Introduction

Dengue, a vector-borne disease transmitted by Aedes aegypti and Aedes albopictus mosquitoes, poses a significant global public health threat, especially in tropical and subtropical regions. Dengue is now endemic in more than 125 countries [1]. Globally, its incidence increased tenfold between 2000 and 2019 to more than 5 million reported cases, with 6.5 million cases reported in 2023 [2]. During the first nine months of 2024, an unprecedented dengue upsurge occurred, with more than 13 million reported cases [3]. The reported rate of dengue likely represents an underestimate due to testing patterns, and it has been estimated that there were 59 million cases in 2021 [4]. The global spread of dengue has been driven by urbanization, globalization, with its associated increase in human mobility, and climate change [5–8]. The global economic cost of dengue was estimated to be US$8.9 billion in 2013 [9].

Most dengue cases are subclinical or present as a non-specific febrile illness with flu-like symptoms [10,11]. However, 2–5% of symptomatic cases progress to severe dengue, a life-threatening condition associated with bleeding and shock [10]. Treatment is supportive, including oral or fluid replacement depending on illness severity [10]. Vector control and preventing mosquito bites historically have been the main approaches to dengue prevention and control [10,11]. Novel vector control methods, including the release of sterile, genetically modified, or Wolbachia-infected mosquitoes, have been developed but are not yet in widespread use [10,11]. Three different tetravalent attenuated virus vaccines, targeting the four serotypes of dengue, have shown efficacy in randomized controlled trials [11–13], and two of these have been licensed [11,12]. However, vaccines are not yet widely used due to a number of challenges [10,11].

Vietnam, where dengue is endemic, is a high-incidence country, although its annual incidence tends to be cyclical [13]. In 2023, the number of reported dengue cases in Vietnam was 369,000, the second highest in the world [1], with an estimated 1.1 million actual cases in 2021 when adjusting for underreporting [14]. The Mekong Delta Region (MDR) of Vietnam has been particularly affected, with 40–50% of reported cases in Vietnam occurring in the MDR during 2000–2016 [13]. Drivers of dengue incidence in Vietnam include urbanization, ambient temperature, and hydrometeorological variables [15]. The economic cost of dengue in Vietnam was estimated to be US$94 million in 2013 [9]. Vietnam’s current dengue prevention and control strategies, as outlined in the Ministry of Health’s Decision No. 3711/QD-BYT dated 19/9/2014, rely predominantly on vector control, such as entomological monitoring, community mobilisation to eliminate breeding sites, and insecticide spraying [16]. However, these interventions are often reactive, only implemented after cases have been reported, which could be too late to preemptively reduce the impact of dengue outbreaks.

A great deal of work has been devoted to developing models to predict dengue incidence and outbreaks [17]. Such models could be used to develop early warning systems to enable proactive interventions. Investigators recently developed an ensemble of probabilistic dengue models to predict dengue incidence one to six months in advance at the provincial level -- the largest geographical and administrative unit -- in Vietnam [18]. While this approach offers a valuable tool for improving dengue surveillance and planning preventive interventions, a critical limitation is that the predictions are made at the provincial level, which may not be sufficient for localised action. In Vietnam, dengue prevention and control measures are implemented at the district level, with a typical province in the MDR comprising seven to 15 districts, making it necessary to refine models to provide district-level predictions that align more closely with the administrative and operational realities of dengue control programs.

This study aims to evaluate the performance of a range of complementary probabilistic models for forecasting the incidence of dengue 1–3 months in advance at the district level in the MDR, incorporating meteorological, epidemiological, sociodemographic, and preventive intervention data. To enhance the accuracy and robustness of forecasts, we integrated five models into an ensemble. In future work, we aim to incorporate our ensemble model into user-friendly software for an early warning system to assist districts in proactively implementing interventions to mitigate dengue outbreaks.

Our study presents three novelties: first, we generate probabilistic forecasts for 112 districts, aligning with the operational scale at which Vietnam’s Ministry of Health implements vector-control measures. Second, by combining the best‐performing hierarchical, semi‐mechanistic (hhh4), and supervised PCA models into an ensemble, we achieve greater accuracy and calibration across diverse climatic and demographic contexts. Third, our district-level, 1–3-month lead-time predictions can be used in cluster-randomized proactive intervention trials (e.g., targeted larval control or community education only in districts where outbreaks are predicted), thereby evaluating the efficacy of interventions under real-world conditions. This work could allow for more proactive dengue prevention efforts and improved resource allocation in Vietnam.

Methods

Overview of approach

The goal for these analyses was to generate and evaluate probabilistic forecasts for dengue cases and outbreaks at the district level, one to three months in the future, and to combine the best of these models into an ensemble model. We evaluated three distinct modelling approaches – hierarchical spatiotemporal Bayesian models, semi-mechanistic spatiotemporal (hhh4) models, and a machine learning supervised principal components analysis regression -- each with various sets of variables. The data were split into three periods: the initial set covering the years 2004–2011, the set from 2012-2016 for time series cross validation, and the evaluation set from 2017-2022 to assess the performance of the ensemble model out-of-sample (Fig 1). Using a rolling scheme, the entire 2004–2016 dataset participates in cross‐validation: 2004–2011 from the initial window, and from January 2012 onward, we continually enlarge the training set by one month at each step, always holding out the next three months for testing. The model is fitted only to data up to a specified date, say January 2012. Once fitted, it makes three forecasts: one month ahead (February 2012), two months ahead (March 2012), and three months ahead (April 2012). We compare each forecast to the actual values in those months and record CRPS, bias, and diffuseness. Next, we expand the fitting window to span from 2004 to February 2012, and we reserve March–May 2012 for testing. We fit again, forecasting those three months. Repeating this every month from January 2012 through December 2016 lets us use the full 2004–2016 record. This rolling approach mirrors real-time forecasting and gives the out-of-sample evaluation for every month in 2012–2016.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Time series of dengue cases from 2004 to 2022, highlighting the initial training period (red), the time-series cross validation (TSCV) phase (blue), and the evaluation period (grey).

The initial training data is used for model calibration, followed by TSCV for model validation, and concluding with an evaluation period for assessing forecast accuracy.

https://doi.org/10.1371/journal.pntd.0013571.g001

Models

We evaluated three types of modelling approaches:

1. 1]. Hierarchical Bayesian spatiotemporal models. These models tested lags of different covariates, including meteorological variables (e.g., temperature, precipitation), sociodemographic variables (e.g., urbanization, income), and preventive measures/entomologic indices (e.g., active spray, communications or training, Breteau index). The models also included 1–3 month lags of observed monthly dengue cases and spatiotemporal random effects with different structures to account for different sources of correlation and ensure accurate statistical inference (i.e., autoregressive correlated terms modelled independently across each district, spatial random effects, spatiotemporal random effects). This work builds upon the approach developed by Colón-González [18] for monthly data at the provincial level in Vietnam, with a couple of simplifications. First, we used weather variables and the observed monthly dengue case variable lagged by the amount of the maximum forecast horizon (e.g., for models predicting up to 3 months in the future, the minimum lag of observed cases and weather variables was 3 months) — this avoided the need to recursively use forecasts of the weather variables as predictors in the model. There are numerous possible combinations of covariates and random effects structures, resulting in the evaluation of 60 models (see S1 Table). To ensure a thorough exploration of potential covariates, we tested at least one representative covariate from each covariate type (meteorological, sociodemographic, preventive measures/entomologic indicators, monthly dengue cases) for the different sets of covariates. To address high correlations between some variables, we avoided including highly correlated variables in the same model (see S1 Table).
2. 2]. hhh4 models. These semi-mechanistic spatiotemporal models account for variations in baseline incidence, autocorrelation, and spatial spread. Each of these three components is modelled as a function of covariates. The endemic and epidemic components used 3-month lagged temperature, precipitation, and cumulative incidence as covariates. The spatial spread component assumed that the correlation between neighbouring districts followed a power-law structure. Each of the components was modelled using a random effect, which allowed the effects to vary by district. These models were fitted using the hhh4 function in the surveillance package in R [26].
3. 3]. Supervised (“Y-aware”) principal components regression [27,28] that models the time series of monthly dengue cases from each district as a function of the lags in all the districts (summarised with principal components), along with seasonality and lags of the weather variables from the same district. A matrix of lagged, standardized log-incidence was generated, including lags from 3-5 months for all districts. We first fit univariate regression for each of the covariates and scaled the variables based on the effect estimates; this effectively provides more weight to variables that are more relevant to the outcome. We then performed principal components analysis (PCA) [27,28] using these rescaled covariates. Based on the variance plots, the top 10 principal components were used in a regression model along with harmonic terms to capture seasonality, an autoregressive correlated random intercept, and lagged temperature and rainfall. These models were fit separately for each district using the INLA package in R.

See S3 Text for further details of the three modelling approaches.

Time series cross validation and model selection.

Considering different covariate combinations and random effect structures, the performance of 72 models was evaluated using time series cross validation, including 60 hierarchical Bayesian spatiotemporal models, 9 hhh4 models, and 3 PCA models, across 60 forecast periods (i.e., months) from January 2012 to December 2016.

Each of these 72 models was evaluated using time series cross validation by moving forward the end of the training period by one month at a time. Several forecasting performance measures were applied to evaluate the forecasting performance for a forecast horizon of 1–3 months. The Continuous Ranked Probability Score (CRPS), bias, and diffuseness were used to assess the accuracy of forecasting dengue incidence [18], and the Brier score was used to assess the accuracy of forecasting dengue outbreaks [29]. The lower the CRPS, the more accurate the forecast, with a value of 0 indicating a perfect forecast; there is no upper bound to CRPS. Bias values can range from -1 to 1, with a score of zero reflecting no bias. The Brier score can range from 0 to 1, with a lower value indicating more accurate outbreak classification. We evaluated diffuseness (i.e., the width of the predictive distribution) for the top five models with the best performance on the other three metrics. Diffuseness scores closer to 0 are desirable, indicating a narrow (i.e., sharp) distribution, with poor forecasts having higher diffuseness scores. We used the scoringutils package in R to calculate these four metrics [30].

Based on model performance using this time series cross validation approach, we combined a subset of the best models into an ensemble forecast. Models were selected based on three criteria. First, we evaluated how the models performed district by district and used hierarchical clustering to identify sets of models that performed similarly in terms of performance across districts (the heat map by district and by month can be found in Figs B and C in S1 Text) We included at least one representative of each of these clusters in the ensemble. This clustering approach ensured that each “type” of model only had one representative, preventing an imbalance in the ensemble. Second, if multiple models performed similarly in terms of CRPS, we preferred the model with fewer covariates and simpler random effects structures. Finally, to include a diversity of modelling approaches, we ensured that at least one model from each of the three methods (Bayesian spatiotemporal, hhh4, PCA) was included in the ensemble. This process resulted in five models being included in the final ensemble. The weights for the five models were determined based on the CRPS scores and remained constant over time and across districts (see Fig G in S1 Text)

To create the ensemble model, we first generated 10,000 samples derived from the predicted distribution of dengue cases for each of the five models included in the final ensemble. We then calculated the CRPS for each of the five models for the time series cross validation period by comparing the sample predictions against observed dengue cases. Next, we determined weights for each model using (1/CRPS)/sum(1/CRPS). Samples were then taken from the predictive distributions of each of the five individual models, with the sampling proportional to the weights. The 2.5th, 50th, and 97.5th percentiles of the resulting distribution were calculated [18]. A summary of our ensemble method steps can be found in S3 Text.

Reference model.

A basic reference model was developed to estimate the average monthly dengue incidence for each district without covariates, lagged incidence, or spatiotemporal correlation. Let present dengue cases in the district at time where is the total number of districts, and the time steps (months). The model assumes a Poisson likelihood for observed cases and is defined as:

is the population of the district during year , included as an offset to adjust case counts by population. The model incorporates a global intercept , assumes that the mean number of monthly cases across different districts is random and uncorrelated. Additionally, the model includes a cyclic random walk for the calendar month to ensure that seasonal trends are captured while allowing for variations in these trends by district. In line with an earlier study [18], we used the Continuous Ranked Probability Skill Score (CRPSS) to quantify the improvement of a model’s predictive performance relative to this reference model. It is defined as

A positive CRPSS indicates better performance than the reference, with a CRPSS of 1 indicating a perfect model. A CRPSS close to zero suggests no improvement, and a negative CRPSS suggests that the model underperforms relative to the reference.

Availability of code.

All code used in the analysis was written in R version 4.2.2 and is available at Github.com/e-dengue/dengue_District_HPC

Results

Fit of the models

We tested 72 models in total, representing a broad range of model structures and covariates (see detailed composition of the models in the S1 Table. We developed an ensemble model that incorporated five of the top-performing individual models, including three hierarchical Bayesian spatiotemporal models, one hhh4 model, and one PCA model (Table 1). The ensemble model demonstrated a lower CRPS (3.2) than any of the 72 individual models, indicating more accurate probabilistic forecasts compared to any individual model. Among the 72 individual models, spatio-temporal model 3 and spatio-temporal model 2 exhibited the lowest CRPS, with values of 3.38 and 3.99, respectively. In contrast, the seasonal reference model had the highest CRPS of 12.2 compared to all 72 individual models, reflecting the lowest predictive accuracy. Among the five models included in the ensemble, the hhh4 model showed the lowest bias at 0.13, followed by Spatio-temporal 3 with a bias of 0.23. Although the ensemble model had lower bias compared to the baseline model, its bias metric was not superior to that of its five component individual models.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Summary of model performance metrics and features for dengue incidence prediction three months in advance for the baseline model, the ensemble model, and the five individual models included in the ensemble, aggregated across the Mekong Delta Region districts, 2012-2016. Lower CRPS indicates better accuracy, while bias closer to zero suggests closer alignment with observed values, and lower diffuseness reflects more precise predictions. Weight is the weight each of the five individual models contributed to the ensemble.

https://doi.org/10.1371/journal.pntd.0013571.t001

The performance of all models tended to decline as the forecast horizon increased from 1 to 3 months, with CRPSS decreasing and diffuseness increasing with longer lead times, although bias tended to be consistent across time horizons (Fig 2). The decreasing predictive performance of the ensemble model with increasing time horizon is further illustrated in Fig 3, which shows the ensemble forecast mean and its 95% predictive interval compared to the observed dengue cases. For example, during the 2012 dengue wave, the 3-month-ahead prediction (blue solid line) underestimated the magnitude of the outbreak, whereas as we moved closer to the actual date (2-month and 1-month predictions), the forecasts became more accurate, with the 1-month-ahead prediction (orange solid line) tracking closest to the observed cases and the uncertainty intervals narrowing. Nevertheless, despite the lower accuracy of the 3-month predictions, these longer-term forecasts offer early indications of rising cases well in advance of the actual outbreak, giving public health officials time to prepare. For instance, for the period June through September 2012, the 3-month horizon prediction starts showing the upward trend in dengue cases well before the outbreak fully occurs (Fig D in S1 Text). The spatial distribution of forecasts from each of the five models alongside the observed cases. can be found in Figs A-E in S2 Text.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Comparison of performance metrics for the baseline (reference model), the ensemble model, and the five individual models included in the ensemble across forecast horizons of 1 to 3 months, aggregated across the Mekong Delta Region districts, 2012-2016.

Panel (a) shows CRPSS, with higher values indicating a better forecast compared to the baseline model. Panel (b) displays bias, with values closer to zero reflecting more accurate predictions. Panel (c) presents diffuseness, which measures the spread of the predicted probabilities, with lower values indicating greater precision. Panel (d) presents Brier score, using Mean+2 SD, with a lower value indicating more accurate outbreak classification.

https://doi.org/10.1371/journal.pntd.0013571.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Observed vs. forecasted dengue cases (mean and 95% predictive interval) using the ensemble model, by time horizon, aggregated across the Mekong Delta Region districts, 2012-2016.

The solid red line represents observed cases, the solid blue line forecasted cases with a 3-month horizon, the solid green line forecasted cases with a 2-month horizon, and the solid yellow line forecasted cases with a 1-month horizon. The dashed lines correspond to the 95% credible intervals for the forecasts.

https://doi.org/10.1371/journal.pntd.0013571.g003

The ensemble model’s ability to classify outbreak and non-outbreak periods was evaluated using the Brier score and four different outbreak threshold definitions -- mean + 2 standard deviations, 95^th percentile, Poisson threshold, and fixed thresholds based on incidence rates, as described in the Methods. Each threshold involves trade-offs, and the choice of which to use should align with the surveillance system’s programmatic goals. For example, a high fixed threshold will have low sensitivity, meaning fewer outbreaks are detected. However, this also reduces the likelihood of false positives. The highest classification accuracy was achieved using a stringent outbreak threshold set at 50 cases per 100,000 population, followed by the 95^th percentile threshold, and then the mean plus two standard deviations threshold (Fig 4). The accuracy was generally higher in the summer months when case counts were higher. A spatial map comparing outbreaks based on the Mean + 2 SD threshold over the last five years (excluding high epidemic years) with probabilistic ensemble model predictions can be found in Figs F- I in S2 Text.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Brier score averaged by month across the three time horizons using the ensemble model with five different threshold definitions, aggregated across the Mekong Delta Region districts, 2012-2016.

The Brier score can range from 0 to 1, with 0 representing the optimal value.

https://doi.org/10.1371/journal.pntd.0013571.g004

In addition to using the Brier score, the accuracy of the ensemble model for classifying outbreaks can be evaluated by comparing the predicted probability of an outbreak with the frequency of outbreaks in different bins of predicted probabilities. If the points align along the diagonal, this indicates a well-calibrated model (Fig E in S1 Text). In general, the model underestimated the risk of outbreaks at lower predicted probabilities, while overestimating the risk at higher predicted probabilities. The accuracy varied depending on the threshold used, with the fixed thresholds generally showing better calibration.

To explore geographic heterogeneity in accuracy, we computed the Brier score for each district under our main outbreak‐threshold model (Mean + 2 SD) and mapped these values. District‐level Brier scores ranged from approximately 0.05 to 0.22. Southwestern districts, particularly those along the Cambodian border (e.g., An Giang, Kien Giang) exhibited higher Brier values (0.15–0.22), indicating lower accuracy in outbreak prediction. In contrast, central delta districts such as Hau Giang and Vinh Long tended to have lower Brier scores (<0.08), indicating more reliable forecasts in those areas. A handful of coastal districts (e.g., Bac Lieu, Soc Trang) showed intermediate Brier values (0.10–0.15) (Fig G in S1 Text). The spatial pattern of Brier scores suggests that local factors may influence predictive accuracy. Higher Brier scores in southwestern border areas could stem from greater environmental heterogeneity (e.g., fluctuating rainfall patterns and cross‐border movement of people), as well as variable completeness in case reporting. Conversely, the lower Brier scores in central provinces might reflect more stable climatic conditions and more consistent surveillance systems. Further, weighting CRPS by district or by month did not improve results (Fig H in S1 Text).

Model performance out-of-sample

We evaluated the ensemble model’s performance, along with that of its five individual model components and the baseline model, during the 2017–2022 out-of-sample evaluation period (Fig 5). We found lower classification accuracy (higher Brier score) than during the cross-validation period, but still with reasonable accuracy for 1-, 2-, and 3-month horizons. The relative performance of the models across all metrics was similar to what was seen during the cross-validation period, with the lowest Brier score obtained from the ensemble model. As an example of the ensemble model’s out-of-sample prediction performance, Fig F in S1 Text. shows that the model captured the overall seasonal dynamics; however, misalignments are evident in the years 2019 and 2022.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Performance metrics for the baseline (Reference) model, the ensemble model, and the five individual models included in the ensemble across forecast horizons of 1 to 3 months, Mekong Delta Region, during the 2017-2022 out-of-sample evaluation period.

Panel (a) shows CRPSS, with higher values indicating a better forecast compared to the baseline model. Panel (b) displays bias, with values closer to zero reflecting more accurate predictions. Panel (c) presents diffuseness, which measures the spread of the predicted probabilities, with lower values indicating greater precision. Panel (d) represents the Brier score, using Mean+2 SD, which can range from 0 to 1, with 0 representing the optimal value.

https://doi.org/10.1371/journal.pntd.0013571.g005

In 2022, during the post-COVID recovery phase, the ensemble model struggled to predict dengue outbreaks accurately. As seen in Fig 6, sharp increases in observed cases during certain periods of 2022 were not matched by the model’s predictions, or the model predictions only increased in step with the observed cases. As a result of such misalignments, the accuracy of predicting epidemics was compromised, with notable variations observed across different thresholds and years. For instance, the overall accuracy fluctuated considerably between years, achieving the highest performance in 2017 (77%) and 2018 (73%), followed by a decline in 2019 (65%). Although a slight recovery was noted in 2020 (72%) and 2021 (71%), the accuracy dropped again in 2022 (56%). These inconsistencies highlight the ongoing challenges in balancing the detection of outbreaks and minimising false alarms, particularly when applying the model to distinct groups of years with varying outbreak dynamics.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Time-series plot of predicted dengue cases from the top five individual models compared to observed cases (2017-2022).

This plot illustrates the temporal patterns and predictive performance of different models over the years. Notably, there is a visible misalignment in predictions for the years 2019 and 2022, indicating challenges in model accuracy during these periods.

https://doi.org/10.1371/journal.pntd.0013571.g006

Discussion

In this study, we have demonstrated that an ensemble of probabilistic models can predict dengue outbreaks at 1- to 3-month lead times, specifically at the district level in the MDR region of Vietnam. By combining individual spatial temporal, semi-mechanistic (hhh4), and supervised PCA models, our ensemble outperformed individual models. The ensemble model’s strength lies in its ability to integrate key features of the five individual models included in the ensemble, such as temporal smoothing through autoregressive correlation, district-level random effects, meteorological factors (i.e., temperature and precipitation), and lagged cases from various periods. The ability to generate probabilistic predictions can enhance decision-making for proactive public health interventions, providing timely early warnings for targeted dengue control measures [31]. Beyond shifting to the district level, our contributions include combining three modelling families into a weighted ensemble and adopting a decision-focused evaluation (CRPS, Brier, calibration, sharpness) with multiple outbreak thresholds, so model probabilities translate directly into actions. Disruptions to historical seasonal patterns in recent years have proven a challenge to the accuracy of the models, and further monitoring is needed to determine if model performance improves as epidemic patterns settle back into a more typical seasonality.

The Bayesian spatio-temporal models and hhh4 frameworks already represent well-established time-series forecasting approaches. Bayesian space-time models have a long history of probabilistic prediction in epidemiology, and they have proven themselves repeatedly for infectious disease counts. We focused on these epidemiologically grounded, probabilistic approaches, including lagged-incidence and PCA-based models, as they yield full predictive distributions [18]. Furthermore, our ensemble consistently delivers strong and reliable performance across every forecast period, district, and horizon. While a single model might outperform the others in a handful of district-month combinations, none remains consistently ahead for all 72 forecasts, 112 districts, and three lead times. By weighting each model according to its accuracy (giving more weight to those with lower CRPS), the ensemble balances complementary strengths: some models provide more accurate short-term forecasts in normal seasons, others cope better with unusual rainfall, or capture spatial links between neighbouring areas. In out-of-sample tests (2017–2022), the 95% prediction intervals of the ensemble were consistently narrower and better calibrated than those of any individual component, resulting in fewer false alarms and missed outbreaks, aligning with what health teams need when resources are limited. Even in years with atypical seasonality or reporting changes after COVID-19, the ensemble remained the most robust early-warning tool, smoothing over the under- or over-predictions that tripped up single models.

Recent dengue forecasting efforts have employed a variety of methods, ranging from single-model approaches, such as ARIMA and mechanistic SIR models, to machine-learning and super-ensemble frameworks at the provincial scale, but have often lacked both probabilistic uncertainty quantification and acceptable spatial resolution. For example, Colón-González et al. [18] built a super ensemble of seasonal models to predict province-level dengue in Vietnam, achieving modest CRPSS gains over a reference model but without modelling district-level data. Shi and colleagues [32] used three-month-ahead time-series models in Singapore to support early warning, demonstrating the utility of real-time forecasts while also reporting wider predictive intervals when scaled to finer administrative units. In contrast, our district-level ensemble uniquely combines Bayesian spatio-temporal, semi-mechanistic, and supervised PCA approaches, yielding a median CRPS of 3.2, which is substantially lower than the CRPS typically seen in provincial-scale studies, which typically exceeds 4, and consistently narrower, better-calibrated 95% prediction intervals. By integrating complementary strengths (local autoregression, spatial smoothing, lagged covariates) and weighting models by CRPS, we both improve probabilistic accuracy and align forecasts with the scale at which public health actions are deployed. Additionally, we have included the plots of the Residuals vs. Fitted (Posterior Mean). All ensemble models exhibit residuals clustered around zero, indicating that the predictions are unbiased (see S5 Text).

The performance of the ensemble model varied across geographic locations, forecast horizons, and months of the year. Our analysis demonstrated the model’s accuracy in predicting spatiotemporal variations in dengue cases and the occurrence of outbreaks with forecast horizons of up to 3 months. However, as lead time increased, the model’s predictive performance deteriorated, and the associated uncertainty grew. This trend aligns with other findings from Vietnam and from similar studies in other settings and diseases, where longer forecast horizons tend to result in reduced accuracy due to increasing variability and complexity in the underlying dynamics [32]. Outbreaks are challenging to predict, as assessed using scoring rules across various types of outbreak thresholds [33].

The findings of this study have direct implications for improving dengue prevention and control efforts in the Mekong Delta. The ability to forecast dengue outbreaks at the district level enables local public health officials to take timely action to reduce the effects of dengue outbreaks, including proactive vector-control interventions such as larval control, insecticide spraying, or community mobilization. The use of probabilistic forecasting provides decision-makers with credible intervals around potential outbreak scenarios, enabling them to plan interventions based on risk variability. For instance, if an area shows a high probability of an outbreak, preventive measures can be applied immediately, while regions with lower risks can adopt a more monitored approach, optimizing the use of limited resources.

The district-level focus of the model is particularly valuable for Vietnam’s health system. The model’s ability to provide fine-scale forecasts ensures that local health departments can tailor their interventions based on the specific epidemiological and environmental context of each district, rather than relying on broader, less targeted provincial-level forecasts. In addition to supporting proactive interventions, the model aligns well with Vietnam’s efforts to adapt to climate change, which is a recognized driver of vector-borne diseases, such as dengue. The government is increasingly seeking tools that can incorporate climate predictions into public health planning. Our model, which leverages seasonal climate data, can serve as a component of an early warning system that links environmental data with health outcomes.

A previous study by Colón-González et al. [18] generated seasonal dengue forecasts at the provincial level in Vietnam; our approach, however, offers more localized forecasts, which are crucial for on-the-ground interventions. Moreover, our ensemble incorporated three types of modelling approaches, which provide more robust predictions across a range of different districts. Importantly, our contribution goes beyond simply using districts. We set out clear, simple steps and show that district-level forecasts preserve key local differences, avoid the smoothing inherent in province level, and often provide earlier warnings in the right locations. the district-level forecasts achieved similar or better accuracy than province-level runs while flagging high-risk districts sooner and delivering clearer, more useful warnings for district-level action.

To ensure our work aligns with real-world public health needs, we emphasize that our models are predictive and operational rather than purely explanatory. Each specification generates a full probabilistic forecast of dengue cases one to three months ahead, from which we derive the risk probability of exceeding a prespecified outbreak threshold. We plan to establish a system in which these probabilities feed into a simple, monthly dashboard for district- and province-level teams. When a district’s risk surpasses a set action level (for example, 80% probability), an automatic alert triggers pre-emptive vector-control measures, such as larval source reduction or focal spraying. By prioritizing calibration and sharpness over statistical inference alone, we ensure the forecasts reliably guide intervention timing and resource allocation. Moreover, these same probabilistic outputs form the backbone of a forthcoming cluster-randomized trial, in which districts will be randomized to receive threshold-triggered interventions versus standard reactive control, thereby embedding our early-warning system in a rigorous evaluation of its impact.

Despite its strengths, this study also has limitations. Our modelling framework integrates key factors influencing dengue incidence, including meteorological variables. However, it does not directly account for other determinants of disease incidence, such as variations in serotypes, immunity, and human mobility. Additionally, the COVID-19 pandemic may have impacted dengue reporting during 2020–2022. In our analysis, spatial and spatiotemporal random effects are incorporated in each model within the ensemble to account for unmeasured variability associated with these factors. While we were able to forecast at the district level, scaling down to even finer resolutions, such as the commune (the third-level administrative unit below the province and district), would be beneficial for planning public health responses, but remains a challenge due to the sparsity and inconsistency of data at these levels. We also found that different epidemic thresholds yielded varying results. This variability highlights the sensitivity of outbreak detection to the choice of threshold. Adjusting thresholds based on historical data, such as previous outbreak patterns, and expert knowledge of local epidemiological trends could help improve the model’s accuracy in identifying true outbreaks while minimizing false alarms. For example, thresholds could be dynamically set to reflect seasonal trends or differences between high- and low-transmission years, allowing for more context-specific detection.

Another limitation of this study is that the observed misalignments in 2019 and 2022 underscore the ensemble model’s inadequate ability to adapt to rapidly changing conditions. During the post-pandemic recovery in 2022, changes in mobility patterns, environmental conditions, and transmission dynamics led to underpredicted outbreaks. These challenges highlight the need for greater flexibility in the ensemble approach. Introducing dynamic weighting [34], where the influence of component models adjusts based on real-time data and emerging trends, could improve the model’s responsiveness.

Conclusion

This study shows that using an ensemble of models can effectively forecast dengue outbreaks at the district level in the Mekong Delta Region of Vietnam with lead times of one to three months. However, its performance during atypical years, such as 2019 and 2022, revealed limitations in adapting to systemic disruptions and evolving transmission dynamics, underscoring the need for more flexible approaches. Incorporating dynamic weighting, which adjusts model contributions in real time based on changing conditions, could enhance the model’s adaptability and robustness. The ensemble’s district-level focus is especially important for local interventions, and we plan to incorporate the ensemble into software for an early warning system that will integrate real-time meteorological and dengue surveillance data to be used by districts to head-off dengue outbreaks through proactive interventions. The ensemble model development approach we used can also be adapted for use in other countries and for predicting other diseases like malaria or Zika. While the model works well at the district level, further improvements are needed for more detailed predictions at the commune level. Incorporating additional data, such as dengue serotypes and human movement patterns, could make the system even more accurate.

Supporting information