Inter national J our nal of Inf ormatics and Communication T echnology (IJ-ICT) V ol. 15, No. 3, September 2026, pp. 1066 ∼ 1077 ISSN: 2252-8776, DOI: 10.11591/ijict.v15i3.pp1066-1077 ❒ 1066 Pr ognosis of v ector bor ne dengue disease outbr eak in urban ar eas using multi v ariate analysis Pratik S. Machchar 1 , Pur vi N. Ramanuj 2 , Rajan P atel 3 , Jitendra Bhatia 4 , K untesh J ani 5 1 Gujarat T echnological Uni v ersity , Ahmedabad, India 2 V ishw akarma Go v ernment Engineering Colle ge, Ahmedabad, India 3 Kalol Institute of T echnology and Research Center , India 4 Nirma Uni v ersity , Ahmedabad, India 5 L.D. Colle ge of Engineering, Ahmedabad, India Article Inf o Article history: Recei v ed August 22, 2025 Re vised March 14, 2026 Accepted May 21, 2026 K eyw ords: Deep learning Dengue fe v er Machine learning T ime series forecasting V ector -borne diseases ABSTRA CT V ector borne disease lik e dengue continues to pose a signicant climate-sensiti v e public health challenge in tropical re gions such as Brazil, Per u, and India. This study e xamines the feasibility of predicting dengue outbreaks using weekly mul- ti v ariate time-series data from San Juan (SJ), Puerto Rico and Iquitos (IQ), Peru. Dengue incidence w as analyzed alongside meteorological, en vironmental, and v e getation-based v ariables to capture k e y climatic inuences. Se v eral machine learning and deep learning approaches were e v aluated, including LightGBM. Model performance w as assessed using root m ean square error (RMSE) and mean absolute error (MAE). The results sho w that LightGBM achie v ed the lo w- est RMSE/MAE, indicating strong short-te rm predicti v e accurac y and e xcellent interpretability . Feature importance analysis and principal component analy- sis (PCA) identied precipitation, de w point tempera ture, and humidity as the most inuential predictors of dengue incidence. The study demonstrates that ad- v anced machine learning models can serv e as reliable early w arning systems for v ector -borne diseases. While this research focuses on dengue, the methodology is adaptable to other v e ctor -bone datasets and diseases, of fering a e xible tool for public health authorities to predict and mitig ate outbreaks in di v erse urban conte xts. This is an open access article under the CC BY -SA license . Corresponding A uthor: Pratik S. Machchar Gujarat T echnological Uni v ersity Ahmedabad, Gujarat, India Email: pratikmachchar@gmail.com 1. INTR ODUCTION Dengue fe v er continues to pose a signicant and gro wing public health threat in tropical and subtropi- cal re gions w orldwide. T ransmitted primarily by Aedes ae gypti and Aedes albopictus mosquitoes, dengue no w places nearly half of the global population at risk of infection. In 2024, the W orld Health Or g anization (WHO) reported o v er 14.4 million dengue cases and more than 11,200 deaths across all six WHO re gions, marking one of the highest annual global b urdens on record [1]. In India , national surv eillance data indicate a substantial increase in dengue incidence, with o v er 233,000 reported cases and nearly 300 associated deaths in 2024, con- tinuing a trend of high disease b urden in recent years [1]. These gures reect the e xpanding geographic spread and climate-sensiti v e dynamics of dengue transmission, underscoring its persistent public health impact. J ournal homepage: http://ijict.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 1067 The four dengue serotypes (DENV1-4) cause illness ranging from mild fe v er to life-threatening hem- orrhagic complications [2]. Urbanization and climate change ha v e e xpanded mosquito habitats. These v ectors also transmit malaria, Zika, chikungun ya, and yello w fe v er . En vironmental f actors lik e temperature, rainf all, and humidity af fect mosquito breeding and virus replication [3]-[5]. Higher temperatures accelerate virus de- v elopment, rainf all creates breeding sites [6], [7], and v e getation density pro vides ideal mosquito habitats [8]. Despite adv ances in s urv ei llance, predicting dengue outbreaks remains dif cult [9]-[11]. Exis ting methods struggle to capture interactions between climate, human beha vior , and mosquito dynamics [12], lim- iting timely public health responses [13]-[15]. This study e v aluates multi v a riate time series forecasting for dengue prediction using data from San Juan (SJ) and Iquitos (IQ). Most pre vious w ork focuses on uni v ariate statistical/deep learning methods, or often with limited predictors [9], [16], [17]. In this study , six multi v ariate models —SARIMAX, LightGBM, Seq2Seq, RNN, CNN-BiLSTM, and Stack ed LSTM— comparing statistical, machine learning, and deep learn- ing approaches under identical conditions [11], [18]-[26]. F or each model, principal component analysis (PCA) identies k e y c limate dri v ers of dengue i nci- dence, addressing the interpretability g ap in ensemble methods [27], [28]. W eekly forecasts align with opera- tional surv eillance needs, unlik e annual or long-term studies [18], [29]. This w ork contrib utes: - An inte grated frame w ork combining climate and epidemiological v ariables. - Systematic cross-paradigm model e v aluation comparing classical statistical methods (SARIMAX), machine learning approaches (LightGBM), and deep learning architectures (Seq2Seq, RNN, CNN-BiLSTM, Stack ed LSTM) under identical e xperimental conditions to identify the most ef fecti v e forecasting approach - Identication of inuential predictors/v ariables that contrib ute most to dengue incidence. - Comparison of weekly forecast ability for practical disease surv eillance. The results sho w that LightGBM outperforms other approaches, with precipitation, de w point tem- perature, and humidity identied as the most inuential predictors. While demonstrated using dengue, the frame w ork is broadly applicable to other v ector -borne diseases in urban settings. 2. MA TERIALS AND METHODS 2.1. Dataset The dataset used for this study focuses on v ector -borne disease predictions, such as dengue, and includes detailed climate and en vironmental data for tw o cities: SJ and IQ [30] (Figure 1). It spans se v eral years on a weekly timescale, capturing the corelation between pre v ailing climat e v ariables and dengue outbreaks. K e y features include dengue case counts alongside v arious climate metrics such as temperature, precipitation, and humidity , all of which are kno wn to inuence the transmission of v ector - borne diseases. 2.2. Data pr epr ocessing Ef fecti v e data preprocessing is crucial for impro ving the accurac y of models predicting v ector -borne disease outbreaks. In this study , multiple steps were tak en to clean and prepare the dataset before applying forecasting models. - Data formatting: The date column w as con v erted into a standard DateT ime format. Abnormal or erroneous dates (e.g., 1994-02-31) were corrected by replacing them with corresponding dates from the pre vious year to maintain temporal consistenc y . - Handling missing and incorrect v alues: Incorrect or f alse v alues, including null v alues, were systematically identied and remo v ed from the dataset to pre v ent distortion in model predictions. - Outlier detection and replacement: Outliers were detect ed based on statistical thresholds and replaced with NaN v alues to pre v ent sk e wed data points from inuencing the analysis. - Interpolation missing v alues, including those introduced by outlier replacement, were lled using interpola- tion methods, ensuring a continuous dataset without introducing bias. - Data scaling: T o standardize the dataset and a v oid an y bias due to dif fering v alue ranges across features, data scaling w as applied, ensuring uniformity across all v ariables for model training. These preprocessing steps were essential for rening the dataset, enabling accurate and reliable dengue outbreak predictions using v arious forecasting models. The dataset incorporates weather station data from Pr o gnosis of vector borne dengue disease outbr eak in urban ar eas using ... (Pr aktik S. Mac hc har) Evaluation Warning : The document was created with Spire.PDF for Python.
1068 ❒ ISSN: 2252-8776 NO AA, such as maximum, minimum, and a v erage temperatures, total precipi tation, and diurnal temperature range. Additionally , reanalysis data includes relati v e and specic humidity , air temperature, de w point temper - ature, and precipitation, pro viding a comprehensi v e understanding of climatic f actors. Satellite-based precipi- tation (PERSIANN) and v e getation indices (ND VI) of fer further insights into en vironmental conditions. These features enable a multi v ariate analysis of ho w f actors lik e temperature, rainf all, and humidity impact dengue transmission dynamics in urban areas. Figure 1. Study locations: San Juan (Puerto Rico) and Iquitos (Peru) 2.3. A v ailable parameters f or pr edicting dengue cases The dataset comprises 24 en vironmental and climatic features alongside the tar get v ariable (total dengue cases), pro viding a comprehensi v e multi v ariate frame w ork for outbreak prediction. Before analyz- ing the impact of indi vidual climate parameters, it is important to understand ho w each v ariable contrib utes to dengue transmissi on . The follo wing sections detail the main climate and en vironmental features considered in this study , starting with temperature, which plays a crucial role in mosquito de v elopment and virus replication. 2.3.1. T emperatur e - station max temp c, station min temp c, station a vg temp c: W eather station data capturing daily tempera- ture uctuations, critical because Aedes mosquitoes thri v e within certain temperature ranges. Higher temper - atures speed up mosquito life c ycles and virus replication. - reanalysis max air temp k, reanalysis min air temp k, reanalysis a vg temp k: Reanalysis data pro vides broader scale temperature insights, important for capturing re gional climate patterns that af fect mosquito habitats. - reanalysis tdtr k, station diur temp rng c: Diurnal temperature range indicates the dif ference b e tween day and night temperatures. This range af fects mosquito acti vity , biting beha vior , and survi v al, directl y inuenc- ing dengue transmission rates. 2.3.2. Pr ecipitation - station precip mm, precipitation amt mm, reanalysis precip amt kg per m2, reanalysis sat precip amt mm: Precipitation creat es breeding grounds for Aedes mosquitoes, particularly in stagnant w ater . The amount of rainf all inuences mosquito population density , while hea vy rains may disrupt breeding c ycles by w ashing a w ay larv ae. Monitoring precipitation is essential for understanding v ector proliferation and disease risk. Int J Inf & Commun T echnol, V ol. 15, No. 3, September 2026: 1066–1077 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 1069 2.3.3. Humidity - reanalysis relati v e humidity percent, reanalysis specic humidity g per kg: Humidity plays a pi v otal role in mosquito survi v al and longe vity . Higher humidity increases the lifespan of mosquitoes, allo wing them to bite multiple humans and thus spread the disease, while relati v e humidity measures the percentage of saturation. - reanalysis de w point temp k: De w point temperature reects moisture le v els in the atmosphere. Higher de w points create f a v orable conditions for mosquito survi v al and virus transmission. 2.3.4. Diur nal temperatur e range - station diur temp rng c: A smaller diurnal temperature range indicates stable temperatures, which f a v or consistent mosquito acti vity . W ider ranges m ay disrupt m osquito acti vity b ut can also shorten the virus incubation period within the v ector . 2.3.5. V egetation index (ND VI) - ndvi se, ndvi sw , ndvi ne, ndvi nw: ND VI measures v e getation density around the city . Dense v e getation pro vides ideal breeding and resting places for mosquitoes. Higher ND VI v alues generally correspond to higher mosquito populations, increasing the risk of v ector -borne disease transmission in those areas. Moni- toring v e getation helps identify hotspots for v ector acti vity . Each of these parameters of fers unique insights into the en vironmental and climatic conditions con- ducti v e to dengue transmission. Their combined analysis enables more accurate predictions of v ector -borne disease outbreaks. 2.4. Principal component analysis Since the en vironmental, epidemiological, and meteorological features are highly correlated, PCA w as conducted to reduce dimensionality , thereby impro ving model stability and prediction performance [8], [12], [21]. The dataset initially contained 24 climatic, en vironmental , and epidemiological features. PCA w as con- ducted to re du c e dimensionality and identify the most informati v e predictors. As sho wn in Figure 2, through multiple trials and e xperiments, four k e y features were nalized: precipitation amount (precipitation amt mm), de w point temperature (reanalysis de w point temp k) specic humidity (reanalysis specic humidity g per kg) and historical total cases (total cases). These features consistently demonstrated the strongest contrib ution to e xplaining v ariance in dengue incidence patterns. The usa of these four features across dif ferent models is sho wn in Figure 2. Figure 2. PCA analysis sho wing feature importance across all models 2.5. Hyper parameter tuning Hyperparameter tuning met hod s lik e Bayesian [13] and grid search [29] are popular methods for h y- perparameter optimization. All models were ne-tuned using Optuna [13]. F or each model, a predened search space w as e xplored o v er multiple trials, where model performance w as e v aluated using v alidation loss. The Pr o gnosis of vector borne dengue disease outbr eak in urban ar eas using ... (Pr aktik S. Mac hc har) Evaluation Warning : The document was created with Spire.PDF for Python.
1070 ❒ ISSN: 2252-8776 tuning process w as carried out in a consistent manner across models to identify stable parameter settings. The best-performing h yperparameter congurations were then selected and used for the nal training and e v aluation of each model. Details of the optimized h yperparameters for each model are pro vided in Appendix A. 3. RESUL TS AND DISCUSSION In this study , multiple machine learning and deep learning models were e v aluated to forecast dengue outbreaks using multi v ariate time series data from San Juan (sj) and Iquitos (iq). Six models were tested, in- cluding LightGBM, Stack ed LSTM, CNN-BiLSTM, Seq2Seq, RNN, SARIMAX as sho wn in T able 1. The performance w as assessed using root mean squared error (RMSE) and mean absolute error (MAE) metrics. RMSE w as selected as the primary metric because it hea vily penalizes lar ge prediction errors, which is critical for dengue forecasting where missing outbreak spik es has serious public health consequences. MAE comple- ments RMSE by pro viding the a v erage typical error magnitude, and together these metrics enable assessment of both w orst-case scenarios and consistent performance—essential for epidemiological forecasting where out- break timing and magnitude are paramount. R² w as e xcluded as it can be misleadingly high in time series with strong seasonal patterns and does not adequately penalize temporal lag or phase shifts in outbreak predictions. Based on comprehensi v e e v aluation, LightGBM achie v ed the best performance with an RMSE of 16.94 and MAE of 11.42. This section presents an in-depth analysis of this model, while detailed results of all other models are pro vided in Appendix B for reference. T able 1. Performance comparison of all models Model RMSE MAE LightGBM 16.94 11.42 Stack ed LSTM 18.75 18.01 CNN-BiLSTM 22.13 19.38 Seq2Seq 22.23 13.44 RNN 25.27 13.98 SARIMAX 34.90 18.37 3.1. LightGBM-best model The LightGBM model achie v ed the best performance with an RMSE of 16.94 and MAE of 11.42, demonstrating the ef fecti v eness of gradient boosting for dengue forecasting. The model sho wed strong predic- ti v e capability and e xcellent interpretability (see Figure 3). Figure 3. Actual vs predicted v for LightGBM model 3.1.1. Pr edicti v e perf ormance analysis The LightGBM model’ s predic tions demonstrate strong alignment with actual outbreak patterns using a multi v ariate approach that inte grates four k e y features: preci pitation amount, de w point temperature, total cases, and specic humidity . The model e xcels at capturing o v erall outbreak trends with minimal noise in predictions and successfully predicts the magnitude of outbreak peaks , though occasionally underestimates Int J Inf & Commun T echnol, V ol. 15, No. 3, September 2026: 1066–1077 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 1071 e xtreme e v ents. The gradient boosting frame w ork with 245 estimat ors ef fecti v ely learns comple x interactions between climate and epidemiological v ariables. 3.1.2. F eatur e importance analysis The LightGBM model w as trained using four k e y features selected through feature selection and PC A: precipitation amount (preci pitation amt mm), specic humidity (reanalysis specic humidity g per kg), de w point temperature (reanalysis de w point temp k), and historical total cases (total cases). The model’ s feature importance analysis re v ealed that precipitation emer ged as the dominant predictor among these v ariables, indi- cating that rainf all patterns create f a v orable breeding conditions for mosquito v ectors. De w point temperature and specic humidity , both moisture-related v ariables, signicantly inuenced predictions by af fecting v ec- tor survi v al and reproduction rates. The tree-based structure of LightGBM pro vides e xcellent interpretability , allo wing the model to capture non-linear interactions between these climate and epidemiological v ariables. 3.2. Comparati v e analysis LightGBM, with an RMSE of 16.94 and MAE of 11.42, pro vides e xcellent interpretability and cap- tures comple x non-linear relationships ef fecti v ely . LightGBM is suited for model interpretability and under - standing of feature interactions, making it ideal for public health applications where transparenc y in decision- making is crucial. 3.3. Principal component analysis and featur e selection acr oss models PCA w as conducted to identify the most inuential climate and en vironmental features for dengue pre- diction. The analysis re v ealed that three moisture-related v ariables precipitation (precipi tation amt mm), de w point temperature (reanalysis de w point temp k) and specic humidity (reanalysis specic humidity g per kg) collecti v ely accounted for the majority of v ariance in dengue incidence, conrming their critical role in disease transmission dynamics. Precipitation emer ged as the most frequently selected feature, appearing in four of six models (Light- GBM, RNN, Seq2Seq, and SARIMAX), conrming its fundamental role in dengue prediction. The best- performing models utilized comprehensi v e moisture-related v ariables-LightGBM (RMSE 16.94) incorporated precipitation, de w point tempera ture, and specic humidity alongside historical case data. While multi v ariate models generally outperform ed uni v ariate approaches, Stack ed LSTM (RMSE 18.75) demonstrated compet- iti v e perform ance using only historical case patterns, indicating that temporal autocorrelation in disease in- cidence contains substantial predicti v e information. The PCA-guided feature selection strate gy successfully identied inuential predictors while reducing model comple xity , pro viding strong e vidence for climate-dri v en dengue transmission patterns. 3.4. Simplicity vs complexity trade-off The deep learning models, including Seq2Seq, RNN, CNN-BiLSTM, and Stack ed LSTM, pro vided moderate to good performance, with Stack ed LSTM performing the best among them. Ho we v er , the LightGBM model, which is relati v ely simpler in architecture, outperformed the deep learning models. This outcome chal- lenges the assum p t ion that deeper , more comple x architectures necessarily yield better results in multi v ariate time series forecasting tasks. The superior performance of traditional machine learning approaches suggests that for the gi v en dataset size and feature space, model comple xity does not al w ays translate to impro v ed prediction accurac y . 3.5. LightGBM’ s superior perf ormance interpretability through its tree-based structure and gradient boosting frame w ork. Its ability to pro vide clear feature importance rankings and capture non-linear interactions between climate v ariables mak es it v aluable for understanding the dri v ers of dengue outbreaks. Its transparent decision-making process is cruc ial for public health applications where understanding model predictions is as important as accurac y . The model’ s com- putational ef cienc y and strong predicti v e capability mak e it ideal for operational forecasting and strate gic planning. 3.6. Deep lear ning limitations The moderate performance of deep learning models (Stack ed LSTM, CNN-BiLSTM, Seq2Seq, and RNN) compared to LightGBM highlights potential limitations when applied to this type of dengue dataset. Pr o gnosis of vector borne dengue disease outbr eak in urban ar eas using ... (Pr aktik S. Mac hc har) Evaluation Warning : The document was created with Spire.PDF for Python.
1072 ❒ ISSN: 2252-8776 While these architectures e xcel at capturing long-term dependencies and comple x temporal patterns, the char - acteristics of dengue outbreak data—including irre gular periodicity , sparse e xtreme e v ents, and climate-disease interaction comple xities—may pose challenges that traditional machine learning approaches handle more ef- ciently . The computational cost and relati v ely higher RMSE suggest that deep learning models need further renement or additional training data to compete with traditional approaches in this forecasting conte xt. 3.7. Classical model constraints The SARIMAX model’ s poor performance underscores the limitations of classical time series models when applied to multi v ariate forecasting tasks. SARIMAX’ s statistical frame w ork, whi le ef fecti v e for uni v ari- ate analysis, struggles to adequately model the comple x interactions between multiple climate v ariables that inuence dengue outbreaks. This limitation emphasizes the need for machine learning approaches that can capture non-linear relationships and feature interactions in epidemiological forecasting. The ndings ha v e practical implications for dengue surv eillance systems. The models pro vide a cri ti- cal timeframe for implementing pre v enti v e interv entions and resource allocation. Feature importance analysis consistently identied precipi tation, de w point temperature, and humidity as k e y predictors, which can guide tar geted mosquito control ef forts during high-risk climatic conditions. LightGBM is ideal for both real-time forecasting and strate gic planning decisions due to its balance of accurac y and interpretability . Current limitations include the dataset being restricted to San Juan (SJ) and Iquitos (Iq), focus on climate v ariables without socioeconomic f actors, and weekly resolution that may miss ner temporal dynamics. Future research should e xplore ensemble approaches combining model s trengths, inte gration of additional data sources such as satellite imagery and mobi lity data, de v elopment of city-specic models, and in v estig ation of climate change scenarios for long-term outbreak predictions. Note: Detailed r esults for all other e valuated models (Seq2Seq, RNN, CNN-BiLSTM, Stac k ed LSTM, and SARIMAX) ar e pr o vided in Appendix B, including performance metrics, visualizations, and model-specic analyses. 4. CONCLUSION This research demonstrates the ef fecti v eness of using multi v ariate analysis and time series forecasting techniques for predicting outbreaks of v ector -borne diseases, with a specic focus on dengue as a demonstra- tion case. The models e xplored in this study , including Seq2Seq, RNN, CNN-BiLSTM , Stack ed LSTM, and LightGBM. Pro vides a di v erse set of approaches for capturing the comple x interactions between en vironmental f actors and disease transmission. The results highlight the potential of adv anced deep learning models, such as Stack ed LSTM and CNN-BiLSTM, and machine learning models lik e LightGBM SARIMAX, which strug- gled to handle the dataset’ s comple xity . These ndings align with recent comparati v e studies demonstrating the superiority of machine learning approaches for dengue forecasting. These deep learning models were able to model the intricate temporal and seasonal patterns that are critical in disease forecast ing, thus of fering more accurate predictions. The study reinforces that using a range of models—spanning from deep learning to sta- tistical approaches—of fers a comprehensi v e me thod for forecasting v ector -borne disease outbreaks, enabling public health authorities to deplo y proacti v e measures to control the spread of disease in urban areas. 5. FUTURE W ORK The future w ork of this research aims to enhance the model’ s capabilities by e xploring more ad- v anced models and comparing their perf ormance in predicting v ector -borne diseases. Additionally , e xpanding the dataset to include di v erse geographical re gions and other diseases lik e malaria, chikungun ya, and Zika will allo w the model to generalize better . Future research may also suggest ne w , impro v ed models while incorporat- ing additional v ariables, such as socio-economic and en vironmental f actors. Furthermore, inte grating real-time data streams could impro v e prediction accurac y and assist in the de v elopment of early w arning systems for outbreak mitig ation. Lar ge language models may also be e xplored for enhanced forecasting capabilities. FUNDING INFORMA TION Authors state that no funding w as in v olv ed. Int J Inf & Commun T echnol, V ol. 15, No. 3, September 2026: 1066–1077 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 1073 A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the Contrib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib u- tions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu Pratik S. Machchar ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Purvi N. Ramanuj ✓ ✓ Rajan P atel ✓ Jitendra Bhatia ✓ K untesh Jani ✓ C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administration V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting CONFLICT OF INTEREST ST A TEMENT Authors declare no conict of interest. D A T A A V AILABILITY Deri v ed data supporting the ndings of this study , including dengue case records and climate v aria bles, and the code used for model de v elopment is a v ailable from the author [P .S.M.] upon request. Some data may be subject to institutional or pri v ac y restrictions. REFERENCES [1] W orld Health Or g anization (WHO), “Dengue - Global situation, ” W eekly Epidemiological Record , v ol. 100, no. 52, pp. 665–678, 2024, [Online]. A v ailable: https://www .who.int/publications/i/item/who-wer10052-665-678. [2] P . Silitong a, B. De wi, A. Bustamam, and H. Al-Ash, “Ev aluation of dengue model performances de v eloped using articial neural netw ork and random forest classiers, ” Procedia Computer Science , v ol. 179, pp. 135–143, 2021, doi: 10.1016/j.procs.2020.12.018. [3] C. Caminade et al . , “Impact of recent and future climate change on v ector -borne diseases, ” Annals of the Ne w Y ork Academy of Sciences , v ol. 1436, no. 1, pp. 157–173, 2019, doi: 10.1111/n yas.13950. [4] P . Arcari, N. T apper , and S. Pfueller , “Re gional v ariability in relationships between climate and dengue/DHF in Indonesia, ” Sing a- pore Journal of T ropical Geograph y , v ol. 28, pp. 251–272, 2007, doi: 10.1111/j.1467-9493.2007.00300.x. [5] Y . L. Hii, H. Zhu, N. Ng, L. C. Ng, and J. Rockl ¨ ov , “F orecast of dengue incidence using temperature and rainf all, ” PLOS Ne glected T ropical Diseases , v ol. 6, no. 11, p. e1908, 2012, doi: 10.1371/journal.pntd.0001908. [6] S. Bal and S. Sodoudi, “Modeling and prediction of dengue occurrences in K olkata, India, based on climate f actors, ” International Journal of Biometeorology , v ol. 64, no. 8, pp. 1379–1391, 2020, doi: 10.1007/s00484-020-01918-9. [7] M. Gharbi et al. , “T ime series ana lysis of dengue incidence in Guadeloupe, French W est Indies: forecasting models using climate v ariables as predictors, ” BMC Infectious Diseases , v ol. 11, p. 166, 2011, doi: 10.1186/1471-2334-11-166. [8] ´ A. G. Mu ˜ noz et al. , “ AeDES: a ne xt-generation monitoring and forecasting system for en vironmental suitability of Aedes-borne disease transmission, ” Scientic Reports , v ol. 10, no. 1, p. 12640, 2020, doi: 10.1038/s41598-020-69625-4. [9] S. S. Deshmukh and S. Deb, “ A surv e y of statistical and machine learning methods of quantile re gression in time series and their suitability in predicting dengue outbreaks, ” Japanese Journa l of Sta tistics and Data Science , pp. 1–49, 2025, doi : 10.1007/s42081- 025-00297-y . [10] S. Gambhir , S. K. Malik, and Y . K umar , “The diagnosis of dengue disease: an e v aluation of three machine learning approaches, ” International Journal of Healthcare Information Systems and Informatics , v ol. 13, no. 3, pp. 1–19, 2018. [11] P . Siriyasatien, S. Chadsuthi, K. Jampachaisri, and K. K esorn, “Dengue epidemics prediction: A surv e y of the state-of-the-art based on data science processes, ” IEEE Access , v ol. 6, pp. 53757–53795, 2018, doi: 10.1109/A CCESS.2018.2871241. [12] R. Jain, S. Sontisiri kit, S. Iamsiritha w orn, and H. Prendinger , “Prediction of dengue outbreaks based on disease surv eillance, mete- orological and socio-economic data, ” BMC Infectious Diseases , v ol. 19, no. 1, p. 272, 2019, doi: 10.1186/s12879-019-3874-x. [13] S. G. Kakarla et al. , “W eather inte grated multiple machine learning models for prediction of dengue pre v alence in India, ” Interna- tional Journal of Biometeorology , v ol. 67, no. 2, pp. 285–297, 2023, doi: 10.1007/s00484-022-02405-z. [14] W . Anggraeni et al. , “Prediction of dengue fe v er outbreak based on climate f actors using fuzzy-logistic re gression, ” in Proc. 2020 International Seminar on Intelligent T echnology and Its Applications (ISITIA) , 2020, pp. 199–204, doi: 10.1109/ISI- TIA49792.2020.9163708. [15] O. S. Baquero, L. M. R. Santana, and F . Chiara v alloti-Neto, “Dengue forecasting in S ˜ ao P aulo city with generalized additi v e models, articial neural netw orks and seasonal autore gressi v e inte grated mo ving a v erage models, ” PLoS One , v ol. 13, no. 4, p. e0195065, 2018, doi: 10.1371/journal.pone.0195065. Pr o gnosis of vector borne dengue disease outbr eak in urban ar eas using ... (Pr aktik S. Mac hc har) Evaluation Warning : The document was created with Spire.PDF for Python.
1074 ❒ ISSN: 2252-8776 [16] P . Guo et al. , “De v eloping a dengue forecast m odel using machine learning: a case study in China, ” PLOS Ne glected T ropical Diseases , v ol. 11, no. 10, p. e0005973, 2017, doi: 10.1371/journal.pntd.0005973. [17] J . Xu et al. , “F orecast of dengue cases in 20 Chinese cities based on the deep learning method, ” International Journal of En viron- mental Research and Public Health , v ol. 17, no. 2, p. 453, 2020, doi: 10.3390/ijerph17020453. [18] E. Mussumeci and F . C. Coelho, “Lar ge-scale multi v ariate forecasting models for Dengue - LSTM v ersus random forest re gression, ” Spatial and Spatio-temporal Epidemiology , v ol. 35, p. 100372, 2019, doi: 10.1016/j.sste.2020.100372. [19] A. L. Ramadona, L. Lazuardi, Y . L. Hii, A. Holmner , H. K usnanto, and J. Rockl ¨ ov , “Prediction of dengue outbreaks based on disease surv eillance and meteorological data, ” PLoS One , v ol. 11, no. 3, p. e0152688, 2016, doi: 10.1371/journal.pone.0152688. [20] T . K. Y amana, S. Kandula, and J. Shaman, “Superensemble forecasts of dengue outbreaks, ” Journal of the Ro yal Society Interf ace , v ol. 13, no. 123, p. 20160410, 2016, doi: 10.1098/rsif.2016.0410. [21] M . A. Mobin, “Multi v ariate forecasting of dengue infection in Bangladesh: e v aluating the inuence of data do wnscaling on machine learning predicti v e accurac y , ” BMC Infectious Diseases , v ol. 25, no. 1, p. 761, 2025, doi: 10.1186/s12879-025-11159-z. [22] X. Chen and P . Morag a, “ Assessing dengue forecasting methods: a comparati v e study of statistical models and machine learning techniques in Rio de Janeiro, Brazil, ” T ropical Medicine and Health , v ol. 53, no. 1, p. 52, 2025, doi: 10.1186/s41182-025-00723-7. [23] A. E. Laureano-Rosario et al. , “ Application of articial neural netw orks for dengue fe v er outbreak predictions in the northwest coast of Y ucatan, Me xico and San Juan, Puerto Rico, ” T ropical Medicine and Infectious Disease , v ol. 3, no. 1, p. 5, 2018, doi: 10.3390/tropicalmed3010005. [24] I. Sutsk e v er , O. V in yals, and Q. V . Le, “Sequence to sequence learning with neural netw orks, ” arXi v , 2014, [Online]. A v ailable: https://arxi v .or g/abs/1409.3215. [25] S. Hochreiter and J. Schmidhuber , “Long short-term memory , ” Neural Computation , v ol. 9, pp. 1735–1780, 1997, doi: 10.1162/neco.1997.9.8.1735. [26] G. K e et al. , “LightGBM: a highly ef cient gradient boosting decision tree, ” in Proc. 31st International Conference on Neural Information Processing Systems (NIPS’17) , 2017, pp. 3149–3157. [27] M. Cabrera et al. , “Dengue prediction in Latin America Using machine learning and the one health perspecti v e: A literature re vie w , ” T ropical Medicine and Infectious Disease , v ol. 7, p. 322, 2022, doi: 10.3390/tropicalmed7100322. [28] H. Barle vi, Z. E. R. Cordero, F . J. Colon-Gonzalez, and R. Lo we, “ A reproducible ensemble machine learning approach to forecast dengue outbreaks, ” Scientic Reports , v ol. 14, no. 1, p. 3807, 2024, doi: 10.1038/s41598-024-52796-9. [29] Z. I. Lai, W . K. Fung, and E. Che w , “Machine learning models for dengue forecasting in Sing apore, ” arXi v preprint arXi v:2407.00332 , 2024, doi: 10.48550/arXi v .2407.00332. [30] Y . Qian, “DengAI Dataset, ” Kaggle , [Online]. A v ailable: https://www .kaggle.com/datasets/qian yig ang129/deng ai-dataset. APPENDICES APPENDIX A. Model h yper parameters This subsection summarizes the main h yp e rparameters and settings for each model as used in this study . A.1. LightGBM Input features: [precipitation amt mm, reanalysis de w point temp k, total cases, reanalysis specic humidity g per kg] (Multi v ariate). Learning rate: 0.00475, L1 re g: 0.0981, L2 re g: 0.0141, Num lea v es: 245, Num estimators: 245, Feature fraction: 0.8150, Max depth: 8, Min data in leaf: 74, RMSE: 16.94. A.2. Stack ed LSTM Input features: [total cases] (Uni v ariate), K ernel initializer: He normal, Number of layers: 1, Neurons: 69, Dropout rate: 0.105, Epochs: 97, Batch size: 16, Optim izer: Adam, Acti v ation function: ReLU, Early stopping patience: 16, RMSE: 18.75. A.3. CNN-BiLSTM Input features: [total cases] (Uni v ariate), Number of con v olutional lters: 39, K ernel size: 3, BiLSTM neurons: 127, Epochs: 19, Batch : 16, Optimizer: SGD, Acti v ation: ELU, Early stopping patience: 11, RMSE: 22.13. A.4. Seq2Seq Input features: [precipitation amt mm, reanalysis specic humidity g per kg, total cases] (Multi v ari- ate), W indo w length: 8, Neurons: 47, Dropout rate: 0.174, Epochs: 61, Batch size: 5, Optimizer: Nadam, Early stopping patience: 17, RMSE: 22.23. A.5. RNN Input features: [precipitation amt mm, reanalysis de w point temp k, total cases, reanalysis specic humidity g per kg] (Multi v ariate). K ernel init: Glorot uniform, W indo w length: 5, Num layers: 1, Neurons: 48, Dropout: 0.111, Epochs: 47, Batch size: 10, Optimizer: RMSprop, Acti v ation: ReLU, Early stopping patience: 14, RMSE: 25.27. Int J Inf & Commun T echnol, V ol. 15, No. 3, September 2026: 1066–1077 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Inf & Commun T echnol ISSN: 2252-8776 ❒ 1075 A.6. SARIMAX Input (e xogenous) features: [precipitation amt mm, reanalysis a vg temp k], T ar get v ariable: total cases, T ype: Multi v ariate, F orecasting horizon: 4 steps ahead, RMSE: 34.90. APPENDIX B. Detailed r esults of other models This subsection pro vides comprehensi v e results for the remaining v e models e v aluated in the study: Seq2Seq, RNN, CNN-BiLSTM, Stack ed LSTM, and SARIMAX. B.1. Seq2Seq model The Seq2Seq multi output model achie v ed an RMSE of 22.23. This model’ s moderate performance indicates its capability to capture sequential dependencies, though it did not perform optimally compared to the best models. See Figure 4 the Seq2Seq model sho ws general alignment with observ ed outbreaks b ut e xhibits signicant uctuations during sharp changes. Figure 4. Actual vs predicted graph of Seq2Seq model B.2. RNN model The RNN multioutput model produced an RMSE of 25.27. The shallo w structure with 1 layer and 48 neurons limited its ability to fully capture the nuances of the multi v ariate data. See Figure 5 the RNN model sho ws signi cant de viation from actual data during periods of rapid change, f ailing to predict higher peaks accurately . Figure 5. Actual vs predicted graph of RNN model B.3. CNN-BiLSTM model The CNN-BiLSTM model achie v ed an RMSE of 22.13, combining con v olutional layers for spatial feature e xtraction with bidirectional LSTMs for temporal dependencies. See Figure 6 the CNN-BiLSTM sho ws reasonable alignment with actual data though struggles with certain peaks. Pr o gnosis of vector borne dengue disease outbr eak in urban ar eas using ... (Pr aktik S. Mac hc har) Evaluation Warning : The document was created with Spire.PDF for Python.