Implementing federated learning for privacy-preserving emotion detection in educational environments

F on ie s in A i icial In elligence 01 on ie sin.o g
Implemen ing ede a ed lea ning
o p i acy-p ese ing emo ion
de ec ion in educa ional
en i onmen s
RommelGu ié ez
1, WilliamVillegas-Ch
1* and
Se gioLuján-Mo a
2
1 Escuela de Ingenie ía en Cibe segu idad, FICA, Uni e sidad de Las Amé icas, Qui o, Ecuado ,
2 Depa amen o de Lenguajes y Sis emas In o má icos, Uni e sidad de Alican e, Alican e, Spain
Emo ion de ec ion has become an essen ial ool in educa ional se ings, whe e
unde s anding and esponding o s uden s’ emo ions is c ucial o imp o ing hei
engagemen , academic pe o mance, and emo ional well-being. Howe e , adi ional
emo ion de ec ion sys ems, such as DeepFace, and hyb id ans o me -based
models ace signi ican da a p i acy and scalabili y limi a ions. These models
ely on ans e ing sensi i e da a o cen al se e s, comp omising s uden
con iden iali y and making deploymen in la ge o di e se popula ions di icul .
In his wo k, wep opose a ede a ed lea ning-based model designed o de ec
emo ions in educa ional se ings, p ese ing da a p i acy by p ocessing hem
locally on s uden s’ de ices (sma phones, able s, and lap ops). The model was
in eg a ed in o he Moodle pla o m, allowing i s e alua ion in a con en ional
educa ional en i onmen . Ad anced anonymiza ion and p ep ocessing echniques
we e implemen ed o ensu e he secu i y o emo ional da a and op imize i s
quali y. The esul s demons a e ha he p oposed model achie es a p ecision
o 87%, a ecall o 85%, and an F1-sco e o 86%, main aining i s pe o mance
unde ad e se condi ions, such as low ligh ing and ambien noise. In addi ion, a
15% inc ease in academic pa icipa ion and a 12% imp o emen in he a e age
academic pe o mance o s uden s we e obse ed, highligh ing he sys em’s
posi i e impac on educa ional dynamics. This inno a i e me hod combines p i acy,
scalabili y, and pe o mance, posi ioning i sel as a iable and sus ainable solu ion
o emo ion de ec ion in con empo a y educa ional en i onmen s.
KEYWORDS
ede a ed lea ning, emo ion de ec ion, da a p i acy, educa ional en i onmen s,
a i icial in elligence
1 In oduc ion
Emo ion de ec ion has eme ged as a c i ical a ea in de eloping in elligen sys ems,
pa icula ly in educa ional con ex s, whe e emo ions play a pi o al ole in s uden lea ning
and beha io (Mu awa and Hassouneh, 2024; Shmelo a e al., 2024). Unde s anding and
esponding o s uden emo ions can signi ican ly imp o e he pe sonaliza ion o eaching
s a egies, op imize academic engagemen and pe o mance, and con ibu e o he o e all
emo ional well-being o s uden s (Elisondo e al., 2024). Howe e , he implemen a ion o
emo ion de ec ion sys ems aces signi ican challenges ela ed o da a p i acy, scalabili y, and
in eg a ion in o di e se educa ional se ings (Wang A. e al., 2024).
Cen alized models, such as DeepFace by An e al. (2023) acial ea u es ha e been widely
used o emo ion de ec ion due o hei high pe o mance on me ics such as p ecision and
OPEN ACCESS
EDITED BY
Ri a O ji,
Dalhousie Uni e si y, Canada
REVIEWED BY
Oladapo Oyebode,
Dalhousie Uni e si y, Canada
G ace A aguba,
Dalhousie Uni e si y, Canada
Da io Di Da io,
Uni e si y o Sale no, I aly
*CORRESPONDENCE
William Villegas-Ch
[email p o ec ed]
RECEIVED 13 June 2025
ACCEPTED 28 Oc obe 2025
PUBLISHED 07 No embe 2025
CITATION
Gu ié ez R, Villegas-Ch W and
Luján-Mo a S (2025) Implemen ing ede a ed
lea ning o p i acy-p ese ing emo ion
de ec ion in educa ional en i onmen s.
F on . A i . In ell. 8:1644844.
doi: 10.3389/ ai.2025.1644844
COPYRIGHT
© 2025 Gu ié ez, Villegas-Ch and
Luján-Mo a. 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 (CC BY). The
use, dis ibu ion o ep oduc ion in o he
o ums is pe mi ed, p o ided he o iginal
au ho (s) and he copy igh owne (s) a e
c edi ed and ha he o iginal publica ion in
his jou nal is ci ed, in acco dance wi h
accep ed academic p ac ice. No use,
dis ibu ion o ep oduc ion is pe mi ed
which does no comply wi h hese e ms.
TYPE O iginal Resea ch
PUBLISHED 07 No embe 2025
DOI 10.3389/ ai.2025.1644844
Gu ié ez e al. 10.3389/ ai.2025.1644844
F on ie s in A i icial In elligence 02 on ie sin.o g
F1-sco e. Howe e , hese sys ems equi e he ans e o sensi i e da a
o ex e nal se e s o aining and in e ence, which poses signi ican
p i acy isks. In con as , models based on Takase and Kiyono (2023)
ans o me s ha e demons a ed hei abili y o in eg a e mul iple
modali ies, such as ex , audio, and images, o e ing a mo e obus
app oach. Despi e hei ad an ages, high compu a ional cos s and
con igu a ion complexi y limi hei implemen a ion in
educa ional se ings.
In educa ion, p i acy and accessibili y a e c ucial ac o s.
T ans e ing s uden s’ emo ional da a o ex e nal se e s comp omises
con iden iali y and poses e hical and legal challenges in handling
sensi i e in o ma ion (Lee e al., 2024). A he same ime, accessibili y
e e s o he abili y o emo ion de ec ion sys ems o unc ion e ec i ely
ac oss di e se educa ional con ex s, including ins i u ions wi h limi ed
in as uc u e o s uden s wi h a ying le els o echnological access.
Sys ems ha equi e high-pe o mance compu ing o s able
connec i i y may exclude pa o he s uden popula ion, ein o cing
educa ional inequali y. Despi e hese limi a ions, ew s udies ha e
add essed hese issues by de eloping emo ion de ec ion sys ems
speci ically designed o bein eg a ed in o lea ning pla o ms, such as
Moodle, ensu ing bo h da a p o ec ion and adap abili y o he
echnological eali ies o educa ional ins i u ions (Labidi e al., 2021;
Woodwa d e al., 2024).
This wo k in oduces a model based on ede a ed lea ning
designed o emo ion de ec ion in educa ional se ings, which
add esses hese c i ical challenges. Fede a ed lea ning allows aining
models o be un di ec ly on use s’ local de ices, elimina ing he need
o ans e sensi i e da a o cen al se e s (Sengup a e al., 2024). In
addi ion o model de elopmen , his wo k emphasizes he p ac ical
in eg a ion o he ede a ed emo ion de ec ion sys em in o eal
educa ional en i onmen s, assessing i s in luence on s uden
engagemen and academic ou comes. This ea u e imp o es p i acy
and enables g ea e scalabili y by allowing he sys em o ope a e on
la ge and he e ogeneous s uden popula ions (Wang e al., 2024a).
The p oposed me hodology includes a mul i-s age app oach,
s a ing wi h he collec ion o emo ional da a h ough images, audio,
and ex gene a ed du ing academic ac i i ies. The da a was
p ep ocessed using ad anced anonymiza ion and ea u e ex ac ion
echniques, such as andom acial poin mapping and p osody analysis
in speech. Subsequen ly, he ede a ed model was ained locally on
de ices such as sma phones, able s, and lap ops, using ede a ed
a e aging algo i hms o combine he model upda es on a cen al
se e wi hou comp omising da a p i acy (Do iguzzi-Co in and
Si acusa, 2024).
The esul s o his app oach show ha he p oposed model
achie es compe i i e me ics in e ms o p ecision 87%, ecall 85%,
and F1-sco e 86%, which posi ions i as a obus al e na i e o
cen alized sys ems such as DeepFace and comme cial solu ions such
as A ec i a SDK (Hammann e al., 2022). Fu he mo e, obus es s
pe o med unde ad e se condi ions, such as a ia ions in ligh ing
and en i onmen al noise, demons a ed ha he model main ains
consis en pe o mance, ou pe o ming cen alized models in simila
scena ios. Fo example, he model’s p ecision in low ligh ing
condi ions was 80%, compa ed o 75% o cen alized models
e alua ed unde he same condi ions.
In eg a ing he model in o Moodle, a widely used lea ning
managemen sys em, enabled us o e alua e i s p ac ical applicabili y
in a con en ional educa ional en i onmen (Shched ina e al., 2021).
This p ocess demons a ed he sys em’s ease o adop ion and
highligh ed i s posi i e impac on s uden beha io . The esul s
indica e ha posi i e emo ions, such as mo i a ion, de ec ed by he
sys em a e associa ed wi h a 15% inc ease in academic engagemen
and a 12% imp o emen in s uden s’ a e age academic pe o mance.
In con as , al hough i is mo e challenging o de ec nega i e
emo ions, such as s ess and us a ion, i p o ides aluable da a o
adjus educa ional s a egies and p o ide a ge ed emo ional suppo .
Despi e hese ad ances, he model aces limi a ions inhe en o he
ede a ed app oach, such as dependence on he e ogeneous de ices
and sensi i i y o he quali y o ne wo k connec ions du ing he model
agg ega ion p ocess. Al hough signi ican , hese limi a ions do no
comp omise he sys em’s iabili y; ins ead, hey highligh he need o
u u e esea ch o op imize i s pe o mance in en i onmen s wi h
limi ed echnological in as uc u e.
This s udy’s main con ibu ion lies in combining p i acy,
scalabili y, and pe o mance in an emo ion de ec ion sys em
speci ically designed o educa ional en i onmen s. I aims o
de e mine he ex en o which a ede a ed lea ning model can
accu a ely iden i y s uden s’ emo ional s a es, bo h explici and
nuanced, using da a om undamen al academic in e ac ions ac oss
mul iple modali ies. The wo k u he explo es how decen alized
aining a ec s model eliabili y unde eal-wo ld cons ain s,
including limi ed in as uc u e and di e se emo ional exp ession
pa e ns. Unlike exis ing solu ions, he p oposed app oach ensu es he
con iden iali y o emo ional da a while p o iding an adap able and
p ac ical ool o academic ins i u ions.
The emainde o his a icle is s uc u ed as ollows: Sec ion 2
p esen s a li e a u e e iew on emo ion de ec ion sys ems, highligh ing
he cu en limi a ions in e ms o p i acy and scalabili y. Sec ion 3
desc ibes he ma e ials and me hods, including he da a collec ion
p ocess, p ep ocessing echniques, and he design o he ede a ed
lea ning a chi ec u e. Sec ion 4 p esen s he expe imen al esul s,
including pe o mance compa isons, obus ness e alua ions, and
assessmen s o eal-wo ld impac . Sec ion 5 discusses he indings
abou exis ing li e a u e, add esses limi a ions, and ou lines u u e
esea ch di ec ions. Finally, Sec ion 6 summa izes he main
con ibu ions and conclusions o he s udy.
2 Li e a u e e iew
Emo ion de ec ion has been he subjec o nume ous s udies
examining a ious app oaches o iden i ying human emo ions in
di e se con ex s. Among hese app oaches, cen alized sys ems such
as DeepFace (An e al., 2023) ha e demons a ed high pe o mance in
emo ion classi ica ion based on acial ea u es (Anand and Babu,
2024). DeepFace uses highly ained con olu ional neu al ne wo ks
(CNNs) o p ocess images on cen al se e s (Zhang Y. e al., 2024),
achie ing p ecision le els o up o 90% in emo ion de ec ion asks.
Howe e , his cen alized model aces signi ican c i icism due o he
need o ans e sensi i e pe sonal da a o ex e nal se e s,
comp omising use p i acy. This aspec limi s i s applicabili y in
educa ional se ings, whe e da a p o ec ion is a p io i y.
Ano he p ominen app oach is hyb id ans o me -based
models, such as hose p esen ed by Teng e  al. (2024), which
combines image, audio, and ex p ocessing o achie e mo e obus
emo ion de ec ion. These sys ems can analyze mul iple modali ies,
Gu ié ez e al. 10.3389/ ai.2025.1644844
F on ie s in A i icial In elligence 03 on ie sin.o g
in eg a ing con ex ual and empo al ea u es in o hei a chi ec u e.
Al hough ans o me s o e ad an ages in e ms o lexibili y and
pe o mance, hei high compu a ional cos and eliance on la ge
olumes o da a limi hei deploymen in esou ce-cons ained
educa ional se ings. Mo eo e , like cen alized sys ems, hese models
o en equi e ans e ing da a o ex e nal se e s, posing simila
isks o p i acy.
Comme cial sys ems, such as he A ec i a SDK, ha e been
speci ically designed o p ac ical applica ions in ma ke ing and
beha io al analysis (Kulke e al., 2020). This so wa e u ilizes ad anced
compu e ision echniques o iden i y acial emo ions in eal- ime
and is op imized o comme cial pla o ms. A ec i a s ands ou o i s
ease o use and compe i i e pe o mance (Shwe Sin and Khin, 2022),
wi h accu acies anging om 85 o 88%. Howe e , i s closed app oach
and high licensing cos s hinde i s adop ion in educa ional se ings,
whe e budge s a e o en limi ed, and cus om con igu a ions a e
essen ial o in eg a e in o exis ing pla o ms such as Moodle.
Fede a ed lea ning eme ges as an inno a i e solu ion o add ess
he p i acy and scalabili y limi a ions o cen alized models. Huang
e al. (2023) demons a ed he po en ial o ede a ed a chi ec u es o
emo ion de ec ion; howe e , hei amewo k exhibi ed limi a ions in
de ice he e ogenei y, equi ing uni o m clien capabili ies and s able
communica ion channels. These cons ain s limi ed scalabili y and
educed e ec i eness in dynamic educa ional en i onmen s, whe e
de ice esou ces and connec i i y a y conside ably. Mo eo e , hei
wo k did no include p ac ical in eg a ion wi h lea ning pla o ms,
which limi ed i s pedagogical impac and eal- ime applicabili y
wi hin class oom sys ems.
Compa ed o he e iewed models, he ede a ed app oach
p oposed in his wo k s ands ou o i s abili y o balance pe o mance,
p i acy, and scalabili y. I explici ly add esses he echnical challenges
no ed by Huang e al. (2023) by in oducing adap i e p ep ocessing
echniques ha ole a e de ice a iabili y, op imizing local aining o
cons ained de ices, and in eg a ing di ec ly wi h Moodle and o he
lea ning managemen pla o ms. This educes deploymen complexi y
and suppo s ins i u ions wi h limi ed in as uc u e (Muk a e al.,
2024). Despi e ad ances in ede a ed lea ning, exis ing li e a u e has
ye o explo e i s comp ehensi e implemen a ion in hyb id academic
en i onmen s ha combine eal use s, pla o m in eg a ion, and
p i acy-by-design p inciples. This wo k seeks o add ess his
sho coming by p esen ing an in eg a ed and deployable sys em
designed o emo ion de ec ion in educa ional se ings.
3 Ma e ials and me hods
3.1 Desc ip ion o he es en i onmen
The ede a ed lea ning-based emo ion de ec ion sys em was
implemen ed in a uni e si y educa ional en i onmen , speci ically in
he Facul y o Technologies, which includes app oxima ely 650
s uden s. This en i onmen is cha ac e ized by a hyb id educa ion
modali y, meaning ha s uden s a end classes bo h in pe son and
online. This hyb id modali y p esen s an in e es ing challenge o
implemen ing emo ion de ec ion echnologies, as s uden s in e ac
wi h con en and eache s in mul iple ways—ei he in he physical
class oom o h ough digi al pla o ms—enabling he collec ion o
emo ional da a in di e se con ex s.
The Facul y o Technologies o e s aining p og ams in disciplines
ela ed o compu e science, elec onic enginee ing, and
communica ion ne wo ks. This academic p o ile makes he ede a ed
lea ning app oach pa icula ly sui able, as mos s uden s a e amilia
wi h using echnological ools and a e ac i e use s o sma de ices,
which acili a es he adop ion o he p oposed echnology o
emo ion de ec ion.
A o al o 150 s uden s we e selec ed o pa icipa e in he s udy,
ep esen ing app oxima ely 23% o he acul y’s o al s uden
popula ion. This g oup was chosen andomly bu ep esen a i ely,
ensu ing he inclusion o s uden s om di e en majo s wi hin he
acul y and cap u ing a di e se sample o emo ions. In addi ion, 20
eache s ac i ely pa icipa ed in he s udy, allowing o he moni o ing
o s uden s’ emo ional well-being h oughou he cou se, bo h in ace-
o- ace and online classes.
The selec ed sample consis ed o unde g adua e s uden s wi h an
a e age age o 21.2 yea s (SD = 1.7), anging om 18 o 25. Gende
dis ibu ion was app oxima ely 56% male and 44% emale. S uden s
came om h ee main academic p og ams: Compu e Science,
Elec onic Enginee ing, and Communica ion Ne wo ks. All
pa icipan s we e en olled in hyb id cou ses ha combined in-pe son
and i ual componen s, ensu ing a wide ange o in e ac ion
modali ies wi h he sys em. This di e si y suppo s he gene alizabili y
and obus ness o he expe imen al indings.
The implemen a ion occu s in an online and hyb id educa ion
en i onmen , p o iding an ideal oppo uni y o collec ing emo ional
da a in eal- ime and asynch onous in e ac ions (Pi one e al., 2021).
S uden s in e ac wi h he sys em h ough a ious de ices, ei he
du ing online classes, emo e exams, o discussion o ums and
ac i i ies wi hin he Moodle pla o m, which se ed as he Lea ning
Managemen Sys em (LMS) in his pilo es .
The emo ional da a collec ion p ocess is pe o med h ough
a ious sma de ices, such as sma phones, able s, and lap ops,
which a e s anda d in he acul y and in eg a ed in o he s uden s’
daily ac i i ies. These de ices cap u e emo ional da a h ough acial
exp ession analysis, emo ion de ec ion h ough one o oice du ing
o al in e ac ions, and ex analysis in w i en esponses on LMS
pla o ms, mainly in o um ac i i ies and assessmen asks.
Each de ice ac s as a node in he ede a ed sys em, whe e he
emo ional da a cap u ed on each one is p ocessed locally o p ese e
he p i acy o he s uden s (Ribei o Junio and Kamienski, 2024). The
s uden s’ de ices p ep ocess he emo ional da a h ough applica ions
de eloped speci ically o his es , ex ac ing ele an ea u es om
acial images, ocal one, and ex ual esponses. The emo ion de ec ion
model is ained locally on hese de ices, using he da a collec ed in
eal- ime, wi hou such da a lea ing he de ice (Almalki e al., 2024).
S uden s can pa icipa e in he sys em h ough he mobile app and
on hei desk op de ices wi hou equi ing cons an di ec in e ac ion
wi h he sys em, hus allowing he ede a ed lea ning model o adap
o a ia ions in emo ions h oughou he educa ional day. Teache s
can access he emo ion epo s gene a ed wi hou comp omising
s uden s’ p i acy and use hese epo s o adjus hei pedagogical
s a egies in eal ime, especially ega ding s uden wo kload and
s ess du ing classes and assessmen s.
The sys em in as uc u e is based on ede a ed a chi ec u e,
whe e s uden de ices ain he emo ion de ec ion model
independen ly. Communica ion be ween he de ices and he cen al
se e is limi ed o model upda es only, ensu ing ha sensi i e da a is
Gu ié ez e al. 10.3389/ ai.2025.1644844
F on ie s in A i icial In elligence 04 on ie sin.o g
no sha ed a any ime (Zhou e al., 2024)Model upda ing employs
echniques such as ede a ed a e aging, which enables he cen al
se e o agg ega e upda es o local models wi hou equi ing he
o iginal da a o each s uden o becen alized.
Figu e1 p esen s he a chi ec u e o he p oposed sys em o
emo ion de ec ion using ede a ed lea ning. I demons a es how local
s uden de ices, such as sma phones, lap ops, and able s, in e ac
wi h he sys em. Each de ice pe o ms local p ep ocessing o he
emo ional da a, ex ac ing ele an ea u es om acial exp essions,
one o oice, and ex ual in e ac ions. Once he egional models a e
ained on he de ices, he model upda es a e sen o he cen al se e
o agg ega ion h ough a ede a ed a e aging p ocess. This p ocess
allows he cen al se e o combine he model upda es wi hou
cen alizing he o iginal s uden da a, ensu ing da a p i acy (Zhang
H. e al., 2024). Fu he mo e, agg ega ed epo s o s uden emo ions
a e isualized h ough he Teache Dashboa d, p o iding eache s
wi h aluable in o ma ion abou he class’s emo ional well-being
wi hou equi ing access o indi idual pe sonal da a. The connec ion
o Moodle enables he cap u e o s uden s’ academic opinions in eal-
ime, while model upda es con inually imp o e as mo e emo ional
da a is collec ed.
3.2 Da a collec ion
3.2.1 Emo ion cap u e me hod
Th ee speci ic echniques a e used o emo ion de ec ion: acial
exp ession analysis, oice one de ec ion, and ex analysis. These
echniques a e applied complemen a ily o ensu e a comple e and
accu a e assessmen o s uden s’ emo ional s a es du ing educa ional
in e ac ions. Facial exp ession analysis is based on he p emise ha
human emo ions a e eliably e lec ed in acial mo emen s, which
a e de ec ed and classi ied wi h high p ecision using compu e
ision echniques (Lyu, 2023). This p ocess is ca ied ou using a
CNN-based model, which allows he iden i ica ion o key acial
ea u es such as eye, mou h, and eyeb ow mo emen s. Th ough he
on - acing came as o he de ices, he sys em cap u es he s uden s’
acial exp essions in eal- ime. The da a ob ained is p ocessed locally
on each de ice o ex ac he ele an emo ional ea u es, allowing
he de ec ion o emo ions such as happiness, sadness, ange ,
su p ise, con emp , and disgus , which co espond o he basic
emo ions iden i ied by Ekman e al. (1998). The applica ion o his
model is ca ied ou con inuously du ing he s uden ’s in e ac ions
wi h he academic en i onmen , ensu ing he accu a e cap u e o
emo ions in a ious si ua ions. Howe e , i is acknowledged ha
some o hese in e ac ions may occu ou side he co e academic
pla o m (e.g., Moodle), we e ex e nal, non-educa ional ac o s
could in luence emo ional a ia ion. These ac o s lie beyond he
eache ’s con ol and could in oduce biases in in e p e ing s uden s’
emo ional s a es, a limi a ion also highligh ed in ecen e hical
s udies on emo ion ecogni ion in educa ional se ings (Di Da io
e al., 2024).
Voice pi ch de ec ion is ano he c ucial me hod in emo ion
de ec ion. This p ocess in ol es cap u ing and analyzing a ia ions in
he acous ic ea u es o he oice, such as undamen al equency,
in ensi y, du a ion, and p osody (Jiang e al., 2023). These ea u es
indica e emo ional a ia ions in speech, as ocal pi ch and hy hm
change in esponse o he indi idual’s emo ional s a e. Mic ophones
in he de ices pick up he s uden ’s oice du ing o al in e ac ions.
Using audio signal p ocessing algo i hms, such as acous ic ea u e
analysis and ime-sequence modeling, a ia ions in speech a e
analyzed o iden i y emo ions, including s ess, con usion, o
sa is ac ion. A model based on Recu en Neu al Ne wo ks (RNN),
speci ically Long Sho -Te m Memo y (LSTM), is used, which can
iden i y emo ional pa e ns h oughou oice sequences, allowing
accu a e de ec ion in dynamic si ua ions (Chen e al., 2017).
FIGURE1
A chi ec u e o he ede a ed lea ning-based emo ion de ec ion sys em.
Gu ié ez e al. 10.3389/ ai.2025.1644844
F on ie s in A i icial In elligence 05 on ie sin.o g
Tex analysis examines he emo ional con en o s uden s’ w i en
esponses on pla o ms such as Moodle, especially in o um
in e ac ions, assignmen s, and exams. The sys em iden i ies linguis ic
pa e ns ela ed o emo ions using na u al language p ocessing (NLP)
echniques (Me i y e  al., 2018). Ad anced models, such as
Bidi ec ional Encode Rep esen a ions om T ans o me s (BERT),
a e applied, which can analyze he seman ic con ex o wo ds and
ph ases wi hin ex s. This analysis allows he classi ica ion o
unde lying emo ions, such as anxie y, mo i a ion, o con usion, om
w i en in e ac ions (Subak i e al., 2022). Tex da a is p ocessed in eal
ime, assessing s uden s’ emo ions based on hei w i en exp essions
du ing academic ac i i ies.
3.2.2 De ices and senso s
Emo ional da a is collec ed using a ious de ices and senso s
embedded in s uden s’ de ices, speci ically sma phones, able s, and
lap ops. Each de ice has speci ic echnologies ha accu a ely cap u e
emo ional da a based on he in e ac ion modali y.
The came as cap u e acial exp essions, enabling he eal- ime
isual analysis o emo ions. The came as can iden i y and classi y
acial pa e ns ela ed o basic emo ions, such as happiness, sadness,
ange , su p ise, and o he s.
On he o he hand, he mic ophones buil in o he de ices allow
o cap u ing he acous ic cha ac e is ics necessa y o analyze oice
one. These mic ophones a e designed o cap u e sounds a an
app op ia e equency, enabling he de ec ion o a ia ions in pi ch,
olume, and speech a e ha indica e di e en emo ional s a es.
Th ough hese mic ophones, he sys em can iden i y whe he he
s uden is expe iencing emo ions such as s ess o sa is ac ion, which
co ela es wi h he one and dynamics o hei oice.
The Moodle pla o m is used o collec ing ex ual da a. S uden s
in e ac on he pla o m h ough o ums, assignmen s, and exams,
gene a ing w i en esponses ha a e hen p ocessed o assess he
unde lying emo ions in hei con en . The sys em analyzes he
wo ds, ph ases, and ex s uc u e using na u al language p ocessing
models o iden i y emo ional s a es ela ed o he con en o he
esponses, such as anxie y, mo i a ion, o con usion. Table 1
summa izes he de ices, senso s, and pla o ms employed, along
wi h hei espec i e unc ions in he emo ional da a
collec ion p ocess.
All de ices used o da a collec ion we e he pe sonal p ope y o
he s uden s. The emo ion de ec ion sys em was no p e-ins alled;
ins ead, i was accessed en i ely h ough he Moodle Lea ning
pla o m, which p o ided a seamless in e ace o da a cap u e and
analysis. This in eg a ion ensu ed ha no addi ional so wa e needed
o beins alled on s uden de ices, he eby minimizing in usi eness
and main aining use au onomy. Addi ionally, he sys em’s a chi ec u e
ensu es ha all da a is p ocessed locally on he de ice, aligning wi h
he p inciples o p i acy-by-design.
3.3 Da a p ep ocessing
In he il e ing and anonymiza ion p ocess, speci ic echniques a e
applied o p o ec sensi i e da a, especially s uden s’ acial and audio
ea u es (Hanisch e al., 2024). Facial exp ession da a is p ocessed o
emo e backg ounds and ligh ing a ia ions i ele an o emo ion
de ec ion. This il e ing is pe o med by a ace segmen a ion algo i hm
using he OpenCV lib a y, which de ec s he exac loca ion o he ace
wi hin he image and c ops only he egion o in e es . A Gaussian
smoo hing il e is applied o he ace egion o educe backg ound
noise and ensu e ha only ele an acial ea u es a e p ocessed
(Nandan e al., 2024).
Da a anonymiza ion is pe o med by modi ying he de ec ed
acial poin s so he indi idual canno beiden i ied. In he case o acial
analysis, key landma ks, such as he eyes, eyeb ows, and mou h, a e
eplaced by gene ic poin s ha do no co espond o a speci ic iden i y.
This echnique uses acial mapping algo i hms ha andomly eloca e
acial ea u es wi hin a ange o s anda d acial pa ame e s (Wang,
2024). In addi ion, he da a is no s o ed in i s o iginal o m; ins ead,
only he model upda es a e sen , implying ha he acial images ne e
lea e he local de ices and do no con ain iden i iable in o ma ion.
In enc yp ion, he Ad anced Enc yp ion S anda d (AES-256)
enc yp s he model pa ame e s when hey a e sen om he local
de ices o he cen al se e (Ajagbe e al., 2024; Mish a e al., 2024).
This enc yp ion ensu es ha e en i he da a is in e cep ed, i canno
bedec yp ed wi hou he p ope key, p o ec ing he s uden s’ p i acy
du ing model communica ion. The ea u e ex ac ion p ocess o each
da a ype ( acial exp essions, oice, and ex ) is ca ied ou using
speci ic algo i hms designed o each modali y.
To ensu e cla i y and ep oducibili y, he emo ional s a es a ge ed
by each modali y we e explici ly de ined and consis en ly applied
h oughou he aining and e alua ion phases. Each modali y was
associa ed wi h a dis inc subse o emo ional labels based on he
na u e o he da a and he capabili ies o he co esponding model.
These labels we e selec ed om well-es ablished emo ional axonomies
ha ha e been adap ed o educa ional se ings. Table2 summa izes
he exac emo ions de ec ed by acial exp essions, oice signals, and
ex ual con en .
3.3.1 Facial exp ession de ec ion
In acial exp ession analysis, Dlib’s acial poin de ec ion algo i hm
iden i ies c i ical poin s on he ace, such as he con ou s o he eyes,
nose, and mou h. The ma hema ical p ocess unde lying his algo i hm
is based on supe ised lea ning and nonlinea eg ession echniques.
Once hese poin s a e de ec ed, he Ac i e Shape Model (ASM) is used
o model he a iabili y in he shape o he ace (Ala i e al., 2024).
The ASM can be desc ibed ma hema ically by an elas ic
de o ma ion model ha adjus s pa ame e s o cap u e acial a ia ion.
The wa ping algo i hm uses a ine ans o ma ion ma ices, whe e
TABLE1 De ices, senso s, and pla o ms used o collec ing emo ional da a.
De ice/senso Func ion Associa ed echnique
Sma phone/ able /lap op Cap u e o emo ional da a h ough a came a and a mic ophone Facial analysis, oice de ec ion, ex analysis
Came a Cap u e o acial exp essions o eal- ime analysis Facial exp ession analysis
Mic ophone Cap u e o one o oice, a ia ions in equency, and ampli ude o emo ion de ec ion Voice one analysis
Moodle (LMS) A pla o m o collec ing ex ual esponses h ough educa ional in e ac ions Tex analysis

Gu ié ez e al. 10.3389/ ai.2025.1644844
F on ie s in A i icial In elligence 06 on ie sin.o g
each ans o ma ion T is ep esen ed by a pa ame e ma ix Ɵ ha
adjus s he posi ion o each acial poin p(x, y) in he image acco ding
o he wa ping model o he Equa ion (1):
( ) ( ) ( )
θ
=
′
⋅
,,
p xy T p xy (1)
T is he ans o ma ion ma ix, which desc ibes how acial poin s
a e adjus ed acco ding o emo ional a ia ions. In addi ion, he
Euclidean dis ance be ween he de ec ed acial poin s is calcula ed o
measu e he deg ee o change in acial exp essions as shown in
Equa ion (2):
( ) ( )
= − +−
22
21 21
d xx yy
(2)
This allows he quan i ica ion o de o ma ion in he ace ela ed o
emo ions such as su p ise o sadness.
Speech signal analysis is based on acous ic ea u es such as he
undamen al equency (F
0
) ex ac ed using he Fas Fou ie
T ans o m (FFT). Ma hema ically, he FFT decomposes a signal x( )
in o i s equency componen s, ep esen ing he signal in he
equency domain as a sum o sinusoids, as shown in Equa ion (3):
( ) ( )
π
∞
−
−∞
=
∫
2j
X x e d
(3)
X( ) is he equency domain ep esen a ion o he signal, is he
equency, and x( ) is he ime domain signal.
The undamen al equency (F
0
) is he lowes componen o he
audio signal and is ela ed o he pi ch o he oice. This pa ame e is
ex ac ed o measu e emo ional a ia ions in he oice pi ch, such as
when ange o joy is de ec ed. Addi ionally, p osody analysis is
employed, which examines he in ensi y and hy hm o he oice.
Ma hema ically, hy hm can bemeasu ed in e ms o syllable du a ion
and speech a e, and in ensi y is e alua ed as he ampli ude o he
audio signal in each ime window using ene gy measu emen
o mulas, as shown in Equa ion (4):
( ) ( )
=
= +
∑2
0
N
n
E x n
(4)
whe e E( ) is he ene gy in a ime window, x( + n) is he alue o
he audio signal a ime + n, and N is he numbe o samples wi hin
he ime window.
P osody analysis is hen used o eed he LSTM model, which
applies backp opaga ion h ough ime (BPTT) o upda e he neu al
ne wo k weigh s and model emo ions based on speech’s pi ch and
empo al a iabili y. LSTMs use ac i a ion unc ions such as sigmoid
o anh, which classi y emo ions based on he empo al con en o
he signal.
3.3.2 Tex analysis
Tex analy ics is based on ans o me models, such as BERT,
designed o cap u e he bidi ec ional con ex o wo ds wi hin a
sen ence (Ko wal e al., 2022). Ma hema ically, his model is a wo d
embedding, which maps wo ds o high-dimensional ec o s in a
ec o space, using unc ions such as so max o gene a e
classi ica ion p obabili ies.
In ma hema ical e ms, he embedding p ocess is desc ibed by a
p ojec ion o each wo d w
i
in o a d-dimensional ea u e space, as
shown in Equa ion (5):
( )
=
ii
w
(5)
whe e i is he ea u e ec o o he wo d wi, and is he p ojec ion
unc ion lea ned du ing aining.
Using sel -a en ion, he BERT model cap u es con ex ual
ela ionships be ween wo ds, which compu es he weigh ed
ela ionship be ween wo ds wi hin a gi en con ex . A en ion is
ma hema ically de ined in Equa ion (6):
( )

=



,,
T
k
QK
A en ion Q K V so max V
d
(6)
whe e Q, K, and V a e he que y, key, and alue ma ices,
espec i ely, and d
k
is he dimension o he keys. This a en ion
mechanism enables he model o cap u e long- ange dependencies
wi hin he ex , allowing i o de ec complex emo ions such as
us a ion o mo i a ion.
Once he ec o ep esen a ions o he wo ds a e ob ained, hey
a e used o classi y he emo ions associa ed wi h he ex h ough a
deep neu al ne wo k ha adjus s he weigh s using he
backp opaga ion algo i hm and he so max ac i a ion unc ion o
ob ain he p obabili y o each emo ion, as shown in Equa ion (7):
( )
=∑
i
j
z
iz
j
e
P emo ion
e
(7)
whe e z
i
a e he ne wo k ou pu s o each emo ional class, and
P(emo ioni) is he p obabili y ha he emo ion is p esen in he ex .
3.3.3 Model aining by modali y and da ase
desc ip ion
Fo emo ion de ec ion in educa ional con ex s, h ee specialized
models we e de eloped, each adap ed o a di e en modali y: acial
images, oice signals, and w i en ex . These models we e ained
using public da ase s widely alida ed in he li e a u e, ensu ing hei
a ailabili y and alidi y o emo ion classi ica ion asks. Fu he mo e,
in asi e collec ion p ocesses o hose dependen on sensi i e
in o ma ion we e a oided, aligning wi h he p i acy p inciples de ined
in he o e all sys em design.
Fo emo ion de ec ion using acial exp essions, he JAFFE da ase
was u ilized, which comp ises 213 images wi h a esolu ion o 48 × 48
TABLE2 Emo ional s a es de ec ed by each modali y.
Modali y Emo ional s a es de ec ed
Facial exp ession Happiness, sadness, ange , su p ise, disgus , con emp
Voice (audio) S ess, con usion, sa is ac ion, bo edom, engagemen
Tex ual con en Mo i a ion, anxie y, con usion, us a ion, cu iosi y
Gu ié ez e al. 10.3389/ ai.2025.1644844
F on ie s in A i icial In elligence 07 on ie sin.o g
pixels, ca ego ized in o se en emo ional s a es: happiness, sadness,
ange , su p ise, ea , disgus , and neu ali y. These images we e
cap u ed in uncon olled scena ios, enabling he model o gene alize
mo e e ec i ely in eal-li e condi ions. Fo speech-based de ec ion,
he Rye son Audio-Visual Da abase o Emo ional Speech and Song
(RAVDESS) da ase was used. I consis s o 1,440 audio clips eco ded
by p o essional ac o s exp essing eigh emo ions (calm, happiness,
sadness, ange , ea , disgus , su p ise, and neu al). Finally, o he
ex ual modali y, he Emo ionX da ase , which ocuses on undamen al
con e sa ional in e ac ions, was u ilized. This co pus includes b ie
esponses manually labeled wi h emo ions such as happiness, sadness,
ange , mo i a ion, us a ion, o su p ise. All da ase s we e openly
accessible, and no p i a e da a was included; addi ionally, no manual
labeling was pe o med.
Each model was designed o espond o he speci ic cha ac e is ics
o i s modali y. The acial image model employed a VGG-13-based
a chi ec u e, comp ising wo con olu ion blocks wi h 64 and 128
il e s, espec i ely, ollowed by max pooling ope a ions and dense
laye s o 512 and 128 uni s, be o e he so max ou pu laye . ReLU
ac i a ion unc ions and a d opou alue o 0.5 we e used o p e en
o e i ing. Fo speech modali y, he model was buil on an LSTM
ne wo k wi h 256 hidden uni s, ollowed by a dense laye wi h 64
neu ons and a so max ou pu uned o eigh classes. The inpu
consis ed o sequences o MFCC coe icien s ex ac ed om 25-ms
segmen s. Ba ch no maliza ion, a d opou o 0.3, and he ca ego ical
c oss-en opy loss unc ion we e employed. Fo ex , he BERT model
(uncased e sion, 110 million pa ame e s) was implemen ed, o which
a dense laye wi h 128 neu ons and a so max ou pu o six classes was
added. Fine- uning was pe o med only on he las ou laye s o he
ans o me o p ese e he p e- ained seman ic capaci y.
The h ee models we e ained using a amilia hype pa ame e
se ing, wi h sligh a ia ions ailo ed o he compu a ional needs o
each modali y. A ba ch size o 32 was used o images and speech, and
16 o ex . The ini ial lea ning a e was 0.0001, wi h he Adam
op imize and a weigh decay penal y o 1e-5. The maximum numbe
o epochs was se o 50, wi h an ea ly s opping mechanism ac i a ed
i no imp o emen was obse ed in alida ion a e 10 i e a ions. In
all cases, he se s we e di ided in o 70% o aining, 15% o
alida ion, and 15% o es ing, ollowing a consis en p o ocol
ac oss modali ies.
The aining en i onmen consis ed o no ebooks de eloped in
Py hon 3.9 using PyTo ch 2.0, HuggingFace T ans o me s, and he
lib osa lib a y o acous ic ea u e ex ac ion. The expe imen s we e
conduc ed on Google Colab P o+ wi h access o a 16 GB Tesla T4
GPU and 52 GB o RAM, enabling e icien and ep oducible aining.
I is essen ial o cla i y ha hese models we e no ained di ec ly on
s uden da a, bu a he p e- ained on he da ase s abo e and
subsequen ly deployed in a ede a ed a chi ec u e. The ede a ed
p ocess in ol ed h ee o i e local ine- uning cycles pe de ice,
enabling he models o g adually specialize acco ding o he emo ional
cha ac e is ics o he eal-li e educa ional en i onmen , while
p ese ing use p i acy.
To complemen hese public da ase s, he models we e no
deployed in hei p e- ained o m only. Once in eg a ed in o he
ede a ed en i onmen , each modali y was ine- uned locally
using anonymized eco ds de i ed om undamen al s uden
in e ac ions wi hin he Moodle pla o m, including o um
messages, oice pa icipa ion, and acial exp essions cap u ed
du ing hyb id sessions. This local ine- uning p ocess ensu ed
ha he models adap ed o he speci ic linguis ic, acous ic, and
beha io al cha ac e is ics o he a ge educa ional popula ion,
while espec ing p i acy cons ain s. Impo an ly, no aw
in e ac ion da a was cen alized; only model upda es we e
ansmi ed ollowing he p inciples o ede a ed lea ning. In his
way, he aining s a egy combined he obus ness o publicly
alida ed da ase s wi h he con ex ual speci ici y o eal-wo ld
da a, ensu ing me hodological consis ency and ecological alidi y.
To add ess he misma ch be ween he a ge emo ional ca ego ies
and he labels p esen in he p e- aining co po a, addi ional open
da ase s and a ha moniza ion s a egy we e inco po a ed. Fo acial
modali y, supplemen a y co po a such as A ec Ne (Mollahosseini
e al., 2019) we e used o include classes no co e ed by FER2013
(San oso and Kusuma, 2022), pa icula ly con emp , while s ill elying
on FER2013 as he baseline o basic acial emo ions. In he audio
modali y, RAVDESS was expanded wi h esou ces like EMO-DB,
IEMOCAP, and RECOLA (Joudeh e al., 2023; Khu ana e al., 2024;
Ong e al., 2024), which p o ide ca ego ies closely aligned wi h s ess,
con usion, and bo edom. P osodic dimensions om hese co po a,
mapped along alence–a ousal axes, enabled he de i a ion o
sa is ac ion and engagemen - ela ed cues. Fo ex , da ase s such as
GoEmo ions and educa ion-speci ic co po a we e in eg a ed, ensu ing
co e age o s a es like mo i a ion, anxie y, and cu iosi y h ough
seman ic mapping and weak supe ision echniques (Demszky
e al., 2020).
This p ocess ollowed a label-space ha moniza ion app oach in
which seman ically equi alen o p oxima e ca ego ies om di e en
sou ces we e me ged in o a uni ied axonomy. Mapping was suppo ed
by dis ibu ional simila i y measu es and embedding-based alignmen
o main ain consis ency ac oss modali ies. When labels we e absen
om he p e- aining co po a bu p esen in supplemen al ones,
ans e lea ning mechanisms we e employed o ans e
ep esen a ions in o he ede a ed ine- uning s age.
Rega ding FER2013, only he 32,298 publicly a ailable images
we e used, as he emaining po ion o he o iginal co pus is
es ic ed and inaccessible. The da ase is dis ibu ed in o a
p ede ined aining spli o 28,709 images and a es spli o 3,589
images. To in oduce a alida ion s age consis en wi h he
70/15/15 s a egy applied ac oss modali ies, we u he
pa i ioned he aining spli by ealloca ing 15% o i s samples
(≈4,307 images) as a alida ion subse , while e aining 24,402
images o aining. The o iginal es spli o 3,589 images was
p ese ed wi hou modi ica ion o se e as he inal e alua ion
se . This p ocedu e ensu ed me hodological uni o mi y ac oss
modali ies while main aining compa ibili y wi h he canonical
FER2013 e alua ion p o ocol, he eby a oiding he use o
non-public da a and ein o cing he ep oducibili y o
he expe imen s.
Finally, speci ic high-le el a ec i e cons uc s, such as
engagemen , we e no di ec ly p edic ed by a single classi ie bu
in e ed h ough mul imodal usion. In hese cases, he sys em
combined audio-p osodic indica o s, acial ac i a ion le els, and
beha io al aces om LMS in e ac ions o de i e a composi e s a e.
This ensu ed ha all emo ional ca ego ies de ined in he s udy we e
echnically g ounded, ei he h ough explici da ase co e age,
mapped p oxies, o composi e modeling s a egies aligned wi h he
ede a ed a chi ec u e.
Gu ié ez e al. 10.3389/ ai.2025.1644844
F on ie s in A i icial In elligence 08 on ie sin.o g
3.3.4 De ini ion o emo ion classes by modali y
The de ini ion and ca ego iza ion o he emo ional s a es a ge ed
in his s udy we e es ablished o gua an ee cla i y, ep oducibili y, and
echnical igo . Each modali y— acial exp essions, oice signals, and
ex ual con en —was associa ed wi h a se o emo ional labels aligned
wi h alida ed axonomies in a ec i e compu ing and educa ional
psychology. This s uc u ed de ini ion ensu ed ha he classi ica ion
asks we e consis en ac oss modali ies and ha he e alua ion o he
ede a ed sys em was aceable and compa able wi h exis ing models.
Fo acial modali y, he axonomy p oposed by Ekman and
Rosenbe g was adop ed, as i p o ides a obus ounda ion o iden i ying
emo ions ha a e consis en ly exp essed h ough acial mo emen s
(Ekman e  al., 1998). Emo ions such as happiness, sadness, ange ,
su p ise, con emp , and disgus we e selec ed because hey p esen
dis inc i e isual cues ha can bequan i ied using con olu ional neu al
ne wo ks. These ca ego ies ha e been epea edly alida ed in emo ion
ecogni ion s udies, allowing o a eliable mapping be ween obse able
acial ea u es and unde lying a ec i e s a es.
In he oice modali y, he classes o s ess, con usion, and
sa is ac ion we e de ined, gi en hei s ong co ela ion wi h a ia ions
in p osodic ea u es such as pi ch, in ensi y, and hy hm. These
emo ions a e adequa ely ep esen ed in co po a like RAVDESS and a e
c i ical in educa ional con ex s, whe e ocal modula ion o en e lec s
cogni i e load and a ec i e esponses o lea ning asks. By ocusing on
hese speci ic s a es, he sys em cap u es meaning ul indica o s o
s uden s’ emo ional dynamics in o al in e ac ions (Bilo i e al., 2024).
In addi ion o s ess, con usion, and sa is ac ion, he model also
inco po a ed wo de i ed a ec i e s a es—bo edom and engagemen .
These s a es we e no di ec ly anno a ed in he base RAVDESS co pus.
S ill, hey we e ob ained h ough he in eg a ion o IEMOCAP and
RECOLA da ase s, whe e p osodic pa e ns we e mapped along he
alence–a ousal plane. Bo edom was associa ed wi h low a ousal and
neu al- o-nega i e alence speech segmen s, while engagemen
co esponded o high a ousal and posi i e alence p osodic pa e ns.
These de i ed s a es we e inco po a ed h ough label ha moniza ion
and alida ed du ing he ede a ed ine- uning phase, allowing he
model o in e mo i a ional in ensi y om oice cues.
In he ex ual modali y, emo ions such as anxie y, mo i a ion, and
us a ion we e p io i ized. These ca ego ies a e highly ele an in
w i en academic in e ac ions, whe e s uden s equen ly exp ess hei
a ec i e s a es indi ec ly h ough language. Using ans o me -based
seman ic embeddings, pa icula ly BERT, he sys em was able o analyze
he bidi ec ional con ex o w i en esponses, cap u ing sub le a ia ions
in meaning ha e lec s uden s’ a ec i e condi ions (Sayeed e al., 2023).
Beyond mo i a ion, anxie y, and us a ion, wo addi ional a ec i e
s a es—cu iosi y and con usion—we e in eg a ed h ough seman ic
mapping using he GoEmo ions co pus and educa ion-speci ic ex
samples. Cu iosi y was iden i ied h ough linguis ic cons uc ions
e lec ing posi i e explo a o y in en (e.g., in e oga i e o ms combined
wi h posi i e sen imen ). In con as , con usion eme ged as a composi e
ca ego y de i ed om us a ion and unce ain y labels h ough weak
supe ision. These ca ego ies we e e ained du ing ine- uning as hey
equen ly occu in lea ning con ex s, enabling mo e accu a e modeling
o cogni i e-a ec i e dynamics in s uden w i ing. The axonomy,
o ganized by modali y, aligns wi h Table2, whe e basic emo ions (e.g.,
happiness, sadness) coexis wi h de i ed and con ex -speci ic s a es (e.g.,
engagemen , cu iosi y).
This h ee old de ini ion o emo ional classes p o ides a igo ous
amewo k o he ede a ed model, ensu ing ha each modali y
con ibu es in a complemen a y manne o he global de ec ion
p ocess. The ca e ul alignmen o modali ies wi h dis inc emo ional
ca ego ies a oids o e laps, educes ambigui y in classi ica ion, and
ein o ces he in e p e abili y o he esul s ob ained in wo ld
educa ional en i onmen s.
3.4 De elopmen o he emo ion de ec ion
model
3.4.1 Emo ion de ec ion models
Di e en ypes o deep lea ning models a e used o add ess
emo ion de ec ion in s uden s, ailo ed o he speci ic cha ac e is ics
o each da a modali y: acial images, audio, and ex . These models
ha e been selec ed o hei abili y o lea n complex, high-le el
ep esen a ions o emo ional da a, and each one specializes in he ype
o da a i is p o ided wi h.
Fi s , CNNs a e employed o acial exp ession analysis, which can
ex ac spa ial ea u es om acial images. CNNs a e especially
e ec i e in compu e ision asks due o hei abili y o iden i y
hie a chical pa e ns o in o ma ion, anging om simple ea u es
such as edges and ex u es o complex pa e ns, including emo ions
exp essed on he ace. The model is ained using high- esolu ion
acial images, whe e he ne wo k lea ns o iden i y spa ial ela ionships
be ween key poin s on he ace.
CNNs ope a e by applying con olu ional il e s o images, whe e
each il e Wk gene a es a ea u e map Ck as de ined in Equa ion (8):
= ∗
kk
C WI
(8)
whe e I is he inpu image and * deno es he con olu ion ope a ion.
These ea u e maps a e combined o ex ac emo ions such as
happiness, sadness, ange , o su p ise.
An RNN and an LSTM a e used o oice one analysis and a e
ideal o p ocessing empo al da a sequences such as audio signals.
LSTMs a e designed o cap u e long- e m dependencies in audio
sequences, which is c ucial o iden i ying emo ions ha e ol e in a
con e sa ion o speech (Hashmi and Yayilgan, 2024). LSTM
pa ame e s, such as inpu , ou pu , and o ge ga es, allow he ne wo k
o emembe and o ge in o ma ion based on he empo al
cha ac e is ics o he signal. Ma hema ically, he LSTM model is
de ined by he ollowing Equa ions (9)– (13):
( )
σ
−
=⋅+


1,
WhX b
(9)
( )
σ
−
=⋅+


1,
i i
i Wh X b
(10)
( )
−
=⋅+


1
ˆ
an ,
c c
C hW h X b (11)
−
= ∗ +∗
1ˆ
C C ic
(12)
( )
= ∗ anh
ho C
(13)
Gu ié ez e al. 10.3389/ ai.2025.1644844
F on ie s in A i icial In elligence 09 on ie sin.o g
whe e
, i
, and o
a e he o ge , inpu , and ou pu ga es, x
is he
empo al inpu (audio signal), h
is he hidden s a e, and C
is he cell s a e.
A ans o me -based model, speci ically BERT, is used o ex
analysis, which is highly e ec i e a p ocessing ex bidi ec ionally.
The BERT model can unde s and he ull con ex o a wo d wi hin a
sen ence, as i examines bo h he p eceding and ollowing con ex s o
he wo d. This app oach ou pe o ms adi ional one-way models and
is pa icula ly use ul o unde s anding complex emo ions in language.
Tex analysis in BERT is done h ough wo d embedding, whe e each
wo d is ans o med in o a high-dimensional ec o ha cap u es
i s con ex .
3.4.2 Fede a ed model
The ede a ed lea ning model implemen ed in his s udy enables
emo ion de ec ion models o be ained in a decen alized manne ,
i.e., wi hou cen alizing sensi i e da a on a se e . This app oach is
c ucial o ensu ing s uden p i acy, as only model upda es a e sha ed,
no he o iginal da a.
Each s uden de ice uses da a o ain he emo ion de ec ion
model du ing local aining. This p ocess is conduc ed locally,
meaning ha each s uden ’s emo ional da a emains on hei de ice.
The model on each de ice is con inuously uned and imp o ed as
mo e emo ional da a is collec ed om he s uden ’s in e ac ions wi h
educa ional con en .
The local model pe o ms pa ame e upda es using he g adien
descen algo i hm. Since he da a is no cen alized, aining is
ca ied ou in pa allel on each de ice wi hou sha ing in o ma ion
abou he s uden s’ da a. The model pa ame e s, which a e weigh
ec o s wi, a e upda ed based on he local e o calcula ed a each
de ice, and he upda e ollows he s anda d g adien ule, as
exp essed in Equa ion (14):
η
= − ⋅∇ i
ii w
ww L
(14)
whe e
η
is he lea ning a e, and
∇i
wL
is he g adien o he loss
unc ion L conce ning he pa ame e s wi.
To in eg a e emo ional in o ma ion ob ained om he h ee da a
modali ies, acial images, oice eco dings, and ex inpu s, he sys em
employs a la e usion s a egy. Each modali y is p ocessed
independen ly on he s uden ’s de ice using he espec i e specialized
models: a CNN o acial exp essions, an LSTM o oice one, and a
ans o me (BERT) o ex ual da a. Each model ou pu s a p obabili y
dis ibu ion o e he p ede ined se o emo ional classes. These h ee
dis ibu ions a e hen combined using a weigh ed a e age, whe e he
weigh s we e empi ically uned du ing he de elopmen phase o
op imize o e all classi ica ion pe o mance. The inal emo ional
p edic ion co esponds o he class wi h he highes combined
p obabili y. This modula app oach enables lexible p ocessing e en
in scena ios whe e one o mo e modali ies a e empo a ily una ailable
(e.g., no audio inpu ), ensu ing he obus ness and adap abili y o he
ede a ed lea ning sys em.
3.4.3 Model agg ega ion
Once he local model has been ained on each de ice, he model
pa ame e upda es a e sen o he cen al se e , which combines hem
using an agg ega ion p ocess. In ede a e lea ning, his is done using
he ede a ed a e aging algo i hm. This me hod allows he cen al
se e o combine model upda es wi hou accessing he o iginal
lea ne da a (Ren e al., 2024).
Ma hema ically, ede a ed a e aging can be exp essed as a
weigh ed a e age o he local model upda es, deno ed as he local
model upda es
∆i
w
om each de ice i as shown in Equa ion (15):
=
∆= ∆
∑
1
1N
i
i
ww
N
(15)
whe e N is he o al numbe o de ices pa icipa ing in he
aining, he cen al se e calcula es he weigh ed a e age o he
pa ame e upda es
∆
i
w
and ine- unes he global model, which is
hen dis ibu ed back o he de ices o con inue he aining p ocess.
This local aining and ede a ed agg ega ion p ocess enables
con inuous imp o emen o he emo ion de ec ion model wi hou
equi ing cen aliza ion o da a. I ensu es ha lea ne s’ p i acy is
p ese ed while he model con inues o lea n collec i ely. The esul is a
mo e accu a e and obus global model ha can de ec emo ions in eal-
ime, wi h sensi i e da a ne e sha ed ou side local de ices.
3.5 E alua ing model p ecision
Se e al s anda d me ics a e used in machine lea ning o e alua e
he pe o mance o he emo ion de ec ion model. These me ics a e
essen ial o unde s anding how he model iden i ies emo ions, bo h
in e ms o p ecision and ecall, and o gaining a comp ehensi e iew
o i s pe o mance. The me ics used in his s udy a e as ollows.
P ecision: P ecision measu es he p opo ion o co ec
p edic ions o a posi i e class (e.g., he “happy” emo ion) among all
p ojec ions o ha class. Ma hema ically, i is exp essed as:
P ecision is de ined as shown in Equa ion (16):
=
+
TP
P ecision
TP FP
(16)
whe e:
• TP (T ue Posi i es) a e he co ec p edic ions o he posi i e class.
• FP (False Posi i es) a e he inco ec p edic ions o he
posi i e class.
Recall measu es he abili y o he model o de ec all posi i e
ins ances (speci ic emo ions) in he da a. I is calcula ed as shown in
Equa ion (17):
=
+
TP
Recall
TP FN
(17)
whe e:
• FN (False Nega i es) a e he posi i e ins ances ha he model
inco ec ly classi ied as nega i e.
Gu ié ez e al. 10.3389/ ai.2025.1644844
F on ie s in A i icial In elligence 16 on ie sin.o g
Vi ual en i onmen s demons a e excellen p ecision and
s abili y, sugges ing ha a wholly digi al en i onmen , wi hou ace-
o- ace in e ac ion, enables he model o wo k wi h mo e consis en
da a. In compa ison, hyb id en i onmen s, which combine in-pe son
and i ual in e ac ions, p esen mo e a iabili y in esul s, likely due
o social in e ac ions and a ia ions in non e bal communica ion,
which in oduce addi ional noise in o he emo ion de ec ion p ocess.
4.6 Impac o he sys em on s uden
beha io
The emo ion de ec ion sys em implemen ed in he educa ional
en i onmen has a signi ican impac on s uden s’ emo ional and
academic beha io . The esul s demons a e how he de ec ed
emo ions impac s uden s’ academic engagemen and pe o mance
in educa ional ac i i ies. In Figu e7, h ee g aphs clea ly illus a e
how he de ec ed emo ions impac a ious aspec s o s uden
beha io . Figu e 7A shows he empo al e olu ion o s uden s’
engagemen and academic pe o mance be o e and a e ecei ing
emo ional eedback. A gene al imp o emen in bo h pa ame e s is
obse ed a e eedback, especially in hose s uden s wi h posi i e
emo ions, such as mo i a ion. Howe e , he a iabili y o he esul s
sugges s ha nega i e emo ions, such as s ess and us a ion, ha e
an une en impac on academic beha io , esul ing in less
consis en ou comes.
Figu e7B analyzes he ela ionship be ween he de ec ed emo ions
and he le els o academic engagemen . The esul s indica e ha
posi i e emo ions, such as mo i a ion, a e associa ed wi h ema kably
high le els o engagemen . In con as , emo ions such as s ess and
FIGURE5
E olu ion o model p ecision du ing he dynamic i ing p ocess.
FIGURE6
Model pe o mance analysis in eal-wo ld condi ions. (A) De ice compa ison. (B) Model p ecision o complex emo ions. (C) Pe o mance compa ison
in i ual s. hyb id en i onmen s.

Gu ié ez e al. 10.3389/ ai.2025.1644844
F on ie s in A i icial In elligence 17 on ie sin.o g
us a ion a e ela ed o lowe engagemen in academic ac i i ies.
This beha io highligh s he di ec impac o emo ions on s uden s’
deg ee o engagemen .
Figu e7C illus a es he impac o de ec ed emo ions on s uden s’
academic pe o mance. The esul s e lec ha posi i e emo ions a e
s ongly associa ed wi h be e g ades, while nega i e emo ions, such
as s ess and us a ion, con ibu e o lowe academic pe o mance.
This analysis unde sco es he impo ance o posi i e emo ions in
o e all academic pe o mance and highligh s he need o mi iga e he
nega i e impac o emo ions.
The abula ed da a o e s a quan i a i e analysis ha complemen s
he esul s displayed in he cha s. Table6 de ails he le els o academic
engagemen acco ding o he de ec ed emo ions. Posi i e emo ions,
such as mo i a ion, a e associa ed wi h highe engagemen in
educa ional ac i i ies. In con as , emo ions such as s ess and
us a ion a e associa ed wi h signi ican ly lowe le els o
pa icipa ion, indica ing ha hese emo ions nega i ely impac he
deg ee o s uden in ol emen in educa ional ac i i ies.
To p o ide a mo e g anula unde s anding o he ela ionship
be ween emo ional s a es and academic ou comes, Figu e8 in oduces
wo addi ional isualiza ions. These g aphs complemen he indings
p esen ed in Figu e7 by disagg ega ing he da a a he s uden le el
and explo ing he dis ibu ion o academic pe o mance ac oss
de ined sco e b acke s and emo ional ca ego ies.
Figu e8A p esen s a sca e plo o academic pe o mance o each
s uden , g ouped by he dominan emo ion de ec ed. This
indi idualized analysis e eals ha s uden s who a e consis en ly
mo i a ed achie e high sco es, clus e ing be ween 80 and 100%. In
con as , s uden s unde emo ional s a es such as s ess, anxie y, o
us a ion display mo e dispe sed ou comes, wi h us a ion being
mos associa ed wi h sco es below 70%. This isualiza ion con i ms he
gene al ends obse ed in he agg ega ed esul s, while also
highligh ing ou lie s and in e -indi idual a iabili y, which emphasizes
he impo ance o emo ional p o iling o pe sonalized in e en ions.
Figu e8B p o ides a his og am o pe o mance dis ibu ion, whe e
s uden s a e g ouped in o sco e b acke s (e.g., 40–49, 50–59, …,
90–100) and ca ego ized by emo ional s a e. This analysis e eals a high
concen a ion o mo i a ed s uden s in he op wo b acke s (80–89 and
90–100), while us a ed s uden s a e p ima ily ound in he 50–69
ange. The anxie y and s ess g oups p esen a b oade dis ibu ion,
ein o cing he no ion o emo ional he e ogenei y in academic
con ex s. The his og am o e s a equency-based pe spec i e,
suppo ing he in e p e a ion ha posi i e emo ions no only imp o e
pe o mance a e ages bu also educe a iabili y in s uden ou comes.
I is essen ial o cla i y ha he da ase used in Figu e8 encompasses
a b oade academic popula ion (N = 150) han he g oup in ol ed in
he sys em’s ield alida ion (N = 58). While he 58-s uden subse was
used o e alua e he sys em’s e ec i eness in a eal deploymen , he
ex ended analysis in Figu e8 was designed o assess he a iabili y and
dis ibu ion o academic pe o mance ac oss di e en emo ional
p o iles. This allows o a mo e obus s a is ical explo a ion o how
dis inc emo ional s a es co ela e wi h pe o mance b acke s, wi hou
con lic ing wi h he empi ical alida ion phase.
Table 7 p esen s he esul s ela ed o academic pe o mance
acco ding o he emo ions de ec ed. S uden s who expe ience posi i e
emo ions end o exhibi highe academic pe o mance han hose
who ace nega i e emo ions, such as s ess and us a ion.
Fu he mo e, g ade a iabili y, measu ed h ough s anda d de ia ion,
is highe in s uden s wi h nega i e emo ions, sugges ing mo e di e se
esponses in his g oup. This highligh s he need o a ge ed
in e en ions o suppo s uden s expe iencing hese complex
emo ions and enhance hei academic pe o mance.
To assess he p ac ical e ec o he emo ion de ec ion sys em on
s uden s’ engagemen and academic pe o mance, a compa ison was
conduc ed be ween pa icipa ion eco ds and academic ou comes
egis e ed be o e and a e he sys em’s implemen a ion. Speci ically,
in e ac ion logs om Moodle and educa ional eco ds om he
FIGURE7
Impac o de ec ed emo ions on academic beha io . (A) Time e olu ion o pa icipa ion and academic pe o mance. (B) Rela ionship be ween
emo ions and le els o educa ional pa icipa ion. (C) Rela ionship be ween emo ions and academic pe o mance.
TABLE6 Le els o academic pa icipa ion acco ding o he emo ions
de ec ed.
Emo ion Pa icipa ion
le el (%)
A e age pa icipa ion
pe s uden (%)
S ess 60% 62%
Anxie y 65% 63%
F us a ion 55% 58%
Mo i a ion 85% 82%
Gu ié ez e al. 10.3389/ ai.2025.1644844
F on ie s in A i icial In elligence 18 on ie sin.o g
ins i u ional pla o m we e analyzed o e wo equi alen academic
pe iods, each co e ing a whole semes e .
The esul s show a 15% ela i e inc ease in a e age academic
engagemen , measu ed by he equency o s uden in e ac ions in
o ums, assignmen submissions, and eedback eques s. Likewise, he
a e age academic pe o mance, based on inal cou se g ades,
imp o ed by 12% a e he in eg a ion o he emo ion-awa e eedback
sys em. These indings sugges a posi i e shi in bo h beha io al and
academic ou comes associa ed wi h he sys em’s deploymen .
Addi ionally, s uden s’ esponses o he adap ed TPA ques ionnai e
(Jian e al., 2000; Scha owski e al., 2025) e ealed highly pe cei ed
use ulness (M = 4.2, SD = 0.6), eliabili y (M = 4.1, SD = 0.7), and
p i acy con idence (M = 4.4, SD = 0.5). These esul s sugges ha he
sys em’s ede a ed design made a posi i e con ibu ion o i s
accep ance and usabili y. The a o able pe cep ion o p i acy
p o ec ion likely enhanced s uden s’ us in he sys em, ein o cing
hei engagemen and ecep i eness o emo ion-awa e eedback
du ing he academic pe iod.
While he imp o emen s obse ed in s uden engagemen and
academic pe o mance a e s ongly aligned wi h he implemen a ion
o he emo ion-awa e eedback sys em, i is essen ial o acknowledge
ha isola ing he speci ic con ibu ion o he ede a ed lea ning
componen emains a me hodologically complex ask. Howe e , he
success ul deploymen o he p i acy-p ese ing sys em in a eal
educa ional se ing, wi hou deg ading pe o mance o usabili y,
ein o ces he p ac ical iabili y o ou app oach. These indings
sugges ha p i acy-p ese ing, eal- ime emo ional eedback can
e ec i ely suppo s uden engagemen and lea ning ou comes, e en
in he e ogeneous de ice en i onmen s. The e alua ion o such
in eg a ed sys ems o e mo e ex ended pe iods and in mo e di e se
lea ning con ex s will bekey o u he con i ming hese bene i s.
Table8 summa izes he compa ison.
4.7 Compa ison wi h o he emo ion
de ec ion models
Compa ing he p oposed model and o he exis ing app oaches o
emo ion de ec ion is essen ial o highligh i s ad an ages and a eas o
imp o emen . The pe o mance analysis is based on p ecision, ecall,
and F1-sco e me ics, compa ing ou ede a ed lea ning-based
app oach wi h cen alized models such as DeepFace and hyb id
ans o me -based sys ems. Rega ding p i acy, we e alua e how
cen alized models ely on ans e ing sensi i e da a, whe eas ou
p oposal p ocesses da a locally. Fu he mo e, scalabili y is analyzed
based on he sys em’s abili y o manage la ge s uden popula ions
wi hou comp omising pe o mance, highligh ing he lexibili y o ou
solu ion o adap o pla o ms such as Moodle.
Table 9 summa izes he p oposed model’s main ea u es and
esul s in compa ison o well-known sys ems, including DeepFace,
hyb id ans o me -based models, and he comme cial A ec i a SDK
sys em. The able includes c i ical pe o mance me ics, as well as
aspec s o p i acy, scalabili y, and ease o in eg a ion in o educa ional
pla o ms. In e ms o pe o mance, he p oposed model shows
compe i i e esul s in p ecision, ecall, and F1-sco e, app oaching he
alues ob ained by sys ems such as DeepFace. Howe e , i ou pe o ms
cen alized models by main aining da a p i acy and a oiding he
ans e o da a o cen al se e s. Fu he mo e, i s ede a ed app oach
enables g ea e scalabili y, allowing i o handle la ge s uden
popula ions wi hou signi ican pe o mance deg ada ion.
Rega ding in eg a ion, he p oposed model s ands ou o i s
abili y o in eg a e di ec ly wi h he pla o m. One such ea u e is
Moodle, which is no ye ully a ailable in comme cial sys ems such
as he A ec i a SDK. This acili a es i s adop ion in educa ional
FIGURE8
Rela ionship be ween emo ional s a es and academic pe o mance. (A) Indi idual academic pe o mance g ouped by p edominan emo ion.
(B) F equency o s uden s by g ade ange acco ding o de ec ed emo ion.
TABLE7 Academic pe o mance acco ding o he emo ions de ec ed.
Emo ion A e age a ing
(%)
A e age
de ia ion (%)
S ess 70% 7%
Anxie y 75% 6%
F us a ion 65% 8%
Mo i a ion 85% 5%
Gu ié ez e al. 10.3389/ ai.2025.1644844
F on ie s in A i icial In elligence 19 on ie sin.o g
en i onmen s, educing con igu a ion cos s and adap ing o he
speci ic needs o ins i u ions.
5 Discussion
The esul s ob ained in his s udy show ha he ede a ed lea ning-
based model o emo ion de ec ion p esen s signi ican ad an ages in
e ms o p i acy, scalabili y, and pe o mance, aligning wi h exis ing
li e a u e in se e al key aspec s. Compa ed o DeepFace and hyb id
ans o me -based models, ou app oach achie es compe i i e me ics
o p ecision (0.87), ecall (0.85), and F1-sco e (0.86) while main aining
da a p i acy by p ocessing i locally on use s’ de ices. This inding is
consis en wi h hose o Kai ouz e al. (2021), who demons a ed ha
ede a ed lea ning could p ese e p i acy wi hou comp omising
pe o mance. Howe e , he obse ed a iabili y in model pe o mance
unde di e en condi ions, such as changes in illumina ion and
ambien noise, sugges s ha adap ing he sys em o dynamic
en i onmen s emains challenging, as poin ed ou by p e ious wo k
on cen alized models by Saha e al. (2024).
The me hodological p ocess in ol ed a design ha combined
emo ional da a collec ion using mul iple modali ies (images, audio, and
ex ) wi h p ep ocessing o anonymize he da a be o e local aining. This
app oach add esses he need o p o ec use s’ iden i ies and op imizes he
quali y o he p ocessed da a (Ribei o Junio and Kamienski, 2024).
Implemen ing he ede a ed model enabled he upda ing o pa ame e s
on local de ices wi hou ans e ing sensi i e in o ma ion,
demons a ing i s iabili y in educa ional en i onmen s wi h high
p i acy s anda ds. Howe e , he eliance on de ices wi h he e ogeneous
capabili ies in oduced challenges in pe o mance uni o mi y,
pa icula ly on pla o ms wi h limi ed esou ces. This p oblem is
inhe en o ede a ed models and has been iden i ied in he li e a u e as
an a ea equi ing u he op imiza ion (B iguglio e al., 2024).
In p ac ical e ms, he in eg a ion wi h Moodle acili a ed eal-wo ld
adop ion and assessmen o he model. The esul s ob ained in he hyb id
lea ning en i onmen , encompassing bo h ace- o- ace and online
classes, demons a e ha posi i e emo ions, such as mo i a ion, a e
associa ed wi h inc eased engagemen and imp o ed academic
pe o mance. Fu he mo e, he p oposed app oach add esses a c i ical
need in emo ion de ec ion: he abili y o ope a e on a scale wi hou
comp omising s uden s’ p i acy. This ep esen s a signi ican
imp o emen o e comme cial sys ems such as A ec i a SDK, which,
al hough e icien in e ms o pe o mance, do no o e he same da a
p o ec ion o cus omized in eg a ion wi h educa ional pla o ms (Kulke
e al., 2020). This ad ance has di ec implica ions o he design and
deploymen o scalable and e hically esponsible educa ional echnologies.
Despi e i s con ibu ions, he wo k p esen s limi a ions ha mus
bediscussed o con ex ualize he indings app op ia ely. One o he
main es ic ions is he dependence on he quali y o he de ices he
s uden s use. Al hough ede a ed lea ning is highly scalable, i s
pe o mance can be a ec ed by de ices wi h limi ed p ocessing
capabili ies, pa icula ly in e ms o la ency and p ecision. This ac o
could bias he esul s in popula ions wi h unequal access o echnology,
posing equi y challenges in implemen ing he sys em ac oss di e en
educa ional ins i u ions (Mohapa a e  al., 2024). Fu he mo e,
al hough he model main ains high le els o p i acy by p ocessing da a
locally, a iabili y in he quali y o ne wo k connec ions could in luence
he e ec i eness o model upda es agg ega ed a he cen al se e ,
especially in en i onmen s wi h inconsis en ne wo k in as uc u e.
Addi ionally, al hough he de ices used—sma phones, able s,
and lap ops—we e he e ogeneous and e lec ed ypical s uden
ha dwa e, no s a i ied benchma king was pe o med o assess
model beha io ac oss di e en de ice ypes. The sys em was
designed o beligh weigh and pla o m-independen ; howe e ,
a ia ions in CPU, memo y, o senso esolu ion could ha e
in oduced mino disc epancies in in e ence ime o p edic ion
accu acy. Fu u e wo k should include pe o mance audi s ac oss
de ice ca ego ies o be e unde s and and op imize eal-wo ld
deploymen s in di e se educa ional se ings. Finally, an addi ional
key limi a ion o his s udy is ha use pe cep ions ega ding he
abili y o he ede a ed model o p ese e hei p i acy we e no
e alua ed. While he echnical design ensu es ha sensi i e da a
TABLE8 Academic indica o s be o e and a e sys em deploymen .
Indica o Be o e implemen a ion A e implemen a ion Rela i e change
A g. engagemen sco e (%) 67.5 77.6 +15.0%
A g. academic pe o mance (%) 71.2 79.7 +12.0%
S d. de . o pe o mance 7.4 6.1 —
The engagemen sco e was compu ed om no malized in e ac ion me ics ( o um ac i i y, ask comple ion, and ime-on- ask). Pe o mance da a we e d awn om inal cou se g ades in bo h
e ms.
TABLE9 Compa ison o he p oposed ede a ed model wi h o he emo ion de ec ion sys ems.
Fea u e P oposed model
( ede a ed)
DeepFace
(cen alized)
T ans o me -based
hyb id
A ec i a SDK
(comme cial)
P ecision 87% 90% 85% 88%
Recall 85% 88% 84% 86%
F1-sco e 86% 89% 84.5% 87%
P i acy Local da a p o ec ed Cen alized da a Pa ially localized da a Cen alized da a
Scalabili y Highly adap able o medium-
o la ge-sized popula ions
Low, limi ed o cen alized
en i onmen s
Mode a e, compu a ionally
dependen Low, designed o small g oups
Ease o in eg a ion wi h LMS Di ec in eg a ion wi h Moodle Requi es ad anced con igu a ion Pa ial suppo No speci ied
Gu ié ez e al. 10.3389/ ai.2025.1644844
F on ie s in A i icial In elligence 20 on ie sin.o g
emains on local de ices, he ac ual us and accep ance o such
mechanisms by s uden s and educa o s emain unexplo ed. Fu u e
wo k will inco po a e use -cen e ed s udies o assess hese
pe cep ions, complemen ing echnical alida ion wi h empi ical
e idence o usabili y and us wo hiness.
Ano he signi ican limi a ion is he sys em’s sensi i i y o ad e se
condi ions, such as sudden changes in ligh ing o high ambien noise
le els. Al hough he p ep ocessing me hods imp o ed he sys em’s
obus ness agains hese condi ions, less common emo ions, such as
us a ion o anxie y, p esen ed highe e o a es in hei de ec ion.
This could bedue o insu icien examples o hese emo ions in he
da ase used o ini ial aining, a p oblem widely documen ed in he
li e a u e on emo ion de ec ion (Wang e al., 2024b). Add essing his
limi a ion will equi e expanding he da ase o include mo e
ep esen a i e examples o hese emo ions and implemen ing ans e
lea ning echniques o imp o e he model’s gene aliza ion.
F om a me hodological pe spec i e, he model assumes ha
emo ions de ec ed in educa ional ac i i ies a e consis en wi h
s uden s’ emo ional s a es. Howe e , his assump ion migh bein alid,
as ex e nal ac o s un ela ed o he academic en i onmen may
in luence he emo ional exp essions de ec ed. This aspec may
in oduce biases in in e p e ing he esul s, especially i used o assess
s uden s’ emo ional well-being o pe sonalize educa ional eedback.
Mi iga ing his p oblem will equi e a mo e holis ic app oach ha
combines emo ion de ec ion wi h o he con ex ual me ics, such as
cogni i e load o social in e ac ion.
The s udy demons a es he iabili y and po en ial o ede a ed
emo ion de ec ion sys ems o eal-wo ld educa ional se ings, while
iden i ying key challenges ha mus beadd essed o widesp ead
implemen a ion. The indings p o ide a ounda ion o u u e esea ch
o imp o e he obus ness, equi y, and con ex ual awa eness o
emo ion-awa e lea ning echnologies.
The esul s show a 15% ela i e inc ease in a e age academic
engagemen , measu ed by he equency o s uden in e ac ions in
o ums, assignmen submissions, and eedback eques s. Likewise, he
a e age academic pe o mance, based on inal cou se g ades,
imp o ed by 12% a e he in eg a ion o he emo ion-awa e eedback
sys em. Howe e , i is essen ial o no e ha hese indings a e based on
desc ip i e analysis. No s a is ical signi icance es s, such as eg ession
models o pai ed hypo hesis es s, we e applied o de e mine whe he
hese di e ences a e s a is ically signi ican o a ibu able solely o he
sys em’s deploymen . As such, he esul s sugges a po en ial posi i e
shi in bo h beha io al and academic ou comes, bu do no es ablish
a causal ela ionship. Fu u e wo k should include in e en ial s a is ical
analysis o alida e he obse ed imp o emen s.
The success ul in eg a ion o he sys em in o Moodle highligh s
i s p ac ical applicabili y and o e s se e al implica ions o la ge-scale
educa ional deploymen . Unlike comme cial sys ems such as A ec i a
SDK (Kulke e al., 2020), which o en equi e p op ie a y en i onmen s
and lack educa ional cus omiza ion, ou open a chi ec u e acili a es
di ec alignmen wi h exis ing lea ning pla o ms. Gi en i s p i acy-
p ese ing design and low compu a ional equi emen s, he sys em
can beadop ed in ins i u ions wi h a ied in as uc u e le els wi hou
signi ican echnical cons ain s. Howe e , e ec i e implemen a ion
depends on ins i u ional policies and he eadiness o educa o s. P io
s udies (Huang e al., 2023) he e is a need o add ess eache aining
in in e p e ing emo ional analy ics and o es ablish go e nance
amewo ks ha egula e e hical use. Ou indings emphasize ha
eache capaci y-building and policy alignmen a e necessa y o
ansla e emo ional insigh s in o meaning ul pedagogical ac ions.
Mo eo e , unlike ans o me -based sys ems (Teng e al., 2024) ou
sys em suppo s decen alized scalabili y, which is pa icula ly
bene icial o applica ions ha equi e ex ensi e cloud esou ces,
sugges ing easibili y o na ional o mul i-ins i u ional deploymen s
wi h minimal cos and s ong alignmen o educa ional alues.
6 Conclusions and u u e wo k
This s udy demons a es ha a ede a ed lea ning-based app oach
o emo ion de ec ion is bo h e ec i e and p ac ical in educa ional
en i onmen s. The model achie ed high p ecision, ecall, and F1-sco e
alues while p ese ing s uden da a p i acy and enabling scalabili y.
I s in eg a ion in o Moodle con i med ha he sys em can ope a e in
eal academic se ings wi h minimal ic ion, o e ing eal- ime
emo ional eedback wi hou comp omising con iden iali y.
The implemen a ion showed a measu able impac on s uden
ou comes: s uden s whose posi i e emo ions we e de ec ed and
esponded o exhibi ed a 15% inc ease in academic engagemen and
a 12% imp o emen in pe o mance. These esul s suppo he
e ec i eness o he app oach and i s alue as a ool o emo ionally
adap i e lea ning.
Howe e , he sys em s ill aces limi a ions. I s pe o mance
depends on he he e ogenei y o use de ices, and de ec ing complex
emo ions like us a ion o anxie y emains challenging due o hei
unde ep esen a ion in he da ase . These cons ain s did no
comp omise he sys em’s iabili y, bu ins ead highligh ed a eas o
u u e imp o emen .
Fu u e wo k will ocus on op imizing he model o low- esou ce
de ices and inco po a ing syn he ic da a and ans e lea ning
echniques o enhance he di e si y o emo ions. Addi ionally,
explo ing u he beha io al and cogni i e indica o s will help e ine
emo ional in e ence and expand he sys em’s pedagogical impac .
Da a a ailabili y s a emen
The da a analyzed in his s udy is subjec o he ollowing licenses/
es ic ions: he da ase used in his s udy con ains sensi i e emo ional
in o ma ion collec ed om s uden s wi hin a uni e si y en i onmen .
Due o p i acy conside a ions and ins i u ional egula ions, he da ase is
no publicly a ailable. Access is es ic ed o au ho ized esea che s unde
da a-sha ing ag eemen s ha ensu e compliance wi h e hical guidelines
and p i acy laws. Reques s o da a access may beconside ed on a case-
by-case basis and equi e app o al om he ins i u ional e hics commi ee
and he da a con olle . Reques s o access hese da ase s should
bedi ec ed o william. il[email p o ec ed].
E hics s a emen
The s udies in ol ing humans we e app o ed by “Gami icación
educa i a po enciada po in eligencia a i icial”. File numbe
UA-2025-05-24. The s udies we e conduc ed in acco dance wi h
he local legisla ion and ins i u ional equi emen s. W i en
in o med consen o pa icipa ion was no equi ed om he
Gu ié ez e al. 10.3389/ ai.2025.1644844
F on ie s in A i icial In elligence 21 on ie sin.o g
pa icipan s o he pa icipan s’ legal gua dians/nex o kin
because Acco ding o Minis e ial Ag eemen No. 0005-2022 o he
Minis y o Public Heal h o Ecuado , w i en in o med consen
is no equi ed o s udies ha do no in ol e di ec in e en ion
on human beings, use o biological samples, pa icipa ion o
ulne able popula ions, o access o con iden ial pe sonal da a.
This s udy complied wi h all hese condi ions, as i was limi ed o
he use o anonymized emo ional da a collec ed h ough
non-in asi e educa ional in e ac ions.
Au ho con ibu ions
RG: Da a cu a ion, Fo mal analysis, In es iga ion,
Me hodology, So wa e, Valida ion, Visualiza ion, W i ing –
o iginal d a . WV-C: Concep ualiza ion, Fo mal analysis,
In es iga ion, Me hodology, Supe ision, Valida ion,
Visualiza ion, W i ing – e iew & edi ing. SL-M:
Concep ualiza ion, Supe ision, Valida ion, Visualiza ion,
W i ing– e iew & edi ing.
Funding
The au ho (s) decla e ha no inancial suppo was ecei ed o
he esea ch and/o publica ion o his a icle.
Con lic o in e es
The au ho s decla e ha he esea ch was conduc ed in he
absence o any comme cial o inancial ela ionships ha could
becons ued as a po en ial con lic o in e es .
Gene a i e AI s a emen
The au ho s decla e ha no Gen AI was used in he c ea ion o
his manusc ip .
Any al e na i e ex (al ex ) p o ided alongside igu es in his
a icle has been gene a ed by F on ie s wi h he suppo o a i icial
in elligence and easonable e o s ha e been made o ensu e accu acy,
including e iew by he au ho s whe e e possible. I youiden i y any
issues, please con ac us.
Publishe ’s no e
All claims exp essed in his a icle a e solely hose o he au ho s
and do no necessa ily ep esen hose o hei a ilia ed
o ganiza ions, o hose o he publishe , he edi o s and he
e iewe s. Any p oduc ha may bee alua ed in his a icle, o
claim ha may bemade by i s manu ac u e , is no gua an eed o
endo sed by he publishe .
Re e ences
Ajagbe, S. A., Adeniji, O. D., Olayiwola, A. A., and Abiona, S. F. (2024). Ad anced
enc yp ion s anda d (AES)-based ex enc yp ion o nea ield communica ion (NFC)
using Hu man comp ession. SN Compu . Sci. 5:156. doi: 10.1007/s42979-023-02486-6
Ala i, H., Sei i, M., Rouhollahei, M., Ra a i, M., and A ab a d, M. (2024). De elopmen
o local so wa e o au oma ic measu emen o geome ic pa ame e s in he p oximal
emu using a combina ion o a deep lea ning app oach and an ac i e shape model on
X- ay images. J. Imaging In o m. Med. 37, 633–652. doi: 10.1007/s10278-023-00953-3
Almalki, J., Alshah ani, S. M., and Khan, N. A. (2024). A comp ehensi e secu e sys em
enabling heal hca e 5.0 using ede a ed lea ning, in usion de ec ion and blockchain.
Pee J Compu . Sci. 10:e1778. doi: 10.7717/pee j-cs.1778
An, Y., Lee, J., Bak, E. S., and Pan, S. (2023). Deep acial emo ion ecogni ion using
local ea u es based on acial landma ks o a secu i y sys em. Compu . Ma e . Con in.
76:1817–1832. doi: 10.32604/cmc.2023.039460
Anand, M., and Babu, S. (2024). Mul i-class acial emo ion exp ession iden i ica ion
using DL-based ea u e ex ac ion wi h classi ica ion models. In . J. Compu . In ell. Sys .
17:2–17. doi: 10.1007/s44196-024-00406-x
Bilo i, U., Bisogni, C., De Ma sico, M., and T amon e, S. (2024). Mul imodal emo ion
ecogni ion ia con olu ional neu al ne wo ks: compa ison o di e en s a egies on wo
mul imodal da ase s. Eng. Appl. A i . In ell. 130:1–13. doi:
10.1016/j.engappai.2023.107708
B iguglio, W., Youse , W. A., T ao e, I., and Mamun, M. (2024). Fede a ed supe ised
p incipal componen analysis. IEEE T ans. In . Fo ensics Secu . 19, 646–660. doi:
10.1109/TIFS.2023.3326981
Chen, Q., Ling, Z., Jiang, H., Zhu, X., Wei, S., and Inkpen, D. (2017). Enhanced LSTM
o na u al language in e ence, in ACL 2017– 55 h Annual Mee ing o he Associa ion
o Compu a ional Linguis ics, P oceedings o he Con e ence (Long Pape s). doi:
10.18653/ 1/P17-1152
Demszky, D., Mo sho i z-A ias, D., Ko, J., Cowen, A., Nemade, G., and Ra i, S.
(2020). GoEmo ions: a da ase o ine-g ained emo ions, in P oceedings o he Annual
Mee ing o he Associa ion o Compu a ional Linguis ics. doi:
10.18653/ 1/2020.acl-main.372
Di Da io, D., Pen angelo, V., Colella, M. I., Palomba, F., and G a ino, C. (2024).
Collec ing and implemen ing e hical guidelines o emo ion ecogni ion in an
educa ional me a e se, in Adjunc p oceedings o he 32nd ACM con e ence on use
modeling, adap a ion and pe sonaliza ion, (New Yo k, NY, USA: Associa ion o
Compu ing Machine y), 549–554. doi: 10.1145/3631700.3665232
Do iguzzi-Co in, R., and Si acusa, D. (2024). Flad: adap i e ede a ed lea ning o
DDoS a ack de ec ion. Compu . Secu . 137:103597. doi: 10.1016/j.cose.2023.103597
Ekman, P., Rosenbe g, E. L., and Helle , M. (1998). Wha he ace e eals. Basic and
applied s udies o spon aneous exp ession using he acial ac ion coding sys em (FACS).
Psycho hé apies 18, 179–180. doi: 10.1177/00030651221107681
Elisondo, R. C., Suá ez-Lan a ón, B., and Ga cía-Pe ales, N. (2024). Uni e si y
s uden s´ emo ions in o ced emo e educa ion. An explo a o y s udy in Spain and
A gen ina. Mul idiscip. J. Educ. Res. 14:59–78. doi: 10.17583/ emie.10251
Hammann, T., Schwa ze, M. M., Zen el, P., Schlomann, A., E en, C., Wahl, H. W.,
e al. (2022). The challenge o emo ions—an expe imen al app oach o assess he
emo ional compe ence o people wi h in ellec ual disabili ies. Disabili ies 2, 611–625.
doi: 10.3390/disabili ies2040044
Hamon, R., Junklewi z, H., Sanchez, I., Malgie i, G., and De He , P. (2022). B idging
he gap be ween AI and explainabili y in he GDPR: owa ds us wo hiness-by-design
in au oma ed decision-making. IEEE Compu . In ell. Mag. 17, 72–85. doi:
10.1109/MCI.2021.3129960
Hanisch, S., Tod , J., Pa ino, J., E ans, N., and S u e, T. (2024). A alse sense o p i acy:
owa ds a eliable e alua ion me hodology o he anonymiza ion o biome ic da a.
P oc. P i . Enhancing Technol. 2024, 116–132. doi: 10.56553/pope s-2024-0008
Hashmi, E., and Yayilgan, S. Y. (2024). Mul i-class ha e speech de ec ion in he
No wegian language using FAST-RNN and mul ilingual ine- uned ans o me s.
Complex In ell. Sys . 10, 4535–4556. doi: 10.1007/s40747-024-01392-5
Hellmann, F., Me es, S., Benouis, M., Hus inx, A., Hsieh, T.-C., Cona i, C., e al.
(2024). GANonymiza ion: a GAN-based ace anonymiza ion amewo k o p ese ing
emo ional exp essions. ACM T ans. Mul imedia Compu . Commun. Appl. 21, 1–27. doi:
10.1145/3641107
Huang, F., Yang, N., Chen, H., Bao, W., and Yuan, D. (2023). Dis ibu ed online mul i-
label lea ning wi h p i acy p o ec ion in in e ne o hings. Appl. Sci. 13:2713. doi:
10.3390/app13042713
Jian, J.-Y., Bisan z, A. M., and D u y, C. G. (2000). Founda ions o an empi ically
de e mined scale o us in au oma ed sys ems. In . J. Cogn. E gon. 4, 53–71. doi:
10.1207/S15327566IJCE0401_04
Jiang, W., Li, M., Shabaz, M., Sha ma, A., and Haq, M. A. (2023). Gene a ion o oice
signal one sandhi and melody based on con olu ional neu al ne wo k. ACM T ans.
Asian Low-Resou . Lang. In . P ocess. 22, 1–13. doi: 10.1145/3545569

Gu ié ez e al. 10.3389/ ai.2025.1644844
F on ie s in A i icial In elligence 22 on ie sin.o g
Joudeh, I. O., C e u, A. M., Boucha d, S., and Guimond, S. (2023). P edic ion o
con inuous emo ional measu es h ough physiological and isual da a. Senso s 23:5613.
doi: 10.3390/s23125613
Kai ouz, P., McMahan, H. B., A en , B., Belle , A., Bennis, M., Bhagoji, A. N., e al.
(2021). Ad ances and open p oblems in ede a ed lea ning. Found. T ends® Mach.
Lea n. 14, 1–210. doi: 10.1561/2200000083
Kaiss, W., Mansou i, K., and Poi ie , F. (2023). P e-e alua ion wi h a pe sonalized
eedback con e sa ional agen in eg a ed in Moodle. In . J. Eme g. Technol. Lea n. 18,
177–189. doi: 10.3991/ije . 18i06.36783
Ke man, N. T., Banihashem, S. K., Ka ami, M., E , E., an Ginkel, S., and No oozi, O.
(2024). Online pee eedback in highe educa ion: a syn hesis o he li e a u e. Educ. In .
Technol. (Do d ) 29, 763–813. doi: 10.1007/s10639-023-12273-8
Khu ana, Y., Gup a, S., Sa hya aj, R., and Raja, S. P. (2024). Robinne : a mul imodal
speech emo ion ecogni ion sys em wi h speake ecogni ion o social in e ac ions.
IEEE T ans. Compu . Soc. Sys . 11, 478–487. doi: 10.1109/TCSS.2022.3228649
Ko wal, K., Bha acha jee, S., Abbe , P., Mos aani, Z., Wei, H., Wenkang, X., e al.
(2022). Domain-speci ic adap a ion o CNN o de ec ing ace p esen a ion a acks in
NIR. IEEE T ans. Biom. Beha . Iden i y Sci. 4, 135–147. doi:
10.1109/TBIOM.2022.3143569
Kulke, L., Feye abend, D., and Schach , A. (2020). A compa ison o he A ec i a
iMo ions acial exp ession analysis so wa e wi h EMG o iden i ying acial exp essions
o emo ion. F on . Psychol. 11:1–9. doi: 10.3389/ psyg.2020.00329
Labidi, A., Ouamani, F., and Saoud, N. B. B. (2021). An on ology based ex app oach
o cul u e awa e emo ion mining: a Moodle plugin, in 2021 In e na ional Con e ence
o Women in Da a Science a Tai Uni e si y, WiDSTai 2021. doi:
10.1109/WIDSTAIF52235.2021.9430232
Lee, J. P., Jang, H., Jang, Y., Song, H., Lee, S., Lee, P. S., e al. (2024). Encoding o mul i-
modal emo ional in o ma ion ia pe sonalized skin-in eg a ed wi eless acial in e ace.
Na . Commun. 15:530. doi: 10.1038/s41467-023-44673-2
Lyu, S. (2023). Facial exp ession ecogni ion based on Mini_Xcep ion. Highl. Sci. Eng.
Technol. 39, 1178–1187. doi: 10.54097/hse . 39i.6726
Me i y, S., Keska , N. S., and Soche , R. (2018) Regula izing and op imizing LSTM
language models, in 6 h In e na ional Con e ence on Lea ning Rep esen a ions, ICLR
2018– Con e ence T ack P oceedings
Mish a, S. K., Kuma , N. S., Rao, B., B ahmend a, and Teja, L. (2024). Role o ede a ed
lea ning in edge compu ing: a su ey. J. Au on. In ell. 7:1–22. doi: 10.32629/jai. 7i1.624
Mohapa a, R. K., Jolly, L., Lyngdoh, D. C., Mou ya, G. K., Changaai Mangalo e, I. A.,
Alam, S. I., e al. (2024). A comp ehensi e su ey o s udy he u ili ies o image
segmen a ion me hods in clinical ou ine. Ne w. Modeling Anal. Heal h In o m.
Bioin o m. 13:2–26. doi: 10.1007/s13721-023-00436-z
Mollahosseini, A., Hasani, B., and Mahoo , M. H. (2019). A ec Ne : a da abase o
acial exp ession, alence, and a ousal compu ing in he wild. IEEE T ans. A ec .
Compu . 10, 18–31. doi: 10.1109/TAFFC.2017.2740923
Muk a, M. S. H., Ahmad, J., Zaman, A., and Islam, S. (2024). A en ion and me a-
heu is ic based gene al sel -e icacy p edic ion model om mul imodal social media
da ase . IEEE Access 12, 36853–36873. doi: 10.1109/ACCESS.2024.3373558
Mu awa, A. M., and Hassouneh, A. (2024). Mul imodal eal- ime pa ien emo ion
ecogni ion sys em using acial exp essions and b ain EEG signals based on machine lea ning
and log-sync me hods. Biomed. Signal P ocess. Con ol. 91:2–6. doi: 10.1016/j.bspc.2023.105942
Nandan, D., Kanungo, J., and Mahajan, A. (2024). An e o -e icien Gaussian il e
o image p ocessing by using he expanded ope and decomposi ion loga i hm
mul iplica ion. J. Ambien . In ell. Humaniz. Compu . 15:1045–1052. doi:
10.1007/s12652-018-0933-x
Ong, K. L., Lee, C. P., Lim, H. S., Lim, K. M., and Alqah ani, A. (2024). MaxMViT-
MLP: mul iaxis and mul iscale ision ans o me s usion ne wo k o speech emo ion
ecogni ion. IEEE Access 12, 18237–18250. doi: 10.1109/ACCESS.2024.3360483
Ped ycz, W. (2023). Ad ancing ede a ed lea ning wi h g anula compu ing. Fuzzy In .
Eng. 15, 1–13. doi: 10.26599/FIE.2023.9270001
Pi one, C., Va asi, S., Pla ania, G. A., and Cas ellano, S. (2021) Face- o- ace and
online lea ning: he ole o echnology in s uden s’ me acogni ion, in CEUR Wo kshop
P oceedings
Rai, R., and Gup a, S. (2022). A no el lea ning managemen sys em based on
mic ose ice a chi ec u e using Moodle. In . J. C ea . Res. Though s 10:d356–d363.
Ren, H., Anicic, D., and Runkle , T. A. (2023). TinyRep ile: TinyML wi h ede a ed
me a-lea ning, in P oceedings o he In e na ional Join Con e ence on Neu al
Ne wo ks. doi: 10.1109/IJCNN54540.2023.10191845
Ren, H., Deng, J., Xie, X., Ma, X., and Wang, Y. (2024). Fedboos ing: ede a ed lea ning
wi h g adien p o ec ed boos ing o ex ecogni ion. Neu ocompu ing 569:1–12. doi:
10.1016/j.neucom.2023.127126
Ribei o Junio , F. M., and Kamienski, C. A. (2024). Fede a ed lea ning o pe o mance
beha io de ec ion in a og-IoT sys em. In e ne Things 25:101078. doi:
10.1016/j.io .2024.101078
Saha, P. K., A ya, D., and Sekimo o, Y. (2024). Fede a ed lea ning–based global oad
damage de ec ion. Compu . Aided Ci . In . Eng. 39, 2223–2238. doi: 10.1111/mice.13186
San oso, B. E., and Kusuma, G. P. (2022). Facial emo ion ecogni ion on FER2013
using VGGSpinalNe . J. Theo . Appl. In . Technol. 100:2088–2102.
Sayeed, M. S., Mohan, V., and Mu hu, K. S. (2023). BERT: a e iew o applica ions in
sen imen analysis. HighTech Inno . J. 4, 453–462. doi: 10.28991/HIJ-2023-04-02-015
Scha owski, N., Pe ig, S. A. C., on Fel en, N., Aeschbach, L. F., Opwis, K.,
Win e sbe ge , P., e al (2025). To us o dis us AI: a ques ionnai e alida ion s udy,
in P oceedings o he 2025 ACM con e ence on ai ness, accoun abili y, and
anspa ency, (NewYo k, NY, USA: Associa ion o Compu ing Machine y), 361–374.
doi: 10.1145/3715275.3732025
Sengup a, D., Khan, S. S., Das, S., and De, D. (2024). FedEL: ede a ed educa ion
lea ning o gene a ing co ela ions be ween cou se ou comes and p og am ou comes
o in e ne o educa ion hings. In e ne Things 25:101056. doi: 10.1016/j.io .2023.101056
Shched ina, E., Valie , I., Sabi o a, F., and Babaskin, D. (2021). P o iding adap i i y
in Moodle LMS cou ses. In . J. Eme g. Technol. Lea n. 16:95. doi: 10.3991/ije . 16i02.18813
Shmelo a, T., Smolanka, V., Siki da, Y., and Sechko, O. (2024). Real- ime moni o ing
and diagnos ics o he pe son’s emo ional s a e and decision-making in ex eme
si ua ions o heal hca e. Decis. Mak. Anal.. 11–32. doi: 10.55976/dma.22024121911-32
Shwe Sin, T., and Khin, O. (2022). Facial exp essions classi ica ion on and oid
sma phone o a use , in Lec u e no es in ne wo ks and sys ems. doi:
10.1007/978-981-16-1781-2_21
Souali, K., Rahmaoui, O., and Ouzzi , M. (2023). Men o Bo : a aceabili y-based
ecommenda ion Cha bo o Moodle, in Lec u e no es in ne wo ks and sys ems. doi:
10.1007/978-3-031-26384-2_37
Subak i, A., Mu i, H., and Ha iadi, N. (2022). The pe o mance o BERT as da a
ep esen a ion o ex clus e ing. J. Big Da a 9:15. doi: 10.1186/s40537-022-00564-9
Takase, S., and Kiyono, S. (2023). Lessons on pa ame e sha ing ac oss laye s in
ans o me s, in P oceedings o he annual mee ing o he Associa ion o Compu a ional
Linguis ics. doi: 10.18653/ 1/2023.sus ainlp-1.5
Teng, S., Liu, J., Huang, Y., Chai, S., Ta eyama, T., Huang, X., e al. (2024). An in a-
and in e -emo ion ans o me -based usion model wi h homogeneous and di e se
cons ain s using mul i-emo ional audio isual ea u es o dep ession de ec ion. IEICE
T ans. In . Sys . E107.D, 342–353. doi: 10.1587/ ansin .2023HCP0006
Wang, X. (2024). Resea ch on he digi al p ese a ion o in angible cul u al he i age
o olk dance a ca ego y. Appl. Ma h. Nonlinea Sci. 9:1–16. doi:
10.2478/amns.2023.2.01500
Wang, R., Chen, T., Wang, J., Lai, J., Li, J., Zhang, M., e al. (2024a). In eg a ed
comp ehensi e analysis me hod o educa ion quali y wi h ede a ed lea ning, in
Communica ions in Compu e and In o ma ion Science 691. doi:
10.1007/978-981-99-9492-2_37
Wang, A., Meng, Z., Zhao, B., and Zhang, F. (2024). Using social media da a o
esea ch he impac o campus g een spaces on s uden s’ emo ions: a case s udy o
Nanjing campuses. Sus ainabili y 16. doi: 10.3390/su16020691
Wang, R., Song, P., and Zheng, W. (2024b). G aph-di usion-based domain-in a ian
ep esen a ion lea ning o c oss-domain acial exp ession ecogni ion. IEEE T ans.
Compu . Soc. Sys . 11, 4163–4174. doi: 10.1109/TCSS.2024.3355113
Wong, R. Y., Chong, A., and Coope Aspeg en, R. (2023). P i acy legisla ion as
business isks: how GDPR and CCPA a e ep esen ed in echnology companies’
in es men isk disclosu es. P oc. ACM on Hum.-Compu . In e ac . 7, 1–26. doi:
10.1145/3579515
Woodwa d, K., Kanjo, E., and Tsanas, A. (2024). Combining deep lea ning wi h
signal-image encoding o mul i-modal men al wellbeing classi ica ion. ACM T ans.
Compu . Heal hca e 5, 1–23. doi: 10.1145/3631618
Xu, B., Gu, X., Wang, J., Zhang, Y., Zhu, J., and Zhang, W. (2024). De o ma ion con ol
me hod o ac i e shape mo phing la ice s uc u e using opology op imiza ion
app oach. Sma Ma e . S uc . 33:025006. doi: 10.1088/1361-665X/ad1bab
Zhang, Y., Li, Z., Han, X., Ding, S., Li, J., Wang, J., e al. (2024). Pseudo-da a based
sel -supe ised ede a ed lea ning o classi ica ion o his opa hological images. IEEE
T ans. Med. Imaging 43, 902–915. doi: 10.1109/TMI.2023.3323540
Zhang, H., Wu, T., Cheng, S., and Liu, J. (2024). Cc- eda g: compu a ionally
cus omized ede a ed a e aging. IEEE In e ne Things J. 11, 4826–4841. doi:
10.1109/JIOT.2023.3300080
Zhou, J., Pal, S., Dong, C., and Wang, K. (2024). Enhancing quali y o se ice h ough
ede a ed lea ning in edge-cloud a chi ec u e. Ad Hoc Ne w. 156:103430. doi:
10.1016/j.adhoc.2024.103430