Uni e sidad Complu ense de Mad id
Facul ad de In o ma ica
Glucose classi ica ion and
p edic ion sys em wi h
Neu al Ne wo ks
Final Deg ee P ojec
Au ho s:
Alejand o Va ela Lo enzo
Al a o Delgado Gu ie ez
Ad iso : J. Ignacio Hidalgo
June 2020
Con en s
1 Summa y 3
2 In oduc ion 5
2.1 Backg ound ............................ 5
2.2 Goals................................ 7
2.2.1 De elopmen o glucose p edic ion models . . . . . . . 7
2.2.2 Websi e se ice . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Wo k-plan............................. 7
3 S a e o he a 9
4 GlucNe 15
4.1 De ices .............................. 15
4.2 Da asou ces............................ 16
4.3 Da a p ep ocessing . . . . . . . . . . . . . . . . . . . . . . . . 17
4.3.1 S udy o he da a-se . . . . . . . . . . . . . . . . . . . 17
4.3.2 Da a illing ........................ 18
4.3.3 No maliza ion . . . . . . . . . . . . . . . . . . . . . . . 19
4.3.4 Da a o be supe ised . . . . . . . . . . . . . . . . . . 19
4.3.5 The da a ame is p epa ed . . . . . . . . . . . . . . . 21
4.4 Glucose o ecas ing wi h nume ical da a . . . . . . . . . . . . 21
4.5 Glucose o ecas ing wi h wa ele s images . . . . . . . . . . . . 24
4.5.1 In oduc ion o Digi s . . . . . . . . . . . . . . . . . . 25
4.5.2 C ea ing wa ele images . . . . . . . . . . . . . . . . . 25
4.5.3 T aining wi h NVIDIA Digi s . . . . . . . . . . . . . . 27
4.6 GlucNe Websi e ......................... 28
4.6.1 Connec ion wi h da abase . . . . . . . . . . . . . . . . 33
5 Expe imen al Resul s 34
5.1 Me ics............................... 34
5.2 Resul s............................... 35
5.2.1 Fi s execu ion wi h Pa ien 1 . . . . . . . . . . . . . . 35
5.2.2 Desc ip ions o he esul s . . . . . . . . . . . . . . . . 36
5.2.3 T ain wi h Feed Fo wa d . . . . . . . . . . . . . . . . . 37
5.2.4 T ain wi h LSTM . . . . . . . . . . . . . . . . . . . . . 45
5.2.5 T ain wi h NVIDIA Digi s . . . . . . . . . . . . . . . . 53
1
6 Conclusions and Fu u e wo k 55
6.1 Gene al conclusions . . . . . . . . . . . . . . . . . . . . . . . . 55
6.2 Fu u ewo k............................ 56
7 Alejand o Va ela wo k 58
8 Al a o Delgado wo k 60
9 Acknowledgmen 62
10 Bibliog aphy 63
2
1. Summa y
Abs ac
Glucose le els p edic ion is a di icul ask commonly aced by people wi h
diabe es, a ch onic heal h condi ion ha a ec s how a human body syn he-
sizes ood. People wi h diabe es mus ha e an exhaus i e con ol o he le els
o suga in he bloods eam in o de o manage insulin in akes, a p ocedu e
ha is usually done manually and wi hou enough accu acy. Le els o glu-
cose in a human body depends on a lo o di e en ac o s, so he isk o
doing miscalcula ions is always aken by he pa ien .
Nowadays, using new echnologies such as A i icial In elligence (AI) o
Machine Lea ning (ML), hese calcula ions can be suppo ed and eased by
he applica ion o p edic ion sys ems. The ield o AI and ML is e y la ge,
p o iding scien is s wi h se e al di e en ools o c ea e o ecas ing algo-
i hms. Du ing his p ojec , we a e going o ocus on he c ea ion and use o
Neu al Ne wo ks (NN) o glucose le el p edic ion. To c ea e his sys ems,
we a e explo ing di e en ypes o Neu al Ne wo ks (NN), anging om eg-
ula nume ic NN o g aphic NN applica ions, which a e as ly known in he
wo ld o da a scien is s.
Howe e , hese algo i hms a e always a di icul and hidden p ocess o
he eal use s, diabe es pa ien s, so in o de o make a highe impac on
people a ec ed by his disease, we will c ea e an online applica ion ha will
sha e all he powe o NNs wi h he inal pa ien s, using a clea and in ui i e
use in e ace. In he mean ime, pa ien s will collabo a e on he NN aining
p ocess, as long as da a p o ided by hose using he applica ion will allow
us o de elop a mo e hea ily ained NN, hus imp o ing i s e ec i eness in
glucose le els o ecas ing.
Re e ing o he nume ical NN and acco ding o he esul s, we can s a e
ha hei pe o mance o 30 and 60 minu es p edic ion is qui e accu a e
and, o 90 and 120 minu es p edic ion, i h ows p omising esul s.
3
Ins ead, o he g aphical NN, e en hough he app oach is in e es ing
and he s udies a e showing us ha he echnology is e y powe ul, i needs
much mo e in es iga ion un il we ge good esul s.
Keywo ds
Diabe es Melli us, Glucose Le els P edic ion, Neu al Ne wo ks, Fo ecas ing,
Online se ice, Machine Lea ning, NVIDIA Digi s.
4

2. In oduc ion
2.1. Backg ound
Diabe es is a ch onic disease ha is i e e sible and a ec s he no mal ans-
o ma ion o suga in ene gy. I occu s when blood glucose, also called blood
suga , is oo high in he bloods eam. Insulin, a ho mone made by he pan-
c eas, helps glucose om ood ge in o he cells o be used o ene gy. Some-
imes, he body doesn’ make enough insulin o doesn’ use insulin co ec ly,
so glucose s ays in he bloods eam and doesn’ each he cells. O e ime,
ha ing oo much glucose in he blood can cause se e al heal h p oblems. We
can di e en ia e wo main ypes o diabe es:
•T1DM ( ype 1)
This ype is he less common wi hin he pa ien s o diabe es, a en h
o o al. I is an au oimmune illness ha a acks he panc ea ic cells
which gene a e insulin, making hem dys unc ional. Thus, he pa ien s
need o supply hemsel es wi h daily doses o insulin ha eplace hose
he me abolic sys em should gene a e.
•T2DM ( ype 2)
Wi h diabe es ype 2, he me abolic sys em is able o gene a e insulin,
bu he immune sys em c ea es esis ance o i , esul ing in a simila
ou come as in T1DM. Anyways, people wi h T2DM do no always need
o supply hemsel es wi h daily doses o insulin. No mally, i is enough
i hey ha e a heal hy ou ine and die . This ype o diabe es appea s
along he li e ime, and can be p e en ed o delayed wi h a heal hy
li es yle (exe cise, heal hy die s...).
To main ain no mal le els o glucose, he pa ien s need o ha e con ol
o hose le els e e y ime, aking in o accoun he physical e o , die and
o he ac o s such as alcohol in akes, le el o s ess, e c.
To measu e he le els o glucose i is usual o ha e a manual glucose me-
e o a con inuous glucose moni o ing sys em. This is a edious ask ha
is done on a daily basis by he pa ien and ha p o okes an ex a s ess.
Pa ien s need o know he glucose le el o decide i hey need o ake a co -
ec i e ac ion, such as injec ing insulin o ea ing some ood. To help wi h his
5
ou ine, new me hods a e being in es iga ed. One o hem is he p edic ion
o glucose le els, which could po en ially allow pa ien s o an icipa e when
he e is a isk o hypoglycemia o hype glycemia.
Howe e , glucose le els p edic ion is a ex emely di icul ask because
o he g ea amoun o ac o s ha a e in ol ed. Fu he mo e, p edic ions
mus be eally accu a e, since a w ong dose o insulin, aken om a w ong
p edic ion, can cause heal h p oblems on he pa ien .
This ield o s udy is no new. P e iously, nume ous echniques ha e been
ied o glucose o ecas ing. The g ea es aim o his echniques is o achie e
a pe manen solu ion o diabe es, such as he de elopmen o an ”a i icial
panc eas”. Figu e 1 summa izes he di e en au oma izing app oaches o
diabe es [1].
Figu e 1: Compa ison o au oma izing on cu en me hods o diabe es ea men [1]
Th ough an a i icial panc eas wi h an ad anced p edic ion sys em along
wi h measu emen ins umen such as a con inuous glucose moni o , i could
be possible o au oma e he insulin injec ion p ocess in he same way as he
panc eas o a heal hy pe son wo ks.
Some o he echniques used o his p edic ion model a e: Gene ic P o-
g amming, Random Fo es Reg ession, K-Nea es Neighbo s o G amma ical
E olu ion [1] and, o cou se, Neu al Ne wo ks [2].
6
2.2. Goals
2.2.1. De elopmen o glucose p edic ion models
The main goal o he p ojec is o o ecas he le el o glucose ha diabe es
pa ien s a e going o ha e in he nex 30, 60, 90 o 120 minu es based on
hei p e ious da a. Wi h his da a, we’ll ain di e en e sions o models
o ob ain as many esul s as possible. Thus, we can make compa isons be-
ween di e en models, and unde s and which pa ame e s and con igu a ion
a e he bes o he glucose p edic ion. This will be a long- e m s udy in he
py hon en i onmen o ha e a consis en model o ain and p edic glucose
alues.
2.2.2. Websi e se ice
Once we ob ain a p ecise model wi h an op imal con igu a ion o he da a,
we wan o de elop a websi e whe e any diabe es pa ien can access o make
hei glucose p edic ions wi h an in ui i e in e ace. This applica ion mus
be as simple as possible o he use . Wi h his pla o m, no only we p o ide
he use wi h a p edic ion o i s glucose le els, bu also, i he use au ho ize
i , we use his da a o ain ou models. All he esul s and aining will be
sa ed in a da abase, ha we will need o c ea e and adap o he cu en
AbsysG oup da abase. Thus, he e will be a connec ed sys em be ween ou
p ojec and he p e ious one.
Hence, he main goal is o c ea e he websi e sys em ha will co e and
be he inal image o he models de elopmen and he machine lea ning wo k.
2.3. Wo k-plan
Ou wo k-plan o he c ea ion o GlucNe will be sepa a ed in he ollowing
s ages:
•S udy he in o ma ion ela ed o diabe es and he co ela ion be ween
he di e en ac o s and he glucose le el
•Ge in ouch wi h he da a s uc u es
•P e-p ocessing he da a
7
•S a wi h he i s neu al ne wo k (NN) aining
•T y di e en pa ame e s and ypes o NN
•Implemen he websi e
•Join he algo i hm and he websi e
•Connec he da abase wi h he websi e
The es o his documen is o ganized as ollows. Sec ion 3 e iews
p e ious app oaches on glucose p edic ion by Neu al Ne wo ks. Sec ion 4
explains he whole p ocess o he p ojec including:
•Explana ion he da a o ma in sec ions 4.1 and 4.2.
•Da a p e-p ocessing ask in sec ion 4.3.
•C ea ion and aining o nume ical NN in sec ion 4.4.
•Adap a ion o he da a o he g aphical NN and i s aining in N idia
Digi s in sec ion 4.5.
•Implemen a ion o he websi e and i s da abase in sec ion 4.6.
A e wa ds, sec ion 5 explains he me ics used o he esul s and discuss he
esul s om he di e en NN. Finally, on Sec ion 6, we s a e ou conclusions
om he p ojec and p opose some u u e wo k o keep imp o ing he p ojec .
8
4. GlucNe
GlucNe is an online se ice o glucose classi ica ion and p edic ion ha is
based in Neu al Ne wo ks implemen ed in Py hon.
The whole p ocess o ge he inal p oduc includes:
•Wo king wi h he da a and p epa ing i o e e y di e en scena io.
•C ea ing he neu al ne s ha will wo k in he websi e.
•Explo ing a new app oach on glucose p edic ion and classi ica ion based
in G aphic NN.
•C ea ing a websi e whe e NN will be suppo ed.
4.1. De ices
The ollowing senso s a e used o measu e he glucose and insulin alues,
physical ac i i y and ood in akes:
•F ee S yle Lib e-Abbo
•Minimed Med onic CGM
•Med onic Insulin Pump
•Roche Insulin Pump
•Fi bi Ionic
•No es om pa ien s
Please no e, ha no all he pa ien s ha e all he da a because we we e
no able o eco e all he i ems. Using his de ices, we eco ded he ollowing
da a:
•In e s i ial glucose
•No es o es ima ed ca bohyd a e uni s inges ed, aken by each pa ien .
15

•Insulin injec ed using an insulin in use de ice om Med onic and/o
Roche, which egis e s injec ions o bo h basal and bolus insulin e e y
i e minu es.
•Bu ned calo ies and s eps
•O he in o ma ion
4.2. Da a sou ces
In his sec ion we will show he wo di e en sou ces o da a we go o wo k
wi h.
Ou main sou ce o da a and he one ha is going o guide he s udy a e
he iles con aining da a om Pa ien s o he Hospi al Uni e si a io P ´ıncipe
de As u ias de Alcal´a de Hena es in Mad id, om June 13 h, 2018 o, July
17 h, 2019.
Column Meaning Measu ing De ice No e
Basal Ra e (U/h) Basal a e Med onic Pump E e y 5 min i i is ok
Bolus Type Bolus ype Med onic Pump any ime,
Bolus Volume Deli e ed (U) Bolus Uni s Med onic Pump any ime,
BWZ Ca b Ra io (U/Ex) Insulin- o-ca bohyd a e Ra io Med onic Pump any ime,
BWZ Ca b Inpu (exchanges) Amoun o ca bohyd a es Med onic Pump any ime,
Senso Glucose (mg/dL) The senso glucose alue CGM Med onic E e y 5 min i i is ok
Ho a Time Anyone
Calo ias Calo ies om Fi bi Fi bi any ime,
Ri mo Ca diaco HR om Fi bi Fi bi any ime,
Es ado Sleeping s a e om Fi bi Fi bi any ime,
Pasos S eps om i bi Fi bi any ime,
His ´o ico glucosa (mg/dL) Au oma ic glucose le el F ees yle E e y 15 min i i is ok
Glucosa le´ıda (mg/dL) Manual glucose le el om F eeS yleLib e F ees yle any ime,
Insulina de acci´on ´apida sin alo num´e ico Bolus Ma k F ees yle any ime,
Insulina de acci´on ´apida (unidades) Bolus uni s F ees yle any ime,
Alimen os sin alo num´e ico Food in ake F ees yle any ime,
Ca bohid a os ( aciones) Food in ake uni s F ees yle any ime,
Insulina de acci´on len a sin alo num´e ico Slow ac ion insulin imes amp F ees yle any ime,
Insulina de acci´on len a (unidades) Uni s o slow ac ion insulin F ees yle any ime,
No as No es F ees yle any ime,
Glucosa de la i a (mg/dL) Glucose Me e om F eeS yle Lib e F ees yle any ime,
Ce onas (mmol/L) Ke ones om F ees yle Lib e F ees yle any ime,
Insulina comida (unidades) Insulin uni s F ees yle any ime,
Insulina co ecci´on (unidades) Uni s om co ec o bolus F ees yle any ime,
Insulina cambio usua io (unidades) Daily insulin uni s F ees yle Low ideli y
Ho a an e io Time change F ees yle any ime,
Ho a ac ualizada Time change F ees yle any ime,
Senso Calib a ion BG (mg/dL) Glucose calib a ion Ex e n glucome e any ime,
BG Reading (mg/dL) Glucose calib a ion Ex e n glucome e any ime
Table 1: Meaning o each pa ame e om Alcal´a de Hena es pa ien s da ase s
16
Table 1 shows he a iables s o ed on he cs ile wi h hei meanings,
he de ices om whe e da a was e ie ed, an explana ion and no es on he
equency o acquisi ion and/o any o he pa icula i y. We ha e access o
15 da a-se s o di e en pa ien s om his hospi al, each one wi h a ound
3500 ows o da a along wi h he imes amps.
Ou second sou ce o da a a e he iles om 6 di e en pa ien s om
Ohio’s Hospi al. These cs ha e e y di e en columns and almos all o hem
do no ha e enough da a o be use ul (less han 25% o he o al amoun o
ows in he da a-se ). The columns ha will ha e mo e impo ance o us a e
”basale en alue”, ”accele a ione en alue” and ”Glucose le ele en alue”.
F om he i s wo we can ge he equi alen Basal Ra e (U).
4.3. Da a p ep ocessing
As in e e y da a-mining and machine lea ning p ojec , he i s ask a e
analysing he da a is p epa ing i o be used by he algo i hm. In his glu-
cose p edic ion p ojec we a e wo king on da a om single indi iduals based
on ime se ies.
4.3.1. S udy o he da a-se
Fi s ly, when we ake a look a he main sou ce o da a (da a om Pa ien s
o he Hospi al Uni e si a io P ´ıncipe de As u ias de Alcal´a), we no iced ha
he majo i y o he columns do no p o ide use ul in o ma ion as hey ha e
epea ed in o ma ion o no o many alues o wo k wi h. This columns ha
a e no use ul o ha e oo much ”no assigned” (NaN) ha e o be dele ed.
Fu he mo e, as we a e wo king wi h da a based on ime, mo e speci ically
e e y 5 minu es, i is impo an o ea he da a as a ime se ies. The
column o ime is no only a a iable, bu also he alue ha speci ies he
o de in which he da a is ela ed. To speci y his o de we use he ime as
an index, making i easie o ain he da a. In he case o he Basal Ra e,
i is alued as uni s pe hou so i is necessa y o di ide e e y alue in his
column by 12, esul ing in uni s pe 5 minu es. Figu e 2 shows he esul o
d opping he columns om he da a-se .
17
Figu e 2: Resul ing da a-se a e d opping columns
Fu he mo e, he e is ano he da a sou ce, he Ohio’s Hospi al, which
has e y di e en columns compa ed o ou p e ious sou ce. As we wan o
c ea e an uni ied model ha can suppo di e en da a sou ces, we need o
adap his da a-se o ou main s uc u e.
As we did wi h he o he da a-se s om P incipe de As u ias Hospi al,
we need o d op all he columns ha do no ha e su icien in o ma ion. In
his case, he only columns ha ha e good and enough da a o ain a e he
le els o glucose and he Basal Ra e. This da a-se is also o ganized in 5
minu es imes amps.
4.3.2. Da a illing
Once he da a ame has he p ope columns and is con e ed in o a ime
se ies, he nex s ep is o c ea e da a o he non assigned alues, his ask
is commonly known as da a- illing. In his case, i was necessa y o use wo
di e en ypes o da a- illing, in e pola ed and b ill. The in e pola e unc ion
ills he alues be ween 2 di e en numbe s making use o a lineal unc ion,
and he b ill unc ion simply copies he da a om he las seen alue in all
subsequen missing alues.
We use he in e pola e unc ion o ill he in e media e gaps in he nu-
me ical columns such as he le els o glucose o ob ain easonable new alues.
Using in e pola e o he basal a e won’ be accu a e since he alues a e
aken e e y hou . in his case, we simply copy he alue o 5 minu es in
e e y hou . Finally, o he s eps and calo ies columns i is no necessa y o
18
pe o m any da a- illing.
4.3.3. No maliza ion
On he machine lea ning wo k, e e y ype o column usually ha e e y di -
e en anges o alues which a e no homogeneous. No maliza ion is he
bes solu ion o ha e simila anges o wo k wi h. I is some imes equi ed
o no malize he da a o aining wi h some neu al ne wo ks such as Ke as.
In he case o he nume ic neu al ne s, we use he MinMax scale be ween 0
and 1 because he da a-se does no con ain nega i e alues.
Ins ead, when using g aphic neu al ne wo ks, we apply he Zsco e no mal-
iza ion o ep esen he da a g aphically and ain he algo i hm in Digi s.
We apply his di e en me hod o he g aphical da a because wi h Min-
MaxScale he g aphics o he columns wi h low alues did no show he
ange p ope ly. Zsco e applies one no maliza ion o each ype o da a.
4.3.4. Da a o be supe ised
Once we ha e all he columns wi h he igh da a, we need o con e he
da a-se in a supe ised da a-se . This is a equi ed ask because we need all
he da a ela ed o each p edic ion in each line. Mo e deeply, o each ow,
which a ge is 30, 60, 90 o 120 minu es p edic ion, we need he da a ha
a ec s di ec ly o he p edic ion.
As we see in he Figu e 3, i is always necessa y o ha e he alues o
glucose and basal a e om he p e ious 120 minu es and he da a o he
basal a e om 30 o 120 minu es in he u u e.
To do his, we copy he glucose and basal a e da a o he p e ious 24
ows and he da a o he u u e 6-24 ows in each ow. In Figu e 4 he
unc ion ha pe o ms i is shown.
The unc ion ha con e s he da a ame o supe ised da a, using he
unc ion on Figu e 4, depends on he ype o p edic ion you wan o pe o m,
ha ing as pa ame e he minu es o he p edic ion.
19
Figu e 3: Explana ion o a iables needed o supe ised lea ning
Figu e 4: Func ion ha p epa es he da a o supe ised lea ning
20

Taking his a gumen , he unc ion calcula es how many columns a e
needed o c ea e he supe ised da a-se : always 24 columns o he pas 120
minu es o glucose and 24 columns o he las 120 minu es o insulin in ake.
I also c ea es columns o he u u e insulin in akes; his is use ul because i
is one o he main easons why glucose le el can all. I is also needed o c op
he da a-se since, a he beginning, we do no ha e pas alues o insulin,
and a he end, we do no ha e u u e alues.
4.3.5. The da a ame is p epa ed
A e his, he da a ame is al eady p epa ed o wo k wi h neu al ne wo ks,
bu some o he asks could be done o imp o e he pe o mance o he algo-
i hm. One o hem is he Da a augmen a ion, his means c ea ing o ally
new da a based on gi en da a ames. In he Medicine-AI ield his can be
eally use ul since one o he main p oblems is he lack o da a. C ea ing
new da a makes possible o wo k wi h mo e examples, hus c ea ing a mo e
accu a e neu al ne .
4.4. Glucose o ecas ing wi h nume ical da a
In his p ojec we a e dealing wi h a mul i- a iable eg ession p oblem, as
he goal is o p edic he glucose ha a diabe es pa ien can ha e in a de e -
mined ime. Th ough he da a con e sion o ime se ies, we can ans o m
he da a o ha e a supe ised da a-se . This means ha he algo i hm is
ained wi h he da a o he pa ien be o e he consul a ion o he le el o
glucose, and wi h he da a o he pa ien a e ha consul a ion ( igu e 5).
Due o some pa ien s ha e less da a pa ame e s han o he s, we had o
di ide hem and c ea e di e en models depending on he pa ame e s hey
ha e. Thanks o his, we can e alua e he aining pe o mance depending
on he ype o da a, and ge some conclusions.
21
Figu e 5: Columns o he supe ised da a-se
Hence, hese a e he ypes o model depending on he pa ame e s o each
pa ien :
•Pa ien s wi h insulin and basic alues (3 pa ien s)
–Glucose
–S eps
–Hea a e
–Calo ies
–Basal a e
•Pa ien s wi h only basic alues (10 pa ien s; he pa ien s wi h insulin
alues a e also alid o his aining)
–Glucose
–S eps
–Hea a e
–Calo ies
•Pa ien s om Ohio and 3 om Alcal´a de Hena es ha con ains:
–Glucose
–Basal a e
•All he pa ien s, bu only wi h he glucose alues.
–Glucose
22
The ini ial model we chose o ain he da a was a simple eed o wa d
decla ed wi h Ke as o es i he supe ised da a-se gene a ed was well
con igu ed.
Once checked ha he da a p e-p ocessing was p ope ly p epa ed, we
con igu ed he i s ype o neu al ne wo k o ain and es all he pa ien s.
The unc ion ha c ea es his NN can be seen on Figu e 6.
Figu e 6: C ea ion o he FF model
This ini ial neu al ne wo k is con igu ed wi h he inpu laye (da a supe -
ised), one hidden dense laye wi h a hype bolic angen ac i a ion unc ion
ha expec s ows o da a wi h he numbe o columns o he supe ised
da a-se , and one ou pu dense laye wi h one node.
•190 epochs
•Tahn ac i a ion unc ion
•Adam op imize
•Inpu shape: (1, numbe o columns o he da a-se )
•Mean absolu e e o as loss unc ion
A e es ing he models men ioned wi h he eed o wa d neu al ne , he
nex s ep was o modi y he con igu a ion o he model o see i i was pos-
sible o imp o e he esul s.
The e a e di e en a ia ions o neu al ne wo ks acco ding o he ype o
da a. Due o da a-se s a e ime se ies, we decided o es con igu a ions wi h
a ype o ecu en neu al ne wo k called LSTM (Long sho - e m memo y).
23
These ypes o neu al ne wo ks we e c ea ed o gi e a solu ion o he
sho - e m memo y. They ha e an in e nal mechanism called ga es ha can
egula e he low o in o ma ion. These ga es can lea n which da a in a
sequence is impo an o keep o h ow away. By doing ha , i can pass
ele an in o ma ion down he long chain o sequences o make p edic ions.
This is he con igu a ion o he new neu al ne :
model = Sequen ial ()
model . add (LSTM( numbe columns , inpu shape =(1 ,numbe columns ) ,
a c i a io n =’ elu ’ , e u n seq uence s = T ue ))
model . add ( Fla en ( ) )
model . add ( Dense (1 , ac i a io n =’ anh ’ ) )
model . compile ( l o s s =’ mean absolu e e o ’ , op imize =’Adam’ ,
me ics =[”mse ” , ’ accu acy ’ ] )
4.5. Glucose o ecas ing wi h wa ele s images
Fo his di e en app oach o glucose o ecas ing, we ace a classi ica ion
p oblem: we a e going o p edic whe he he pa ien is going o ha e a hy-
poglycemia episode in he nex pe iod based on images wi h da a om 6 o
24 hou s.
These images a e di ided in wo classes. Taking a pe iod om 6h o
24h (depends on he model), we will seek in he ollowing pe iod o alues
o glucose below 70 mg/dL. I ha pe iod has a leas one alue below 70
mg/dL, i will be conside ed as a Hypoglycemic pe iod, o he wise, i will be
conside ed as a No mal pe iod.
This ype o aining wi h images is some hing o ally new in he ield o
diabe es-AI, so i will be an expe imen al wo k.
To p epa e he da a o his sec ion some p e-p ocessing asks a e a oided
such as making he da a se supe ised.
To suppo he neu al ne s and i s aining we a e going o use N idia
Digi s, a pla o m ha eases he wo k wi h images.
24
Figu e 11: T aining page
–I he .cs does no ha e alues o glucose o e e y 5 minu es, he
missing ields need o be emp y in o de o ecognize i as NaN.
–Numbe s need o use decimal do .
•Use page: On he use sc een, shown in Figu e 12, he use pe sonal
in o ma ion is displayed. This includes: name, su name, email and
use name. Also, we display a small summa y o he las 5 aining
esul s ha he use has pe o med.
•Resul s page: This page is he one ha comes a e aining is pe -
o med. I shows he esul s o he aining and p edic ion wi h he
da a uploaded. The in o ma ion displayed, as seen in Figu e 13, is he
RMSE (Roo Mean Squa ed E o ), he Cla ke e o g id along wi h
he numbe o poin s on each sec ion and a g aph ha compa es he
eal a ge wi h ou p edic ion. We also make a single p edic ion o he
las 120 minu es in he da a-se , simula ing how he algo i hm would
wo k in eal li e usage.
31

Figu e 12: Use page
Figu e 13: Resul s page
32
4.6.1. Connec ion wi h da abase
Gi en he ac ha , in his websi e, he use c ea es new models, uploads da a
and he sys em gene a es esul s, i was necessa y o implemen a da abase
o keep s o ed all his in o ma ion. Fu he mo e, due o he e was an im-
plemen ed da abase om o he inal deg ee p ojec ha was unc ioning, we
needed o ollow a es ablished syn ax o c ea e a da abase which was adap -
able o he exis ing one.
The needed ables o inco po a e o he sys em a e he ollowing:
•Use ables ha Django c ea es au oma ically. We a e using he able
’au h-use ’ o use s managemen .
•Files ables whe e he sys em keeps he uploaded iles o he use , wi h
he use id.
•Models able ha sa es he ype o model, he las use ha ained
i , whe he i is pe sonal o gene al and he model RMSE.
•Resul s able o keep he esul s o a p edic ion (RMSE, zones o he
cla ke g aphic, use id).
To c ea e he da abase, we used MySQL Wo kbench and Se e , which
a e eally in ui i e o manage he di e en componen s o he da a s uc u e.
33
5. Expe imen al Resul s
5.1. Me ics
In o de o measu e he e iciency and beha iou o he model, we a e going
o use he ollowing me ics:
•Cla ke e o g id
This Cla ke g aphic isualizes black poin s ha ep esen he eal es
a ge (axis x) and he p edic ed alues (axis y). In a i s ins ance,
he close he poin s a e alloca ed a ound he diagonal (which means
ha p edic ions a e equals o eal es se ), he mo e p ecise he model
will be.
This Cla ke g aphic displays di e en zones: A,B,C,D,E. These zones
ep esen he alidi y o he p ecision. So he a ge o he model is o
ha e mos o he poin wi hin he zones A and B, and as ew poin s
as possible in he zones C, D and E. An example o an emp y Cla ke
E o G id can be seen in Figu e 14.
Figu e 14: Cla ke E o G id
•RMSE
To ha e mo e measu emen s apa om he Cla ke image and he num-
be o poin s in each zone, we calcula e he RMSE i.e oo mean squa e
34
e o , ollowing he o mula in Figu e 15. Due o he RMSE is calcu-
la ed o e glucose alues ha usually ange be ween 40 and 500 ml/dl,
a good RMSE should no exceed 30 mg/dl.
Figu e 15: Roo Mean Squa ed E o o mula
5.2. Resul s
In his sec ion we a e going o documen all he esul s ob ained a e es ing
di e en echniques o p edic glucose alues.
5.2.1. Fi s execu ion wi h Pa ien 1
Ou i s con ac wi h a eal p edic ion was aining he Pa ien 1 om he
Hospi al Uni e si a io P ´ıncipe de As u ias. The pa ien da a-se con ained
3824 ows o da a a e p e-p ocessing, enough o spli i in o a ain and es
se s o y ou a i s execu ion.
The i s con igu a ion o gene a e he supe ised da a wi h ime se ies
(sec ion 4.3.4) was wi h 30 minu es, i.e, he ain da a con ains all he in o -
ma ion o he pa ien 120 minu es be o e he ime 0, and he a ge a e all
he glucose alues a e 30 minu es.
We used he i s con igu a ion o he model (Feed o wa d desc ibed in
sec ion 4.3) o es a i s execu ion. The esul s can be seen in Figu e 16.
In his i s p edic ion o e he es se om he Pa ien 1, mos o he
poin s in he Cla ke e o g id all in zone A, bu some poin s a e in he
c i ical a eas.
Howe e , he conclusion o his i s execu ion is ha , conside ing his is
a simple model and we only ained 70% o he da a-se o one pa ien , he
esul s a e p omising wi h his ype o supe ised da a, wi h he ime se ies
35
Figu e 16: Resul s o he i s aining wi h he supe ised da a-se
and wi h he pa ame e s ained.
This execu ion is documen ed in he ile ’Hello diabe es.pynb’
5.2.2. Desc ip ions o he esul s
To s udy he di e en esul s, we use he ollowing dynamic:
1. C ea e he model adap ed o he ype o da a
#T ained wi h pa ie n s ha con ains i n s u l i n alues
model 30 i nsulin = c ea modeloFF ( numbe columns )
2. T ain he model o each pa ien and sa e he model
execu e (30 , p a 7 i ns uli n , model 30 insulin , T ue )
execu e (30 , p a 8 i ns uli n , model 30 insulin , T ue )
execu e (30 , p a 9 i ns uli n , model 30 insulin , T ue )
model 30 i nsulin . sa e ( ’ Models/ model 30 insulin . h5 ’ )
36

3. C ea e a new model o each pa ien and compa e he esul s o he
new model wi h hose o he p e- ained model
p in (” Pa ien 7 No a in in g ”)
execu e (30 , p a 7 i ns ul in , c ea modeloFF (61) , T ue )
p in (” Pa ien 7 T ained ”)
execu e (30 , p a 7 i ns ul in , ained modedl 30 insulin , False )
5.2.3. T ain wi h Feed Fo wa d
The nex s ep was o y his basic model wi h mo e pa ien s, and s udy i
aining a single model wi h mo e han one di e en pa ien e u ns be e
esul s. The e o e, we needed o au oma e he p ocess o p epa ing he pa-
ien s ( ill NaN alues, cu he da a-se only wi h he ows wi h da a, d op
useless columns...).
All hese es s and s udies a e collec ed in he ile ’Model Dense laye
(o iginal FF).pynb’. We a e using all he pa ien s wi h da a om Hospi al
Uni e si a io P ´ıncipe de As u ias and om Ohio’s Hospi al.
These i s de ailed esul s a e o p edic ion o glucose o 30 minu es:
1. Model wi h insulin and basic alues
Wi h his p ocedu e we we e able o see no able di e ences. In he i s
execu ion wi h pa ien s con aining basic alues and insulin (3 pa ien s),
we could app ecia e imp o emen s:
Fo ins ance, wi h ou pa ien agged as Pa ien 8, we go he ollowing
esul s:
•Model wi h no aining:
Zone A: 880 (76.78883071553228%)
Zone B: 246 (21.465968586387437%)
Zone C: 7 (0.6108202443280977%)
Zone D: 13 (1.1343804537521813%)
Zone E: 0 (0.0%)
RMSE: 32.800755
•Model p e- ained wi h o he pa ien s:
37
Zone A: 925 (80.7155322862129%)
Zone B: 198 (17.277486910994764%)
Zone C: 0 (0.0%)
Zone D: 23 (2.006980802792321%)
Zone E: 0 (0.0%)
RMSE: 23.619013
Figu e 17: New FF model wi h glu-
cose, basic ales and insulin
Figu e 18: P e- ained FF model
wi h glucose, basic ales and insulin
As seen in bo h igu es, he poin s in he igu e 17 (no p e- ained
model) a e mo e sca e ed han in he igu e 18 wi h he p e- ained
model. We can app ecia e ha he p e- ained model ends o g oup
he poin s. O e all, p e- ained models wi h di e en pa ien s wi h in-
sulin and basic alues ob ain mo e accu a e esul s.
2. Model wi h basic alues
Then, we p o ed wi h he same dynamic explained abo e o ain and
p epa e a model wi h pa ien s wi h only basic alues (glucose, s eps,
hea a e and calo ies). This model was ained wi h 10 pa ien s, so
he expec a ions we e highe .
The model had a no able imp o emen ; when compa ing all he Cla ke
38
images gene a ed in he no ebook o 10 pa ien s, da a had isible
changes and he goal o g oup he poin s in he zone A was almos
accomplished.
Le ’s see an example wi h he pa ien agged as Pa ien 5:
Figu e 19: New model o glucose
and basic alues wi h FF
Figu e 20: P e- ained model o glu-
cose and basic alues wi h FF
In Figu e 19 we can see ha , wi h a new model, he e a e many sca -
e ed poin s, bu wi h he p e- ained model ( igu e 20 he dis ibu ion
o he poin s is eally ema kable. The RMSE educes i s alue om
36.3 o 20.8. The conclusion wi h his model is ha he insulin alue
is no as ele an as he numbe o pa ien s and da a o ain. Fu he -
mo e, hese e idences con i m ha i is possible o ha e good esul s
wi h models ained wi h di e en pa ien s when pe sonal models a e
no a ailable.
3. Model wi h glucose and insulin
A e wa ds, we c ea ed a new model o he Ohio’s Hospi al pa ien s
in which, a e p e-p ocessing hei da a, we only kep he glucose and
insulin alues. The da a-se s om hese pa ien s con ain much mo e
da a han he Spanish pa ien s. A e aining he model wi h all he
pa ien s, and compa ing he e iciency o he model wi h he me hod
39
desc ibed be o e, we no iced ha , wi h his da a-se o pa ien s, he e is
no imp o emen s compa ed wi h he o he s. Also, we could app ecia e
ha only a ew ha e pe sonal beha iou s, so he esul s wi h a gene al
model a e almos he same. Fu he mo e, as in he o he s models, we
igu ed ou ha he ull- ained model has he endency o inc ease he
poin s ha all in zone A (Figu e 21 and 22).
Example wi h ou pa ien agged as Pa ien 6 om Ohio’s Hospi al:
Figu e 21: New model o glucose
and insulin wi h FF
Figu e 22: P e- ained model o glu-
cose and insulin wi h FF
RMSE new model = 19.19
RMSE p e− ained model = 14.13
4. Model wi h glucose
Finally, we c ea ed a gene al model ha can be used o any pa ien ,
because i jus equi es glucose alues. This model was ained wi h
all he pa ien s da a-se s we had access o, excep 3 pa ien s ha we
ese ed o es he model. Ins ead o compa ing he esul s wi h a
new model and he ained model o e e y pa ien , we a e going o
es he beha iou o he model wi h h ee pa ien s ha we e no used
o ain he model. This means ha we we e p edic ing he glucose
40
Figu e 29: New model o glucose
and basic alues wi h LSTM
Figu e 30: P e- ained model o glu-
cose and basic alues wi h LSTM
3. Model wi h glucose and insulin
O e all, he esul s wi h his model a e be e han wi h he Feed Fo -
wa d one, bo h he RMSE alue and in he Cla ke image. Howe e , i ’s
impo an o men ion ha a pa ien (pa ien agged as Pa ien 5) o
his se had a peculia conduc o his glucose le el ha s ands ou o e
he o he s; ha ’s he eason why using a new model o his pa ien
e u ns be e esul s han using a p e- ained model. None heless, he
esul s o bo h models a e eally simila .
Resul s o pa ien agged as Pa ien 6:
Wi h new model:
Zone A: 3665 (89.9607265586647%)
Zone B: 396 (9.72017673048601%)
Zone C: 0 (0.0%)
Zone D: 13 (0.31909671084928815%)
Zone E: 0 (0.0%)
RMSE: 16.55
Wi h p e- ained model:
47

Zone A: 3849 (94.47717231222386%)
Zone B: 217 (5.326460481099656%)
Zone C: 0 (0.0%)
Zone D: 8 (0.19636720667648502%)
Zone E: 0 (0.0%)
RMSE: 13.29
4. Model wi h glucose
Finally, a e aining all he pa ien s only wi h he glucose pa ame e ,
he esul s a e a bi be e han wi h he Feed Fo wa d model, bu
no hing ema kable.
Resul s o pa ien agged as Pa ien 14 (Figu e 31):
Figu e 31: Resul s o es ing Pa ien 14 wi h ou 30 min LSTM model ha wo ks only
wi h glucose alues
Zone A: 3394 (88.98793917147351%)
Zone B: 388 (10.17304667016256%)
Zone C: 0 (0.0%)
Zone D: 32 (0.8390141583639223%)
Zone E: 0 (0.0%)
48
RMSE: 17.001036
The conclusion a e compa ing hese wo models is ha he Feed Fo wa d
model has a pa icula beha iou and endency so, a he ime o aining
a pa icula pa ien ha has a da a-se ha s ands ou o e he o he s, he
p edic ions a e wo se. On he o he hand, he LSTM adap s o he ype o
pa ien , making i mo e eliable and p ecise.
A e es ing and explaining in de ail all he di e en models wi h a 30
minu es o ecas ing, we a e going o show he esul s o he glucose p edic ion
o 60, 90 and 120 minu es. In his case, we a e p esen ing only he glucose
model, which e u ns he bes esul s. We a e using ou pa ien agged as
Pa ien 14 o compa e esul s:
•Resul s o he glucose p edic ion o 60 minu es (Figu e 32):
Zone A: 2760 (72.47899159663865%)
Zone B: 922 (24.212184873949578%)
Zone C: 0 (0.0%)
Zone D: 126 (3.308823529411765%)
Zone E: 0 (0.0%)
RMSE: 28.71
49
Figu e 32: Resul s o es ing pa ien 14 wi h ou 60 min LSTM model ha wo ks only
wi h glucose alues
•Resul s o he glucose p edic ion o 90 minu es (Figu e 33):
Zone A: 2245 (59.047869542346135%)
Zone B: 1387 (36.48079957916886%)
Zone C: 8 (0.21041557075223566%)
Zone D: 160 (4.208311415044713%)
Zone E: 2 (0.052603892688058915%)
RMSE: 38.74
•Resul s o he glucose p edic ion o 120 minu es (Figu e 34):
Zone A: 2081 (54.82086406743941%)
Zone B: 1496 (39.409905163329825%)
50
Figu e 33: Resul s o es ing pa ien 14 wi h ou 90 min LSTM model ha wo ks only
wi h glucose alues
Zone C: 6 (0.15806111696522657%)
Zone D: 211 (5.558482613277134%)
Zone E: 2 (0.052687038988408846%)
RMSE: 42.56
O e all, he esul s compa ed o he Feed Fo wa d a e eally simila , bu
he aining ime o all his models is much longe .
51
Figu e 34: Resul s o es ing pa ien 14 wi h ou 120 min LSTM model ha wo ks only
wi h glucose alues
52

5.2.5. T ain wi h NVIDIA Digi s
Gi en he ac ha all he di e en con igu a ions and NN e u ned simila
esul s, we a e going o commen one o hem wi hou speci ic de ails o he
con igu a ion. The NN used o his epo is AlexNe ne wo k wi h Adam
(Adap i e Momen Es ima ion) sol e / RMSp op, aining du ing 50 epochs
wi h a base lea ning a e o 0.5.
Figu e 35: P ocess o aining wi h all he images on N idia Digi s
In i s ins ance, we saw e y p omising esul s wi h a ound 78% o accu-
acy. Howe e , when we wn in o u he de ail, we no iced ha he accu acy
emained always a ound he same alue du ing he whole aining. To p o e
he accu acy o he neu al ne , we ied 2 di e en p edic ions om an Hypo
class image and a No mal class image. Fo bo h images he esul was he
same, 78% o No mal in p edic ion. F om his p edic ions we could obse e
ha he neu al ne had an s a is ic in luence, 78 % o he images a e No mal.
To p o e ha he neu al ne only used s a is ical c i e ia o p edic ion, we
decided o ain ano he neu al ne wo k wi h mo e balanced da a inpu , 64%
o No mal images and 36% o Hypo images.
53
Figu e 36: P ocess o aining wi h image-da a balanced on N idia Digi s
This las esul , shown in Figu e 36 could con i m ou i s hypo hesis:
he NN is no able o ind any cha ac e is ic o di e en ia e be ween he wo
classes.
The e is s ill a lo o wo k o do un il we each good esul s. A e ou
wo k, we can see wo clea ways o ind a unc ional model.
•Explo e o he ways o ep esen he da a wi h images.
•T y ou di e en NN, p e- ained models ocused on medical da a anal-
ysis and apply o he echniques such as ans e ed lea ning.
54
6. Conclusions and Fu u e wo k
6.1. Gene al conclusions
A e ha ing ca ied ou he c ea ion o a NN model o p edic glucose le -
els and a websi e se ice whe e we can sha e ou algo i hm wi h pa ien s
o diabe es o ype 1, we eel sa is ied o ha e de eloped a i s e sion o
a inal se ice ha encompasses bo h a esea ch wo k o glucose p edic ion
(co e ing di e en scopes in his sec o ) and an in ui i e web in e ace o
he use .
In he s age o p e-p ocessing, we lea ned he di e en beha iou s o di-
abe es pa ien s and we amilia ized wi h hei le els o glucose, disco e ing
which pa ame e s we e he bes o an accu a e p edic ion.
We also ealized ha much o he pa ien s pa ame e s gi en we e no
eally necessa y o ob ain good p edic ions o glucose, so basic alues as he
glucose, s eps and ew mo e pa ame e s easy o measu e a e enough o ge a
p ecise esul . The e o e, i he pa ien has an app op ia e sys em o measu e
he le els o glucose, he will be able o ob ain good p edic ions.
Also, as seen on he esul s epo , one concep we lea ned wi h his
s udy is he ollowing quo e: ’ he glucose is he ’mi o ’ o o he diabe es
pa ame e s, and he a ia ions o each pa ame e a e e lec ed in he le els
o glucose’. Hence, i he pa ien has a egula conduc ela ed o physical
exe cise, ood..., he pa ien should ob ain an accu a e p edic ion based only
on he e olu ion o i s le els o glucose.
A e es ing all he models implemen ed wi h Ke as, he idea o p edic -
ing glucose beha iou wi h images came up (J. Ignacio Hidalgo’s idea). We
we e inexpe ienced in his ype o aining using he NVIDIA Digi s and,
al hough we didn’ ob ain good esul s because i ’s a new ield ha needs
u he esea ch, we acqui ed some knowledge abou image aining and ypes
o ep esen a ions o he diabe es pa ame e s. Also, we disco e ed ha i is
a eally good sys em o e u n p edic ion esul s, because his se ice has a
55
eally in ui i e and isual in e ace. Howe e , as we we e mo e amilia wi h
he Py hon sc ip s and we we e incapable o ob aining sa e and eliable e-
sul s wi h NVIDIA Digi s, we decided o use he de eloped models in Py hon.
Ano he aspec o highligh is ha his ull p ojec was ealized a he dis-
ance, because bo h o us had an E asmus s uden ship, and he hole p ojec
was de eloped h ough ideo calls om Leiden, Ne he lands (Al a o’s des i-
na ion) and Mil´an, I aly (Alejand o’s des ina ion). Also we had weekly calls
wi h ou u o J. Ignacio Hidalgo, and we didn’ expe ience p oblems and
ealized ha i is possible o wo k a om each o he .
O e all, we acqui ed aluable knowledge ela ed o di e se ields o com-
pu e science such as: da a om diabe es pa ien s, co ela ion o da a wi h
glucose, ypes o p edic ions o glucose o ecas ing, how o c ea e a websi e
wi h Django and inco po a e he implemen ed sc ip s, how o connec a SQL
da abase o sa e he in o ma ion, e c. In conclusion, we conside ha his
inal deg ee p ojec is he esul and ep esen a ion o ou yea s o lea ning.
6.2. Fu u e wo k
Since he app oach o he p ojec is e y gene al and he main objec i e was
o c ea e a ully wo king online applica ion ha p o ides use ul in o ma ion
o pa ien s, we could no ge deep on each one o he ask in ol ing GlucNe .
Ob iously, he e is a lo o in es iga ion emaining on he s uc u e o
he neu al ne s. Wi h he exis ing echnologies, di e en lib a ies and ypes
o NN i is no possible o y ou e e y op ions. Anyways, his is he wo k
ha needs o be done now on, wi h he suppo o he cu en websi e. Also,
o he echniques in he glucose p edic ion ield can be s udied in o de o
gi e hem suppo on GlucNe .
In he u u e, he idea is o apply di e en o ecas ing and ensembles
echniques such as KNN, andom o es , g adien boos ing... and compu e
mo e me ics o ha e a be e e alua ion o he esul s.
Fu he mo e, he aining and classi ica ion o wa ele s images is a ield
which needs ime o be s udied o accomplish easonable esul s. Also, in
he u u e, he idea is o add he esul s o NVIDIA Digi s o he websi e
56
10. Bibliog aphy
•Tenso - low use guide
•Ke as use guide (h ps://ke as.io/guides/)
•N idia Digi s use guide
•Django use guide
•h ps:// ude .io/op imizing-g adien -descen /index.h ml
•h ps://dia ibe.o g/unde s anding-a e age-glucose-s anda d-de ia ion-
c -and-blood-suga - a iabili y
•h ps:// owa dsda ascience.com/ he-mos ly-comple e-cha -o -neu al-ne wo ks-
explained-3 b6 2367464
•S acko e low
•Mul iple o he sou ces
Re e ences
[1] J. I. Hidalgo, J. M. Colmena , G. K onbe ge , S. M. Winkle , O. Ga -
nica, J. Lancha es, Da a based p edic ion o blood glucose concen a-
ions using e olu iona y me hods, J. Medical Sys ems 41 (2017) 142.
[2] C. P´e ez-Gand´ıa, A. Facchine i, G. Spa acino, C. Cobelli, E. G´omez,
M. Rigla, A. de Lei a, M. He nando, A i icial neu al ne wo k algo-
i hm o online glucose p edic ion om con inuous glucose moni o ing,
Diabe es echnology & he apeu ics 12 (2010) 81–88.
[3] V. T esp, T. B iegel, J. Moody, Neu al-ne wo k models o he blood
glucose me abolism o a diabe ic, IEEE T ansac ions on Neu al ne wo ks
10 (1999) 1204–1213.
[4] F. Allam, Z. Nossai, H. Gomma, I. Ib ahim, M. Abdelsalam, A ecu en
neu al ne wo k app oach o p edic ing glucose concen a ion in ype-
1 diabe ic pa ien s, in: Enginee ing Applica ions o Neu al Ne wo ks,
Sp inge , 2011, pp. 254–259.
63

[5] Q. Sun, M. V. Janko ic, L. Bally, S. G. Mougiakakou, P edic ing blood
glucose wi h an ls m and bi-ls m based deep neu al ne wo k, in: 2018
14 h Symposium on Neu al Ne wo ks and Applica ions (NEUREL),
IEEE, 2018, pp. 1–5.
[6] E. M. Aiello, G. Lisan i, L. Magni, M. Musci, C. To anin, The apy-
d i en deep glucose o ecas ing, Enginee ing Applica ions o A i icial
In elligence 87 (2020) 103255.
[7] T. El Id iss, A. Id i, I. Abnane, Z. Bakkou y, P edic ing blood glu-
cose using an ls m neu al ne wo k, in: 2019 Fede a ed Con e ence on
Compu e Science and In o ma ion Sys ems (FedCSIS), IEEE, 2019, pp.
35–41.
[8] C. Meijne , S. Pe sson, Blood Glucose P edic ion o Type 1 Diabe es us-
ing Machine Lea ning Long Sho - e m Memo y based models o blood
glucose p edic ion, Mas e ’s hesis, 2017.
[9] J. Ma insson, A. Schliep, B. Eliasson, C. Meijne , S. Pe sson, O. Mo-
g en, Au oma ic blood glucose p edic ion wi h con idence using ecu -
en neu al ne wo ks, in: 3 d In e na ional Wo kshop on Knowledge
Disco e y in Heal hca e Da a, KDH@ IJCAI-ECAI 2018, 13 July 2018,
2018, pp. 64–68.
[10] S. Mi sheka ian, R. Bunescu, C. Ma ling, F. Schwa z, Using ls ms o
lea n physiological models o blood glucose beha io , in: 2017 39 h
Annual In e na ional Con e ence o he IEEE Enginee ing in Medicine
and Biology Socie y (EMBC), IEEE, 2017, pp. 2887–2891.
[11] S. Mi sheka ian, H. Shen, R. Bunescu, C. Ma ling, Ls ms and neu al a -
en ion models o blood glucose p edic ion: Compa a i e expe imen s
on eal and syn he ic da a, in: 2019 41s Annual In e na ional Con e -
ence o he IEEE Enginee ing in Medicine and Biology Socie y (EMBC),
IEEE, 2019, pp. 706–712.
[12] M. Mayo, L. Chepulis, R. G. Paul, Glycemic-awa e me ics and o e -
sampling echniques o p edic ing blood glucose le els using machine
lea ning, PloS one 14 (2019).
64
[13] K. Li, J. Daniels, C. Liu, P. He e o, P. Geo giou, Con olu ional ecu -
en neu al ne wo ks o glucose p edic ion, IEEE jou nal o biomedical
and heal h in o ma ics 24 (2019) 603–613.
[14] A. Be achi, L. Biagi, I. Con e as, N. Luo, J. Veh´ı, P edic ion o blood
glucose le els and noc u nal hypoglycemia using physiological models
and a i icial neu al ne wo ks., in: KHD@ IJCAI, 2018, pp. 85–90.
[15] K. S. Eljil, G. Qadah, M. Pasquie , P edic ing hypoglycemia in diabe ic
pa ien s using ime-sensi i e a i icial neu al ne wo ks, In e na ional
Jou nal o Heal hca e In o ma ion Sys ems and In o ma ics (IJHISI) 11
(2016) 70–88.
[16] X. Mo, Y. Wang, X. Wu, Hypoglycemia p edic ion using ex eme lea n-
ing machine (elm) and egula ized elm, in: 2013 25 h Chinese Con ol
and Decision Con e ence (CCDC), IEEE, 2013, pp. 4405–4409.
[17] P. P. San, S. H. Ling, H. Nguyen, e al., E ol able ough-block-based
neu al ne wo k and i s biomedical applica ion o hypoglycemia de ec ion
sys em, IEEE ansac ions on cybe ne ics 44 (2013) 1338–1349.
65