Bi e-Weigh Es ima ion Using Comme cial Ea Buds
Vasileios Papapanagio ou and S e anos Gano akis and Anas asios Delopoulos
Abs ac — While au oma ic acking and measu ing o ou
physical ac i i y is a well es ablished domain, no only in
esea ch bu also in comme cial p oduc s and e e y-day li e-
s yle, au oma ic measu emen o ea ing beha io is signi ican ly
mo e limi ed. Despi e he abundance o me hods and algo i hms
ha a e a ailable in bibliog aphy, comme cial solu ions a e
mos ly limi ed o digi al logging applica ions o sma -phones.
One ac o ha limi s he adop ion o such solu ions is ha hey
usually equi e specialized ha dwa e o senso s. Based on his,
we e alua e he po en ial o es ima ing he weigh o consumed
ood (pe bi e) based only on he audio signal ha is cap u ed by
comme cial ea buds (Samsung Galaxy Buds). Speci ically, we
examine a combina ion o ea u es (bo h audio and non-audio
ea u es) and ainable es ima o s (linea eg ession, suppo
ec o eg ession, and neu al-ne wo k based es ima o s) and
e alua e on an in-house da ase o 8 pa icipan s and 4 ood
ypes. Resul s indica e good po en ial o his app oach: ou
bes esul s yield mean absolu e e o o less han 1 g o 3 ou
o 4 ood ypes when aining ood-speci ic models, and 2.1 g
when aining on all ood ypes oge he , bo h o which imp o e
o e an exis ing li e a u e app oach.
I. INTRODUCTION
In he con ex o die a y moni o ing, a ious wea able
senso s ha e been p oposed in o de o measu e di e en
pa ame e s o ea ing beha io . One o he i s senso s ha
was used is he in-ea mic ophone: he in-ea placemen
enables he cap u ing chewing senso s clea ly as hey a e
ansmi ed h ough he skull du ing mas ica ion [1].
Al e na i e senso s ha e also been s udied in li e a u e.
A piezoelec ic senso has been used in [2]; he senso is
a ached on he skin close o he jaw ha cap u es mus-
cle mo emen du ing mas ica ion. The pe iodic na u e o
chewing is also p esen in he piezoelec ic senso ’s signal
and is used o de ec chewing. Al e na i e placemen s o
he piezoelec ic senso ha e also been examined, such as
a ached o sma glasses o o neck colla s [3] [4]. Su ace
elec omyog aphy (EMG) has also been used o chewing
de ec ion [5] [6], howe e , is cu en ly one o he leas
disc e e solu ions.
Senso s o es ima ing he weigh o a meal (o bi e) ha e
also been p oposed, and achie e ela i ely high e ec i eness
and low e o s. Howe e , he senso s equi e manual place-
men and ac i a ion (e.g. pla e weigh scale [7]) o a e pa o
Au ho s a e wi h he Mul imedia Unde s anding G oup, Dp . o
Elec ical and Compu e Enginee ing, Facul y o Enginee ing, A is o le
Uni e si y o Thessaloniki, G eece [email p o ec ed],
[email p o ec ed],[email p o ec ed]
© 2021 IEEE. Pe sonal use o his ma e ial is pe mi ed. Pe mission om
IEEE mus be ob ained o all o he uses, in any cu en o u u e media,
including ep in ing/ epublishing his ma e ial o ad e ising o p omo ional
pu poses, c ea ing new collec i e wo ks, o esale o edis ibu ion o
se e s o lis s, o euse o any copy igh ed componen o his wo k in
o he wo ks.
he able [8] and hus canno be used in ee-li ing condi ions
such as ea ing ou side o on- he-go.
Mo e ecen ly, he in e es has been shi ing o o - he-
shel solu ions o elimina e he need o specialized ha dwa e
as well as dec ease he senso s’ in usi eness. In pa icula ,
he 3D accele ome e s and gy oscopes ha a e commonly
embedded in comme cial sma -wa ches can be used o
ambien ly de ec ea ing ges u es (i.e. he epea ed mo emen s
o b inging ood o he mou h om a pla e, ay, e c) and
achie e e y p omising esul s in challenging, ee-li ing
condi ions [1] [9] [10] [11]. Al e na i ely, an accele ome e
moun ed on he empo alis [12], [13] has also been used o
de ec muscle con ac ion du ing mas ica ion wi h p omising
e ec i eness.
Addi ionally, analysis o pho os aken wi h sma phones
can p o ide de ailed in o ma ion abou ea ing habi s, includ-
ing ypes o consumed ood, ing edien s, e c ( o example,
he goFOODTM [14] sys em can es ima e he calo ie and
mac o-nu ien con en o a meal based on ei he wo pho os
o he meal o a sho ideo). A single pho o is used in [15]
o pe o m segmen a ion, ecogni ion, and olume es ima ion
o di e en oods, and esul s show simila e ec i eness o
me hods ha equi e mul iple pho os o he meal.
In his wo k, we p opose a me hod o es ima ing bi e
weigh using he audio signal o comme cially a ailable ea
buds. Ou app oach includes ex ac ing ea u es ha a e used
o ain bi e weigh es ima o s, based on anno a ions o s a
and s op ime-s amps o chews and ood ype. We e alua e
di e en ea u e se s and di e en ypes o es ima o s on an
in-house da ase we ha e collec ed, using lea e-one-subjec -
ou (LOSO) aining and es ing. We examine wo cases, one
whe e ood ype in o ma ion is a ailable (co esponding o
a use-case whe e ood in o ma ion is ob ained by asking he
use di ec ly o by some ood- ype ecogni ion sys em such
as [14]) and one whe e i is no (co esponding o a use-case
whe e only chewing ac i i y is de ec ed using some audio-
based au oma ed me hod such as [16]).
II. RELATED WORK
An algo i hm o bi e-weigh es ima ion, also om sound
cap u ed by an in-ea mic ophone, has been p oposed in [17].
Audio was eco ded a 44 kHz; a o al o 8.64 hou s we e
eco ded om eigh indi iduals.
The algo i hm p oposed in [17] used 8 ea u es (Table I
o [17]) ha can be ex ac ed om a sequence o chews.
O hese 8 ea u es, 7can be compu ed solely om he s a
and s op ime-s amps o he chews, and only 1 equi es audio
signal (i.e. mean signal ene gy). Fo each chewing bi e, hese
8 ea u es a e ex ac ed 6 imes: om he en i e chewing
a Xi :2108.00771 1 [eess.AS] 2 Aug 2021
0 0.5 1 1.5
(s)
-0.5
0
0.5
Fig. 1: Example o audio signal: he i s h ee chews o
an apple bi e a e down, along wi h he manual g ound- u h
(g ay boxes).
bou , om he 1s , 2nd, and 3 d hi d o he chewing bou ,
and om he chewing bou s ha consis o he i s 3and 5
chews only, espec i ely. This yields 48 ea u es; wo mo e
ea u es a e compu ed and a inal ec o o 50 ea u es is
p oduced.
A linea eg ession (wi h bias) model is used o es ima e
bi e weigh . A di e en model is ained o each ood ype
(a o al o h ee ood ypes a e used: po a o chip, le uce,
and apple). Two me hods o ea u e selec ion a e examined:
manual selec ion based on Spea man’s co ela ion coe icien
(be ween each ea u e and he bi e-weigh ) and s ep-wise
eg ession i . Au ho s conclude ha bo h me hods yield
simila esul s.
In his wo k, we di e en ia e signi ican ly om [17] by
(a) using comme cially a ailable ea buds, (b) ocusing
on audio-based ea u es and explo ing di e en agg ega ion
me hods, and (c) compa ing di e en eg ession models o
bi e weigh es ima ion.
III. BITE-WEIGHT ESTIMATION ALGORITHM
Ou p oposed app oach aims o es ima e he weigh o a
single bi e. In sho , we i s ex ac a se o ea u es (we
examine bo h non-audio and audio ea u es) which we hen
use o ain an es ima ion model. Models a e ained in he
ypical LOSO scheme, whe e a di e en model is ained
o each subjec o he da ase using he da a om he o he
subjec s each ime.
A. Fea u e ex ac ion
We use wo dis inc se s o ea u es. The i s se (non-
audio ea u es) does no depend on he audio signal, bu only
on he s a and s op ime-s amps o he chews (Figu e 1).
No e ha in his wo k he s a and s op ime-s amps ha e
been de e mined manually. Speci ically, le 1[i]and 2[i]
o i= 1, . . . , n deno e he s a and s op ime-s amps o
a bou o nchews. We compu e he ollowing six ea u es:
numbe o chews (i.e. n), mean and s anda d de ia ion o
chew du a ion (chew du a ion is 2[i]− 1[i]), and mean and
s anda d de ia ion o chewing a e (ins an aneous chewing
a e is es ima ed as 1[i]− 1[i−1]), and ood ype (as is a
ca ego ical a iable).
The second se o ea u es is based on he audio signal.
A challenge lays in he ac ha (a) each chewing bou has
a di e en du a ion and di e en numbe o chews, (b) each
indi idual chew has a di e en du a ion. To o e come his,
we ollow a wo-s ep p ocess: i s , one ea u e ec o is
ex ac ed om each indi idual chew o a single chewing
bou , and hen, all he ea u es ec o s o he chews ha
belong o a single chewing abou a e agg ega ed oge he
o p oduce a inal, single ea u e ec o o he chewing
bou ( his inal ec o is hen used o aining he weigh
p edic o s).
In he i s s ep, he ea u es ha we ex ac om each
indi idual chew a e he ones used in [7], [18]. The ea u es
include signal ene gy in log-scale ene gy bands, highe
o de s a is ics (including skewness and ku osis), and ac al
dimension. Es ima ing each o hose ea u es is independen
o he leng h o he a ailable audio signal (i.e. om chew
du a ion); his allows us o ob ain compa able alues o each
ea u e among chews o a ying du a ion. A e ex ac ing
he ea u es o he en i e da ase we s anda dize hem by
sub ac ing he mean (o each ea u e) and di iding by i s
s anda d de ia ion.
In he second s ep, we agg ega e he ea u es ec o s o
he chews o each chewing bou . We examine wo simila ap-
p oaches o his: bag-o -wo ds (BoW) and ec o s o locally
agg ega ed desc ip o s (VLAD). Cen oids a e ob ained o e
he a ailable aining po ion o he da ase (AIC is used o
selec ing he numbe o cen oids), which a e hen used bo h
on he ain as well as he es po ions o he da ase .
B. Bi e-weigh es ima o s
To es ima e he bi e weigh om he a ailable ea u es,
we examine ou di e en algo i hms. The i s es ima o we
e alua e is LR, simila o [17]. We ha e also expe imen ed
wi h models ha include c oss-p oduc e ms bu ha e ound
ha he o e all e ec i eness is no a ec ed signi ican ly.
The second algo i hm is suppo ec o eg ession (SVR).
We use a adial-basis unc ion (RBF) ke nel, and use a g id
sea ch o hype -pa ame e s Cand γby andomly spli ing
he da a om he m−1subjec s o 70% o aining and 30%
o alida ion. We sea ch o Cin 10i, i =−2,−1,0,1,2
and o γ= 10in−1, i =−1,0,1,2,3, whe e nis he leng h
o he ea u e ec o .
We also examine classic eed- o wa d neu al ne wo ks
(FFNN). We conside a chi ec u es wi h ei he 2o 3hidden
laye s, and 5,10,15, and 20 neu ons pe laye ( hus, a o al
o 8dis inc a chi ec u es, as we do no conside a chi ec u es
wi h di e en numbe o neu ons pe laye ). The choice o
he a chi ec u e is ea ed as he hype -pa ame e o his
model and is selec ed based on a 90% aining and 10%
alida ion spli o he m−1subjec s, and is hus di e en
o each subjec . T aining minimizes he mean absolu e e o
(MAE); lea ning a e is se o 0.01 and he maximum numbe
o epochs is 1,000. We use he BFGS Quasi-New on back-
p opaga ion algo i hm.
Finally, we also examine gene alized eg ession neu al
ne wo ks (GRNN) [19] wi h Gaussian ke nel. Simila ly o
he p e ious models, we also selec he hype -pa ame e σo
he ke nel using a ain- alida ion spli on he m−1subjec s.
IV. DATASET
To e alua e ou app oach we ha e collec ed an in-house
da ase . A o al o 8 pa icipan s we e en olled o he da a
collec ion ials (6males and 2 emales, age 25±1.07 yea s,
body-mass index 25.49 ±3.06). Fou di e en ood ypes
we e consumed: apple, banana, ice, and po a o chips. These
ou ypes we e selec ed as hey ha e a unique combina ion
o c ispiness (apple and po a o chips) and we ness (apple and
banana). The da ase includes 2hou s o ea ing and con ains
a o al o 473 chewing bou s and 7.539 chews.
Audio signals we e collec ed by using comme cially a ail-
able Samsung Galaxy Buds. We ha e c ea ed a cus om
And oid applica ion ha cap u es synch onized audio (a
44.1kHz, 16 bi ) om he ea buds and pla e weigh (a 1Hz)
om a Blue oo h-enabled pla e scale. The pla e weigh scale
has been used o de i e g ound u h alues o bi e weigh .
A ideo eco ding o each session has also been cap u ed o
u he assis us in he anno a ion p ocess.
V. EVALUATION
To e alua e ou app oach we pe o m a ious combina-
ions o ea u e se s and es ima ion models. Gi en he non-
audio based ea u e se and he wo me hods o agg ega ing
he audio based ea u es, we examine he ollowing i e
combina ions: non-audio ea u es ( 1), audio wi h BoW ( 2),
audio wi h VLAD ( 3), combina ion o non-audio and audio
wi h BoW ea u es ( 4), and combina ion o non-audio and
audio wi h VLAD ea u es ( 5). Fo each o hese ea u e
se s we examine ou es ima o s (Sec ion III-B): LR, SVR,
FFNN, GRNN. Finally, we ain i e di e en models pe
combina ion: he i s ou a e ood speci ic, while he i h
is ained on he da a om all ood ypes. Table I shows he
mean absolu e e o pe expe imen , and Table II shows he
mean absolu e pe cen age e o (%) in he same s uc u e.
All expe imen s a e pe o med in LOSO ashion; hus, each
esul is he mean ac oss he 8pa icipan s o ou da ase .
We also p esen e alua ion esul s om ou implemen a-
ion o he algo i hm o [17] (Sec ion II).
Based on he esul s, he combina ion o bo h non-audio
and audio based ea u es imp o es he es ima ion accu acy
(yielding lowe e o ) o mos cases. This is mo e e iden
when aining on all ood ypes oge he (Figu e 2). This
conclusion is also inline wi h he esul s o he algo i hm o
[17] which uses a combina ion o 7non-audio ea u es and 1
audio ea u e, as i is be e om ou non-audio and audio-
only app oaches while sligh ly wo se om ou non-audio
and audio combina ions.
Based on he esul s, FFNN and GRNN a e able o achie e
he bes esul s (lowes e o s) compa ed o LR and SVR.
When aining on a single ood ype, GRNN-based models
wi h 4achie e he lowes MAE (close o o less han 1
g) and simila ly low s anda d de ia ion o absolu e e o s.
The only excep ion is o po a o chips whe e FFNN achie es
he lowes e o s. Howe e , GRNN achie es he second
lowes e o s and he di e ence ( om FFNN)) is e y small:
0.25 (0.4) g o GRNN e sus 0.20 (0.4) o FFNN. When
aining on all ood ypes oge he , FFNN wi h 5seems o
achie e he lowes e o s.
Compa ing he wo di e en ypes o agg ega ing audio
ea u es ( om chews o chewing bou s) he e seems o be
TABLE I: Mean and s anda d de ia ion absolu e e o s pe
algo i hm and ea u e se . Bes esul (lowes e o ) in ed.
Apple Banana Rice Chips All
LR
13.02 (2.3) 5.89 (2.8) 3.80 (2.5) 0.93 (0.7) 4.94 (3.4)
24.13 (3.4) 5.82 (3.9) 3.82 (2.2) 1.23 (0.8) 3.70 (3.2)
34.83 (3,7) 5.57 (3.4) 5.06 (3.2) 1.10 (0.8) 3.86 (3.4)
43.65 (2.5) 5.54 (3.3) 3.86 (3.1) 1.02 (0.7) 3.24 (2.8)
53.15 (2.2) 5.97 (3.3) 3.74 (2.7) 0.92 (0.7) 3.44 (3.0)
SVR
13.18 (2.2) 6.30 (3.2) 3.97 (2.7) 0.97 (0.8) 4.52 (3.3)
24.76 (3.5) 5.01 (3.2) 4.44 (3.0) 1.17 (0.8) 3.66 (3.2)
35.07 (3.3) 5.11 (3.2) 5.02 (3.2) 1.12 (0.8) 3.76 (3.5)
43.57 (2.3) 5.56 (3.2) 3.19 (2.6) 1.05 (0.8) 3.16 (2.9)
53.00 (2.4) 5.44 (3.4) 3.68 (2.7) 1.00 (0.8) 3.02 (2.5)
FFNN
13.70 (2.7) 5.17 (3.6) 4.72 (3.2) 0.93 (0.8) 3.31 (3.0)
24.19 (3.7) 4.68 (3.8) 4.29 (3.4) 1.00 (0.9) 3.86 (3.4)
33.64 (2.6) 4.10 (3.2) 3.52 (2.1) 0.81 (0.6) 4.10 (3.4)
43.56 (2.7) 5.56 (3.5) 4.52 (3.0) 1.22 (0.9) 2.77 (2.7)
52.60 (2.4) 2.55 (3.0) 2.22 (2.4) 0.20 (0.4) 2.12 (2.4)
GRNN
12.39 (1.9) 3.82 (2.3) 3.27 (2.1) 0.82 (0.6) 3.59 (2.8)
23.17 (3.0) 2.85 (2.2) 2.50 (2.4) 0.87 (0.8) 4.14 (3.6)
34.43 (4.0) 4.78 (2.9) 4.01 (2.8) 1.04 (0.7) 6.19 (3.4)
41.10 (1.8) 0.89 (1.1) 0.93 (1.4) 0.25 (0.4) 3.90 (3.4)
54.30 (3.2) 5.66 (3.8) 3.61 (2.9) 1.08 (0.8) 3.80 (3.4)
Am e al.
3.26 (2.7) 5.96 (2.9) 5.22 (3.2) 1.10 (0.8) 3.37 (3.0)
TABLE II: Mean and s anda d de ia ion o absolu e ela i e
(%) e o s pe algo i hm and ea u e se . Bes esul (lowes
e o ) in ed.
Apple Banana Rice Chips All
LR
131 (27) 54 (36) 47 (46) 47 (59) 57 (26)
240 (39) 57 (43) 44 (45) 60 (73) 38 (34)
346 (44) 44 (41) 51 (53) 49 (34) 40 (47)
438 (29) 50 (37) 47 (51) 40 (49) 32 (36)
528 (27) 55 (40) 43 (42) 37 (44) 35 (32)
SVR
128 (24) 56 (38) 48 (48) 47 (47) 53 (59)
247 (40) 42 (38) 43 (40) 54 (43) 38 (33)
344 (36) 42 (38) 47 (48) 50 (39) 40 (43)
437 (27) 52 (39) 38 (46) 40 (49) 31 (32)
527 (25) 51 (38) 42 (41) 48 (39) 31 (36)
FFNN
133 (25) 46 (42) 50 (38) 40 (28) 33 (29)
236 (60) 40 (42) 41 (41) 50 (45) 40 (29)
333 (58) 29 (21) 36 (45) 42 (42) 46 (41)
430 (26) 56 (43) 45 (44) 56 (58) 39 (31)
526 (26) 26 (32) 25 (35) 15 (28) 20 (27)
GRNN
122 (18) 36 (30) 40 (38) 34 (52) 35 (31)
228 (25) 19 (17) 28 (38) 35 (57) 42 (37)
336 (38) 32 (32) 44 (58) 51 (54) 64 (37)
409 (14) 06 (10) 09 (20) 20 (45) 40 (43)
543 (31) 57 (50) 42 (47) 44 (46) 36 (37)
Am e al.
30 (23) 54 (33) 60 (53) 55 (73) 34 (31)
NA BOW VLAD NA+BOW NA+VLAD
2
3
4
5
6
MAE (g)
LR
SVR
FFNN
GRNN
Am e al.
Fig. 2: Mean absolu e e o pe ea u e se and es ima o
model when aining on all ood ypes oge he .
no clea conclusion abou whe he BoW is be e o VLAD.
This holds bo h o when using only audio ea u es (i.e. 2 s.
3), as well as o when combining hem wi h he non-audio
ea u es (i.e. 4 s. 5). The only excep ion is GRNN ha
seems o bene i om he use o BoW; his can be a ibu ed
o he s ic e quan iza ion o he ea u e space ha BoW
applies.
Finally, MAE is qui e lowe o po a o chips compa ed o
he o he ood ypes. Howe e , his is no a esul o “be e ”
ained es ima ion models, bu o he ac ha po a o-chip
bi es a e gene ally ligh e ( han apple bi es o example).
This can be con i med by compa ing MAPE e o s ha a e
shown in Table II.
E alua ion esul s o he algo i hm o Am e al. [17]
a e compa able wi h ou app oach. The a e clea ly su passed
hough by ou FFNN and GRNN based app oaches. O e all,
MAE is highe in ou da ase compa ed o he alues epo ed
by he au ho s in hei o iginal wo k o [17]. This can be a -
ibu ed o he mo e challenging na u e o ou da ase as well
as di e ences in he sound cap u ed by ou comme cially
a ailable ea buds and hei cus om-made senso .
VI. CONCLUSIONS
In his wo k we ha e p esen ed an app oach o es ima ing
bi e weigh om audio signal cap u ed by comme cially
a ailable ea buds. Using comme cially a ailable ha dwa e
is essen ial o enable highe adop ion a es o such die a y
moni o ing app oaches, since hey educe in asi eness and
discom o o he end use .
Ou app oach uses a combina ion o non-audio and audio
ea u es which a e used o ain es ima ion models. We
e alua e on an in-house da ase o app oxima ely 2hou s.
Ou bes esul s a e ob ained by aining ood-speci ic GRNN
models and non- ood-speci ic FFNN models. GRNN models
yield MAE o app oxima ely 1g o less, and FFNN yield
a o al MAE o 2.12 g. We also compa e wi h an exis ing
algo i hm om li e a u e and achie e lowe e o s o all
cases.
An impo an limi a ion o ou app oach is ha i equi es
anno a ions o he s a and s op ime-s amps o indi idual
chews, as well as ood ype anno a ions ( o ood- ype–
speci ic models). Fu u e wo k includes e alua ing on bigge
and mo e di e se da ase s wi h mo e ood ypes and di e en
da a-cap u ing condi ions (close o ee-li ing) as well as
e alua ing in combina ion wi h audio based chewing de ec-
o s and au oma ically de ec ed ood ypes.
ACKNOWLEDGMENTS
The wo k leading o hese esul s has ecei ed unding
om he EU Commission unde G an Ag eemen No.
965231, he REBECCA p ojec (H2020).
REFERENCES
[1] O. Am , H. Junke , and G. T os e , “De ec ion o ea ing and d inking
a m ges u es using ine ial body-wo n senso s,” in ISWC’05, 2005, pp.
160–163.
[2] E. Sazono and J. M. Fon ana, “A senso sys em o au oma ic
de ec ion o ood in ake h ough non-in asi e moni o ing o chewing,”
IEEE Senso s Jou nal, ol. 12, no. 5, pp. 1340–1348, 2012.
[3] H. Kalan a ian, N. Alshu a a, and M. Sa a zadeh, “A wea able
nu i ion moni o ing sys em,” in 2014 11 h BSN, 2014, pp. 75–80.
[4] M. Fa ooq and E. Sazono , “Segmen a ion and cha ac e iza ion o
chewing bou s by moni o ing empo alis muscle using sma glasses
wi h piezoelec ic senso ,” IEEE JBHI, ol. 21, no. 6, pp. 1495–1503,
2017.
[5] O. Am and G. T os e , “On-body sensing solu ions o au oma ic
die a y moni o ing,” IEEE Pe asi e Compu ing, ol. 8, no. 2, pp.
62–70, 2009.
[6] C. Sonoda e al., “Associa ions among obesi y, ea ing speed, and o al
heal h,” Obesi y Fac s, ol. 11, no. 2, pp. 165–175, 2018.
[7] V. Papapanagio ou e al., “A no el chewing de ec ion sys em based
on PPG, audio and accele ome y,” IEEE JBHI, ol. 21, no. 3, pp.
607–618, 5 2017.
[8] R. S. Ma eld, E. R. Mu h, and A. Hoo e , “Measu ing he consump-
ion o indi idual solid and liquid bi es using a able-embedded scale
du ing un es ic ed ea ing,” IEEE JBHI, ol. 21, no. 6, pp. 1711–1718,
2017.
[9] K. Ky i sis, C. Diou, and A. Delopoulos, “Modeling w is mic omo e-
men s o measu e in-meal ea ing beha io om ine ial senso da a,”
IEEE JBHI, ol. 23, no. 6, pp. 2325–2334, 2019.
[10] ——, “A da a d i en end- o-end app oach o in- he-wild moni o ing
o ea ing beha io using sma wa ches,” IEEE JBHI, ol. 25, no. 1,
pp. 22–34, 2021.
[11] K. Ky i sis e al., “Assessmen o eal li e ea ing di icul ies in
pa kinson’s disease pa ien s by measu ing pla e o mou h mo emen
elonga ion wi h ine ial senso s,” Scien i ic Repo s, ol. 11, no. 1, p.
1632, 1 2021.
[12] S. Wang, G. Zhou, L. Hu, Z. Chen, and Y. Chen, “Ca e: Chewing
ac i i y ecogni ion using nonin asi e single axis accele ome e ,” in
UbiComp, se . UbiComp/ISWC’15 Adjunc . New Yo k, NY, USA:
Associa ion o Compu ing Machine y, 2015, p. 109–112.
[13] S. Wang e al., “Ea ing de ec ion and chews coun ing h ough sensing
mas ica ion muscle con ac ion,” Sma Heal h, ol. 9-10, pp. 179–191,
2018, cHASE 2018 Special Issue.
[14] Y. Lu e al., “go ood m: An a i icial in elligence sys em o die a y
assessmen ,” Senso s, ol. 20, no. 15, 2020.
[15] ——, “A mul i- ask lea ning app oach o meal assessmen ,” in
P oceedings o he Join Wo kshop on Mul imedia o Cooking and
Ea ing Ac i i ies and Mul imedia Assis ed Die a y Managemen , se .
CEA/MADiMa ’18. New Yo k, NY, USA: Associa ion o Compu ing
Machine y, 2018, p. 46–52.
[16] V. Papapanagio ou, C. Diou, and A. Delopoulos, “Chewing de ec ion
om an in-ea mic ophone using con olu ional neu al ne wo ks,” in
EMBC 2017, 2017, pp. 1258–1261.
[17] O. Am , M. Kusse ow, and G. T os e , “Bi e weigh p edic ion om
acous ic ecogni ion o chewing,” IEEE TBE, ol. 56, no. 6, pp. 1663–
1672, 2009.
[18] V. Papapanagio ou, C. Diou, J. an den Boe , M. Ma s, and A. De-
lopoulos, “Recogni ion o ood- ex u e a ibu es using an in-ea mi-
c ophone,” in ICPR 2020, Pa V, LNCS 12665 p oceedings. Cham:
Sp inge In e na ional Publishing, 2020 [in p ess].
[19] D. Spech , “A gene al eg ession neu al ne wo k,” IEEE T ansac ions
on Neu al Ne wo ks, ol. 2, no. 6, pp. 568–576, 1991.
