Predicting Dengue Fever Incidence and Disease Dynamics under Climate Change in Southeast Asia
Abstract
Dengue fever is a climate-sensitive vector-borne disease primarily transmitted by Aedes mosquitoes, A. aegypti and A. albopictus. Previous research has analyzed the relationship between climate and disease, with varying outcomes. Temperature and precipitation have been demonstrated as relevant predictors in most studies. The effects of climate change on dengue fever were found to be uncertain, highlighting the need for further study. This study analyzed how environmental variables interact with disease transmission, enabling predictive modeling to forecast dengue incidence. For deployment, climate change simulations were used as a framework to assess the disease’s response to changing environmental factors. The incidence and environmental data for 17 Southeast Asian locations were collected from the national Ministries of Health and the National Oceanic and Atmospheric Administration (NOAA) from 2016-2023. Traditional machine learning and deep learning models were used to forecast dengue incidence based on ten input features of temperature, precipitation, and lagged observations. The predictive ability was evaluated using Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). Deep and machine learning models showed similar results for predicting dengue incidence. The Convolutional Neural Network (CNN) achieved the lowest error with an average MAE of 10.10 and RMSE of 13.61 on the validation set. Models showed varying predictive abilities across locations. Despite extensive data preparation, some locations performed worse on all models, indicating potential issues with initial data quality. Errors were reduced for all models on the test set, with CNN demonstrating superior with an average MAE of 5.06 and RMSE of 7.09. Although errors decreased with additional data, model performance could benefit from additional variables that were not included. Lastly, CNN was deployed to assess the disease’s response to climate change. The predicted model was found to be sensitive to simulated changes in total precipitation and mean temperature. Results show positive and negative changes in the annual incidence rates for both emission scenarios, with a positive linear trend observed for mean temperature.
Full text
Mas e Deg ee P og am in
Da a Science and Ad anced Analy ics
P edic ing Dengue Fe e Incidence and Disease Dynamics unde
Clima e Change in Sou heas Asia
Josephine Lu e
Mas e Thesis
p esen ed as pa ial equi emen o ob aining a Mas e ’s Deg ee in Da a Science and Ad anced Analy ics
NOVA In o ma ion Managemen School
Ins i u o Supe io de Es a ís ica e Ges ão de In o mação
Uni e sidade No a de Lisboa
MDSAA
NOVA In o ma ion Managemen School
Ins i u o Supe io de Es a ís ica e Ges ão de In o mação
Uni e sidade No a de Lisboa
P edic ing Dengue Fe e Incidence and Disease Dynamics unde Clima e Change in
Sou heas Asia
by
Josephine Lu e
Mas e Thesis p esen ed as pa ial equi emen o ob aining he Mas e ’s deg ee in Da a
Science and Ad anced Analy ics, wi h a specializa ion in Business Analy ics
Supe ised by
Robe o Hen iques, PhD, No a In o ma ion Managemen School
July, 2024
i
STATEMENT OF INTEGRITY
I he eby decla e ha ing conduc ed his academic wo k wi h in eg i y. I con i m ha I ha e no
used plagia ism o any o m o undue use o in o ma ion o alsi ica ion o esul s along he
p ocess leading o i s elabo a ion. I u he decla e ha I ha e ully acknowledged he Rules
o Conduc and Code o Hono om he NOVA In o ma ion Managemen School.
[Be lin, July 1s 2024]
ii
ABSTRACT
Dengue e e is a clima e-sensi i e ec o -bo ne disease p ima ily ansmi ed by Aedes
mosqui oes, A. aegyp i and A. albopic us. P e ious esea ch has analyzed he ela ionship
be ween clima e and disease, wi h a ying ou comes. Tempe a u e and p ecipi a ion ha e
been demons a ed as ele an p edic o s in mos s udies. The e ec s o clima e change on
dengue e e we e ound o be unce ain, highligh ing he need o u he s udy.
This s udy analyzed how en i onmen al a iables in e ac wi h disease ansmission, enabling
p edic i e modeling o o ecas dengue incidence. Fo deploymen , clima e change
simula ions we e used as a amewo k o assess he disease’s esponse o changing
en i onmen al ac o s. The incidence and en i onmen al da a o 17 Sou heas Asian loca ions
we e collec ed om he na ional Minis ies o Heal h and he Na ional Oceanic and
A mosphe ic Adminis a ion (NOAA) om 2016-2023. T adi ional machine lea ning and deep
lea ning models we e used o o ecas dengue incidence based on en inpu ea u es o
empe a u e, p ecipi a ion, and lagged obse a ions.
The p edic i e abili y was e alua ed using Mean Absolu e E o (MAE) and Roo Mean Squa ed
E o (RMSE). Deep and machine lea ning models showed simila esul s o p edic ing dengue
incidence. The Con olu ional Neu al Ne wo k (CNN) achie ed he lowes e o wi h an a e age
MAE o 10.10 and RMSE o 13.61 on he alida ion se . Models showed a ying p edic i e
abili ies ac oss loca ions. Despi e ex ensi e da a p epa a ion, some loca ions pe o med
wo se on all models, indica ing po en ial issues wi h ini ial da a quali y. E o s we e educed
o all models on he es se , wi h CNN demons a ing supe io wi h an a e age MAE o 5.06
and RMSE o 7.09. Al hough e o s dec eased wi h addi ional da a, model pe o mance could
bene i om addi ional a iables ha we e no included.
Las ly, CNN was deployed o assess he disease’s esponse o clima e change. The p edic ed
model was ound o be sensi i e o simula ed changes in o al p ecipi a ion and mean
empe a u e. Resul s show posi i e and nega i e changes in he annual incidence a es o
bo h emission scena ios, wi h a posi i e linea end obse ed o mean empe a u e.
KEYWORDS
Dengue Fe e ; Incidence Fo ecas ; Deep Lea ning, Machine Lea ning; Clima e Change
Sus ainable De elopmen Goals (SDG):
iii
TABLE OF CONTENTS
1. In oduc ion .................................................................................................................. 1
2. Li e a u e Re iew.......................................................................................................... 3
2.1. Dengue T ansmission and Vec o .......................................................................... 3
2.2. En i onmen al Va iables ....................................................................................... 5
2.3. Clima e Change and Ex eme Wea he E en s ..................................................... 6
2.4. Summa y ................................................................................................................ 8
3. Me hodology ..............................................................................................................10
3.1. In oduc ion .........................................................................................................10
3.2. Domain Unde s anding .......................................................................................11
3.3. Da a Unde s anding.............................................................................................12
3.4. Da a P epa a ion .................................................................................................14
3.4.1. Da a P ep ocessing .......................................................................................14
3.4.2. Fea u e Enginee ing .....................................................................................19
3.4.3. Fea u e Selec ion ..........................................................................................19
3.5. Modeling ..............................................................................................................20
3.5.1. Model De elopmen .....................................................................................20
3.5.2. Model E alua ion..........................................................................................23
3.6. E alua ion ............................................................................................................23
3.7. Deploymen .........................................................................................................24
4. Resul s and Discussion ................................................................................................25
4.1. En i onmen al D i e s .........................................................................................25
4.2. P edic i e Modeling.............................................................................................25
4.3. Clima e Change Impac Assessmen ...................................................................27
5. Conclusions .................................................................................................................29
6. Limi a ions and Recommenda ions o Fu u e Wo k .................................................30
Re e ences .......................................................................................................................32
Appendix A ......................................................................................................................36
Appendix B.......................................................................................................................37
Appendix C.......................................................................................................................38
Appendix D ......................................................................................................................39
Appendix E .......................................................................................................................40
Appendix F .......................................................................................................................41
Appendix G ......................................................................................................................42
Appendix H ......................................................................................................................43
i
Appendix I ........................................................................................................................44
Appendix J .......................................................................................................................45
Appendix K .......................................................................................................................46
Appendix L .......................................................................................................................48
Appendix M .....................................................................................................................50
LIST OF FIGURES
Figu e 1 - CRISP-DM me hodology ........................................................................................... 10
Figu e 2 - De ailed o e iew o applied me hods .................................................................... 11
Figu e 3 - Domain objec i es .................................................................................................... 12
Figu e 4 - Da a mining goals ..................................................................................................... 12
Figu e 5 - Singapo e: MICE impu a ion o he empe a u e a iables ................................... 16
Figu e 6 - Singapo e: MICE impu a ion o he p ecipi a ion and empe a u e a iables ...... 17
Figu e 7 - Manila, Philippines: Po en ial missing alues .......................................................... 18
Figu e 8 - Holis ic model: Non-aligned ime-based spli .......................................................... 20
Figu e 9 – Holis ic model: Aligned ime-based spli ................................................................. 21
Figu e 10 - Loca ion-speci ic ime-based spli .......................................................................... 22
Figu e 11 - Simula ed changes in mean empe a u e and annual incidence a e ................... 27
Figu e 12 - Simula ed changes in o al p ecipi a ion and annual incidence a e .................... 28
i
LIST OF TABLES
Table 1 - Da a collec ion sou ces ............................................................................................. 14
Table 2 - T ans o med en i onmen al a iables ...................................................................... 15
ii
LIST OF ABBREVIATIONS AND ACRONYMS
A. Aedes
AR5 Fi h Assessmen Repo
CNN Con olu ional Neu al Ne wo k
CRISP-DM C oss-Indus y S anda d P ocess o Da a Mining
DENV Dengue-Vi us
DT Decision T ee
EIP Ex insic incuba ion pe iod
EMA Exponen ial Mo ing A e age
FNN Feed o wa d Neu al Ne wo k
GRU Ga ed Recu en Uni
IPCC In e go e nmen al Panel on Clima e Change
IQR In e qua ile Range
KNN K-Nea es Neighbo
LOCF Las Obse a ion Ca ied Fo wa d
LSTM Long Sho -Te m Memo y
LSTM-ATT Long Sho -Te m Memo y wi h A en ion
MAE Mean Absolu e E o
MCAR Missing Comple ely a Random
MICE Mul i a ia e Impu a ion by Chained Equa ions
MLP Mul ilaye Pe cep on
NOAA Na ional Oceanic and A mosphe ic Adminis a ion
NOCB Nex Obse a ion Ca ied Backwa d
RCP Rep esen a i e Concen a ion Pa hways
RF Random Fo es
RFE Recu si e Fea u e Elimina ion
6
he b eeding p ocess (Benedum e al., 2018). Re iewed li e a u e showcased he disease’s
sensi i i y o clima e a iabili y and he complexi y o concluding disease associa ions.
Small en i onmen al changes a ec disease ansmission, while ising empe a u es expand
he ec o habi a bu cons ain i us i ali y, mosqui o ep oduc ion, and de elopmen in
al eady wa m egions (Li e al., 2018). Seah e al. (2021) examined he impac o maximum
ambien empe a u e and hea wa es in Singapo e. Acco dingly, he ela i e isk o dengue
in ec ions linea ly inc eases wi h ising maximum empe a u es bu exhibi s a non-linea
dec ease beyond a h eshold o 31 °C. Highe empe a u es may enhance ansmission due o
inc eased bi ing a es and a sho e i us EIP. Howe e , p olonged high empe a u es educe
mosqui o li espan and egg- o-adul su i al, leading o a decline in he Aedes popula ion and
lowe dengue ansmission isk o e ime. Simila esul s we e epo ed o hea wa es.
Hea wa es commonly las 2-3 days, wi h 1-2 hea days pe week ha ing no e ec , while 3-6
hea wa e days lead o a dec ease in epo ed dengue in ec ions. Du ing ex eme hea , people
may spend mo e ime indoo s, educing ec o -hos con ac and lowe ing ansmission isk.
Con a y o o he esea ch, hei s udy iden i ied ha he impac o cumula i e p ecipi a ion
on dengue incidence is cons ained due o he signi icance o non-na u al b eeding si es (Seah
e al., 2021). To summa ize, empe a u e has a di ec and indi ec in luence on disease
ansmission.
Findings emphasize he need o a ime se ies analysis wi h lagged obse a ions o unde s and
delayed disease esponse. Wang e al. (2022) explo ed sho - e m associa ions be ween
ex eme wea he e en s and dengue e e in ec ion isk in Sou h and Sou heas Asia.
Ex emely low empe a u es educed ansmission isk wi h a 1-3 week lag bu no o
ex ended lag obse a ions. Ex emely high empe a u es we e associa ed wi h an inc eased
in ec ion isk wi h a lag o 2-3 weeks, depending on he numbe o ex emely ho days in a
week. Ex eme ain all was associa ed wi h a signi ican dec ease in in ec ion isk wi h a lag o
0-4 weeks, wi h he lowes isk a a lag o 2 weeks and an inc easing isk deno ed a a lag o 7
weeks. Thei esea ch concluded he impo ance o unde s anding he sho - e m e ec s o
clima e a iabili y on dengue ansmission (Wang e al., 2022). In con as , long- e m o ecas s
we e associa ed wi h inaccu a e p edic ions due o dec eased accu acy (Nu aini e al., 2021),
deno ing he di e ence in he signi icance and eliabili y o sho - and long- e m analyses.
2.3. CLIMATE CHANGE AND EXTREME WEATHER EVENTS
The clima e-disease ela ionship has been ho oughly esea ched, pa icula ly in Sou heas
Asia, wi h ecommenda ions o u he assess he impac s o clima e change (Kulka ni e al.,
2022). Ha ing examined he geog aphical expansion o dengue associa ed wi h clima e
change, daily mean empe a u e and empe a u e a ia ions s and ou as he main ac o s
in luencing dengue dis ibu ion, ollowed by p ecipi a ion (Ebi & Nealon, 2016).
Unde s anding he signi icance o he changing clima e on dengue e e is ele an o
assessing he disease’s isk and adap ing mi iga ion s a egies (Da is e al., 2021). A ele an
impulse o his esea ch was es ablished by p ojec ing dengue epidemics in opical a eas in
7
Sou h and Sou heas Asia unde di e en emission scena ios. Resul s demons a ed ha
ansmission isk is expec ed o inc ease o all scena ios, showing a peak in dengue e e
epidemic size and ou b eak du a ion (Wang e al., 2023). Mos esea ch inco po a ing clima e
change e e s o he In e go e nmen al Panel on Clima e Change (IPCC). I s Fi h Assessmen
Repo (AR5) p ojec s clima e change, impac , and isk un il 2100 based on ou
Rep esen a i e Concen a ion Pa hways (RCP) scena ios. Each scena io desc ibes di e en
le el o g eenhouse gas emissions and adia i e o cing, wi h a low-emission scena io
(RCP2.6), wo in e media e scena ios (RCP4.5 and RCP6.0), and a high emission scena io
(RCP8.5) (IPCC, 2014).
Clima e change e e s o an al e a ion o ex eme wea he e en s, such as El Niño, yphoons,
loods, d ough s, and hea wa es, in equency and magni ude and a long- e m change in
en i onmen al a iables in luencing he su i al, eplica ion, de elopmen , and dis ibu ion
o he ec o and i us (Li e al., 2018). El Niño e e s o he pe iodic wa ming o sea su ace
empe a u es, a ec ing global wea he pa e ns (Sunda i & K ishnamoo hy, 2019). An
in e ac ion o en i onmen al a ia ions and long- e m clima e change migh os e
ansmission (Ebi & Nealon, 2016). Ex eme wea he e en s a e p edic ed o become mo e
equen (Wang e al., 2022), a ec ing Aedes mosqui oes and ul ima ely p omo ing disease
ansmission (Li e al., 2018). Howe e , ex eme wea he e en s do no exhibi a uni o m
de ini ion. Fo ins ance, Seah e al. (2021) es ed di e en hea wa e de ini ions by adjus ing
he pe cen iles and numbe o consecu i e hea wa e days.
Cheng e al. (2021) explo ed he ela ionship be ween dengue ou b eaks and clima e change-
induced ex eme wea he e en s in China, obse ing a posi i e associa ion and delayed
e ec s. Du ing o immedia ely a e ex eme wea he e en s, he isk o dengue ou b eaks
did no homogenously inc ease, wi h dec eases in some ins ances. A peak in dengue ou b eak
isk was obse ed app oxima ely 1.5 mon hs pos -hea wa es and 1.5-3 mon hs a e ex eme
p ecipi a ion and humidi y, wi h e ec s las ing 2-3 mon hs. The empo al delay showcases
he du a ion o nega i e and posi i e e ec s o en i onmen al condi ions on dengue
ansmission. In conclusion, he impo ance o u he explo ing ex eme wea he e en s in
dengue-p one Asia-Paci ic coun ies and hei ou b eak scale was emphasized (Cheng e al.,
2021).
Bonnin e al. (2022) in eg a ed u u e clima e scena ios in o de eloping a p ocess-based
dengue model o Sou heas Asia. Thei s udy examined seasonal co ela ions, e ealing ha
a u u e inc ease in empe a u e is he main d i e o inc easing seasonal densi ies o Aedes
mosqui oes. The op imal empe a u e o A. aegyp i was iden i ied as 33°C and o A.
albopic us as 29°C. Beyond hese empe a u es, ising empe a u es led o lowe densi ies o
adul emale mosqui oes. P ecipi a ion exhibi s lowe signi icance, ye i s inc ease co ela es
wi h highe adul emale densi ies o bo h mosqui o species. Despi e some a ea-speci ic
di e ences, bo h densi ies a e p ojec ed o inc ease in all u u e clima e scena ios. The s udy
concludes ha e en wi h a igo ous implemen a ion o g eenhouse gas emissions educ ion
8
s a egies, mosqui o densi ies will no decline (Bonnin e al., 2022). Howe e , o he esea ch
a gues ha educed g eenhouse gas emissions would dec ease he sui abili y o ec o
habi a s (K aeme e al., 2019), wi h dis ibu ion sui abili y emaining consis en o sligh ly
dec easing (Da is e al., 2021). The alidi y o hese con adic o y s a emen s is add essed in
he subsequen analysis.
Messina e al. (2019) le e aged IPCC p ojec ions o 2020, 2050, and 2080 o assess dengue
isk based on 2015 using a boos ed eg ession ee. Acco dingly, empe a u e and annual
cumula i e p ecipi a ion we e de e mined as he p ima y ac o s in luencing dengue
sui abili y. P ojec ions un il 2080 indica e a iable shi s in dengue sui abili y’s geog aphical
and empo al dis ibu ion, wi h minimal global changes bu signi ican subna ional a ia ions.
Asian ci ies in coas al eas e n China and Japan a e expec ed o become mo e sui able by 2050.
The model ou come o he isk assessmen depended on he in eg a ed clima e scena io. The
scena io RCP6.0 sugges ed a global inc ease in dengue isk, po en ially a ec ing an addi ional
2.25 billion people be ween 2015 and 2080, encompassing o e 60% o he wo ld's
popula ion. This inc ease would esul om popula ion g ow h in al eady endemic a eas
ins ead o he sugges ed sp ead o dengue in new popula ions. Con a ily, scena io RCP4.5
indica ed a po en ial educ ion in dengue isk be ween 2050 and 2080 (Messina e al., 2019).
Clima e change scena ios a e le e aged in he subsequen analysis o e alua e disease
dynamics in Sou heas Asia u he .
2.4. SUMMARY
The li e a u e e iew p o ides a s uc u ed explo a ion, highligh ing undamen al s udies,
disease cha ac e is ics, and con adic ions. By add essing he complexi y o dengue
ansmission, ec o ecology, and he impac o en i onmen al a iables, seasonali y, and
clima e change, he aim was o es ablish a p o ound unde s anding and add ess knowledge
gaps ela ed o he esea ch ques ion.
Due o he elabo a ed ele ance o disease cha ac e is ics and associa ions, quali a i e
epo ed di e ences, and non-compa able s udy se ings o di e en loca ions, s udy pe iods,
and a iables, he li e a u e e iew did no ocus on compa ing pe o mance me ics. The
e iewed li e a u e buil he ounda ion o he geog aphical ocus, da a collec ion, and
a iable selec ion, which is c ucial o model de elopmen .
De ining he esea ch scope p o ed challenging gi en he a ie y o spa ial co e age anging
om u ban se ings o c oss-bo de egions. The geog aphical ocus o his s udy has been
de e mined h ough a scoping e iew ha emphasized he impac o clima e change on
dengue incidence in Sou heas Asia (Kulka ni e al., 2022). The amewo k o Li e al. (2018)
was an inspi a ion o ollow a mul idimensional app oach by in eg a ing clima e change o
d i e he meaning ulness o he unde lying esea ch. Gi en he epea ed emphasis on he
signi icance o empe a u e and p ecipi a ion on disease ansmission (Bonnin e al., 2022; Ebi
9
& Nealon, 2016; Messina e al., 2019), hese ac o s a e used as p edic o s o he ime se ies
analysis.
10
3. METHODOLOGY
3.1. INTRODUCTION
The esea ch me hodology aligns wi h he CRISP-DM (C oss-Indus y S anda d P ocess o
Da a Mining) amewo k es ablished by Chapman (2000). The schema ollows a s uc u ed
sequence ha o ms a cycle o six phases: (1) Business Unde s anding, (2) Da a Unde s anding,
(3) Da a P epa a ion, (4) Modeling, (5) E alua ion, and (6) Deploymen (Chapman e al., 2000).
The cyclical p ocess is illus a ed in Figu e 1.
Figu e 1 - CRISP-DM me hodology
In he ollowing, Business Unde s anding will be e e ed o as Domain Unde s anding o
be e align wi h he esea ch amewo k. In his phase, he cu en s a e o esea ch was
analyzed o de e mine he esea ch need, di ec ion, and scope o his s udy. Subsequen ly,
he p ojec was ini ialized by de ining he esea ch ques ion and objec i es.
Wi hin he Da a Unde s anding, aw da a was analyzed o e i y sui able o ma and quali y.
Following he da a ex ac ion, a p o ound explo a o y analysis was pe o med o iden i y da a
limi a ions and inconsis encies shaping he Da a P epa a ion.
The Da a P epa a ion phase includes da a p ep ocessing, ea u e enginee ing, and selec ion.
The aim was o clean he da a and add ess da a quali y issues obse ed du ing Da a
Unde s anding. Da a om he di e en loca ions was s anda dized, agg ega ed, and
combined in a o ma sui able o Modeling.
The Modeling encompasses epe i i e model aining and e alua ion. Models we e
de eloped, hype pa ame e uning was conduc ed, and a ious modeling and alida ion
app oaches we e es ed. In addi ion, he pe o mance o applied da a p epa a ion echniques
was assessed o de e mine he bes pe o mance.
Business
Unde s anding
Deploymen
Modeling
E alua ion
Da a P epa a ion
Da a
Unde s anding
Da a
11
Wi hin he E alua ion, he esea ch p ocess was e iewed o p e en i egula i ies, and he
success o he domain objec i es and da a mining goals we e con i med be o e Deploymen .
The Deploymen phase began wi h addi ional esea ch ha p o ided sui able clima e change
simula ions o Sou heas Asia. Finally, he o ecas was deployed in o he selec ed clima e
change amewo k. The p ojec has been examined in de ail o ensu e seamless moni o ing
and main enance in he u u e.
Va ious me hods ha e been applied wi hin hese s ages, p esen ed in Figu e 2 and u he
explained in he subsequen chap e s.
Figu e 2 - De ailed o e iew o applied me hods
3.2. DOMAIN UNDERSTANDING
Dengue e e ep esen s a signi ican heal h bu den in Sou heas Asia, wi h en i onmen al
ac o s such as empe a u e and p ecipi a ion a o ing i s sp ead. Unde s anding disease
associa ions and p edic ing ansmission is c ucial o de eloping p e en ion s a egies.
Building upon he esea ch backg ound and need discussed in he li e a u e e iew, his
p ojec aimed o combine open-sou ce da a om loca ions wi hin Sou heas Asia wi h da a
science ools o enhance he cu en s a e o esea ch. The p ojec used Py hon p og amming
language and suppo ing lib a ies, whose espec i e e sions a e linked in Appendix A.
This esea ch add essed he ques ion: "How can p edic i e modeling using en i onmen al
a iables assess dengue e e incidence and i s esponse o clima e change in Sou heas
Asia?". Th ee domain objec i es we e o mula ed (Figu e 3). The i s objec i e was o iden i y
12
signi ican en i onmen al d i e s o disease ansmission. The second was o de elop he
p edic i e model o dengue incidence using empe a u e and p ecipi a ion a iables. Finally,
he hi d objec i e was o assess how he disease esponds o clima e change.
Figu e 3 - Domain objec i es
F om a echnical pe spec i e, he da a mining goals we e de ined (Figu e 4). Fi s , collec da a
om mul iple Sou heas Asian loca ions. Second, analyze da a as pa o a p o ound
explo a o y analysis. Thi d, p epa e and ans o m aw en i onmen al and incidence da a o
enhance da a quali y and in e p e abili y. Fou h, de elop and op imize selec ed machine and
deep lea ning models in he mul i a ia e ime se ies analysis o c ea e a highly accu a e
dengue incidence o ecas ha gene alizes well on unseen da a. Las ly, e alua e model
pe o mance o ensu e deploymen eadiness.
Figu e 4 - Da a mining goals
3.3. DATA UNDERSTANDING
The geog aphical scope wi hin Sou heas Asia was de e mined by da a a ailabili y and
su iciency. Key insigh s o da a sea ch we e ga he ed h ough he li e a u e e iew.
En i onmen al da a was explo ed on he websi es o na ional me eo ological, en i onmen al,
and clima ological depa men s. Dengue e e incidence da a was collec ed om na ional
heal h minis ies, disease con ol di isions, and WHO coun y o ices. I da a was no publicly
a ailable, i was eques ed indi idually. Un o una ely, digi al eques s o nine o he
Sou heas Asian loca ions did no ecei e a posi i e esponse.
To es ablish a ounda ion o he model de elopmen , cohesi e da a poin s on a egional le el
in daily, weekly, o mon hly in e als we e sough o en i onmen al a iables and dengue
e e incidence. Yea ly da a was no conside ed, as he goal was o cap u e pa e ns and
Iden i ying
en i onmen al d i e s
o disease
ansmission
De eloping he
p edic i e model
Assessing he impac
o clima e change on
dengue incidence
Da a
Collec ion
Da a
Analysis
Da a
P epa a ion
P edic i e
Modeling
Model
E alua ion
13
seasonal ends o e ime o unde s and disease associa ions and e ec i ely add ess he
esea ch ques ion.
Chap e 2 discusses key en i onmen al ac o s in luencing disease sp ead, e ealing a ia ions
in he choice o a iables. Some s udies examined dengue co ela ions wi h en i onmen al
a iables, while o he s inco po a ed ex eme wea he e en s, conside ing hei excessi e
occu ence as a po en ial indica o o clima e change. Fu he mo e, as highligh ed in he
p e ious chap e , analyses yielded incohesi e esul s. While ce ain en i onmen al a iables
eme ged in luen ial in some ins ances, o he s ailed o con i m such asse ions, adding
complexi y o he a iable selec ion. Gi en hei demons a ed ele ance in e iewed
li e a u e, empe a u e and p ecipi a ion ha e been chosen o he model es ablishmen .
Wea he and clima e ex emes a e no di ec ly in eg a ed h ough ules o de ini ions.
Ex emes can be iden i ied h ough pa e ns and cha ac e is ics such as he du a ion and
in ensi y lea ned du ing he aining phase. Dengue e e was no di e en ia ed om mo e
se e e o ms o he i us in ec ion, including dengue hemo hagic e e and dengue shock
synd ome. Da a poin s o hese se e e o ms we e conside ed dengue e e and summed up.
Addi ionally, se o ypes o dengue i us (DENV) DENV 1, DENV 2, DENV 3, and DENV 4 we e
no dis inguished. The seconda y da a was acqui ed om mul iple sou ces a quali y and
accu acy ha canno be gua an eed o e i ied. The da a explo a ion e ealed some
inconsis encies, pa ially isualized in subsequen phases, acili a ing he Py hon lib a ies
Ma plo lib (Hun e , 2007) and Seabo n (Waskom, 2021). The en i e explo a o y analysis, along
wi h he co esponding Jupy e no ebooks, is p o ided in Appendix B.
In conclusion, comp ehensi e da a poin s we e es ablished o 17 loca ions wi hin Malaysia,
he Philippines, and Singapo e. Sequen ial weekly dengue incidence was acqui ed om
na ional depa men s. Daily en i onmen al da a, including p ecipi a ion, maximum
empe a u e, minimum empe a u e, and a e age empe a u e, was ob ained om he
Na ional Oceanic and A mosphe ic Adminis a ion (NOAA) o each coun y, enhancing da a
consis ency (Table 1).
14
Table 1 - Da a collec ion sou ces
Coun y
En i onmen al a iables
Dengue incidence
Time ame
Da a sou ce
Missing
alues
Da a sou ce
Missing
alues
Malaysia
NOAA
Yes
Minis y o
Heal h,
Malaysia
Yes
2017-2022
Philippines
NOAA
Yes
Depa men o
Heal h-
Epidemiology
Bu eau,
Philippines
No
2016-2020
Singapo e
NOAA
Yes
Minis y o
Heal h,
Singapo e
No
2016-2023
3.4. DATA PREPARATION
3.4.1. Da a P ep ocessing
A he beginning o he Da a P epa a ion phase, he da a was ans o med in o a o ma
sui able o he modeling. Fo each loca ion, he Excel ile was s anda dized as ollows:
▪ The i s shee , d _1, con ained daily en i onmen al da a.
▪ The second shee , d _2, con ained weekly incidence da a.
▪ The hi d shee , d _3, combined and agg ega ed he weekly incidence and daily
en i onmen al da a o a s anda d mon hly scale o enhance in e p e abili y.
Consequen ly, he da a quali y and p epa a ion o he i s wo da ase s di ec ly a ec ed he
hi d da ase . Las ly, he mon hly en i onmen al da a was ans o med in o en meaning ul
a iables. Table 2 displays he inal o ma o he hi d da ase , which a iables we e used o
he modeling.
15
Table 2 - T ans o med en i onmen al a iables
The ini ial p ep ocessing ocused on he i s wo da ase s holding aw da a using Pandas (The
pandas de elopmen eam, 2024) and NumPy (Ha is e al., 2020) lib a ies. The main goal o
dealing wi h ime se ies da a was o main ain cohesi e da a poin s conside ing he empo al
o de o obse a ions while cap u ing seasonali y and ends o explain disease in e ac ions.
A i s , missing alues we e impu ed o ensu e consis en da a. The ype o missing alues
was classi ied as Missing Comple ely a Random (MCAR) based on he absence o dis inc
pa e ns a speci ic loca ions, imes, o a iables. The p obabili y o missingness was assumed
o be he same o all samples. Each a iable exhibi ed dis inc pa e ns o missing alues, wi h
some showing con inuous gaps. Consequen ly, mul iple me hods we e explo ed and es ed
indi idually o all a iables. These me hods included Las Obse a ion Ca ied Fo wa d (LOCF)
and Nex Obse a ion Ca ied Backwa d (NOCB), Simple Mo ing A e age (SMA), Exponen ial
Mo ing A e age (EMA), Rolling Mean S a is ics, wi h Seasonal and T end Decomposi ion using
Loess (STL) Decomposi ion, and Mul i a ia e Impu a ion by Chained Equa ions (MICE)
Impu a ion using I e a i eImpu e om Sciki -lea n (Ped egosa e al., 2011).
Va iable
Uni
Desc ip ion
Min_Daily_P cp
Millime e s (mm)
Minimum daily p ecipi a ion
Max_Daily_P cp
Millime e s (mm)
Maximum daily p ecipi a ion
Mon hly_A g_P cp
Millime e s (mm)
Mon hly a e age p ecipi a ion based on daily
empe a u e epo ing
Mon hly_To al_P cp
Millime e s (mm)
Mon hly o al p ecipi a ion based on
cumula ed daily p ecipi a ion
Mon hly_A g_Temp
Deg ees Celsius (°C)
Mon hly a e age empe a u e based on daily
empe a u e epo ing
Min_Daily_Temp
Deg ees Celsius (°C)
Minimum daily p ecipi a ion o he espec i e
mon h
Max_Daily_Temp
Deg ees Celsius (°C)
Maximum daily p ecipi a ion o he espec i e
mon h
Min_A e age_Temp
Deg ees Celsius (°C)
Minimum alue o he a e age daily
empe a u e
Max_A e age_Temp
Deg ees Celsius (°C)
Maximum alue o he a e age daily
empe a u e
N_Raining_Days
Deg ees Celsius (°C)
Cumula i e numbe o ainy days in he
espec i e mon h
22
Figu e 10 - Loca ion-speci ic ime-based spli
The aining phase included a ime se ies c oss- alida ion using TimeSe iesSpli om he
Sciki -lea n lib a y (Ped egosa e al., 2011). He e, he model was i ed o he aining da a and
e alua ed on he alida ion da a in 5 spli s, using an expanding aining window o ensu e
obus gene aliza ion on unseen da a. The spli is displayed in Appendix G using he example
o Singapo e, Loca ion Code 11. The encoded a iable Loca ion Code was d opped o mi iga e
he bias o o dinal anking.
Fo ecas ing models we e de eloped wi h he aim o p ocessing ime se ies eg ession da a
wi h a non-linea ela ionship be ween a ge and independen a iables. Time se ies da a
was ans o med in o a supe ised lea ning p oblem h ough app op ia e da a p epa a ion
and spli ing, making i sui able o adi ional machine lea ning models. Suppo Vec o
Reg ession (SVR), K-Nea es Neighbo (KNN), AdaBoos Reg esso , Decision T ee (DT), and i s
ensemble, RF, we e implemen ed using Sciki -lea n (Ped egosa e al., 2011). XGBoos was
in eg a ed using he XGBoos lib a y (Chen & Gues in, 2016). These models we e ini ially
employed wi h de aul pa ame e s o assess di e en capping app oaches, ea u e selec ion
echniques, and scaling me hods based on he alida ion se . Subsequen ly, hype pa ame e
uning was conduc ed using G idSea chCV om he Sciki -lea n lib a y (Ped egosa e al., 2011)
o op imize model pe o mance by adjus ing pa ame e s while p e en ing o e i ing. Tuned
pa ame e s a e p esen ed in Appendix H.
23
The ini ial esea ch employing deep lea ning models using clima e da a o o ecas long- and
sho - e m dengue incidence and ou b eaks was conduc ed by T an e al. (2022) in Vie nam.
Thei ime se ies analysis e ealed ha Con olu ional Neu al Ne wo k (CNN), Long Sho -Te m
Memo y (LSTM), and Long Sho -Te m Memo y wi h A en ion (LSTM-ATT) we e he bes -
pe o ming models (T an e al., 2022), in o ming he decision o inco po a e deep lea ning
models in o he unde lying analysis o ensu e alignmen wi h he la es esea ch on dengue
e e p edic ion. Deep Lea ning models based on neu al ne wo ks we e implemen ed using
Ke as API wi h Tenso Flow (Abadi e al., 2015). Applied deep lea ning models we e di ided
in o Recu en Neu al Ne wo k (RNN), Feed o wa d Neu al Ne wo k (FNN), and CNN.
SimpleRNN, Ga ed Recu en Uni (GRU), and LSTM we e used o he RNN. A Mul i-Laye
Pe cep on (MLP) wi h wo hidden laye s was implemen ed o he FNN, wi h 1-dimensional
CNN being used o he CNN, passing in o ma ion o wa d. The models we e ained and
e alua ed using ime se ies c oss- alida ion wi h 5 spli s. Ea ly s opping c i e ia we e
implemen ed o a oid o e i ing, which would s op he aining p ocess i no imp o emen
was obse ed a e 5 consecu i e epochs. The con igu a ion o each model is p o ided in
Appendix I.
3.5.2. Model E alua ion
The examined alida ion me hods ha e been discussed in he model de elopmen . In
conclusion, he ain- es spli and ime se ies c oss- alida o wi h 5 spli s we e applied o
assess model pe o mance, ensu ing empo al o de o a oid da a leakage while spli ing
loca ion-speci ic da a. E alua ion me ics o eg ession mean absolu e e o (MAE) (Equa ion
2) and oo mean squa ed e o (RMSE) (Equa ion 3) om Sciki -lea n (Ped egosa e al., 2011)
we e employed o quan i y he models' p edic i e abili y on unseen da a. The RMSE penalizes
la ge e o s h ough he squa e oo , making i pa icula ly ele an o eg ession asks. MAE
was chosen due o i s easy in e p e abili y, di ec ly ep esen ing he a e age e o .
(2) 𝑀𝐴𝐸=1
𝑛∑ |𝑦𝑖−𝑦𝑖|
𝑛
𝑖=1 (3) 𝑅𝑀𝑆𝐸=√1
𝑛∑ (𝑦𝑖−𝑦𝑖)
𝑛
𝑖=1 2
▪ 𝑛 is he numbe o obse a ions.
▪ 𝑦𝑖 is he ac ual alue.
▪ 𝑦𝑖 is he p edic ed alue.
▪ |𝑦𝑖−𝑦𝑖| is he absolu e di e ence be ween he ac ual and p edic ed alues.
▪ (𝑦𝑖−𝑦𝑖)2 is he squa ed di e ence be ween he ac ual and he p edic ed alues.
3.6. EVALUATION
The CRISP-DM amewo k es ablished a sys ema ic and obus ounda ion o add ess he
domain objec i es. En i onmen al ac o s d i ing disease sp ead we e analyzed. The p ojec ’s
aim o de elop a o ecas using dengue incidence and en i onmen al a iables in Sou heas
Asia was achie ed. Subsequen ly, he p ojec p oceeds o he deploymen phase o add ess
he las objec i e.
24
The inal e iew e ealed ha all da a mining goals we e success ully ealized. The da a sea ch
yielded su icien and eliable da a poin s o 17 Sou heas Asian loca ions. Guided by he
ini ial explo a o y analysis, a ious da a p epa a ion echniques we e es ed. Mul iple
machine and deep lea ning models we e de eloped using di e en app oaches, ollowed by
model op imiza ion h ough hype pa ame e uning. Las ly, p edic i e capabili y was
e alua ed h ough epe i i e aining and alida ion using he e alua ion me ics MAE and
RMSE.
3.7. DEPLOYMENT
In he las phase, he o ecas was deployed o assess he disease's esponse o changing
clima e. The in eg a ed clima e change simula ions we e ob ained om Sen ian e al. (2022),
who analyzed clima e change in e ac ions wi h monsoon seasons o assess he impac o
empe a u e and ain all a iabili y in Sou heas Asia. Thei esea ch le e aged he p ojec ed
emission scena ios RCP4.5 and RCP8.5 om he IPCC AR5. Based on hese wo scena ios, hey
simula ed he pe cen age change o Sou heas Asia's mon hly mean o al p ecipi a ion and
mean su ace empe a u e in Janua y and July o 2030, 2050, 2070, and 2100 ela i e o 2013
(Sen ian e al., 2022).
The unde lying esea ch employed hese simula ed changes as a amewo k o p edic
espec i e changes in annual incidence a es. The a iables Mon hly_To al_P cp and
Mon hly_A g_Temp we e modi ied wi h he simula ed changes in he es se , ce e is pa ibus.
I he simula ed changes in o al p ecipi a ion esul ed in nega i e alues, hey we e se o 0.
A e he aining phase, he disease esponse assessmen was pe o med on he en i e es
se comp ising 12 da a poin s, ep esen ing a yea . Gi en he ele ance o lagged
obse a ions, hei simula ed changes we e cap u ed by p edic ing subsequen mon hs.
Finally, he p edic ed changes in annual incidence a es we e compa ed o he ac ual
p edic ions o he a ge a iable o assess disease dynamics unde clima e change gi en wo
scena ios. In conclusion, he inal objec i e was add essed.
25
4. RESULTS AND DISCUSSION
The esea ch ques ion led o h ee domain objec i es.
4.1. ENVIRONMENTAL DRIVERS
The i s objec i e was o explo e he en i onmen al d i e s o disease sp ead. Al hough he
li e a u e discusses he impo ance o en i onmen al a iables, emphasizing empe a u e and
p ecipi a ion, s a is ical echniques did no yield signi ican esul s. In he explo a o y analysis,
independen a iables demons a ed unsa is ac o y ou comes in he co ela ion analysis, wi h
Max_Daily_Temp_lag_4 showing he highes co ela ion o 0.11 wi h he a ge a iable
(Appendix J). Con a y o o he s udies conduc ed in Singapo e (Seah e al., 2021), maximum
daily empe a u e posi i ely co ela ed wi h dengue incidence ac oss all lags (Appendix J).
Howe e , esea ch indings a e consis en wi h Wang e al. (2022), showing a dec eased
co ela ion o minimum daily empe a u e a lag 2, ollowed by an inc ease in co ela ion a
subsequen lags. Simila o hei obse a ion o ex emely high empe a u es being associa ed
wi h inc eased isk a a 2-3 weeks lag (Wang e al., 2022), his esea ch yielded compa able
esul s, showing an inc easing co ela ion be ween maximum empe a u e and incidence un il
lag 4, ollowed by a dec ease in co ela ion (Appendix J). The ou ea u e selec ion me hods
epea edly highligh ed some a iables, indica ing hei signi icance (Appendix F). Las ly,
a ying incidence was obse ed ac oss di e en yea s and loca ions (Appendix D), leading o
no e idence o a causal ela ionship wi h he ainy season as ou lined in p e ious li e a u e (Li
e al., 2018; Wo ld Heal h O ganiza ion, 2024a).
4.2. PREDICTIVE MODELING
Fo he second objec i e, he ini ial app oach was es ablishing one holis ic model ac oss all
loca ions o c ea e a dengue incidence o ecas o Sou heas Asia. The p ima y e alua ion
was based on he alida ion se wi hou pa ame e op imiza ion o main ain a compa able
se ing o analyzing he pe o mance o di e en da a p epa a ion echniques and models. In
he p ep ocessing, wo ou lie capping logics we e applied o mi iga e bias and enhance da a
quali y. The combina ion o bo h app oaches e ealed he bes pe o mance. Consequen ly,
aside om da a agg ega ion and he capping o ela i e ou lie s, he model pe o mance
ad anced om add essing ex emes on a mul i-loca ion le el o he holis ic model.
T aining all a iables, he b oade esea ch amewo k esul ed in unde i ing, wi h low
p edic i e powe o all models. Wi hou hype pa ame e uning, KNN and RF eme ged as he
bes -pe o ming models, wi h an MAE o 17.62 and 18.90 and RMSE o 41.58 and 42.13 ac oss
loca ions, ep esen ing la ge e o s. Al hough ea u e impo ance analysis showed high
ele ance o he a iable Loca ion Code in bo h me hods (Appendix E), he la ge e o
sugges s ha he model could no di e en ia e be ween loca ions e ec i ely. Consequen ly,
he di icul y in lea ning loca ion-speci ic pa e ns a ose, suppo ed by dec eased and a ying
26
aining leng hs (Figu e 9), which migh ha e led o he unde ep esen a ion o some
loca ions.
The independen app oach encompassing mul iple loca ion-speci ic p edic ions showed
be e esul s. E alua ion me ics we e a e aged ac oss 5 spli s and loca ions o de e mine
supe io model pe o mance. Fo each ime se ies model, loca ion-speci ic ou lie capping was
pe o med. A e hype pa ame e uning, AdaBoos achie ed he bes esul on he alida ion
se o scale-in a ian models wi h an a e age MAE o 13.19 and an RMSE o 16.75 using RF
ea u e selec ion. Fo scale- a ian models, SVR eme ged as he bes model wi h an a e age
MAE o 13.03 and an RMSE o 16.35 using Robus Scale . The ea u e selec ion me hods
exhibi ed inc eased pe o mance o ce ain models and me hods. Howe e , all ou me hods
demons a ed in e io esul s compa ed o aining all a iables wi h SVR, which yielded he
bes pe o mance ac oss all machine lea ning models.
Deep lea ning models we e ained, wi h SimpleRNN showing an a e age MAE o 15.88 and
an RMSE o 19.31, ou pe o ming GRU and LSTM. Fo he FNN, MLP yielded an a e age MAE
o 13.25 and an RMSE o 16.83. CNN eme ged as he bes deep lea ning model o p edic ing
dengue incidence, wi h an a e age MAE o 10.10 and an RMSE o 13.61. The inal assessmen
e ealed ha deep lea ning models ou pe o med machine lea ning models. Appendix K
p esen s he loca ion-speci ic MAE and RMSE sco es on he alida ion se s o he ou supe io
models.
All p ojec phases ha e been hough ully pe o med o mi iga e bias, enhance da a quali y,
and yield an e ec i e p edic ion. Howe e , he pe o mance o he p edic i e models showed
consis en ly in e io esul s o ce ain loca ions on he alida ion se (Appendix K). Re iewing
he da a explo a ion did no explain a ying p edic i e abili ies, as nei he a common pa e n,
such as in he case o missing alues, no a coun y-speci ic co ela ion was obse ed.
Unde epo ing, as a gued by Bha e al. (2013), may ha e comp omised he in eg i y o he
aw da a. Consequen ly, p edic ions migh ha e been a ec ed by inaccu acies in he
seconda y open-sou ce da a.
Models showed an o e all dec ease in e o s on he es se , which is a good indica o o
a oiding o e i ing. AdaBoos and SVR showed an MAE o 10.22 and 9.31 and RMSE o 12.49
and 11.34 on he es se . MLP yielded an MAE o 6.82 and an RMSE o 8.87. CNN was chosen
o he deploymen , wi h an MAE o 5.06 and an RMSE o 7.09 on he es se . La ge e o s
we e educed in almos all loca ions, which could be explained by inc eased da a su iciency.
Howe e , a loca ion-speci ic shi in e o s was obse ed in some ins ances. This sugges s ha
eliable p edic ion was no ied o ce ain loca ions. Ins ead, he a ying e o s could ha e
been ela ed o o he ac o s ha we e no included. The loca ion-speci ic e alua ion me ics
on he es se a e de ailed in Appendix L.
27
4.3. CLIMATE CHANGE IMPACT ASSESSMENT
Fo he hi d objec i e, CNN was le e aged o p edic he changes in annual incidence a es
based on he simula ed changes in mean empe a u e and o al p ecipi a ion o Sen ian e al.
(2022). Kuching, Malaysia, was chosen o he deploymen wi h he lowes p edic i e e o
ac oss all loca ions. The pe cen age changes in he annual incidence a es in Kuching,
Malaysia, o bo h emission scena ios a e p esen ed in Appendix M.
Fo bo h scena ios, a posi i e linea end be ween he change in he annual dengue incidence
a e in Kuching, Malaysia, and he mean empe a u e was obse ed (Figu e 11). Howe e , no
clea end was ound be ween he changes in o al p ecipi a ion and he a ge a iable,
showing sca e ed da a poin s in Figu e 12. These obse a ions a e consis en wi h Ebi and
Nealon (2016), who epo ed ha daily mean empe a u e and empe a u e a ia ions a e he
main d i e s o dengue incidence.
In conclusion, assessing disease dynamics based on changes in o al p ecipi a ion, mean
empe a u e, and lagged obse a ions ep esen s a subs an ial achie emen . This success
unde sco es he model's sensi i i y in p edic ing dengue incidence and demons a es i s
abili y o assess he impac o clima e change on disease ansmission.
Figu e 11 - Simula ed changes in mean empe a u e and annual incidence a e
28
Figu e 12 - Simula ed changes in o al p ecipi a ion and annual incidence a e
29
5. CONCLUSIONS
This s udy aimed o align wi h he la es esea ch on dengue e e p edic ion, combining open-
sou ce da a wi h da a science ools. The CRISP-DM schema employed a sys ema ic app oach
and s anda dized s uc u e, enhancing he p ojec ’s obus ness and u u e main enance. The
a ie y o applied me hods con ibu es o success ully answe ing he esea ch ques ion.
CNN eme ged as he bes -pe o ming model wi h an MAE o 5.06 and an RMSE o 7.09 on he
es se , demons a ing ha p edic i e modeling using en i onmen al a iables can e ec i ely
assess dengue e e incidence. Deep lea ning models yielded a sligh dec ease in e o s
compa ed o adi ional machine lea ning models. The educ ion o e o s achie ed
h oughou he p ojec highligh s he necessi y o p o ound da a p epa a ion and he
e ec i eness o he chosen app oach.
Applied models consis en ly showed in e io esul s o ce ain loca ions on he alida ion se .
The e iew o he explo a o y analysis did no explain a ying p edic i e abili ies. Ex ensi e
da a p epa a ion and he mul i ude o applied models led o he assump ion ha high
p edic i e e o s o ce ain loca ions we e due o seconda y da a ob ained. Fo ins ance,
Bha e al. (2013) a gued ha he e is unde epo ing o dengue cap u e. In con as , a
loca ion-speci ic shi in e o s was obse ed on he es se wi h o e all dec eased e o s,
showing a ia ions wi h inc eased e o s in some ins ances. This esul indica ed ha he
models gene alized well wi h inc easing da a su iciency, sugges ing ha a ying e o s we e
ela ed o o he unconside ed ac o s.
The clima e change impac assessmen esul s showed ha o al p ecipi a ion and mean
empe a u e changes enabled a disease esponse, indica ing a sensi i e model. The simula ed
changes showed a ying e ec s on annual incidence a es o bo h scena ios, wi h a posi i e
linea end obse ed o mean empe a u e. In conclusion, he model e ec i ely assessed
disease dynamics unde clima e change.
30
6. LIMITATIONS AND RECOMMENDATIONS FOR FUTURE WORK
This wo k encoun e ed some no able limi a ions. Fi s ly, p e iously applied con ol measu es
ha in luenced egional condi ions and dengue dis ibu ion we e no conside ed, which migh
ha e a ec ed he p edic i e abili y. Addi ionally, limi ed da a a ailabili y could ha e educed
p edic i e powe . Dec eased aining se s due o mul iple independen ime se ies may ha e
limi ed he model's abili y o lea n sequen ial dynamics. As dec eased e o s we e obse ed
on he es se s, inc eased da a olume h ough con inuous main enance could imp o e
u u e loca ion-speci ic model pe o mance.
The independen app oach o loca ion-speci ic ime se ies makes he o ecas egionally
lexible. Howe e , his p ojec was based on en i onmen al a iables ha ha e demons a ed
ele ance o p edic ing dengue e e in Sou heas Asia. The e o e, i is no ecommended o
expand he o ecas globally. Ins ead, u u e wo k could enhance he Sou heas Asian analysis
and build on T an e al. (2022), who demons a ed supe io model pe o mance by adding an
a en ion laye a e he LSTM ne wo k. In addi ion, u he wo k may expand he analysis
wi h addi ional obse a ions. The li e a u e ou lines a posi i e associa ion be ween dengue
e e and socioeconomic ac o s, which migh enhance modeling e o s. Va ying incidences
ac oss loca ions may esul om unique geog aphical condi ions di ec ly a ec ing mosqui o
habi a s. Adding demog aphic and opog aphic a iables could imp o e he p edic ion o
dengue incidence. Mo eo e , imp o ed da a documen a ion migh allow he di e en ia ion
o dengue-incidence-induced Aedes species o enhance he in e p e abili y o disease
dynamics.
The clima e change simula ions and hei applica ion encompass limi a ions. While o al
p ecipi a ion and mean empe a u e changes enabled a disease esponse, hei e ec s migh
be in e ela ed. Fu u e wo k should analyze associa ions be ween independen a iables and
hei in e ela ed e ec s o enhance dengue isk assessmen . In his p ojec , he impac o
clima e change on he annual incidence a es was assessed on he es se and no empo ally
aligned wi h he p ojec ions un il 2100 due o he assump ion o con inuously inc easing e o
o he unde lying p edic ion and equi ed wea he o ecas s. A long- e m goal o he hi d
objec i e would be o in eg a e a egional clima e model a he han elying on seconda y
clima e change simula ions. This way, ime pe iods could be homogenized o inc ease
ep esen a i eness. Fu he mo e, he addi ional emission scena ios RCP2.6 and RCP6.0 could
be inco po a ed. In addi ion, clima e change simula ions we e limi ed o wo a iables a
speci ic imes amps. Since independen a iables in he unde lying analysis a e ela ed,
simul aneous change is expec ed. Simula ing changes o he emaining a iables would
enhance he model's p edic i e abili y, allowing a comp ehensi e assessmen o disease
dynamics. Las ly, clima e change was de ined as a long- e m change in en i onmen al
a iables (Li e al., 2018). The e o e, simula ed changes a he speci ied imes amps do no
ep esen a eal-wo ld scena io.
31
Las ly, unde s anding he disease cycle, ini ia ing ac o s, and iden i ying isk a eas is c ucial
o mi iga e u u e dengue isk. Da a documen a ion and s o age need o be enhanced globally
and made publicly a ailable o acili a e knowledge exchange and enable esea ch. Da a
collec ion e ealed ha na ional incidence epo ing demands mo e s anda diza ion.
Fu he mo e, c oss-na ional collabo a ion could imp o e disease cap u e and analysis. In his
ega d, da a science could unlock g ea po en ial o add ess he u u e disease bu den.
38
APPENDIX C
39
APPENDIX D
40
APPENDIX E
41
APPENDIX F
Fea u e
Me hod
To al
Coun
Selec F om
Model
using RF
Selec F om
Model
using XG
Boos
RFE using
RF
Fo wa d
Fea u e
Selec ion
using RF
Loca ion Code
1
1
1
1
4
Max_Daily_Temp_lag_3
1
1
1
0
3
Mon hly_A g_Temp_lag_5
0
1
1
1
3
Mon hly_A g_Temp_lag_2
1
0
1
1
3
Mon hly_A g_Temp
1
0
1
1
3
Mon hly_To al_P cp_lag_6
1
1
1
0
3
Mon hly_A g_Temp_lag_3
1
1
1
0
3
N_Raining_Days_lag_6
1
1
1
0
3
Max_Daily_Temp_lag_4
0
1
0
1
2
Mon hly_A g_Temp_lag_1
1
1
0
0
2
Mon hly_To al_P cp
0
0
0
1
1
Mon hly_To al_P cp_lag_1
0
0
1
0
1
Mon hly_A g_Temp_lag_6
0
0
0
1
1
Max_Daily_P cp_lag_1
0
1
0
0
1
Max_Daily_Temp_lag_6
1
0
0
0
1
N_Raining_Days_lag_1
0
0
0
1
1
Min_Daily_Temp_lag_2
0
0
0
1
1
N_Raining_Days
0
0
0
1
1
Max_Daily_P cp_lag_2
0
0
1
0
1
No e. 1: Fea u e was selec ed by he co esponding me hod, 0: Fea u e was no selec ed.
42
APPENDIX G
43
APPENDIX H
Model
Hype pa ame e s
DT
{}
RF
{max_dep h: 20, max_ ea u es: 'sq ', min_samples_lea : 4, min_samples_spli :
10}
XGBoos
{colsample_by ee: 0.7, lea ning_ a e: 0.014, max_dep h: 9, min_child_weigh :
2, n_es ima o s: 141, eg_lambda: 3, subsample: 0.956}
AdaBoos
{lea ning_ a e: 0.01, loss: 'linea '}
KNN
{me ic: 'euclidean', n_neighbo s: 10, weigh s: 'dis ance'}
SVR
{C: 100, ke nel: ' b '}
44
APPENDIX I
Model
Model Con igu a ions
SimpleRNN
{Model: Sequen ial, Laye 1: SimpleRNN, 50 uni s, Ou pu Laye : Dense, 1
uni , Loss Func ion: MAE, Op imize : Adam, Epochs: 50, Ba ch Size: 72}
GRU
{Model: Sequen ial, Laye 1: GRU, 50 uni s, Ou pu Laye : Dense, 1 uni , Loss
Func ion: MAE, Op imize : Adam, Epochs: 50, Ba ch Size: 72}
LSTM
{Model: Sequen ial, Laye 1: LSTM, 50 uni s, Ou pu Laye : Dense, 1 uni ,
Loss Func ion: MAE, Op imize : Adam, Epochs: 50, Ba ch Size: 72}
MLP
{Model: Sequen ial, Laye 1: Dense, 64 uni s, ReLU ac i a ion, Laye 2:
Dense, 32 uni s, ReLU ac i a ion, Ou pu Laye : Dense, 1 uni , Loss Func ion:
MAE, Op imize : Adam, Epochs: 50, Ba ch Size: 72}
CNN
{Model: Sequen ial, Laye 1: Dense, 64 uni s, ReLU ac i a ion, Laye 2:
Dense, 31 uni s, ReLU ac i a ion, Ou pu Laye : Dense, 1 uni , Loss Func ion:
MAE, Op imize : Adam, Epochs: 50, Ba ch Size: 72}
45
APPENDIX J
46
APPENDIX K
MAE on he alida ion se
Loca ion Code
Name
AdaBoos
SVR
CNN
MLP
0
Baguio, RP
10.99
11.06
8.56
16.93
1
Dagupan, RP
13.07
10.39
14.51
20.45
2
Ko a Kinabalu
In e na ional, MY
3.75
3.53
2.46
12.47
3
Kuala Lumpu
In e na ional, MY
27.11
29.14
22.09
13.60
4
Kuan an, MY
6.89
7.45
4.14
13.44
5
Kuching, MY
3.18
2.86
2.03
11.83
6
Labuan, MY
5.51
4.10
1.83
14.44
7
Malacca, MY
9.25
8.07
5.05
16.01
8
Manila, RP
8.10
8.36
9.04
18.91
9
Me sing, MY
10.17
9.86
4.66
13.25
10
Penang
In e na ional, MY
12.80
11.37
5.74
12.52
11
Singapo e Changi
In e na ional, SN
19.17
19.99
18.63
18.92
12
Si iawan, MY
6.77
5.57
1.96
11.68
13
Sul an Abdul Aziz
Shah In e na ional,
MY
20.71
22.84
15.65
13.00
14
Sul an Ismail Pe a,
MY
11.63
10.80
7.77
17.44
15
Su igao, RP
21.71
24.01
18.68
23.48
16
Zamboanga, RP
33.48
32.10
28.84
21.36
A e age MAE
13.19
13.03
10.10
13.25
47
RMSE on he alida ion se
Loca ion Code
Name
AdaBoos
SVR
CNN
MLP
0
Baguio, RP
14.56
13.13
11.40
21.62
1
Dagupan, RP
15.73
13.27
17.65
25.91
2
Ko a Kinabalu
In e na ional, MY
4.56
4.26
3.22
16.87
3
Kuala Lumpu
In e na ional, MY
31.21
33.07
26.07
17.32
4
Kuan an, MY
7.73
8.30
4.79
18.20
5
Kuching, MY
3.71
3.26
2.26
15.56
6
Labuan, MY
6.66
4.92
2.49
19.65
7
Malacca, MY
10.80
9.31
6.11
21.79
8
Manila, RP
9.79
10.27
11.43
26.00
9
Me sing, MY
11.62
11.34
5.92
16.83
10
Penang
In e na ional, MY
14.50
13.9
7.60
16.50
11
Singapo e Changi
In e na ional, SN
25.63
25.81
25.05
25.89
12
Si iawan, MY
7.50
6.59
2.69
15.78
13
Sul an Abdul Aziz
Shah In e na ional,
MY
25.34
26.93
21.27
17.14
14
Sul an Ismail Pe a,
MY
13.81
12.84
10.37
23.86
15
Su igao, RP
28.68
30.51
24.66
29.73
16
Zamboanga, RP
52.85
50.16
48.39
29.52
A e age MAE
16.75
16.35
13.61
16.83