EAI Endo sed T ansac ions
on Pe asi e Heal h and Technology Resea ch A icle
Towa ds PTSD Diagnosis Th ough ECG Anomaly
De ec ion based on Au oencode s
Vasileios Ska amagkas1,2,∗, Ioannis Kyp akis1,2, Geo gia S. Ka anasiou3,4, Dimi is I. Fo iadis3,4,
Manolis Tsiknakis1,2
1Biomedical In o ma ics and eHeal h Labo a o y, Dep . o Elec ical and Compu e Enginee ing, Hellenic
Medi e anean Uni e si y, He aklion, 71410 C e e, G eece
2Ins i u e o Compu e Science, Founda ion o Resea ch and Technology Hellas (FORTH), He aklion, 70013
C e e, G eece
3Uni o Medical Technology In elligen In o ma ion Sys ems, Uni e si y o Ioannina, Ioannina, 45110, G eece
4Biomedical Resea ch Ins i u e, Founda ion o Resea ch and Technology Hellas (FORTH), Ioannina, 45110, G eece
Abs ac
INTRODUCTION: Pos -T auma ic S ess Diso de (PTSD) is a debili a ing men al heal h condi ion ha
can de elop a e exposu e o auma ic e en s, o en esul ing in symp oms ha se e ely impai daily
unc ioning. Cu en diagnos ic me hods la gely ely on subjec i e assessmen s, highligh ing he need o
objec i e, non-in asi e ools o imp o e diagnos ic p ecision.
OBJECTIVES: This s udy aims o de elop and alida e an inno a i e deep lea ning app oach using
au oencode neu al ne wo ks o de ec PTSD h ough analysis o elec oca diog aphy (ECG) signals. The goal
is o p o ide a eliable and sophis ica ed diagnos ic me hod ha b idges compu a ional and clinical domains.
METHODS: We employed au oencode neu al ne wo ks o analyze ECG da a collec ed om wea able hea
zone senso s. This unsupe ised lea ning model was ained o de ec sub le anomalies in he ECG signals ha
may se e as bioma ke s o PTSD. The me hodology was e alua ed using da a collec ed om indi iduals wi h
and wi hou PTSD symp oms.
RESULTS: The p oposed model demons a ed s ong po en ial as an objec i e diagnos ic ool, success ully
iden i ying pa e ns in ECG signals associa ed wi h PTSD. The analysis con i med he model’s abili y o
dis inguish PTSD- ela ed anomalies wi h 83% accu acy.
CONCLUSION: This esea ch in oduces a no el, non-in asi e diagnos ic me hodology o PTSD using deep
lea ning and wea able ECG da a. The indings suppo he model’s alue as a po en ial objec i e bioma ke ,
con ibu ing o mo e p ecise psychia ic diagnos ics and expanding he ole o machine lea ning in heal hca e.
Recei ed on 25 May 2024; accep ed on 06 May 2025; published on 02 June 2025
Keywo ds: PTSD Diagnosis, au oencode , anomaly de ec ion, ECG, deep lea ning in heal hca e
Copy igh © 2025 V. Ska amagkas e al., licensed o EAI. This is an open access a icle dis ibu ed unde he e ms o he
CC BY-NC-SA 4.0, which pe mi s copying, edis ibu ing, emixing, ans o ma ion, and building upon he ma e ial in
any medium so long as he o iginal wo k is p ope ly ci ed.
doi:10.4108/ee ph .11.9463
1. In oduc ion
The ield o medical diagnos ics has seen a ans o -
ma i e change wi h he in oduc ion o sophis ica ed
da a analysis me hods, namely in he domain o men al
heal h condi ions like Pos -T auma ic S ess Diso de
(PTSD). PTSD is a men al diso de ha a ises om
a signi ican psychological auma o a h ea ening o
ca as ophic na u e and is cha ac e ised by ecu en
∗Co esponding au ho . Email: [email p o ec ed]
exposu e o componen s o a auma ic inciden , accom-
panied by eelings o wo y, panic, w a h, guil , and
a s ong u ge o a oid s imuli linked o he sou ce
o s ess [1]. I s p e alence is no able among indi-
iduals wi h ch onic diseases, including b eas cance ,
whe e ea men -induced ca diac oxici y can exace -
ba e PTSD symp oms, a ec ing pa ien s’ o e all quali y
o li e [2]. The cu en me hod o diagnosing PTSD
elies on sel - epo s, which may be suscep ible o inac-
cu acies, pa icula ly in indi iduals, including child en
and adul s, who display a oidance signs [3].
1
EAI Endo sed T ansac ions
on Pe asi e Heal h and Technology
| Volume 11 | 2025 |
V. Ska amagkasg
Gi en he in ica e na u e o he p oblem, i is c u-
cial o employ in en i e me hods o achie e p ecise
and unbiased de ec ion. To his goal, machine lea ning
me hods ha e been augh o accu a ely assess he se e -
i y o PTSD by analysing signals, such as elec oen-
cephalog aphy (EEG), and ex ac ing ele an in o -
ma ion [4]. Addi ionally, elec oca diog aphy (ECG),
a eadily accessible and non-in asi e echnology ha
can be ob ained e en om wea able de ices, p o ides
aluable insigh s in o he complex in e ac ion be ween
he hea and he neu ological sys em. In PTSD, hese
a ia ions in ECG hy hms ha e he po en ial o unc-
ion as bioma ke s o he condi ion, p o iding a new
and unbiased diagnos ic ool [5].
Wi hin he apidly g owing domain o deep lea ning
(DL) applica ions in medical diagnos ics [6,7], he e
is a no iceable de iciency: he absence o models
explici ly ailo ed o iden i y PTSD h ough he analysis
o ECG da a, especially in popula ions wi h co-
exis ing condi ions like b eas cance [8]. Al hough
he e has been no able p og ess in c ea ing ad anced
DL models o diagnosing di e en men al diso de s
[9], PTSD has no ecei ed he same amoun o
ocus, especially when i comes o using ECG inpu s.
The absence o ECG analysis in diagnosing PTSD
is signi ican , conside ing i s abili y o unco e he
physiological ounda ions linked o his condi ion [10].
This highligh s a p omising ye unexplo ed app oach
o an objec i e and non-in usi e diagnosis, which
could complemen he exis ing dependence on sel -
epo ed symp oms and clinical e alua ions.
This wo k aims o capi alize on ecen no el
compu a ional app oaches wi h clinical diagnoses o
PTSD, in o de o add ess he exis ing gap. Ou
objec i e is o u ilise au oencode s o analyse ECG
da a and c ea e a model ha can de ec he iny
i egula i ies which can be cha ac e is ic o PTSD,
pa icula ly o b eas cance pa ien s dealing wi h
complex heal h challenges. This echnique no only
p o ides he po en ial o imp o e he p ecision o PTSD
iden i ica ion, bu also makes a aluable con ibu ion
o he wide ield o diagnosing psychia ic diso de s,
whe e he e is a p essing demand o objec i e
bioma ke s.
We in oduce a comple e me hodology ha u ilises
au oencode -based anomaly iden i ica ion o he analy-
sis o ECG da a. Ou app oach is ounded on a ho ough
unde s anding o bo h he clinical aspec s o PTSD
and he echnological complexi ies o machine lea ning
models. Ou model, alida ed wi h an ex ensi e da ase
om wea able de ices, o e s insigh s in o enhancing
PTSD diagnosis in pa ien s wi h b eas cance [11].
2. Rela ed Wo k
2.1. PTSD Diagnosis Based on Machine Lea ning
Machine lea ning app oaches ha e been de eloped
alongside adi ional in e iew-based and e idence-
based diagnos ic me hods o diagnose symp oms o
PTSD [12]. Con olu ional Neu al Ne wo ks (CNN)
ha e been employed o diagnose PTSD by analysing
keywo ds ex ac ed om Twi e [13]. Thei app oach
achie ed 91.00% accu acy wi hin he g oup o cance
su i o s. Lekkas e al. [14] conduc ed an expe imen
on emale auma wi nesses, ocusing on he du a ion o
hei absence om home. By inpu ing he in o ma ion
om he global posi ioning sys em (GPS) in o an XGB
classi ie , hey ob ained AUC alue 0.82 and accu acy
a e 77.00%. Finally, a ecen s udy success ully
employed RF classi ie s o examine medical eco ds o
au oma ed diagnosis o PTSD, esul ing in imp essi e
accu acy a e 99.00% and AUC 0.89 [15].
Mo eo e , mul iple esea ch pape s ha e explo ed
he diagnosis o PTSD by employing machine lea ning
me hods on neu oimaging da a. In [16], esea che s
employed a deep lea ning app oach o achie e
diagnos ic accu acy 71.20%. They u ilised neu al
inge p in s om impo an b ain egions. Zhu e al.
de eloped a deep lea ning g aph- heo e ic app oach o
dis inguish be ween PTSD and auma exposed non-
PTSD g oups. They achie ed an accu acy o 80.00%
by analysing b ain ne wo k g aphs [17]. Fu he mo e,
Gong e al. employed a SVM classi ie using g ay
and whi e ma e me ics. They achie ed an accu acy
o 91.00% in di e en ia ing indi iduals wi h PTSD
om heal hy con ols. Addi ionally, hey achie ed an
accu acy o 67.00% in sepa a ing hose exposed o
auma bu wi hou PTSD. This demons a ed he
e ec i eness o magne ic esonance imaging (MRI) da a
in iden i ying PTSD [18].
2.2. Au oencode s Towa ds he De ec ion o Men al
Diso de s
Au oencode s a e employed in he domain o men al
diso de s o se e al objec i es. In a ecen s udy, a
no el au oencode model ha u ilises a ia ional mode
decomposi ion and mu ual in o ma ion echniques o
iden i y dep ession om speech da a in he DAIC-
WOZ da ase was p esen ed [19]. The p oposed model
achie ed accu acy 78.95% and 1-sco e 0.76. Mo eo e ,
scien is s in [20] u ilised an au oencode , o p ecisely
o ecas missing esponses in dep ession esea ch. The
indings demons a ed he supe io accu acy o his
app oach compa ed o al e na i e models, enabling
i s applica ion in accu a ely o ecas ing pa ien s’
dep ession s a us wi h minimal e o a e 2.85%.
Rega ding o he men al diso de s, Sewani e al. [21]
de eloped a combina ion o unsupe ised au oencode
and supe ised CNN in o de o deli e a p o icien
diagnosis o au ism spec um diso de , pa icula ly
2
EAI Endo sed T ansac ions
on Pe asi e Heal h and Technology
| Volume 11 | 2025 |
Towa ds PTSD Diagnosis Th ough ECG Anomaly De ec ion based on Au oencode s
Figu e 1. A chi ec u e o he buil au oencode model used o ECG anomaly de ec ion.
o child en. Thei model yielded high le el o pe -
o mance, ob aining accu acy 84.05% and AUC alue
0.78. A CNN-based au oencode was also de eloped in
[22] o o ecas indi iduals wi h suicidal endency by
analysing hei s uc u al b ain imaging. The indings
indica ed ha a speci ic a angemen o b ain s uc u es
in a ious a eas can e ec i ely di e en ia e hose wi h
suicidal hough s om hose wi hou such hough s and
heal hy con ols, wi h 85.00% accu acy. The e ec i e-
ness o a 3D con olu ional au oencode (3D-CAE) in
ex ac ing ea u es associa ed wi h schizoph enia, wi h-
ou elying on diagnos ic labels was examined in [23].
The ea u es ob ained using 3D-CAE p ese ed hei
co ela ion wi h clinical da a and i was ound ha he
de eloped model migh po en ially be used o ex ac
ea u es o p edic medica ion dosage and symp om
se e i y in schizoph enia. These s udies illus a e he
e ec i eness o au oencode s in comp ehending and
ackling men al p oblems.
3. Me hodology
3.1. Da ase desc ip ion
ECG signals we e ob ained om 42 pa ien s in wo clin-
ical cen e s in G eece (No. 683/22-11/2022, 31557/27-
12-2022) and one in Cyp us (No. EEBK/EP/2022/58)
du ing he pe iod 06.2023 - 12.2023 in he amewo k
wi hin he CARDIOCARE’s Clinical S udy [11]. The
s udy aims o imp o e pa ien s’ pa icipa ion in hei
ca e p ocess and imp o e hei physical condi ion and
psychological adap a ion o he disease by implemen -
ing an indi idualised ca e plan. This plan is based on
moni o ing he pa ien ’s heal h s a us using a mobile
pla o m including a sma wa ch, a hea zone senso ,
and a mobile phone (mobile-Heal h moni o ing sys-
em), as well as o collec new bioma ke s.
In his con ex , each pa ien was p o ided wi h
a Pola H10 Hea Ra e Senso and was ins uc ed
o pe o m a 30 min. long ECG e e y wo weeks.
ECG signals we e eco ded om he Pola bel a
a 146 Hz and we e uploaded as 5 min. segmen s
on he CARDIOCARE pla o m h ough a mobile
app. Mo eo e , he pa ien s comple ed he IES-R
ques ionnai e du ing hei ac ical isi s a he clinical
cen e s. In o al 5,285 5 min. ECG segmen s we e
uploaded on he pla o m du ing he 6-mon h pe iod
and we e used o he p ep ocessing s age as desc ibed
in 3.2. Addi ionally, based on hei answe s, he 10
pa ien s wi h IES-R sco es o 33 and abo e we e labeled
as class 1: "likely PTSD" and he es 32 wi h sco es
below 33 as class 0: "no o ew PTSD symp oms".
3.2. Da a Acquisi ion and P ep ocessing
We impo ed long- e m ECG signals using he Py hon
da a analysis ool, Pandas. A e loading he da a,
p ep ocessing was essen ial in p epa ing he da ase o
analysis. Ini ially, we u ilised a da a cleaning p ocedu e
and elimina ed noise and he baseline wande om
he ECGs. Nex , we di ided he ECGs om each
pa ien in o sepa a e hea bea s, 100-poin leng h each,
by u ilising neu oki 2 [24]. Subsequen ly, he da ase
was pa i ioned in o wo dis inc subse s: he ea u e
se , which encompasses he ECG signal da a, and he
label se , which encompasses he ela ed labels. To
educe he impac o a ying sizes and dis ibu ions
in he da a, we implemen ed no malisa ion p ocedu es.
The MinMaxScale om Sciki -Lea n was employed o
s anda dise he da a, which is c ucial o op imal model
aining and analysis.
3
EAI Endo sed T ansac ions
on Pe asi e Heal h and Technology
| Volume 11 | 2025 |
V. Ska amagkasg
3.3. Da ase Pa i ioning
The p ep ocessed da a we e di ided in o sepa a e
aining and es ing se s. The da ase was pa i ioned
using Sciki -Lea n’s ain_ es _spli unc ion, wi h 80%
assigned o aining and 20% se aside o es ing. The
andom s a e pa ame e was con igu ed o gua an ee
he eplicabili y o ou indings. The p ocessing s age
included he ca ego iza ion o ou da a in o no mal
(class 0) and anomalous (class 1) g oups o bo h
aining and es ing, acco ding o he label se , as
desc ibed in de ail in sec ion 3.1. Finally, ou da a
comp ised o a o al o 70,095 hea bea samples:
46,025 no mal aining, 11,557 no mal es ing, and
12,513 anomalous es ing samples espec i ely. I is
impo an o men ion ha he no mal samples belong
o he 32 pa ien s wi h no sings o PTSD, whe eas he
anomalous o he es 10 pa ien s who is likely o ha e
PTSD, acco ding o IES-R.
3.4. Model A chi ec u e and T aining
The ocal poin o ou me hodology e ol ed a ound
he de elopmen and aining o an au oencode model.
The au oencode , a ype o neu al ne wo k, was
designed o lea n e icien encodings o he inpu ECG
da a. I s a chi ec u e (Fig. 1) consis s o an encode
and a decode , wi h a c i ical bo leneck laye . The
encode employs a se ies o ou dense laye s, each
consis ing o 64, 32, 16, and 8 neu ons espec i ely,
all u ilising "ReLU" ac i a ions. This p ocess e ec i ely
educes he dimensionali y o he inpu da a om
i s o iginal 100 o 8. Mo eo e , he 8-neu on laye
se es as he bo leneck o he model, cap u ing he
mos compac ep esen a ion o he inpu da a. The
decode subsequen ly es o es he da a by p og essi ely
inc easing i s dimensionali y h ough laye s con aining
16, 32, and 64 neu ons, culmina ing in a inal laye ha
ma ches he inpu size and u ilises "sigmoid" ac i a ion.
The aining p ocess o ou au oencode model was
pe o med using Tenso Flow and Ke as. The model is
con igu ed o use ea ly s opping h ough a callback ha
moni o s he alida ion loss, ceasing aining i he e’s
no imp o emen o wo consecu i e epochs. This is
se up o minimize loss, a common p ac ice o a oid
o e i ing.
3.5. Pe o mance E alua ion
A e he aining session, we assessed he model’s
pe o mance on he anomalous g oup’s da a. The
e alua ion measu es we e cen ed on he model’s
capaci y o p ecisely ec ea e he ECG da a and de ec
anomalies ha a e sugges i e o PTSD. Speci ically,
he model ini ially o ecas s he es o a ion o he
no mal es da a, p oducing esul s ha ideally should
nea ly mi o he o iginal inpu s i he model is
unc ioning e ec i ely. Nex , he mean absolu e e o
(MAE) is calcula ed be ween hese econs uc ions and
he ac ual no mal es esul s. This MAE se es as a
me ic o he econs uc ion e o o each sample.
The h eshold is es ablished by de e mining he 90 h
pe cen ile o hese econs uc ion e o s. Subsequen ly,
his h eshold is employed o iden i y anomalies by
o ecas ing he econs uc ions o he anomalous
da ase and compu ing hei MAE. Ins ances om he
anomalous da ase ha possess a econs uc ion e o
o e he h eshold a e e y p obable o be anomalies,
as hei e o is la ge han ha o he bulk o he
no mal da a. Ul ima ely, he model’s e icacy is assessed
based on i s accu acy in p ope ly de ec ing abno mal
hea bea s as well as he 1-sco e.
4. Resul s
4.1. Impac o E en Scale - Re ised (IES-R)
The IES-R is a commonly employed psychological
ool de eloped o e alua e he pe sonal discom o
esul ing om auma ic si ua ions [25]. The sel - epo
measu e assesses h ee p ima y elemen s o PTSD:
hype a ousal, in usion, and a oidance, encoun e ed
wi hin he p e ious se en days. The scale comp ises 22
i ems, wi h each i em being e alua ed on a i e-poin
scale ha anges om 0 ("no a all") o 4 ("ex emely").
The IES-R yields a o al sco e anging om 0 - 88, whe e
33 is he op imal h eshold o a likely diagnosis o
PTSD.
4.2. Expe imen al se up and implemen a ion
The implemen a ion and expe imen s we e ca ied
ou in a i ual en i onmen using Py hon e sion
3.9.7, ins alled on a pe sonal compu e equipped wi h
a GTX GeFo ce 750 Ti GPU, an In el(R) Co e(TM)
i7-6700 CPU wi h a clock speed o 3.40 GHz, and
32 GB o RAM. To c ea e, ain, and e alua e he
au oencode model, many amewo ks and lib a ies a e
u ilised, such as Tenso Flow-GPU e sion 2.5.0 wi h he
Ke as-GPU on end. The model employs a MAE loss
unc ion and an Adam op imize wi h ini ial lea ning
a e 10−4. The loss o he model is calcula ed and i s
weigh s a e adjus ed du ing aining. Fu he mo e, he
sugges ed model unde wen aining wi h miniba ches
ha ing size 128 o du a ion 8 epochs, esul ing in a
comple ion ime o a ound 2 minu es. Fu he mo e,
NumPy is employed o a mul i ude o ma hema ical
compu a ions, including he manipula ion o a ay
shapes and he conca ena ion.
4.3. Expe imen al Resul s
Fig. 2exhibi s six ECG aces ca ego ised in o wo
g oups: egula (no mal) and abno mal (anomalous)
4
EAI Endo sed T ansac ions
on Pe asi e Heal h and Technology
| Volume 11 | 2025 |
Towa ds PTSD Diagnosis Th ough ECG Anomaly De ec ion based on Au oencode s
Figu e 2. Indica i e no mal and anomalous ECG hea bea s along wi h he model’s econs uc ion. Each ECG belongs o a di e en
pa ien (3 wi h no sings o PTSD, 3 wi h likely PTSD). Due o he esul ing la ge e o , he algo i hm co ec ly de ec s an anomaly
Figu e 3. His og ams o no mal and anomaly losses along wi h
he compu ed h eshold.
ECG hea bea s ha deno e o 6 di e en pa ien s.
Th ee ypical ECGs a e displayed o each g oup. The
g aphs display he o iginal ECG inpu signal in blue
and i s econs uc ion using he model in o ange. The
ligh o ange shaded egion e lec s he disc epancy
be ween he inpu and he decode ’s ou pu , o en
known as econs uc ion e o . In he no mal ECGs, he
econs uc ion ai h ully eplica es he inpu , leading o
a ela i ely minimal econs uc ion e o . Ne e heless,
in he anomalous ECGs, no iceable dispa i ies be ween
he ac ual da a and he econs uc ed da a a e appa en ,
especially a he highes poin s, leading o mo e
p onounced e o ma gins. The e o e, he model’s
capaci y o iden i y anomalies is demons a ed by he
highe econs uc ion e o obse ed in he anomalous
ECG g aphs, which aligns wi h he an icipa ed
beha iou o an anomaly de ec ion me hod.
A e he model aining p ocess, we calcula ed he
h eshold based on he no mal es da a econs uc ion
e o which was ound o be 0.0016, as desc ibed in
Sec ion 3.5. Fig. 3depic s a his og am ha compa es
he dis ibu ion o econs uc ion loss o no mal and
anomalous ECG da a a e being p ocessed by ou
model. The black his og am depic s he loss a ibu ed
o no mal da a, whe eas he g een his og am illus a es
he loss incu ed by anomalies. A e ical ed line,
se es as a h eshold alue ha dis inguishes be ween
he wo. Losses beyond his le el a e likely o be
ega ded as anomalies, gi en ha hey de ia e om
he ypical ange o no mal da a. The g aph indica es
ha he selec ed h eshold success ully di e en ia es
be ween no mal and anomalous da a, as he majo i y o
anomaly losses a e obse ed o exceed his h eshold.
No ably, ou model achie ed 82.64% accu acy in
de ec ing anomalous ECG hea bea s based on he
compu ed h eshold as well as 1-sco e 0.82.
5. Discussion
This s udy’s indings illus a e he e icacy o
au oencode -based models in iden i ying physiological
anomalies linked o PTSD h ough ECG da a. The
au oencode achie ed a de ec ion accu acy o 82.64%
and an 1-sco e o 0.82 in di e en ia ing abno mal
hea bea s om hose o pa ien s de oid o PTSD
symp oms, acco ding o IES-R h esholds. The esul s
indica e ha mino al e a ions in ECG mo phology,
likely indica i e o au onomic dys egula ion equen ly
associa ed wi h PTSD, may be accu a ely de ec ed and
ecognized using unsupe ised lea ning algo i hms.
5
EAI Endo sed T ansac ions
on Pe asi e Heal h and Technology
| Volume 11 | 2025 |
V. Ska amagkasg
The e icacy o he sugges ed me hod con ibu es
o he expanding li e a u e endo sing he applica ion
o machine lea ning in men al heal h diagnos ics,
especially in ins ances when subjec i e sel - epo ing
may p o e inadequa e o inco ec . In con as o
con en ional diagnos ic me hods ha p edominan ly
u ilize ques ionnai es like he IES-R o s uc u ed
in e iews, he ECG-based anomaly de ec ion model
o e s an objec i e, non-in asi e al e na i e ha may
be implemen ed using wea able echnology. This is
pa icula ly signi ican in a - isk g oups, such as
olde b eas cance pa ien s, when como bidi ies,
wea iness, o emo ional dis ess may hinde p ecise
sel -e alua ion.
The a chi ec u e’s dependence on s anda d (non-
PTSD) da a o aining signi ican ly imp o es i s gene -
alizabili y in anomaly de ec ion con ex s, whe e acqui -
ing ex ensi e, well-anno a ed clinical da ase s is e-
quen ly p oblema ic. U ilizing solely he no mal class
o model aining, he me hod builds a physiological
baseline o iden i y a ia ions ha may sugges PTSD.
This anomaly-based app oach is consis en wi h he
clinical goal o ea ly wa ning and moni o ing a he
han de ini i e diagnos ic classi ica ion.
None heless, se e al limi a ions mus be acknowl-
edged. The sample size, while adequa e o p elimina y
assessmen , was compa a i ely limi ed (n=42), exhibi -
ing an imbalance be ween he PTSD (n=10) and non-
PTSD (n=32) coho s. This may cons ain he s a is-
ical powe and gene alizabili y o he indings. The
dependence on IES-R as a benchma k en ails in in-
sic cons ain s linked o subjec i e sel - epo ins u-
men s, no wi hs anding i s clinical alidi y. Subsequen
esea ch should ocus on co obo a ing esul s wi h
la ge , mo e he e ogeneous popula ions and includ-
ing u he objec i e measu es o PTSD, like ho mone
bioma ke s o mul imodal physiological da a (e.g.,
EEG, GSR).
Mo eo e , he model’s exis ing amewo k p esup-
poses a bina y classi ica ion (p esumably PTSD s non-
PTSD), hus o e simpli ying he con inuum o PTSD
se e i y and symp oma ology. In eg a ing con inuous
se e i y sco es o mul i-class ou pu s may imp o e he -
apeu ic ele ance and cus omisa ion. Finally, al hough
anomaly de ec ion is e ec i e o ecognizing anoma-
lies, i s in e p e a ion necessi a es me iculous con ex-
ualiza ion o p e en o e diagnosis, especially in indi-
iduals wi h concomi an ca dio ascula diseases ha
may independen ly a ec ECG pa e ns.
6. Conclusions
In his wo k, we p esen ed a s udy ha demons a es
a obus au oencode model designed speci ically o
de ec ing anomalies in ECG da a, wi h he po en ial o
iden i y i egula i ies connec ed o PTSD. Pa icula ly
ele an o he ocus on elde ly mul imo bid pa ien s
wi h b eas cance , he indings demons a e ha he
model success ully di e en ia es be ween no mal and
anomalous ECG signals, suppo ed by he alida ion
o a dis inc h eshold based on he his og ams o
econs uc ion e o s. This s a egy shows po en ial o
ad ancing he diagnosis o PTSD using non-in asi e
me hods, by aking use o ECG da a, easily ob ained
e en om wea able de ices.
This esea ch adds o he exis ing knowledge in he
ield o using machine lea ning o diagnosing men al
heal h condi ions and highligh s he possibili y o using
i in a wide ange o clinical se ings. Subsequen
esea ch migh p io i ise enhancing he model using
la ge da ase s and in es iga ing eal- ime anomaly
de ec ion, pe haps p o iding subs an ial ad an ages o
p omp ly in e ening and moni o ing PTSD symp oms.
6.1. Acknowledgemen s
This wo k has ecei ed unding om he Eu opean
Union’s Ho izon 2020 esea ch and inno a ion p og am
unde g an ag eemen No 945175 (P ojec : CARDIO-
CARE).
6.2. Copy igh
The Copy igh licensed o EAI.
Re e ences
[1] Vasile a, A.V.,Ka a ae a, T.A. and Radiono ,
D.S. (2023) Pha maco he apy algo i hm o pos -
auma ic s ess diso de . V.M. BEKHTEREV REVIEW
OF PSYCHIATRY AND MEDICAL PSYCHOLOGY
doi:10.31363/2313-7053-2023-741, URL h ps://www.
bekh e e e iew.com/jou /a icle/ iew/841.
[2] Palme , S.C.,Kagee, A.,Coyne, J.C. and DeMichele,
A. (2004) Expe ience o auma, dis ess, and
pos auma ic s ess diso de among b eas
cance pa ien s. Psychosoma ic Medicine 66(2): 258.
doi:10.1097/01.psy.0000116755.71033.10.
[3] F iedman, M.J. (1996) PTSD diagnosis and ea men
o men al heal h clinicians. Communi y Men al Heal h
Jou nal 32(2): 173–189. doi:10.1007/BF02249755, URL
h ps://doi.o g/10.1007/BF02249755.
[4] O hmani, A.,B ahem, B.,Haddou, Y. and Khan,
M. (2023) Machine lea ning-based app oaches
o pos - auma ic s ess diso de diagnosis using
ideo and eeg senso s: a e iew. p ep in .
doi:10.36227/ ech xi .21967115. 1, URL h ps:
//www. ech xi .o g/doi/ ull/10.36227/ ech xi .
21967115. 1.
[5] Alam , A. and A oli, A. (2023) Unsupe ised
ans o me -based anomaly de ec ion in ecg signals.
Algo i hms 16(3): 152. doi:10.3390/a16030152, URL
h ps://www.mdpi.com/1999-4893/16/3/152.
[6] Ska amagkas, V.,Pen a i, A.,Ke alopoulou,
Z. and Tsiknakis, M. (2023) Mul i-modal deep
6
EAI Endo sed T ansac ions
on Pe asi e Heal h and Technology
| Volume 11 | 2025 |
Towa ds PTSD Diagnosis Th ough ECG Anomaly De ec ion based on Au oencode s
lea ning diagnosis o pa kinson’s disease—a
sys ema ic e iew. IEEE T ansac ions on Neu al
Sys ems and Rehabili a ion Enginee ing 31: 2399–
2423. doi:10.1109/TNSRE.2023.3277749, URL
h ps://ieeexplo e.ieee.o g/documen /10129131.
[7] Ska amagkas, V.,Bou a, I.,Ka amanis, G.,Kyp akis,
I.,Fo iadis, D.I.,Ke alopoulou, Z.,Spanaki, C.
e al. (2025) Dual s eam ans o me o medica ion
s a e classi ica ion in Pa kinson’s disease pa ien s
using acial ideos. npj Digi al Medicine 8(1): 1–11.
doi:10.1038/s41746-025-01630-1, URL h ps://www.
na u e.com/a icles/s41746-025-01630-1.
[8] Ge, F.,Yuan, M.,Li, Y. and Zhang, W. (2020)
Pos auma ic s ess diso de and al e a ions in
es ing hea a e a iabili y: a sys ema ic e iew
and me a-analysis. Psychia y In es iga ion 17(1): 9–20.
doi:10.30773/pi.2019.0112, URL h ps://www.ncbi.
nlm.nih.go /pmc/a icles/PMC6992856/.
[9] Ska amagkas, V.,Pen a i, A.,Fo iadis, D.I. and
Tsiknakis, M. (2023) Using he ecu ence plo s
as indica o s o he ecogni ion o Pa kinson’s
disease h ough phonemes assessmen . In 2023
45 h Annual In e na ional Con e ence o he IEEE
Enginee ing in Medicine & Biology Socie y (EMBC):
1–4. doi:10.1109/EMBC40787.2023.10340177, URL
h ps://ieeexplo e.ieee.o g/documen /10340177.
ISSN: 2694-0604.
[10] Khazaie, H.,Saidi, M.R.,Sepeh y, A.A.,Knigh ,
D.C.,Ahmadi, M.,Naja i, F.,Pa izi, A.A. e al.
(2013) Abno mal ecg pa e ns in ch onic pos -wa p sd
pa ien s: a pilo s udy. In e na ional Jou nal o Beha io al
Medicine 20(1): 1–6. doi:10.1007/s12529-011-9197-y,
URL h ps://doi.o g/10.1007/s12529-011-9197-y.
[11] CARDIOCARE CONSORTIUM (2021) AN INTERDIS-
CIPLINARY APPROACH FOR THE MANAGEMENT
OF THE ELDERLY MULTIMORBID PATIENT WITH
BREAST CANCER THERAPY INDUCED CARDIAC
TOXICITY. CORDIS |Eu opean Commission URL h ps:
//co dis.eu opa.eu/p ojec /id/945175.
[12] Wu, Y.,Mao, K.,Denne , L.,Zhang, Y. and Chen, J.
(2023) Sys ema ic e iew o machine lea ning in PTSD
s udies o au oma ed diagnosis e alua ion. npj Men al
Heal h Resea ch 2(1): 1–10. doi:10.1038/s44184-023-
00035-w, URL h ps://www.na u e.com/a icles/
s44184-023-00035-w.
[13] Ismail, N.H.,Liu, N.,Du, M.,He, Z. and Hu, X. (2020) A
deep lea ning app oach o iden i ying cance su i o s
li ing wi h pos - auma ic s ess diso de on Twi e .
BMC Medical In o ma ics and Decision Making 20(4):
254. doi:10.1186/s12911-020-01272-1, URL h ps://
doi.o g/10.1186/s12911-020-01272-1.
[14] Lekkas, D. and Jacobson, N.C. (2021) Using a i icial
in elligence and longi udinal loca ion da a o di e en-
ia e pe sons who de elop pos auma ic s ess diso -
de ollowing childhood auma. Scien i ic Repo s 11(1):
10303. doi:10.1038/s41598-021-89768-2, URL h ps:
//www.na u e.com/a icles/s41598-021-89768-2.
[15] Za a i, H.,Kosowan, L.,Zulke nine, F. and
Signe , A. (2021) Diagnosing pos - auma ic s ess
diso de using elec onic medical eco d da a.
Heal h In o ma ics Jou nal 27(4): 14604582211053259.
doi:10.1177/14604582211053259.
[16] Yang, J.,Lei, D.,Qin, K.,Pinaya, W.H.L.,Suo, X.,
Li, W.,Li, L. e al. (2021) Using deep lea ning
o classi y pedia ic pos auma ic s ess diso de
a he indi idual le el. BMC Psychia y 21(1): 535.
doi:10.1186/s12888-021-03503-9, URL h ps://doi.
o g/10.1186/s12888-021-03503-9.
[17] Zhu, Z.,Lei, D.,Qin, K.,Suo, X.,Li, W.,Li, L.,DelBello,
M.P. e al. (2021) Combining deep lea ning and g aph-
heo e ic b ain ea u es o de ec pos auma ic s ess
diso de a he indi idual le el. Diagnos ics 11(8): 1416.
doi:10.3390/diagnos ics11081416, URL h ps://www.
mdpi.com/2075-4418/11/8/1416.
[18] Gong, Q.,Li, L.,Tognin, S.,Wu, Q.,Pe e sson-
Yeo, W.,Lui, S.,Huang, X. e al. (2014) Using
s uc u al neu oana omy o iden i y auma su i o s
wi h and wi hou pos - auma ic s ess diso de a he
indi idual le el. Psychological Medicine 44(1): 195–203.
doi:10.1017/S0033291713000561.
[19] Wu, H.,Hu, W. and Fu, D. (2023) Au oencode
based on md and mu ual in o ma ion o de ec
dep ession om speech. In P oceedings o he 2022 4 h
In e na ional Con e ence on Robo ics, In elligen Con ol
and A i icial In elligence, RICAI ’22 (New Yo k, NY,
USA: Associa ion o Compu ing Machine y): 424–428.
doi:10.1145/3584376.3584453, URL h ps://doi.o g/
10.1145/3584376.3584453.
[20] Pa k, H.,Raymond Jung, M.W. and Oh, U. (2021) Apd:
an au oencode -based p edic ion model o dep ession
diagnosis. In 2021 IEEE 22nd In e na ional Con e ence
on In o ma ion Reuse and In eg a ion o Da a Science
(IRI): 376–379. doi:10.1109/IRI51335.2021.00058, URL
h ps://ieeexplo e.ieee.o g/documen /9599107.
[21] Sewani, H. and Kashe , R. (2020) An au oencode -based
deep lea ning classi ie o e icien diagnosis o au ism.
Child en 7(10): 182. doi:10.3390/child en7100182, URL
h ps://www.mdpi.com/2227-9067/7/10/182.
[22] Weng, J.C.,Lin, T.Y.,Tsai, Y.H.,Cheok, M.T.,Chang,
Y.P.E. and Chen, V.C.H. (2020) An au oencode and
machine lea ning model o p edic suicidal idea ion wi h
b ain s uc u al imaging. Jou nal o Clinical Medicine
9(3): 658. doi:10.3390/jcm9030658, URL h ps://www.
mdpi.com/2077-0383/9/3/658.
[23] Yamaguchi, H.,Hashimo o, Y.,Sugiha a, G.,Miya a,
J.,Mu ai, T.,Takahashi, H.,Honda, M. e al.
(2021) Th ee-dimensional con olu ional au oencode
ex ac s ea u es o s uc u al b ain images wi h
a “diagnos ic label- ee” app oach: applica ion o
schizoph enia da ase s. F on ie s in Neu oscience 15.
URL h ps://www. on ie sin.o g/a icles/10.
3389/ nins.2021.652987.
[24] (2021), Neu oKi 2: A Py hon oolbox o neu ophysi-
ological signal p ocessing. URL h ps://gi hub.com/
neu opsychology/Neu oKi . O iginal-da e: 2019-10-
29T05:39:37Z.
[25] Weiss, D.S. (2007) The impac o e en scale: e ised. In
Wilson, J.P. and Tang, C.S.k. [eds.] C oss-Cul u al Assess-
men o Psychological T auma and PTSD, In e na ional
and Cul u al Psychology Se ies (Bos on, MA: Sp inge
US), 219–238.
7
EAI Endo sed T ansac ions
on Pe asi e Heal h and Technology
| Volume 11 | 2025 |
