These algo i hms a e complex and
compu a ionally in ensi e, and as such hei po en ial o eal- ime ope a ion is limi ed. Mo e
powe -e icien app oaches such as phase synch oniza ion in mul ichannel eco dings [18],
mul ile el wa ele ans o m [19], Izhike ich neu al ne wo ks [20], au o- anging h eshold
de ec ion [21], o he Wiene algo i hm [22], ha e p o en mo e e ec i e eal- ime seizu e
de ec ion in ha dwa e implemen a ions.; Howe e , he main d awbacks a e associa ed wi h
ene gy and memo y consump ion, as well as educed pe o mance caused by simple imple-
men a ions due o ha dwa e cons ain s.
Designing e icien algo i hms equi es educing h ee main aspec s: compu a ion, mem-
o y, and powe consump ion. By educing he compu a ional cos , we can also dec ease powe
consump ion. This in ol es: (i) educing algo i hm complexi y by op imizing he numbe and
ype o ope a ions pe o med by i , (ii) minimizing memo y access, (iii) inc easing pa allelism,
and (i ) op imizing ha dwa e a chi ec u e. Following hese p inciples, we p opose wo no el
algo i hms ha ocus on op imizing compu a ion and memo y. Based on he p inciples o
op imiza ion, we ha e de eloped wo inno a i e algo i hms ha aim o educe compu a ional
complexi y by minimizing he numbe o ope a ions equi ed and lowe ing memo y usage.
Addi ionally, ou algo i hms add ess he issue o complex and ecu si e compu a ions, which
can be ine icien and ime-consuming. By p io i izing e iciency and e ec i eness, we hope o
PLOS ONE
Time se ies segmen a ion o epilep i o m in- i o MEA eco dings
PLOS ONE | h ps://doi.o g/10.1371/jou nal.pone.0309550 Janua y 24, 2025 2 / 20
GP) and 101070908 (CROSSBRAIN, wi h P incipal
In es iga o BLB).
Compe ing in e es s: The au ho s ha e decla ed
ha no compe ing in e es s exis .
c ea e algo i hms ha can be easily implemen ed and u ilized in a a ie y o se ings. A ime
se ies segmen a ion app oach is used o de ec and classi y epilep i o m e en s wi hin LFPs.
This me hod in ol es di iding a con inuous signal in o disc e e segmen s ha sha e common
cha ac e is ics. The goal is o iden i y common pa e ns, s a es, o beha io s wi hin he da a,
which will p o ide a ying le els o in o ma ion. [23,24]. In he scope o his wo k, ime se ies
segmen a ion can be used o di ide LFPs in o dis inc ca ego ies such as spiking e en s, i ing
pa e ns, o di e en ne wo k s a es. This app oach can help manage a la ge amoun o da a
and acili a e he applica ion o a ious algo i hms o pos -p ocess hese segmen s, such as sei-
zu e p edic ion o con ol. The objec i e o his wo k is o classi y ic al and in e ic al e en s in
LFP da a based on hei empo al and mo phological cha ac e is ics. We also aim o de e mine
he cu en s a e o he b ain ne wo k in eal ime. The p incipal goal is o ensu e ha he algo-
i hms we p opose ha e low compu a ional complexi y, he eby making hem sui able o
ha dwa e implemen a ion. We ha e achie ed his goal by de eloping wo new algo i hms and
alida ing hem h ough mic oelec ode a ay (MEA) eco dings o epilep i o m ac i i y gen-
e a ed by mouse hippocampus-co ex (CTX) slices ea ed wi h he con ulsan d ug 4-amino-
py idine (4AP). This model is widely used o s udy limbic ic ogenesis in i o. [25], as i o e s
se e al pa allelisms wi h he mos common ype o DRE in humans, i.e., mesial empo al lobe
epilepsy (MTLE), including he p ima y o igin o ic al ac i i y in he en o hinal co ex [26–
29].
The wo algo i hms use di e en app oaches; he i s (Zdensi yRODE) u ilizes an adap i e
z-sco e me hod, which uses a sho - e m memo y s a egy o p o ide a smoo h de ec ion o
he b ain’s s a es. The second algo i hm (AMPDE) uses scalog am-based peak de ec ion and
densi y es ima ion in he signal o ex ac ing e en s o in e es de e mined by hei densi y o
exci a ion. Bo h algo i hms inco po a e a pos -p ocessing echnique wi h a cumula i e look-
o wa d ime in eg a ion based on hei densi y and can dis inguish among ic al discha ges,
in e ic al e en s, and baseline. The Zdensi yRODE algo i hm was designed o as e compu a-
ion and educed ope a ions, while he AMPDE was designed o be e adap abili y o long-
e m a ia ions bu a he cos o mo e complex compu a ion.
Ma e ials and me hods
Mic oelec ode a ay eco ding o epilep i o m ac i i y gene a ed by b ain
slices
Ho izon al hippocampus-co ex (CTX) slices (400 μm- hick) we e ob ained om 4-8 weeks
old male CD1 mice. Epilep i o m ac i i y was acu ely induced by ea men wi h 4-aminopy i-
dine (4AP, 250 μM). Ex acellula ield po en ials we e acqui ed using a 6 ×10 plana MEA
wi h TiN elec odes (diame e : 30 μm, in e -elec ode dis ance: 500 μm) and a MEA1060
ampli ie . Signals we e sampled a 2 kHz, low-pass il e ed a hal he sampling equency
be o e digi iza ion, and eco ded o he compu e ’s ha d d i e ia he McRack so wa e. All
eco dings we e pe o med a 32˚C. The ull me hodological de ails desc ibing b ain slice
p epa a ion and main enance can be ound in [30]. Animal p ocedu es we e conduc ed in
acco dance wi h he Na ional Legisla ion (D.Lgs. 26/2014) and he Eu opean Di ec i e 2010/
63/EU, and app o ed by he Ins i u ional E hics Commi ee o Is i u o I aliano di Tecnologia
and by he I alian Minis y o Heal h (p o ocol code 860/215-PR, app o al da e 24/08/2015).
Animals we e moni o ed daily and eu hanized unde deep iso lu ane anes hesia, e i ied as
he lack o e lexes upon ee , paws, and ail pinch, complian wi h FELASA s anda ds. The
equipmen o MEA eco ding and empe a u e con ol was ob ained om Mul ichannel Sys-
ems (MCS), Ge many. Fig 1 highligh s he elec ode mapping ela i e o he b ain slice posi-
ion on he MEA, showing he di e en egions om whe e he LFP signals a e eco ded.
PLOS ONE
Time se ies segmen a ion o epilep i o m in- i o MEA eco dings
PLOS ONE | h ps://doi.o g/10.1371/jou nal.pone.0309550 Janua y 24, 2025 3 / 20
Da ase p epa a ion
A alida ion da ase was gene a ed om 7 b ain slices. B ain slices coupled o 60-channel
MEAs p o ided 11 ±5 (mean ±SD) alid elec odes. A o al o 68 eco ds, each being 30 min-
u es long, we e ex ac ed and hen anno a ed. Epilep i o m e en s we e iden i ied h ough a
semi-au oma ed p ocess. Namely, an au oma ed wa ele -based algo i hm was i s used o
de ec he e en s; hen, he ma ked e en s we e isually inspec ed and con i med o manually
co ec ed as equi ed. This is de ailed in [31]. O e all, he used da ase comp ises a o al o 521
ic al e en s and 6318 in e ic al e en s. LFPs we e labeled as ic al, in e ic al, and baseline, coded
as 2, 1, and 0, espec i ely. To ensu e he homogenei y o he da ase , we es ablished he ollow-
ing inclusion c i e ia: i) uni o m sampling a e o 2 kHz o s anda dize dispa a e sample a es
and ensu e consis en eco ding p ecision; ii) eco ds om he elec odes wi hin he pa ahip-
pocampal co ex (See Fig 1). All da a and me ada a a e publicly a ailable ia he zenodo eposi-
o y (10.5281/zenodo.10154332).
Expe imen al design: Pa ame e sea ch and op imiza ion p ocedu e
All eco ds we e main ained o aining in di e en scena ios, simila o online ope a ion,
unde low and high SNR condi ions, o a oid a possible o e i ing o he algo i hm. Table 1
summa izes he da ase cha ac e is ics o his expe imen . Algo i hms we e ained wi h an
op imiza ion ma ix including a ia ions o each o he pa ame e s o he algo i hms. The
comple e se o combina ions is desc ibed in Table 2, whe e he pa ame e sea ch ame is
de ailed. The expe imen al p o ocol de ised o his s udy is depic ed in Fig 2. The da ase is
di ided in o 70-30 es - alida ion spli s. A e unning he algo i hms in he aining da ase ,
he bes 10 sco es we e a e aged. This pa icula numbe was chosen heu is ically o ob ain a
winning combina ion. Fo eco dings wi h low Signal- o-Noise Ra ios (SNR) (<20 dB),
pa ame e s we e u he ine- uned be o e he alida ion s age, o imp o e he accu acy.
Fig 1. Epilep i o m pa e ns eco ded om 4AP- ea ed hippocampus-CTX slices ia MEA. (A) Mouse b ain slice placed on a plana 6 x 10 MEA g id
(g id sepa a ion o 500 μm). Elec odes a e g ouped by hei posi ion in he b ain s uc u es comp ised wi hin he b ain slice p epa a ion. (B)
Rep esen a i e MEA eco dings om he elec odes ma ked as CTX in panel A. The epilep i o m pa e n comp ises bo h in e ic al and ic al e en s. The
inse s show ep esen a i e ins ances o each e en ype a an expanded ime scale.
h ps://doi.o g/10.1371/jou nal.pone.0309550.g001
PLOS ONE
Time se ies segmen a ion o epilep i o m in- i o MEA eco dings
PLOS ONE | h ps://doi.o g/10.1371/jou nal.pone.0309550 Janua y 24, 2025 4 / 20
Z-sco e Densi y-based Robus Ou lie De ec ion Es ima ion
(ZDensi yRODE)
Since he LFP signal in he used da ase p ima ily consis s o baseline ac i i y, hen we may
p esume ha any de ia ion om he signal’s end could indica e an in e ic al o ic al e en .
By analyzing he ou lie ’s empo al du a ion, p oximi y o o he ou lie s, and magni ude, we
could somewha es ima e he ype o e en i may ep esen . The Zdensi yRODE algo i hm u i-
lizes he s a is ical Z-sco e es o de e mine ou lie s in da a ends using he dis ibu ion’s his-
o ic mean and s anda d de ia ion. The Z- es is a eliable measu e o a iabili y ha can be
used o iden i y ou lie s. By using his o ical da a, he algo i hm can adjus o changes in ampli-
ude caused by ac o s like elec ode deg ada ion. Once he ou lie s a e iden i ied h ough a z-
sco e es , hey a e s o ed o la e spike densi y analysis. The densi y o hese e en s is e alu-
a ed using a lookup window ha mee s speci ic sea ch c i e ia and hen classi ied based on
hei empo al cha ac e is ics and densi y.
This algo i hm can be con igu ed using ou pa ame e s lag (λ), h eshold (τ), in luence (ι),
and del a (Δ). The lag pa ame e de e mines how many pas samples a e conside ed o com-
pu ing he his o ic a e age and s anda d de ia ion. A longe lag p o ides s abili y agains
apid changes bu inc eases memo y load, while a sho e lag enables quicke adap a ion o
new in o ma ion bu inc eased a iabili y. The h eshold pa ame e se s he z-sco e le el a
which a da a poin is deemed signi ican , allowing o mo e p ecise de ec ions based on signal
ea u es like he SNR. The in luence pa ame e ( anging om 0 o 1) con ols he weigh gi en
o incoming samples when ecalcula ing a e ages and s anda d de ia ions, wi h 0 dis ega ding
new da a and 1 p o iding high adap abili y. Las ly, he del a pa ame e de ines he e ac o y
ime be ween peaks, c ucial o pa e n classi ica ion du ing he pos -p ocessing analysis o
peak candida es. This pa ame e ega ds he du a ion o e en s and hei sepa a ion in ime.
The algo i hm wo k low is depic ed in Fig 3. The pseudocode o his algo i hm is also
Table 2. Algo i hm op imiza ion expe imen ma ix. Op imiza ion ma ix o he wo algo i hms. Combina ions o
each o he pa ame e s o he algo i hm and expe imen s.
Algo i hm Op imiza ion Pa ame e s
ZDensi yRODE Th eshold [σ]: 4, 5, 6
Del a con olu ional il e [seconds]: 3, 4, 5
Lag [seconds]: 0.125, 0.25, 0.5
AMPD Th eshold [σ]: 3, 4, 5, 6
Del a Time In eg a ion [σ]: 1.5, 2, 2.5
h ps://doi.o g/10.1371/jou nal.pone.0309550. 002
Table 1. B ain slice MEA elec ophysiology da ase desc ip ion.
Pa ame e Value
To al numbe o signals 68
Numbe o elec odes pe b ain slice 11 ±5
Numbe o ic al e en s pe b ain slice 5 ±1
Numbe o in e ic al e en s pe b ain slice 80 ±50
B ain egion Pa ahippocampal co ex
Sampling a e 2 kHz
To al eco ding ime (a e age) 1800 s
Signal- o-noise a io (SNR) om 20 ±10 dB
h ps://doi.o g/10.1371/jou nal.pone.0309550. 001
PLOS ONE
Time se ies segmen a ion o epilep i o m in- i o MEA eco dings
PLOS ONE | h ps://doi.o g/10.1371/jou nal.pone.0309550 Janua y 24, 2025 5 / 20

p o ided in Code 1. The algo i hm’s complexi y was assessed, esul ing in a O(n+m) com-
plexi y, whe e n is he inpu da a size and m is he numbe o ele an ou lie s ound in he sig-
nal. I he signal con ains no ou lie s o hose a e e y limi ed, he spike densi y classi ie is
much as e as he e a e ewe e en s o compa e.
Code 1.Pseudo code o Z-sco e Densi y-based Robus Ou lie De ec ion Es ima ion
algo i hm.
Algo i hm 1 Zdensi yRODE Peak Classi ica ion
1: Inpu : signal, lag (λ), in luence (ι), h eshold (τ), del a (Δ)
2: Ou pu : classi ied egions
3: De ine an in e media y signal o il e ed inpu alues: Y
4: o x
i
in signal do
5: i T aining s age hen upda e bu e s μand σbu e s
6: else
7: i (x
i
−μ
i−1
) > h eshold �σ
i−1
hen
8: x
i
is an ou lie
9: Y ≔(ι�x
i
) + ((1 − ι)*U
i−1
);
Fig 2. Expe imen design desc ip ion. A) Compendium o MEA eco ding da ase s. B) Anno a ed signals showing he h ee di e en pa e ns, highligh ed
in di e en colo s: baseline (g een), in e ic al (yellow) and ic al ( ed). C) Tes / ain spli o he da ase . D) Expe imen design block diag am wi h he
sequence o s ages chosen o he expe imen al s age.
h ps://doi.o g/10.1371/jou nal.pone.0309550.g002
PLOS ONE
Time se ies segmen a ion o epilep i o m in- i o MEA eco dings
PLOS ONE | h ps://doi.o g/10.1371/jou nal.pone.0309550 Janua y 24, 2025 6 / 20
10: else U≔x
i
11: end i
12: a g ≔a e age(Ge Chunk o Y o las λ samples);
13: s d ≔S anda d de ia ion(Ge Chunk o Y o las λ samples);
14: o peak s o ed in p ocessed window do
15: Accoun numbe o peaks in a Δ window
16: i N peaks in Δ ime-lapse �Th eshold_A hen
17: Classi y in e al as A label
18: i Th eshold_B�N peaks in Δ window �Th eshold_A hen
19: Classi y in e al as B label
The algo i hm s a s wi h a aining s age, in which du ing he i s λsamples, he algo i hm
does no p o ide any de ec ion bu ills he bu e s and calcula es he ini ial end o he da a.
Only a no ch il e a 50 Hz o emo e powe line noise (he e, 50 Hz) is used in he signal con-
di ioning s age. Du ing his s age, which las s as long as he lag ime pa ame e , ini ial mean
(μ), and s anda d de ia ion (σ) bu e s a e illed (See Fig 3). Once he algo i hm is ained, i
s a s he ope a ion mode by p ocessing each incoming sample in h ee s eps: (i) a z-sco e es
Fig 3. ZDensi yRODE algo i hm lowcha . Block diag am wi h he wo k low o he Z-sco e Densi y-based Robus Ou lie De ec ion Es ima ion
algo i hm.
h ps://doi.o g/10.1371/jou nal.pone.0309550.g003
PLOS ONE
Time se ies segmen a ion o epilep i o m in- i o MEA eco dings
PLOS ONE | h ps://doi.o g/10.1371/jou nal.pone.0309550 Janua y 24, 2025 7 / 20
ela i e o μand σ, which a e es ima ed alues om p e ious samples in he signal (Eq 1); (ii)
local maxima e alua ion by i s -de i a i e es (Eq 2); (iii) e ac o y pe iod e alua ion.
z¼xnm
sð1Þ
xn1<xn�xnþ1<xnð2Þ
Whe e x
n
is he cu en sample, μis he his o ic mean and σis he his o ic s anda d de ia ion.
Bo h his o ic μand σa e implemen ed as FIFO, s o ing he λlas samples and ecalcula ing on
each sample wi h he in luence pa ame e depending on he e alua ion o he z-sco e. he μ
and σ il e s a e upda ed ollowing he exp ession in Eq 3. When he z-sco e exceeds he
h eshold, i indica es a signi ican and p ominen peak, hus, he his o ical da a is upda ed
based on he in luence de ined by ι. I he z-sco e es is nega i e, implying noise o i ele an
in o ma ion, he il e s μand σa e upda ed wi hou conside ing he in luence pa ame e , ha
is p o iding less impo ance o he cu en sample. This emo es he in luence o his alue on
he adap a ion o he h eshold.
xn i�xi ð1iÞ�xi1
n¼0;1;...;lð3Þ
When de ec ing ou lie s, hese a e ans o med in o in e als by obse ing he densi y o
spikes o e a con inuous pe iod o ime. The numbe o e en s is eco ded wi h a imes amp,
which will be used o de e mine he ype o e en . The ime in e als a e measu ed based on
he e en leng h, he up-spike and down-spike, and hei p oximi y o he nex peak, using a
simple inc emen al coun e . I he e a e one o mo e peaks ha occu wi hin he in eg a ion
ime in e al Δ, he de ec ion window is ex ended un il he peaks a e sepa a ed by a ime
g ea e han he e ac o y pe iod. When his condi ion is no me , he in eg a ion is s opped
and he in e al is classi ied. In e al classi ica ion is de e mined by du a ion, ampli ude, and
spike densi y. These cha ac e is ics p o ide in o ma ion abou e en equency and in ensi y.
Au oma ic mul iscale-based peak de ec ion (AMPDE)
The scalog am echnique has p o en o be e ec i e and eliable in iden i ying peaks, especially
in scena ios wi h noise and a ious peak shapes [32]. This algo i hm uses a mul iscale decom-
posi ion based on scalog ams and a ious windows, ollowed by peak iden i ica ion in each
ow o he scalog am ma ix. As demons a ed by Scholkmann e al. [33], his me hod shows
ema kable pe o mance in de ec ing ele an e en s. The algo i hm comp ises six s ages: Sig-
nal Condi ioning, calcula ion o Local Maxima Scalog am (LMS), de e mina ion o op imal
scale, LMS escaling, peak de ec ion, and es ima ion o peak densi y o b ain ac i i y pa e n
classi ica ion. A comp ehensi e block diag am o he algo i hm is p o ided in Fig 4, o e ing a
Fig 4. AMPD enhanced algo i hm lowcha . Block diag am wi h he wo k low o he Au oma ic mul iscale-based peak de ec ion (AMPDE) algo i hm. In
he i s s age, a bandpass il e be ween 0.5 and 50 Hz is se . Then an LMS calcula ion is pe o med o each window in he signal. A summa ion o each
ow in he LMS e u ns he ec o γ, which a e a escaling ob ains he inal ec o o peaks candida e, il e ed wi h an e alua ion me hod o iden i y he
peaks on he ec o .
h ps://doi.o g/10.1371/jou nal.pone.0309550.g004
PLOS ONE
Time se ies segmen a ion o epilep i o m in- i o MEA eco dings
PLOS ONE | h ps://doi.o g/10.1371/jou nal.pone.0309550 Janua y 24, 2025 8 / 20
isual ep esen a ion o each s ep wi hin he p ocessing pipeline and inpu /ou pu (I/O) delin-
ea ion. The algo i hm’s pseudocode is also a ailable in Code 2. I s complexi y was e alua ed,
esul ing in an O(n
2
) ime complexi y. This is because o he nes ed loops used in he ma ix
ope a ions ha signi ican ly inc ease he algo i hm’s complexi y.
Code 2.Pseudo code o AMPDE algo i hm.
Algo i hm 2 Au oma ic Mul iscale-based Peak De ec ion Enhanced (AMPDE)
1: Inpu : Signal (x), Scale ange (k), h eshold (α), del a (Δ)
2: Ou pu : De ec ed peaks p
3: Calcula e Linea De ended Signal x0by sub ac ing he leas -
squa es i o a s aigh line
4: o kin 2
1
, 2
2
,. . ., 2
L
whe e L=dN/2e− 1 do
5: o iin k+ 2, k+ 3, . . .,N−k+ 1 do
6: i +αwhe e is a numbe 2[0, 1] and α= 1 hen
7: mk;i¼0;i xi1>xik1^xi1>xiþk1
þa;o he wise
(
8: else
9: m
k,i
= 0
10: end i
11: end o
12: end o
13: Compu e Local Maxima Scalog am (LMS) ma ix Musing m
k,i
alues
14: Row-wise summa ion o ge gk¼PN
i¼1mk;i o each scale k
15: Find global minimum λ = a gmin
k
γ
k
16: Reshape ma ix M
by emo ing elemen s m
k,i
o k> λ
17: o iin 1, 2, . . .,Ndo
18: o jin 1, 2, . . ., λ do
19: Calcula e σ
i
using Eq (7) column-wise s anda d de ia ion
o mula
20: end o
21: end o
22: De ec peaks: p= {ijσ
i
= 0}
23: Peak so ing ask
24: o each peak, coun all consecu i e peaks wi hin a del a (Δ) ime
do
25: i nex peak exceeds he in eg a ion (Δ) window hen
26: i Numbe o peaks exceeds classi ica ion h eshold o ic al
hen
27: Ma k he e en as Ic al
28: else
29: i Wi hin in e ic al kind o e en s hen
30: Ma k he e en as In e ic al
31: end i
32: end i
33: end i
34: end o
=0
Le us now conside a uni a ia e signal, x(i), wi h N uni o mly sampled poin s, 1 �i�N,
con aining pe iodic o quasi-pe iodic peaks. The i s s ep in ol es p ep ocessing he inpu
signal o emo e he o se and elimina e he powe line in e e ence (he e, 50 Hz) by applying
a no ch il e wi h a quali y ac o o 90. Then he LMS is compu ed by analyzing all inpu al-
ues (x=x
1
,x
2
,x
3
,x
4
,. . .,x
n
) using a mo ing window app oach. The LMS uses he scalog am
o he de ec ion o peaks in he ime se ies. The scalog am is he signal’s equency dis ibu-
ion o e ime, analogous o he spec og am. This ep esen a ion enables he ex ac ion o sig-
ni ican ea u es such as peaks, edges, and pa e ns. In his wo k, we a e in e es ed in he
PLOS ONE
Time se ies segmen a ion o epilep i o m in- i o MEA eco dings
PLOS ONE | h ps://doi.o g/10.1371/jou nal.pone.0309550 Janua y 24, 2025 9 / 20
ep esen he co esponding esul s o each da ase , while B) and D) depic he SNR anges
o each eco ding session. The pe o mance o he algo i hms indeed co ela es wi h he SNR.
This a iabili y is akin o selec ing he MEA channels based on hei ana omical posi ion and
assessing signal quali y o SNR h eshold compliance. Despi e his a iabili y, he algo i hm
exhibi s an ou s anding o e all pe o mance. Besides, he co ec choice o pa ame e s highly
de e mines he algo i hm pe o mance as illus a ed in Fig 8E and 8F.
Discussion
We ha e p oposed a no el me hodology o de ec ing epilep i o m ac i i y, add essing he
p oblem h ough ime se ies segmen a ion u ilizing he Zdensi yRODE and AMPDE algo-
i hms. The e ec i eness o bo h algo i hms was e i ied by es ing hem on epilep i o m ac i -
i y eco ded om 4AP- ea ed oden b ain slices ia MEA. Bo h algo i hms demons a ed
ema kable pe o mance in de ec ing ic al ac i i y. I should be no ed ha hese algo i hms a e
capable o de ec ing in e ic al e en s as well, as hey pe o m a ime se ies segmen a ion ask.
The de ec ion and classi ica ion o ic al and in e ic al e en s is ypically based on h ee p i-
ma y app oaches: (i) mo phological ea u es, such as maximum ampli ude, spike slope, abso-
lu e mean o windowed e en s, and successi e spike densi y [7,14,42,43]; (ii) spec al
ea u es, including powe con en ac oss di e en equency bands [7,15,18,43–45]; and (iii)
s a is ical ea u es, encompassing a ia ions in median and a iance [7,14,22], ku osis, and
skewness [42].
To imp o e epilepsy diagnosis and ea men , se e al seizu e de ec ion algo i hms ha e
been p oposed. In [18], a magni ude and phase synch oniza ion be ween elec odes could
achie e a 66% ue posi i e a e (TPR). Yoo e al [7], used SVM yielding a TPR o 82.7% o sei-
zu e classi ica ion. Shoeb e al [46] p oposed a machine lea ning echnique h ough a ea u e
ec o o he de ec ion o seizu es, achie ing a 96% o sensi i i y and a 2/24h FPR. Ku osis,
skewness, and coe icien o a ia ion om decima e Disc e e Wa ele Decomposi ion we e
p oposed by [47], achie ing a p ecision o 92.66%. Ronchini e al [38], uses en opy- and-spec-
um ea u es o seizu e de ec ion algo i hm, boas ing an accu acy o 97.8%. In [44], seizu e
de ec ion is done by decima ing disc e e wa ele decomposi ion and in e qua ile ange and
mean absolu e de ia ion, ob aining a p ecision o 84.2%, a sensi i i y o 98.5% and la ency o
1.76 seconds.
Table 4 summa izes he pe o mance and algo i hmic app oaches o he wo ks men ioned
abo e in a compa a i e able. Many o hose implemen a ions we e ailo ed a ound he single
pa ien a he han being designed as pa ien -agnos ic app oaches. In his ega d, he e is a
g ea deba e on he ad an ages o one app oach o e he o he [48]. On he one hand, pe son-
alized algo i hms o e a b oad ange o adjus able pa ame e s and in ol e a aining p ocess;
on he o he hand, pa ien -agnos ic app oaches employ a ixed-pa ame e de ec o wi hou he
need o addi ional aining [48,49]. This decision equi es a ade-o be ween de ec ion accu-
acy, simplici y, and speed. Following any o hose app oaches de e mines he accu acy o he
algo i hm and i s applicabili y o di e en pa ien s and scena ios. In his s udy, we ha e p io i-
ized he pe spec i e de elopmen o a seizu e de ec ion me hod ha is no pa ien -speci ic
bu , a he same ime, implemen s minimum le els o cus omiza ion o imp o ed accu acy.
The iden i ica ion o ele an bioma ke s in LFP signals poses a signi ican challenge in he
epilepsy ield, because o he complex dynamics and di e en mo phologies o ic al and in e -
ic al e en s. Fu he , he SNR signi ican ly in luences he pe o mance o e en de ec ion algo-
i hms. In a ious applica ions, signal quali y assessmen echniques imp o e algo i hm
pe o mance by p e en ing unnecessa y p ocessing unde un a o able condi ions [50,51]. In
PLOS ONE
Time se ies segmen a ion o epilep i o m in- i o MEA eco dings
PLOS ONE | h ps://doi.o g/10.1371/jou nal.pone.0309550 Janua y 24, 2025 16 / 20

his wo k, a low SNR pa icula ly challenged accu a e de ec ion by Zdensi yRODE, whe eas
AMPDE p o ed obus o addi i e noise, excelling in analyzing equency bu s s.
Conclusion
We ha e in oduced a TSS app oach as a no el s a egy o seizu e de ec ion, owa d imp o ed
seizu e con ol ia closed-loop b ain s imula ion. The de ec ion o pa hological bioma ke s in
LFPs eco ded om epilep ic pa ien s is undamen al o he imely de ec ion (o be e , p e-
dic ion) o seizu es o imp o e b ain s imula ion s a egies o epilepsy ea men . Fu he -
mo e, such elec og aphic bioma ke s can p o ide addi ional in o ma ion o ad ance ou
unde s anding o b ain s a es ela ed o epilepsy. Zdensi yRODE and AMPDE ha e p o ed
e icien in bioma ke s de ec ion, while using minimal and s aigh o wa d compu a ion, as
opposed o complex algo i hmic app oaches, such as neu al ne wo ks. We belie e ha hese
algo i hms p o ide a no el pe spec i e o he ield and es ablish he ounda ion o a g owing
subse o algo i hms o he classi ica ion o epilep i o m e en s and he p edic ion o seizu es.
Au ho Con ibu ions
Concep ualiza ion: Gab iel Galeo e-Checa, Gab iella Panuccio, Angel Canal-Alonso, Be nabe
Lina es-Ba anco, Te esa Se ano-Go a edona.
Da a cu a ion: Gab iel Galeo e-Checa, Gab iella Panuccio.
Fo mal analysis: Gab iel Galeo e-Checa, Be nabe Lina es-Ba anco, Te esa Se ano-
Go a edona.
Table 4. Compa a i e analysis o MEA e en de ec ion algo i hms.
Pape Yea Algo i hm app oach P ecision [%] Sensi i i y [%]
[39] 2016 NLSVM - 95.7
[18] 2011 CORDIC - 66
[7] 2013 SVM - 82.7%
[15] 2018 App oxima e En opy + FFT 97.8 -
[14] 2015 KNN, SVM, Nai e Bayes, logis ic eg ession 80 95.24
[16] 2020 SVM - 100
[36] 2018 CNN - 79.2
[38] 2022 STDP 97.8 85.4
[46] 2010 Fea u e ec o ex ac ion - 96
[47] 2012 Ku osis, skewness and coe icien o a ia ion om he decima e DWT - 100
[4] 2013 Single window coun , Mul iple window coun , spec al en opy - 94
[44] 2014 Decima e DWT and in e qua ile ange and mean absolu e de ia ion 84.2 98.5
Zdensi yRODE 2023 Z-sco e es wi h memo y bu e ing s a egy (ic al) 93, (in e ic al) 42 93
AMDPEC 2023 Mul iscale peak de ec ion h ough DWT (ic al) 96, (in e ic al) 54 90
NLSVM: Non-Linea suppo ec o machines
CORDIC: Coo dina e o a ion digi al compu e
FFT: Fas Fou ie ans o m
SVM: Suppo ec o machines
STDP: Spike-Timing Dependen Plas ici y
DWT: Disc e e wa ele ans o m
CNN: Con olu ional neu al ne wo ks
h ps://doi.o g/10.1371/jou nal.pone.0309550. 004
PLOS ONE
Time se ies segmen a ion o epilep i o m in- i o MEA eco dings
PLOS ONE | h ps://doi.o g/10.1371/jou nal.pone.0309550 Janua y 24, 2025 17 / 20
Me hodology: Gab iel Galeo e-Checa.
Supe ision: Gab iella Panuccio, Angel Canal-Alonso, Be nabe Lina es-Ba anco, Te esa Se -
ano-Go a edona.
Valida ion: Angel Canal-Alonso.
Visualiza ion: Gab iel Galeo e-Checa.
W i ing – o iginal d a : Gab iel Galeo e-Checa.
W i ing – e iew & edi ing: Gab iel Galeo e-Checa, Gab iella Panuccio, Be nabe Lina es-Ba -
anco, Te esa Se ano-Go a edona.
Re e ences
1. Falco-Wal e J. Epilepsy–De ini ion, Classi ica ion, Pa hophysiology, and Epidemiology. Semina s in
neu ology. 2020; 40(6):617–623. h ps://doi.o g/10.1055/s-0040-1718719 PMID: 33155183
2. Galeo e-Checa G, Nabaei V, Das R, Heida i H. Wi elessly powe ed and modula lexible implan able
de ice. ICECS 2020—27 h IEEE In e na ional Con e ence on Elec onics, Ci cui s and Sys ems, P o-
ceedings. 2020.
3. Ray TR, Choi J, Bandodka AJ, K ishnan S, Gu u P, Tian L, e al. Bio-in eg a ed wea able sys ems: A
comp ehensi e e iew. Chemical Re iews. 2019; 119:5461–5533. h ps://doi.o g/10.1021/acs.
chem e .8b00573 PMID: 30689360
4. Ho JS, Kim S, Poon AS. Mid ield wi eless powe ing o implan able sys ems. P oceedings o he IEEE.
2013; 101(6):1369–1378. h ps://doi.o g/10.1109/JPROC.2013.2251851
5. Galeo e-Checa G, Panuccio G, Lina es-Ba anco B, Se ano-Go a edona T. Baseline Fea u es Ex ac-
ion om Mic oelec ode A ay Reco dings in an in i o model o Acu e Seizu es using Digi al Signal
P ocessing o Elec onic Implemen a ion. 2021 IEEE In e na ional Con e ence on Omni-Laye In elli-
gen Sys ems (COINS). 2021; p. 1–6.
6. Bu on A, Obaid SN, Va
´zquez-Gua dado A, Schmi MB, S ua T, Cai L, e al. Wi eless, ba e y- ee
subde mally implan able pho ome y sys ems o ch onic eco ding o neu al dynamics. P oceedings o
he Na ional Academy o Sciences o he Uni ed S a es o Ame ica. 2020; 117:2835–2845. h ps://doi.
o g/10.1073/pnas.1920073117 PMID: 31974306
7. Yoo J, Yan L, El-Damak D, Al a MAB, Shoeb AH, Chand akasan AP. An 8-Channel Scalable EEG
Acquisi ion SoC Wi h Pa ien -Speci ic Seizu e Classi ica ion and Reco ding P ocesso . IEEE Jou nal o
Solid-S a e Ci cui s. 2013; 48(1):214–228. h ps://doi.o g/10.1109/JSSC.2012.2221220
8. McGlynn E, Nabaei V, Ren E, Galeo e-Checa G, Das R, Cu ia G, e al. The Fu u e o Neu oscience:
Flexible and Wi eless Implan able Neu al Elec onics. Ad anced Science. 2021; 8. h ps://doi.o g/10.
1002/ad s.202002693 PMID: 34026431
9. Gelle EB, Ska paas TL, G oss RE, Goodman RR, Ba kley GL, Bazil CW, e al. B ain- esponsi e neu o-
s imula ion in pa ien s wi h medically in ac able mesial empo al lobe epilepsy. Epilepsia. 2017; 58
(6):994–1004. h ps://doi.o g/10.1111/epi.13740 PMID: 28398014
10. Zijlmans M, Ji uska P, Zelmann R, Leij en FS, Je e ys JG, Go man J. High-F equency Oscilla ions as a
New Bioma ke in Epilepsy. Annals o Neu ology. 2012; 71(2):169–178. h ps://doi.o g/10.1002/ana.
22548 PMID: 22367988
11. Fishe R, Salano a V, Wi T, Wo h R, Hen y T, G oss R, e al. Elec ical s imula ion o he an e io
nucleus o halamus o ea men o e ac o y epilepsy. Epilepsia. 2010; 51(5):899–908. h ps://doi.
o g/10.1111/j.1528-1167.2010.02536.x PMID: 20331461
12. Ch ojka J, Kudlacek J, Chang WC, No ak O, Tomaska F, O ahal J, e al. The ole o in e ic al dis-
cha ges in ic ogenesis—a dynamical pe spec i e. Epilepsy & Beha io . 2021; 121:106591. h ps://doi.
o g/10.1016/j.yebeh.2019.106591 PMID: 31806490
13. Topalo ic U, Ba clay S, Ling C, e al. A Wea able Pla o m o Closed-Loop S imula ion and Reco ding
o Single-Neu on and Local Field Po en ial Ac i i y in F eely Mo ing Humans. Na u e Neu oscience.
2023; 26:517–527. h ps://doi.o g/10.1038/s41593-023-01260-4 PMID: 36804647
14. Page A, Sagedy C, Smi h E, A a an N, Oa es T, Mohsenin T. A Flexible Mul ichannel EEG Fea u e
Ex ac o and Classi ie o Seizu e De ec ion. IEEE T ansac ions on Ci cui s and Sys ems II: Exp ess
B ie s. 2015; 62(2):109–113. h ps://doi.o g/10.1109/TCSII.2014.2385211
15. Cheng CH, Tsai PY, Yang TY, Cheng WH, Yen TY, Luo Z, e al. A Fully In eg a ed 16-Channel Closed-
Loop Neu al-P os he ic CMOS SoC Wi h Wi eless Powe and Bidi ec ional Da a Teleme y o Real-
PLOS ONE
Time se ies segmen a ion o epilep i o m in- i o MEA eco dings
PLOS ONE | h ps://doi.o g/10.1371/jou nal.pone.0309550 Janua y 24, 2025 18 / 20
Time E icien Human Epilep ic Seizu e Con ol. IEEE Jou nal o Solid-S a e Ci cui s. 2018; 53
(11):3314–3326. h ps://doi.o g/10.1109/JSSC.2018.2867293
16. Pa k YS, Cosg o e GR, Madsen JR, Eskanda EN, Hochbe g LR, Cash SS, e al. Ea ly De ec ion o
Human Epilep ic Seizu es Based on In aco ical Mic oelec ode A ay Signals. IEEE T ansac ions on
Biomedical Enginee ing. 2020; 67(3):817–831. h ps://doi.o g/10.1109/TBME.2019.2921448 PMID:
31180831
17. Aa abi A, He B. Seizu e p edic ion in pa ien s wi h ocal hippocampal epilepsy. Clinical Neu ophysiol-
ogy. 2017; 128(7):1299–1307. h ps://doi.o g/10.1016/j.clinph.2017.04.026 PMID: 28554147
18. Abdelhalim K, Smolyako V, Geno R. Phase-Synch oniza ion Ea ly Epilep ic Seizu e De ec o VLSI
A chi ec u e. IEEE T ansac ions on Biomedical Ci cui s and Sys ems. 2011; 5(5):430–438. h ps://doi.
o g/10.1109/TBCAS.2011.2170686 PMID: 23852175
19. Im eld K, Maccione A, Gandol o M, Ma inoia S, Fa ine PA, Koudelka-Hep M, e al. Real- ime signal p o-
cessing o high-densi y mic oelec ode a ay sys ems. Wiley Online Lib a y. 2009; 23:983–998. h ps://
doi.o g/10.1002/acs.1077
20. Ahmadi-Fa sani J, Lina es-Ba anco B, Se ano-Go a edona T. Digi al-signal-p ocesso ealiza ion o
izhike ich neu al ne wo k o eal- ime in e ac ion wi h elec ophysiology expe imen s. 2019 26 h IEEE
In e na ional Con e ence on Elec onics, Ci cui s and Sys ems, ICECS 2019. 2019; p. 899–902.
21. Kim C, e al. Sub-μV ms-Noise Sub-μW/Channel ADC-Di ec Neu al Reco ding Wi h 200-mV/ms
T ansien Reco e y Th ough P edic i e Digi al Au o anging. IEEE Jou nal o Solid-S a e Ci cui s. 2018;
53(11):3101–3110. h ps://doi.o g/10.1109/JSSC.2018.2870555
22. Rajde P, Wa d MP, Rickus J, Wo h R, I azoqui PP. Real- ime seizu e p edic ion om local ield po en-
ials using an adap i e Wiene algo i hm. Compu e s in Biology and Medicine. 2010; 40(1):97–108.
h ps://doi.o g/10.1016/j.compbiomed.2009.11.006 PMID: 20022319
23. Ca mona-Poya o A, Fe nandez-Ga cia NL, Mad id-Cue as FJ, Du an-Rosal AM. A new app oach o
op imal ime-se ies segmen a ion. Pa e n Recogni ion Le e s. 2020; 135:153–159. h ps://doi.o g/10.
1016/j.pa ec.2020.04.006
24. Chung FL, Fu TC, Ng V, Luk RWP. An e olu iona y app oach o pa e n-based ime se ies segmen a-
ion. IEEE T ansac ions on E olu iona y Compu a ion. 2004; 8(5):471–489. h ps://doi.o g/10.1109/
TEVC.2004.832863
25. A oli M, D’an uono M, Lou el J, Ko
¨hling R, Biagini G, Pumain R, e al. Ne wo k and pha macological
mechanisms leading o epilep i o m synch oniza ion in he limbic sys em in i o. P og ess in Neu obiol-
ogy. 2002; 68(3):167–207. h ps://doi.o g/10.1016/S0301-0082(02)00077-1 PMID: 12450487
26. McIn y e DC, Gilby KL. Mapping seizu e pa hways in he empo al lobe. Epilepsia. 2008; 49:23–30.
h ps://doi.o g/10.1111/j.1528-1167.2008.01507.x PMID: 18304253
27. Ba olomei F, Khalil M, Wendling F, Son heime A, Re
´gis J, Ranje a JP, e al. En o hinal co ex in ol e-
men in human mesial empo al lobe epilepsy: an elec ophysiologic and olume ic s udy. Epilepsia.
2005; 46(5):677–687. h ps://doi.o g/10.1111/j.1528-1167.2005.43804.x PMID: 15857433
28. Ba olomei F, Chau el P, Wendling F. Epilep ogenici y o b ain s uc u es in human empo al lobe epi-
lepsy: a quan i ied s udy om in ace eb al EEG. B ain. 2008; 131(7):1818–1830. h ps://doi.o g/10.
1093/b ain/awn111 PMID: 18556663
29. Spence SS, Spence DD. En o hinal-hippocampal in e ac ions in medial empo al lobe epilepsy.
Epilepsia. 1994; 35(4):721–727. h ps://doi.o g/10.1111/j.1528-1157.1994. b02502.x PMID:
8082614
30. Panuccio G, Colombi I, Chiappalone M. Reco ding and modula ion o epilep i o m ac i i y in oden
b ain slices coupled o mic oelec ode a ays. JoVE (Jou nal o Visualized Expe imen s). 2018;(135):
e57548. h ps://doi.o g/10.3791/57548 PMID: 29863681
31. Ca on D, Canal-Alonso A, Panuccio G. Mimicking CA3 Tempo al Dynamics Con ols Limbic Ic ogen-
esis. Biology. 2022; 11(3). h ps://doi.o g/10.3390/biology11030371 PMID: 35336745
32. Bishop SM, E cole A. Mul i-Scale Peak and T ough De ec ion Op imised o Pe iodic and Quasi-Pe i-
odic Neu oscience Da a. Ac a Neu ochi Suppl. 2018; 126:189–195. h ps://doi.o g/10.1007/978-3-319-
65798-1_39 PMID: 29492559
33. Scholkmann F, Boss J, Wol M. An E icien Algo i hm o Au oma ic Peak De ec ion in Noisy Pe iodic
and Quasi-Pe iodic Signals. Algo i hms. 2012; 5:588–603. h ps://doi.o g/10.3390/a5040588
34. Chung NC, Miasojedow B, S a ek M, Gambin A. Jacca d/Tanimo o simila i y es and es ima ion me h-
ods o biological p esence-absence da a. BMC Bioin o ma ics. 2019; 20(Suppl 15):644. h ps://doi.o g/
10.1186/s12859-019-3118-5 PMID: 31874610
35. G egg NM, Ma ks VS, Sladky V, Lunds om BN, Klassen B, Messina SA, e al. An e io nucleus o he
halamus seizu e de ec ion in ambula o y humans. Epilepsia. 2021; 62(10):e158–e164. h ps://doi.o g/
10.1111/epi.17047 PMID: 34418083
PLOS ONE
Time se ies segmen a ion o epilep i o m in- i o MEA eco dings
PLOS ONE | h ps://doi.o g/10.1371/jou nal.pone.0309550 Janua y 24, 2025 19 / 20
36. T uong ND, Nguyen AD, Kuhlmann L, Bonyadi MR, Yang J, Ippoli o S, e al. Con olu ional neu al ne -
wo ks o seizu e p edic ion using in ac anial and scalp elec oencephalog am. Neu al Ne wo ks.
2018; 105:104–111. h ps://doi.o g/10.1016/j.neune .2018.04.018 PMID: 29793128
37. Ronchini M, Zamani M, Huynh HA, Rezaeiyan Y, Panuccio G, Fa khani H, e al. A CMOS-based neu o-
mo phic de ice o seizu e de ec ion om LFP signals. Jou nal o Physics D: Applied Physics. 2021; 55
(1):014001. h ps://doi.o g/10.1088/1361-6463/ac28bb
38. Ronchini M, Rezaeiyan Y, Zamani M, Panuccio G, Mo adi F. NET-TEN: a silicon neu omo phic ne wo k
o low-la ency de ec ion o seizu es in local ield po en ials. Jou nal o Neu al Enginee ing. 2023; 20
(3):036002. h ps://doi.o g/10.1088/1741-2552/acd029
39. Bin Al a MA, Yoo J. A 1.83 μJ/Classi ica ion, 8-Channel, Pa ien -Speci ic Epilep ic Seizu e Classi ica-
ion SoC Using a Non-Linea Suppo Vec o Machine. IEEE T ansac ions on Biomedical Ci cui s and
Sys ems. 2016; 10(1):49–60. h ps://doi.o g/10.1109/TBCAS.2014.2386891 PMID: 25700471
40. de Cu is M, Je e ys JG, A oli M. In e ic al epilep i o m discha ges in pa ial epilepsy. Jaspe ’s Basic
Mechanisms o he Epilepsies [In e ne ] 4 h edi ion. 2012;.
41. S aba RJ, S ead M, Wo ell GA. Elec ophysiological Bioma ke s o Epilepsy. Neu o he apeu ics. 2014;
11(2):334–46. h ps://doi.o g/10.1007/s13311-014-0259-0 PMID: 24519238
42. Khan YU, Fa ooq O, Sha ma P. Au oma ic de ec ion o seizu e onse in pedia ic EEG. In e na ional
Jou nal o Embedded Sys ems and Applica ions. 2012; 2(3):81–89. h ps://doi.o g/10.5121/ijesa.2012.
2309
43. Shoa an M, Haghi BA, Tagha i M, Fa i a M, Emami-Neyes anak A. Ene gy-E icien Classi ica ion o
Resou ce-Cons ained Biomedical Applica ions. IEEE Jou nal on Eme ging and Selec ed Topics in Ci -
cui s and Sys ems. 2018; 8(4):693–707. h ps://doi.o g/10.1109/JETCAS.2018.2844733
44. Ahammad N, Fa hima T, Joseph P, e al. De ec ion o epilep ic seizu e e en and onse using EEG.
BioMed esea ch in e na ional. 2014; 2014. h ps://doi.o g/10.1155/2014/450573 PMID: 24616892
45. Huang SA, Chang KC, Liou HH, Yang CH. A 1.9-mW SVM p ocesso wi h on-chip ac i e lea ning o
epilep ic seizu e con ol. IEEE Jou nal o Solid-S a e Ci cui s. 2019; 55(2):452–464. h ps://doi.o g/10.
1109/JSSC.2019.2954775
46. Shoeb AH, Gu ag JV. Applica ion o machine lea ning o epilep ic seizu e de ec ion. In: P oceedings o
he 27 h in e na ional con e ence on machine lea ning (ICML-10); 2010. p. 975–982.
47. Das K, Daschaklada D, Roy PP, Cha e jee A, Saha SP. Epilep ic seizu e p edic ion by he de ec ion o
seizu e wa e o m om he p e-ic al phase o EEG signal. Biomedical Signal P ocessing and Con ol.
2020; 57:101720. h ps://doi.o g/10.1016/j.bspc.2019.101720
48. De Cooman T, Vandecas eele K, Va on C, Hunyadi B, Clee en E, Van Paesschen W, e al. Pe sonaliz-
ing hea a e-based seizu e de ec ion using supe ised SVM ans e lea ning. F on ie s in Neu ology.
2020; 11:145. h ps://doi.o g/10.3389/ neu .2020.00145 PMID: 32161573
49. Bi jand alab J, Ja male VN, Nou ani M, Ha ey J. Impac o Pe sonaliza ion on Epilep ic Seizu e P edic-
ion. In: 2019 IEEE EMBS In e na ional Con e ence on Biomedical & Heal h In o ma ics (BHI); 2019.
p. 1–4.
50. He J, Liu D, Chen X. Wea able exe cise elec oca diog aph signal quali y assessmen based on uzzy
comp ehensi e e alua ion algo i hm. Compu e Communica ions. 2020; 151:86–97. h ps://doi.o g/10.
1016/j.comcom.2019.12.051
51. Sa ija U, Ramkuma B, Manikandan MS. A Re iew o Signal P ocessing Techniques o Elec oca dio-
g am Signal Quali y Assessmen . IEEE Re iews in Biomedical Enginee ing. 2018; 11:36–52. h ps://
doi.o g/10.1109/RBME.2018.2810957 PMID: 29994590
PLOS ONE
Time se ies segmen a ion o epilep i o m in- i o MEA eco dings
PLOS ONE | h ps://doi.o g/10.1371/jou nal.pone.0309550 Janua y 24, 2025 20 / 20