Dengue hospitalizations in Brazil: Forecasting with climatic and physicians' digital search data under real-world reporting delays.

Dengue hospitalizations in Brazil: Forecasting with climatic and physicians’ digital search data under real-world reporting delays - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice PLOS Digit Health . 2026 May 29;5(5):e0001206. doi: 10.1371/journal.pdig.0001206 Dengue hospitalizations in Brazil: Forecasting with climatic and physicians’ digital search data under real-world reporting delays Dayanna Quintanilha Palmer 1, 2, * , Marcela Motta 2 , Eduardo Moura 2 , Danielly Xavier 2 , Guilherme Schittine 2 , Angélica Caseri 3 , Ronaldo Gismondi 1, 2 Editor: Charles B Delahunt 4 Author information Article notes Copyright and License information 1 Department of Clinical Medicine, Universidade Federal Fluminense (UFF), Niterói, Rio de Janeiro, Brazil 2 Research and Innovation Center, Afya, São Paulo, Brazil 3 Institute of Mathematical and Computer Sciences (ICMC), Universidade de São Paulo (USP), São Paulo, Brazil 4 University of Washington, UNITED STATES OF AMERICA ✉ * E-mail: dayquintan@gmail.com I have read the journal’s policy and the authors of this manuscript have the following competing interests: DQP, ECM, MM, DX, RG, and GS are employees or collaborators of Afya, the company that owns the clinical platform evaluated in this study. DQP also receives support from Google Research through the 2025 Google PhD Fellowship. AC declares no competing interests. These relationships did not influence the study design, data analysis, decision to publish, or preparation of the manuscript. Roles Dayanna Quintanilha Palmer : Conceptualization, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Visualization, Writing – original draft, Writing – review & editing Marcela Motta : Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review & editing Eduardo Moura : Conceptualization, Formal analysis, Funding acquisition, Methodology, Resources, Supervision, Writing – review & editing Danielly Xavier : Formal analysis, Methodology, Visualization, Writing – review & editing Guilherme Schittine : Formal analysis, Writing – original draft, Writing – review & editing Angélica Caseri : Supervision, Writing – review & editing Ronaldo Gismondi : Supervision, Validation, Writing – review & editing Charles B Delahunt : Editor Received 2026 Jan 13; Accepted 2026 May 8; Collection date 2026 May. © 2026 Palmer et al This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. PMC Copyright notice PMCID: PMC13221015  PMID: 42213709 Abstract Timely forecasting of dengue hospitalizations is essential for public health preparedness but is frequently limited by delays in official reporting systems. While climatic variables are known to influence dengue transmission and can be obtained in near-real time, hospitalization data often become available only weeks after patient admission, reducing their value for early response. Digital information generated during clinical practice, such as physicians’ search patterns, may provide a complementary and more timely signal of emerging disease activity. This study evaluates whether integrating climate data with real-time records of physicians’ searches for dengue-related information improves short-term forecasts of dengue hospitalizations in Brazil under both ideal and realistic reporting conditions. Three complementary data sources were combined to generate forecasts across multiple geographic regions: weekly hospitalization counts, climatic indicators, and anonymized physician search records from a widely used clinical decision-support platform. Model performance was compared under two scenarios: one assuming immediate availability of hospitalization data and another incorporating typical reporting delays. When hospitalization data were timely, simpler model configurations — particularly those relying on hospitalization history alone or combined with climate — achieved the highest predictive accuracy, indicating that the temporal structure of the outcome itself carried substantial forecasting value. Under realistic reporting delays, however, models incorporating physicians’ search behavior consistently outperformed all other approaches across most regions. In several regions, increases in physician search activity preceded or coincided with rises in hospital admissions, indicating early clinical engagement with dengue cases. These findings indicate that physician search behavior constitutes a valuable real-time indicator of dengue activity. Integrating digital clinical behavior with climate data enhances forecasting performance under real-world reporting constraints and may strengthen early-warning systems and public health decision-making for dengue and other climate-sensitive diseases. Author summary Dengue is a major public health challenge in Brazil, where large outbreaks place sudden pressure on health services. Although climate conditions influence dengue transmission, public health responses often rely on hospitalization data that become available only weeks or months after patients are admitted, limiting the ability to act early. In this study, we explored whether real-time digital information generated by physicians could help overcome this delay. We combined weather data with anonymized records of physicians’ searches for dengue-related information within a widely used clinical decision-support platform in Brazil. We then tested whether these digital signals could improve short-term forecasts of dengue hospitalizations across different regions of the country, especially when official hospital data were delayed. We found that when hospitalization data were available without delays, simpler models based on hospitalization history and climate patterns performed well. However, under realistic reporting delays, models that incorporated physicians’ search behavior produced more accurate forecasts across most regions. These findings show that digital clinical behavior can provide early insight into rising disease activity and support more timely public health responses. 1. Introduction Dengue virus (DENV) is an arbovirus with four distinct serotypes (DENV-1 to DENV-4), primarily transmitted by Aedes mosquitoes, especially Aedes aegypti . While primary infection may be asymptomatic or present with mild symptoms, severe cases can progress to dengue hemorrhagic fever and dengue shock syndrome, characterized by coagulopathy and increased vascular permeability, which can ultimately be fatal if not promptly recognized and managed [ 1 , 2 ]. In 2024, Brazil recorded approximately 6.42 million probable dengue cases, representing a 327% increase compared to 2023, when 1.51 million probable cases were reported [ 3 ]. Several factors may contribute to the rise in dengue cases, with climatic conditions playing a central role in mosquito-borne disease dynamics. Temperatures exceeding regional threshold levels accelerate the life cycle of dengue vectors ( Aedes aegypti , and Aedes albopictus ), increase dengue virus proliferation and mosquito bite frequency, and shorten the extrinsic incubation period [ 4 , 5 ]. Higher humidity and precipitation have been linked to increased dengue transmission, as they create favorable conditions for virus replication and spread [ 4 , 6 ]. Furthermore, socio-economic factors play an important role in shaping dengue transmission patterns, contributing to a complex system that influences the probability of future outbreaks [ 7 , 8 ]. Accurately mapping geographic distribution and predicting outbreaks is essential to guide resource allocation and enable timely public health interventions [ 6 , 9 ]. Several predictive models have been proposed, but most focus exclusively on climate variables [ 8 ]. In addition, forecasting efforts are also constrained by reporting delays in official hospitalization and case-notification systems, which reduce the timeliness and operational usefulness of surveillance data [ 10 ]. Moreover, digital behavioral data from physicians—such as clinical search activity within medical decision-support platforms—have not yet been explored in the context of dengue modeling in Brazil, despite their potential to provide early indicators of clinical demand and contribute to real time decision making. Integrating meteorological and non-climatic factors may substantially improve predictive accuracy [ 11 , 12 ]. In this context, digital health data provides a novel opportunity for syndromic surveillance [ 13 , 14 ]. Patterns of access to medical applications can reflect real-time shifts in medical concerns and population health trends. In Brazil, the Afya Whitebook application, a widely used digital health tool among physicians, captures search behaviors, including queries related to dengue. In a previous study, these digital traces showed significant correlation with hospitalizations for dengue, supporting their use as a proxy of clinical demand [ 15 ]. Building upon this rationale, the present study proposes a predictive framework to forecast weekly dengue hospitalizations in Brazil by integrating climatic variables, modeled with lagged predictors to capture delayed environmental effects, and digital health indicators derived from physicians’ search behavior within the Afya Whitebook platform. To address hospitalization reporting delays, we developed two modeling scenarios: an ideal-data model, assuming real-time data availability, and a real-world model, which incorporates the typical delay observed in hospitalization record updates. By comparing these model configurations, we aim to evaluate how incorporating realistic data latency and temporal dependencies can enhance the accuracy and operational relevance of short-term dengue forecasting. 2. Methods 2.1. Ethics statement The research complies with the ethical principles established by Resolution CNS 466/12 of the Brazilian National Health Council. All data used were anonymized and derived from public or internal databases, with no individual-level identifiers or direct participant interaction. The research protocol was reviewed and approved by the Research Ethics Committee of the Centro Universitário Presidente Tancredo de Almeida Neves (UNIPTAN), under CAAE 81123624.4.0000.9667 and Opinion number 6.944.607. The Committee granted a waiver of informed consent given the exclusive use of anonymized secondary data. 2.2. Study design We developed a forecasting framework based on Long Short-Term Memory (LSTM) neural networks to predict dengue-related hospitalizations in Brazil at the Immediate Geographic Region (IGR) level. Fig 1 provides an overview of the analytical workflow, which includes: Fig 1. Workflow of the analytical pipeline, including preprocessing of weather station data, weekly aggregation, imputation of missing values, IGR-level aggregation, derivation of climate variables, LSTM-based variable selection, and SHAP analyses for both ideal-data and real-world scenarios. Open in a new tab (i) integration of climate, physicians` clinical search, and hospitalization datasets from public service; (ii) preprocessing steps such as weekly aggregation, imputation of missing values, and consolidation at the IGR level; (iii) derivation of additional climate indicators; (iv) variable screening using LSTM-based relevance evaluation; and (v) model training and interpretability using SHAP (SHapley Additive Explanations) values. Given the temporal nature of the data and the need to capture both short and long-term dependencies in dengue transmission, we employed LSTM networks, a type of recurrent neural network specifically designed to handle sequential data with temporal correlations [ 16 ]. LSTM models have been successfully applied in infectious disease forecasting, including dengue [ 11 ], demonstrating strong performance in predicting case trends based on climatic and other exogenous variables. The modeling strategy was designed to generate 8-week-ahead forecasts, a horizon chosen to reflect both epidemiological relevance and operational usefulness for public health planning. To capture the intrinsic temporal structure of dengue transmission, we implemented a rolling sliding-window approach, in which recent historical sequences were used to predict subsequent hospitalization counts. This setup enabled the model to learn meaningful temporal patterns and produce actionable short- to medium-term forecasts. 2.3. Study site According to the Brazilian Institute of Geography and Statistics (IBGE), an IGR corresponds to a group of neighboring municipalities organized around one or more urban centers that function as local hubs for goods, services, and labor. These regions reflect short-distance functional relationships, particularly commuting flows and the daily circulation of people and economic activities. The current territorial division, comprises 510 IGRs, that represent an intermediate geographic scale between municipalities and states, ensuring spatial coherence for socioeconomic, environmental, and health analyses [ 17 ]. 2.4. Data sources 2.4.1. Hospitalizations data. The SIH/SUS database, maintained by the Department of Informatics of the Unified Health System (DATASUS), provides anonymized, publicly available information on hospital admissions throughout Brazil [ 18 ]. Data are updated monthly, with a typical release delay of one to two months after hospital care. In accordance with Portaria SAES nº 1.110/2021 of the Brazilian Ministry of Health, the final consolidation of hospitalization data within the SIH/SUS may occur up to four months after the reference month, due to the administrative cycles of data submission, verification, and correction of hospitalization authorizations [ 10 ]. 2.4.2. Clinical search data. Afya Whitebook is a widely used clinical decision-support platform among healthcare professionals in Brazil. The platform provides structured, evidence-based content such as disease overviews, diagnostic algorithms, therapeutic guidelines, drug dosing information, and clinical calculators. Optimized for mobile and desktop use, it supports real-time clinical decision-making across a variety of care settings. Although its primary audience is licensed physicians, the user base also includes residents, medical students, and other allied health professionals. For this study, only search data generated by verified licensed physicians with active professional credentials were included. In Brazil, there were 597,428 practicing physicians as of January 2024, according to the recent edition of Demografia Médica [ 19 ]. In comparison, the Afya Whitebook platform registered approximately 150,000 monthly active physician users by the end of the same year [ 20 ]. However, user density varies considerably across regions, with mean annual active physicians per IGR in the study sample ranging from approximately 200 (Pirapora) to over 68,000 (São Paulo) ( S1 Table ). This heterogeneity is addressed by training independent models per IGR and is further discussed in the Limitations. The Afya Whitebook research database contains metadata on user search behavior, collected in real time and accessible retrospectively for research purposes via a structured query interface. All data are anonymized to ensure confidentiality and compliance with ethical standards, preventing the identification of individual healthcare professionals. 2.4.3. Climate data. Daily observations of precipitation, maximum, mean, and minimum temperature, and mean relative humidity were obtained from the Brazilian National Institute of Meteorology (Instituto Nacional de Meteorologia – INMET) automatic weather stations distributed throughout Brazil [ 21 ]. 2.5. Data pre-processing Climate data underwent cleaning and harmonization: dates were converted to datetime format, variable names standardized, and station metadata (name, geographic coordinates, operational period) appended to each record. Only observations between 2021 and 2024 were retained. The selection of analytical units followed the sequential criteria summarized in Fig 2 . Of the 573 INMET weather stations operating during the study period, 54 met the completeness threshold (no gaps exceeding 30 consecutive days in any climatic variable). Fig 2. Flowchart of weather station and Immediate Geographic Region (IGR) selection. Open in a new tab From the 573 INMET weather stations initially available, sequential exclusions were applied based on climatic data completeness (>30 consecutive missing days), one-station-per-IGR selection criteria, minimum weekly hospitalization counts in the training and test sets (≥5), and model performance (R² ≥ 0). The final analytical sample comprised 27 IGRs. INMET, Instituto Nacional de Meteorologia; IGR, Immediate Geographic Region; R², coefficient of determination. For these 54 stations, daily measures were aggregated into weekly series — averages for temperature and humidity, and weekly sums for precipitation. Missing weekly values were imputed using a centered moving average computed with two weeks forward and two weeks backward (minimum of one available neighbor), ensuring temporal continuity of the time series. Each Immediate Geographic Region (IGR) was then represented by a single station; when more than one eligible station fell within the same IGR, the one with the highest percentage of complete observations was retained, yielding 46 IGRs. These 46 IGRs were cross-referenced with dengue hospitalization records from the SIH/SUS system (510 IGRs with reported hospitalizations between 2021 and 2024) and with clinical search data from the Afya Whitebook platform. IGRs with fewer than 5 weekly hospitalizations in either the training or test set were excluded, reducing the sample to 28 IGRs. After preliminary modeling, one additional IGR was removed because its baseline LSTM model produced a non-positive R², indicating insufficient temporal structure for forecasting. The final analytical sample comprised 27 IGRs: Alegre, Belo Horizonte, Campina Grande, Campos dos Goytacazes, Catalão, Cruz Alta, Distrito Federal, Frederico Westphalen, Ijuí, Juiz de Fora, Linhares, Marília, Maringá, Oliveira, Passo Fundo, Passos, Pirapora, Porto Alegre, Ribeirão Preto, Rio de Janeiro, Salvador, Santa Cruz do Sul, Santa Maria, São Miguel do Oeste, São Paulo, Uberaba, and Uberlândia. These regions presented sufficient data density across all domains to enable robust multi-step forecasting. 2.6. Study variables The primary outcome of interest was the weekly number of hospitalizations for dengue fever extracted from the SIH/SUS database. Predictor variables encompassed both meteorological and digital health domains. Meteorological predictors included lagged indicators of precipitation, temperature, and humidity (1–4 weeks), alongside a structured derivation process that resulted in a final set of 42 climate-related variables, including weekly change flags, interaction terms, extreme climate indicators, seasonal dummies, temperature categories, and precipitation occurrence and patterns. In addition, clinical search data were incorporated through weekly counts of physician searches for dengue in Afya Whitebook, restricted to verified licensed physicians. Together, these variables were selected to capture the combined influence of climatic fluctuations and real-time medical information-seeking behavior on dengue-related hospitalizations. All derived variables are summarized in Table 1 , and their detailed definitions and computation procedures are provided in the Supplementary Material ( S2 Table ). Table 1. Summary of candidate predictors and variable categories considered for inclusion in the models. Category Variable Type Concise Description Representative Unit(s) Temperature Mean, Lagged, Quartiles, Change, Extremes Weekly mean temperature, lagged values (1–4 weeks), quartile indicators, and binary indicators for percentile changes and extremes (≥95th percentile). °C, Binary (0/1) Precipitation Sum, Count, Lagged, Change, Extremes Weekly total precipitation, rainy-day count, lagged values (1–4 weeks), and binary indicators for weekly percentile change and extremes (≥95th percentile). mm, Count (days), Binary (0/1) Humidity Mean, Lagged, Change, Extremes Weekly mean relative humidity, lagged values (1–4 weeks), and binary indicators for percentile changes and extremes (≥95th percentile). %, Binary (0/1) Season Seasonal Indicators Binary indicators for the four climatic seasons (Summer, Autumn, Winter, Spring). Binary (0/1) Clinical Access Normalized search rate Weekly number of clinical searches performed by physicians on the Afya Whitebook platform, normalized by the number of active physicians in the IGR in that year. Searches per 10,000 active physicians Open in a new tab Source: Climate data were obtained from INMET; clinical search data were obtained from Afya Whitebook. Full construction rules are provided in S2 Table (supplement). 2.7. Feature selection Given the presence of 42 climatic input variables derivations, we sought to reduce model dimensionality and avoid multicollinearity by selecting only one variable from each climatic category—humidity, temperature, precipitation, and seasonality. To achieve this, we employed a variable-by-variable evaluation approach, in which the model was trained using the hospitalization variable alongside a single climatic predictor at a time. For each category, the climatic variable yielding the lowest RMSE was retained. As a result, each IGR retained four climatic input variables, one from each category. This strategy was intended to reduce redundancy among correlated climatic predictors before model fitting and to improve the interpretability of subsequent SHAP analyses. Feature importance was further assessed using SHAP values, which enabled interpretation of the most influential predictors per IGR and quantification of the relative contribution of climatic versus digital health variables. 2.8. Data analysis Hospitalization, digital clinical search, and climatic datasets were merged at the IGR–week level to form a unified analytic panel. Exploratory analyses included visualization of temporal trends and descriptive statistics across regions. Forecasting was performed using a LSTM neural network trained on weekly climatic and digital health predictors. Forecasting was performed using a LSTM neural network trained on weekly climatic and digital health predictors. The complete dataset spans 2021–2024, totaling 208 epidemiological weeks. Because the model relies on lagged predictors (1–4 weeks), the first 4 weeks were discarded, leaving 204 usable weeks (SE 5/2021 to SE 52/2024). These were split chronologically in a 70/30 ratio: the training set comprised the first 143 weeks (SE 5/2021 to SE 43/2023, late October 2023), and the holdout test set comprised the remaining 61 weeks (SE 44/2023 to SE 52/2024). Input–output pairs were constructed using a sliding-window approach: a window of 24 consecutive weeks served as the model input, and the immediately following 8 weeks as the prediction target. The window advanced one week at a time along the time series, generating successive overlapping input–output pairs. This procedure was applied independently to the training and test sets. The model was trained exclusively on the training set, and the test set was reserved solely for performance evaluation. All reported results therefore correspond to forecasts on data the model never saw during training. All data preprocessing and modeling were conducted in a Python-based analytical environment [ 22 ]. Hyperparameters were optimized using a grid search evaluating different architectural configurations, including the number of hidden layers (1–4) and hidden units (20 or 40). The final model employed an input window of 24 weeks (seq_len = 24), a prediction horizon of 8 weeks (output_size = 8), a learning rate of 0.01, and a MSE loss function. Each configuration was trained in triplicate to account for stochastic variability during optimization. Model performance was assessed using the RMSE, the MSE, the Mean Absolute Error (MAE), and the coefficient of determination (R²). These metrics were computed as follows: R M S E = ( ( 1 / n ) · Σ ( ( y t − y ^ t ) 2 ) ) M S E = ( 1 / n ) · Σ ( ( y t − y ^ t ) 2 ) M A E = ( 1 / n ) · Σ | y t − y ^ t | To compute the coefficient of determination, we used the Residual Sum of Squares (RSS) and the Total Sum of Squares (TSS): R S S = Σ ( ( y t − y ^ t ) 2 ) T S S = Σ ( ( y t − y ¯ ) 2 ) R 2 = 1 − ( RSS / TSS ) These quantities quantify both the magnitude of prediction errors and the proportion of variance in weekly hospitalizations explained by the forecasting model. To evaluate the relative contribution of different data domains, five model structures were compared: Hospitalization-only model (LSTM: Hospitalization): Trained exclusively on past hospitalization counts to capture the intrinsic temporal dynamics of dengue without external predictors. Clinical-search model (LSTM: Hospitalization + Clinical Search): Trained using hospitalization time series combined with Afya Whitebook physician search data. Climate model (LSTM: Hospitalization + Climate): Trained using hospitalization data and climatic predictors, selecting the best-performing variable from each domain (temperature, humidity, precipitation, season). Integrated model (LSTM: Hospitalization + Clinical Search + Climate): Combines hospitalization data, climatic indicators, and real-time physician search behavior, restricted to the most informative variable per domain. SARIMAX model (SARIMAX: Hospitalization): A classical benchmark trained only on hospitalization counts, used as a baseline for comparative performance assessment [ 23 ]. In addition, two complementary versions were implemented: Ideal-data model: predictors and hospitalization data were temporally aligned. Real-world model: hospitalization data were shifted by up to eight weeks (approximately two months). The two-month delay window was empirically defined based on the average reporting lag observed in the hospitalization dataset, ensuring that the model reflected the actual latency pattern present in official records. Pairwise differences in RMSE between models were assessed across the triplicate runs of each scenario. Normality of the paired differences was first evaluated using the Shapiro–Wilk test; when normality was supported, a paired t-test was applied, and the Wilcoxon signed-rank test was used otherwise. Statistical significance was reported at p < 0.05, with additional thresholds at p < 0.01 and p < 0.001. Model adequacy was assessed through residual diagnostics. The Autocorrelation Function (ACF) of model residuals was evaluated up to lag 8 to verify whether residuals behaved as white noise (autocorrelations within ±1.96/√n). The Ljung-Box test (α = 0.05) was applied to formally test the absence of serial autocorrelation in the residuals. The Time-Lagged Cross-Correlation (TLCC) between predicted and observed series was computed to assess temporal alignment, with a satisfactory fit defined by a peak correlation at lag 0. 2.8.1. LSTM model architecture. LSTM networks are a special type of recurrent neural network (RNN) designed to overcome the vanishing and exploding gradient problem in long sequences [ 16 ]. An LSTM unit contains a memory cell with a constant error carousel (CEC) that preserves information across time steps, while multiplicative gates regulate information flow. Gers, Schmidhuber, and Cummins (2000) extended the model with the forget gate, allowing adaptive resetting of memory contents [ 24 ]. In our study, the LSTM architecture was adapted to model the temporal dynamics of dengue hospitalizations in Brazil. The input, forget, and output gates controlled the flow of information across time, while the memory cell preserved long-range dependencies, allowing the model to integrate both climatic fluctuations and digital health signals. Following this principle, we structured the implementation to use weekly lagged climate indicators, extreme event markers, and physician search behavior from Afya Whitebook as inputs. Through a sliding-window training strategy, the LSTM was able to capture short-term variations and longer seasonal patterns, providing forecasts at the IGR level that aligned with the theoretical design of the architecture. 3. Results 3.1. Spatiotemporal trends in hospitalization and clinical search data (2021–2024) Table 2 presents dengue hospitalization counts and rates across IGRs for the 2021–2024 period. Smaller-population regions such as Frederico Westphalen, Oliveira, São Miguel do Oeste, and Uberlândia exhibited high hospitalization rates, frequently surpassing 200–600 per 100,000 inhabitants in their peak year, while large metropolitan centers, including São Paulo and Rio de Janeiro, reported low hospitalization rates despite contributing some of the highest absolute numbers. Table 2. Hospitalizations and rates per 100000 inhabitants in 27 selected Brazilian Immediate Geographic Regions, 2021–2024. IGRs Population 2021 2022 2023 2024 N Rate N Rate N Rate N Rate Alegre 193,922 4 2.1 19 9.8 166 85.6 245 126.3 Belo Horizonte 5,045,254 46 0.9 117 2.3 996 19.7 8,119 160.9 Campina Grande 938,502 129 13.7 275 29.3 50 5.3 319 34.0 Campos dos Goytacazes 631,164 4 0.6 27 4.3 321 50.9 726 115.0 Catalão 170,345 45 26.4 215 126.2 36 21.1 279 163.8 Cruz Alta 138,860 1 0.7 4 2.9 81 58.3 112 80.7 Distrito Federal 2,817,381 411 14.6 1,292 45.9 1,005 35.7 7,408 262.9 Frederico Westphalen 119,099 32 26.9 228 191.4 18 15.1 742 623.0 Ijuí 203,825 2 1.0 27 13.2 55 27.0 373 183.0 Juiz de Fora 711,921 9 1.3 17 2.4 122 17.1 1,052 147.8 Linhares 333,129 62 18.6 27 8.1 174 52.2 288 86.5 Marília 380,834 78 20.5 182 47.8 110 28.9 245 64.3 Maringá 811,915 43 5.3 477 58.8 278 34.2 950 117.0 Oliveira 115,926 1 0.9 62 53.5 93 80.2 469 404.6 Passo Fundo 278,694 3 1.1 7 2.5 49 17.6 108 38.8 Passos 258,889 3 1.2 97 37.5 349 134.8 671 259.2 Pirapora 133,939 1 0.7 11 8.2 146 109.0 296 221.0 Porto Alegre 2,997,642 6 0.2 595 19.8 310 10.3 1,619 54.0 Ribeirão Preto 1,446,059 19 1.3 191 13.2 260 18.0 1,101 76.1 Rio de Janeiro 11,760,550 21 0.2 151 1.3 1,061 9.0 3,095 26.3 Salvador 3,485,014 11 0.3 167 4.8 749 21.5 369 10.6 Santa Cruz do Sul 364,672 80 21.9 27 7.4 12 3.3 149 40.9 Santa Maria 461,725 1 0.2 4 0.9 130 28.2 157 34.0 São Miguel do Oeste 172,124 11 6.4 533 309.7 34 19.8 406 235.9 São Paulo 20,731,920 125 0.6 141 0.7 207 1.0 7,436 35.9 Uberaba 454,392 19 4.2 106 23.3 231 50.8 206 45.3 Uberlândia 961,202 42 4.4 456 47.4 1,404 146.1 2,902 301.9 Open in a new tab Source: Hospitalization data from the Brazilian Hospital Information System (SIH/DATASUS); population estimates from IBGE. N = annual number of dengue-related hospitalizations. Rate = (N ÷ IGR population) × 100,000. Despite regional variability, most IGRs exhibited a progressive increase in hospitalizations over the study period, with the most pronounced growth occurring between 2023 and 2024. However, several IGRs deviated from this general upward trend, particularly between 2022 and 2023. São Miguel do Oeste experienced the most pronounced decline, dropping from a rate of 309.7 to 19.8 per 100,000—a reduction of over 90%—following one of the highest hospitalization rates observed across all IGRs during the study period. Similar post-peak declines were observed in Frederico Westphalen (191.4 to 15.1 per 100,000), Catalão (126.2 to 21.1 per 100,000), and Campina Grande (29.3 to 5.3 per 100,000). All four regions subsequently rebounded in 2024. Santa Cruz do Sul displayed a distinct non-linear pattern, with an early peak in 2021, a progressive decline through 2023 (21.9 to 3.3 per 100,000), and a renewed increase in 2024. Other regions, such as Passo Fundo and Santa Maria, began with very low hospitalization rates in 2021–2022 but rose to moderate levels by 2023–2024, reflecting a delayed but evident escalation. As shown in Table 3 , most IGRs displayed higher dengue-related search activity in the later years of the study period compared with 2021. Although the trajectory was not strictly linear, many regions showed a general upward tendency over time, with notable increases in 2022 and again in 2024. This temporal pattern is consistent with the national hospitalization profile, which also registered higher counts in these two years. Table 3. Dengue-related searches and access rates per 10000 physicians in 27 Brazilian IGRs, 2021–2024. IGRs 2021 2022 2023 2024 N Rate N Rate N Rate N Rate Alegre 10 324 44 1,232 56 1,436 96 2,866 Belo Horizonte 2,296 716 5,502 1,651 9,302 4,102 11,802 5,127 Campina Grande 176 742 277 1,234 102 426 289 1,903 Campos dos Goytacazes 22 218 56 549 90 1,266 119 1,288 Catalão 14 560 69 2,491 37 1,085 139 3,861 Cruz Alta 12 531 65 2,234 97 2,948 149 3,296 Distrito Federal 2,586 2,324 4,927 3,675 3,153 2,594 4,808 3,883 Frederico Westphalen 15 630 89 3,739 47 1,506 199 7,210 Ijuí 27 572 147 2,442 206 3,527 297 5,756 Juiz de Fora 91 351 147 572 309 964 711 3,468 Linhares 47 641 53 566 147 1,557 197 3,031 Maringá 172 635 664 1,733 501 1,582 753 2,454 Marília 65 617 186 1,662 188 1,366 325 1,822 Oliveira 18 418 71 1,170 140 2,541 129 2,198 Passo Fundo 30 174 158 1,808 229 2,700 283 3,406 Passos 10 220 116 1,629 171 2,627 198 3,225 Pirapora 6 392 33 1,803 228 10,411 81 3,389 Porto Alegre 930 606 5,181 2,486 3,171 2,483 5,421 3,534 Ribeirão Preto 189 308 442 799 500 1,113 1,028 1,988 Rio de Janeiro 2,653 656 5,161 1,309 8,161 2,422 11,130 2,927 Salvador 1,335 1,111 2,635 2,207 2,860 2,362 3,882 3,275 Santa Cruz do Sul 83 797 184 1,785 107 1,269 276 3,977 Santa Maria 61 550 158 1,352 282 2,730 580 5,390 São Miguel do Oeste 11 420 92 2,636 71 2,407 137 4,774 São Paulo 8,556 1,167 11,669 1,652 11,012 1,712 28,384 4,228 Uberaba 64 339 260 1,527 310 2,248 355 2,647 Uberlândia 146 474 620 1,990 965 2,855 1,075 3,161 Open in a new tab Source: Afya Whitebook Access and Engagement Data. Internal company platform. N = number of dengue-related searches. Rate = (N ÷ number of physicians active on the Afya Whitebook platform in the corresponding year) × 10000. Annual physician counts per IGR are provided in S1 Table . 3.2. Correlation analysis between hospitalizations, climate variables, and digital clinical search activity The resulting correlation matrix revealed a clear contrast in the strength and consistency of associations across variable categories, as illustrated in Fig 3 . Afya Whitebook access demonstrated the strongest and most frequent positive correlation with dengue hospitalizations across IGRs. Fig 3. Correlation matrix showing pairwise correlations between dengue hospitalizations, Afya Whitebook search activity, and climate-related predictors across the 27 Immediate Geographic Regions, computed prior to feature selection. Open in a new tab Afya Whitebook access demonstrates the strongest association with hospitalizations. In contrast, climatic predictors displayed predominantly weak and heterogeneous correlations with dengue hospitalizations across IGRs. Humidity and temperature indicators showed low to moderate associations, with no consistent pattern across regions. Precipitation-related variables exhibited minimal and highly variable correlations, without any clear or recurrent signal in most regions. Time-series curves of Afya Whitebook search rates and dengue hospitalizations further supported these findings, as illustrated in Fig 4 . In nearly all IGRs, peaks in dengue-related search activity preceded or coincided with increases in hospitalizations. Fig 4. Weekly dengue hospitalizations (orange line) and Afya Whitebook dengue-related search rate per 10,000 active physicians (blue line) across Immediate Geographic Regions from 2021 to 2024. Open in a new tab Six of the 27 IGRs were selected as representative examples for visualization. Peaks in clinical search activity typically precede or coincide with hospitalization increases, indicating early clinical engagement with dengue and supporting its use as a real-time digital indicator. 3.3. Selected variables across immediate geographic regions (IGRs) During the feature-selection procedure, we ensured that each IGR retained at least one predictor from the four predefined climatic categories—humidity, temperature, precipitation, and seasonality—reflecting their established relevance in dengue transmission dynamics. Despite this harmonized structure, the LSTM model selected markedly different climatic signatures across regions. Some IGRs were primarily influenced by short-term humidity fluctuations or temperature lags, whereas others showed stronger contributions from precipitation extremes, or accumulated rainfall. The full set of predictors selected for each IGR, grouped by climatic category, is provided in S3 Table . 3.4. Predictive performance of climate, digital, and integrated models To improve visual interpretability, we highlighted 12 selected Immediate Geographic Regions (IGRs) representing distinct forecasting contexts: major urban centers with high epidemiological relevance, regions with clearer gains from clinical-search data, regions with marked performance shifts after reporting delays were introduced, and atypical cases illustrating geographic heterogeneity. The full set of other 15 IGRs is presented in S1 Fig . Under the ideal-data scenario ( Fig 5 , left panels), simpler model configurations — particularly the Hospitalization-only and Climate models — more often achieved the lowest prediction errors, suggesting that when hospitalization records are available without reporting delays, the temporal structure of the outcome variable itself provides substantial predictive information. Fig 5. Distribution of root mean squared error (RMSE) across triplicate model runs, by Immediate Geographic Region (IGR), scenario, and feature set. Open in a new tab Boxplots show the distribution of RMSE values obtained from three independent training runs of the LSTM model for each of the 12 IGRs included in the study, under two evaluation scenarios: Ideal-data (left panel of each IGR pair) and Real-world (right panel). For each scenario, four feature-set configurations were compared: hospitalizations only (blue), hospitalizations + clinical search data (orange), hospitalizations + climatic variables (green), and hospitalizations + clinical search + climatic variables (red). Boxes represent the interquartile range (IQR), horizontal lines indicate the median, whiskers extend to the minimum and maximum values within 1.5 × IQR, and dots denote outliers. Brackets and asterisks above each panel indicate statistically significant pairwise differences between the hospitalizations-only baseline and the feature-augmented models, based on a paired t-test (when paired differences were normally distributed by the Shapiro–Wilk test) or a Wilcoxon signed-rank test (otherwise) (* p < 0.05, ** p < 0.01, *** p < 0.001). Overall, the incorporation of clinical search data and/or climatic variables reduced RMSE relative to the hospitalizations-only baseline in most IGRs, with the magnitude and direction of improvement varying across regions and between the ideal-data and real-world scenarios. After introducing real-world reporting delays ( Fig 5 , right panels), the relative ranking of models shifted. Models incorporating clinical-search data — either alone or combined with climate — more frequently emerged among the best-performing approaches, indicating that real-time physician search behavior helped compensate for the reduced temporal informativeness of delayed hospitalization records. Taken together, these results indicate that the predictive value of each data domain depended on data availability conditions. Simpler models performed well when hospitalization data were timely, whereas models incorporating physician digital behavior became more advantageous under real-world reporting delays — the scenario most representative of operational public health conditions. Detailed RMSE values (mean ± standard error) and pairwise statistical comparisons for all 27 IGRs are provided in S4 – S7 Tables . 3.5. Model diagnostics Across the 216 LSTM model fits (four configurations × 27 IGRs × two scenarios), residual diagnostics confirmed adequate model behavior. ACF analysis showed that residuals fell within the white-noise band (±1.96/√n) for all 216 series, indicating that systematic temporal patterns were captured by the models. The Ljung-Box test corroborated this finding, with all p-values > 0.05, supporting the absence of serial autocorrelation. The TLCC analysis revealed peak correlation at lag 0 — indicating no systematic temporal shift between predicted and observed series — in 92 of 108 Ideal-data models (85%) and 79 of 108 Real-world models (73%). The slightly lower proportion under real-world conditions is consistent with the eight-week reporting delay incorporated in this scenario. 3.6. SHAP When comparing SHAP patterns across scenarios ( Fig 6 ; Supplementary S2 Fig and S3 Fig ), a consistent shift emerged: the predictive influence of hospitalization counts decreased markedly in the Real-world model, reflecting the loss of temporal alignment introduced by reporting delays. As hospitalization features became less informative, clinical-search indicators gained prominence, emerging as the most influential domain under delayed conditions and effectively compensating for the degraded hospitalization signal. Climatic predictors retained moderate relevance across all scenarios, though their contribution varied substantially across regions—some IGRs were driven by humidity-related changes, others by temperature categories or precipitation lags—highlighting strong geographic heterogeneity in the role of environmental covariates. Fig 6. SHAP feature importance for the Rio de Janeiro IGR under Ideal-data and Real-world LSTM scenarios. Open in a new tab Mean absolute SHAP values for the six predictors of the integrated model. The Ideal-data model (left) assumes temporal alignment between predictors and hospitalization counts; the Real-world model (right) incorporates an eight-week reporting lag in hospitalization data. 4. Discussion This study yielded three main findings. First, dengue-related searches on the Afya Whitebook platform showed strong temporal alignment with hospitalizations, with search peaks often preceding or coinciding with increases in admissions. Second, climatic predictors varied substantially across IGRs, reflecting regional heterogeneity in environmental drivers. Third, under realistic reporting delays, models incorporating physician digital search behavior — either alone or combined with climate — outperformed approaches relying solely on hospitalization history or climate information, indicating that physician search behavior provides timely epidemiological information that compensates for the loss of predictive signal caused by delayed hospitalization data. Between 2021 and 2024, clinical searches and hospitalizations increased across regions, with strong temporal alignment between the two. This indicates that physicians rapidly adjust their information-seeking behavior in response to changes in clinical demand, consistent with evidence that digital engagement can serve as an early indicator of infectious disease activity [ 13 , 14 ]. Prior studies have shown that public digital indicators—such as Twitter activity—can forecast dengue trends by several weeks [ 25 ]. By using physician search data rather than public-facing digital behavior, our study likely provides a more precise and clinically relevant digital signal, directly connected to medical decision-making and more suitable for operational forecasting. However, not all IGRs followed this general upward trajectory. Several regions, most notably São Miguel do Oeste, Frederico Westphalen, and Catalão, showed marked declines in hospitalizations after high-incidence years, a pattern that may reflect temporary depletion of susceptible individuals following intense local transmission. In São Miguel do Oeste, for example, the hospitalization rate peaked at 309.7 per 100,000 inhabitants in 2022 and dropped sharply in 2023. This interpretation is consistent with previous studies showing that the size of successive dengue waves can be shaped by the interaction between herd immunity and seasonal transmission [ 26 ]. Changes in local climatic dynamics may also have contributed to these temporal fluctuations. These regional deviations further underscore the importance of accounting for local dynamics [ 11 ], consistent with our modeling strategy. Climatic variables contributed mainly through lagged effects, aligning with findings from Chen & Moraga (2025) that forecasting accuracy improves when models explicitly incorporate temporal delays in environmental exposures [ 11 ] In our study, contemporaneous associations between climate and hospitalizations were modest, but the LSTM consistently selected lagged temperature, humidity, and precipitation indicators as relevant inputs, and SHAP analyses confirmed their predictive influence. These results reinforce that climate is a key determinant of dengue dynamics, but its effects become detectable only when temporal structure and non-linear relationships are modeled explicitly. Model comparisons further supported these patterns. Under ideal-data conditions, simpler configurations — particularly the Hospitalization-only model — most frequently achieved the lowest prediction errors, indicating that when hospitalization records are available without delays, their temporal structure alone carries substantial forecasting value. Climate-based models also performed well in a subset of regions, consistent with the recognized environmental dependence of dengue transmission [ 4 , 5 ]. When realistic reporting delays were introduced, this pattern shifted markedly: models incorporating clinical-search data dominated, achieving the best performance in the majority of regions. Digital search data gained importance by providing real-time information that compensated for the diminished value of delayed hospitalization counts, while climatic predictors retained a complementary role. SHAP analyses clarified the interplay between predictors across scenarios. In the absence of reporting delays, climatic variables and recent hospitalization history dominated model contributions, whereas under delayed-reporting conditions, the influence of hospitalization history decreased sharply and digital search activity rose in relative importance. Together, these findings demonstrate that medium-range climatic signals and real-time behavioral indicators provide complementary information, enhancing the accuracy and operational value of forecasting models. This has practical implications for environmental and health surveillance: in settings where reporting delays are structural, incorporating digital search behavior into forecasting frameworks can improve situational awareness, support earlier public health responses, and strengthen regional early-warning systems. Similar benefits have been observed in other hybrid environmental–digital approaches used for dengue nowcasting in Brazil [ 27 ], reinforcing the potential of integrating clinical digital signals into surveillance tools. Several limitations should be considered when interpreting the generalizability of our findings. First, because the analysis was based on aggregated data at the IGR level rather than individual patient records, the models were designed to predict regional hospitalization counts and cannot be used to estimate individual-level risk of dengue. Accordingly, the ecological design restricts inference to population-level associations and precludes causal interpretation at the individual level. Second, SIH/SUS hospitalization data may be affected by underreporting and reporting delays, potentially distorting temporal dengue patterns. In addition, hospitalization records capture only the most severe cases requiring inpatient care, whereas mild or asymptomatic infections—which account for most dengue cases—are not represented. The proportion of dengue cases progressing to hospitalization may also vary across IGRs due to differences in healthcare access, distance to facilities, and regional clinical practices, which may lead to underestimation of disease burden in some settings. However, because each IGR was modeled independently, the LSTM was able to learn region-specific relationships between predictors and local hospitalization patterns. Third, uneven coverage of INMET meteorological stations led to the exclusion of regions with substantial data gaps, limiting nationwide applicability. Moreover, each IGR was represented by a single weather station selected on the basis of data completeness, which may not fully capture climatic heterogeneity within larger or geographically diverse regions. Fourth, Afya Whitebook search data, although high-volume, represent only a subset of clinicians and were unevenly distributed across regions, with mean annual active users per IGR ranging from approximately 200 in smaller regions to more than 68,000 in large urban centers. This likely reflects an urban–rural gradient in digital platform adoption. Although the analysis was restricted to verified physician accounts, these data may still overrepresent younger, urban, or more digitally engaged clinicians. In addition, concurrent outbreaks of other febrile illnesses with overlapping clinical presentations, such as chikungunya or Zika, may have increased dengue-related searches as part of the differential diagnosis, potentially inflating the clinical-search signal independently of true dengue activity. Although dengue was the dominant arboviral disease in Brazil during the study period, this remains a possible source of confounding. Fifth, the absence of entomological or virological indicators, such as vector density or circulating serotypes, may have limited the model’s sensitivity to short-term transmission shifts. In addition, although SHAP analyses improved interpretability, LSTM models remain only partially interpretable. Hyperparameter tuning was also performed using the test set, because the limited length of the time series (204 usable weeks) precluded the creation of a separate validation partition without substantially reducing the training data. This strategy may have introduced optimistic bias in absolute performance estimates, although it is less likely to have affected the relative ranking of models, since all architectures were subjected to the same selection procedure. Future studies based on longer time series should adopt dedicated validation sets or nested cross-validation strategies. Finally, several authors are affiliated with Afya, the company that owns the clinical platform evaluated in this study. Although these relationships were fully disclosed and Afya had no role in the study design, analysis, interpretation, manuscript preparation, or decision to publish, this context should be considered when interpreting the findings. To reduce these potential sources of bias, we restricted Afya Whitebook data to verified physician accounts, applied standardized cleaning procedures to the SIH/SUS time series, and limited the analysis to IGRs with adequate climatic data coverage. In addition, results were interpreted at the IGR level, where temporal aggregation helps reduce random fluctuations and reporting noise. Nevertheless, these precautions do not eliminate residual bias, and the findings should be interpreted in light of these limitations. This study demonstrates that clinician search behavior within a point-of care medical decision tool platform can serve as a timely and reliable digital indicator of dengue activity in Brazil. By integrating climatic conditions with real-time physician search data into our models, we consistently improve the forecasting of dengue hospitalizations, particularly when accounting for reporting delays. This integrated approach provides a more rapid and clinically grounded signal to support outbreak detection and public health decision-making. The next steps are to assess this framework’s generalizability across diverse settings and then evaluate its suitability for operational early-warning use, paving the way for its real-world application. Supporting information S1 Table. Number of physicians with access to the Afya Whitebook platform per IGR and year 2021–2024. (DOCX) pdig.0001206.s001.docx (40.9KB, docx) S2 Table. Climatic and clinical variables used in the forecasting models. Description of all climatic, seasonal, and physician clinical-search variables, including units, derivation, and lagged indicators. (DOCX) pdig.0001206.s002.docx (41.9KB, docx) S3 Table. Predictors selected across Immediate Geographic Regions. Climatic and digital predictors selected by the machine learning model across all Immediate Geographic Regions. (DOCX) pdig.0001206.s003.docx (40.2KB, docx) S4 Table. Model performance under ideal-data conditions. Root mean squared error values across regions for all predictive model configurations. (DOCX) pdig.0001206.s004.docx (41.4KB, docx) S5 Table. Statistical comparison of models under ideal-data conditions. Pairwise statistical tests comparing predictive performance. (DOCX) pdig.0001206.s005.docx (41.6KB, docx) S6 Table. Model performance under real-world reporting conditions. Root mean squared error values across regions accounting for reporting delays. (DOCX) pdig.0001206.s006.docx (41.9KB, docx) S7 Table. Statistical comparison of models under real-world reporting conditions. Pairwise statistical tests comparing predictive performance under reporting delays. (DOCX) pdig.0001206.s007.docx (43.5KB, docx) S8 Table. Model diagnostics under ideal-data conditions. (CSV) pdig.0001206.s008.csv (25.9KB, csv) S9 Table. Model diagnostics under Real-world conditions (eight-week reporting delay). (CSV) pdig.0001206.s009.csv (26KB, csv) S1 Fig. Distribution of root mean squared error (RMSE) across triplicate model runs — complementary set of Immediate Geographic Regions (IGRs). (TIF) pdig.0001206.s010.tif (9.1MB, tif) S2 Fig. SHAP-based feature importance across all 27 IGRs under the Ideal-data scenarios. (TIF) pdig.0001206.s011.tif (12.9MB, tif) S3 Fig. SHAP-based feature importance across all 27 IGRs under the Real-world scenarios. (TIF) pdig.0001206.s012.tif (12.9MB, tif) Acknowledgments We thank Beatriz Landiosi Teixeira for supporting the visual development of the study’s illustrative materials. Data Availability All data and code used in this study are publicly available. Hospitalization data from the Brazilian Hospital Information System (SIH/SUS), meteorological data from the National Institute of Meteorology (INMET), and weekly aggregated clinical search data from the Afya Whitebook platform, together with all code used for data processing, model development, and analysis, are deposited at https://github.com/marcelamottadecastro/dengue-hospitalization-brazil-forecasting . The clinical search dataset comprises de-identified, weekly aggregated counts of dengue-related queries per IGR, with no information on individual physicians, patients, or query content. The full analytical protocol is also available at https://www.protocols.io/view/lstm-based-forecasting-of-dengue-hospitalizations-q26g7ojk8vwz/v1 . Funding Statement This work was supported by Afya Educação Médica ( https://www.afya.com.br ) — institutional research funding to DQP, ECM, MM, DX, GS, and RG (no specific grant number assigned) — and by Google.org ( https://blog.google/company-news/outreach-and-initiatives/google-org/phd-fellowship-program-2025/ ) through the 2025 Google PhD Fellowship awarded to DQP (no grant number assigned; competitive fellowship program). AC declares no financial support relevant to this work. DQP, ECM, MM, DX, GS, and RG are salaried by Afya. The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. References 1. Khan MB, Yang Z-S, Lin C-Y, Hsu M-C, Urbina AN, Assavalapsakul W, et al. Dengue overview: an updated systemic review. J Infect Public Health. 2023;16(10):1625–42. doi: 10.1016/j.jiph.2023.08.001 [ DOI ] [ PubMed ] [ Google Scholar ] 2. Messina JP, Brady OJ, Scott TW, Zou C, Pigott DM, Duda KA, et al. Global spread of dengue virus types: mapping the 70 year history. Trends Microbiol. 2014;22(3):138–46. doi: 10.1016/j.tim.2013.12.011 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Ministério da Saúde B. TABNET/SINAN – Notificações de Dengue no Brasil (2001–2025) [Internet]. 2025 [cited 2025 Oct 19]. Available from: http://tabnet.datasus.gov.br/cgi/tabcgi.exe?sinannet/cnv/denguebbr.def 4. Colón-González FJ, Sewe MO, Tompkins AM, Sjödin H, Casallas A, Rocklöv J, et al. Projecting the risk of mosquito-borne diseases in a warmer and more populated world: a multi-model, multi-scenario intercomparison modelling study. Lancet Planet Health. 2021;5(7):e404–14. doi: 10.1016/S2542-5196(21)00132-7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Damtew YT, Tong M, Varghese BM, Anikeeva O, Hansen A, Dear K, et al. Effects of high temperatures and heatwaves on dengue fever: a systematic review and meta-analysis. EBioMedicine. 2023;91:104582. doi: 10.1016/j.ebiom.2023.104582 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Messina JP, Brady OJ, Golding N, Kraemer MUG, Wint GRW, Ray SE, et al. The current and future global distribution and population at risk of dengue. Nat Microbiol. 2019;4(9):1508–15. doi: 10.1038/s41564-019-0476-8 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Teurlai M, Menkès CE, Cavarero V, Degallier N, Descloux E, Grangeon JP. Socio-economic and climate factors associated with dengue fever spatial heterogeneity: a worked example in New Caledonia. PLoS Negl Trop Dis. 2015;9(12):e0004211. doi: 10.1371/journal.pntd.0004211 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Leung XY, Islam RM, Adhami M, Ilic D, McDonald L, Palawaththa S, et al. A systematic review of dengue outbreak prediction models: Current scenario and future directions. PLoS Negl Trop Dis. 2023;17(2):e0010631. doi: 10.1371/journal.pntd.0010631 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Bhatt S, Gething PW, Brady OJ, Messina JP, Farlow AW, Moyes CL, et al. The global distribution and burden of dengue. Nature. 2013;496(7446):504–7. doi: 10.1038/nature12060 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Ministério da Saúde. Portaria SAES n o 1.110, de 18 de novembro de 2021. 2021. 11. Chen X, Moraga P. Forecasting dengue across Brazil with LSTM neural networks and SHAP-driven lagged climate and spatial effects. BMC Public Health. 2025;25(1):973. doi: 10.1186/s12889-025-22106-7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Roster K, Connaughton C, Rodrigues FA. Machine-learning–based forecasting of dengue fever in Brazilian cities using epidemiologic and meteorological variables. Am J Epidemiol. 2022;191(10):1803–12. doi: 10.1093/aje/kwac090 [ DOI ] [ PubMed ] [ Google Scholar ] 13. Brownstein JS, Freifeld CC, Madoff LC. Digital disease detection--harnessing the Web for public health surveillance. N Engl J Med. 2009;360(21):2153–5, 2157. doi: 10.1056/NEJMp0900702 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Salathé M. Digital epidemiology: what is it, and where is it going? Life Sci Soc Policy. 2018;14(1):1. doi: 10.1186/s40504-017-0065-7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Quintanilha D, Moura E, Xavier D. Hospitalisations in Brazil: an ecological time series analysis of the impact of medical decision support data as an exogenous variable. BMC Public Health. 2025;25(1):2827. doi: 10.1186/s12889-025-24188-9 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Hochreiter S, Schmidhuber J. Long Short-Term Memory. Neural Comput. 1997;9(8):1735–80. doi: 10.1162/neco.1997.9.8.1735 [ DOI ] [ PubMed ] [ Google Scholar ] 17. Instituto Brasileiro de Geografia e Estatística. Divisão Regional do Brasil em Regiões Geográficas Imediatas e [Internet]. [cited 2025 Oct 16]. Available from: https://www.ibge.gov.br/apps/regioes_geograficas/pdfs-extra/O%20recorte%20das%20Regi%C3%B5es%20Geogr%C3%A1ficas%20Imediatas%20e%20Intermedi%C3%A1rias%20de%202017.pdf 18. Ministério da Saúde. DATASUS – Departamento de Informática do Sistema Único de Saúde [Internet]. 2025 [cited 2025 Oct 19]. Available from: https://datasus.saude.gov.br/ 19. Conselho Federal de Medicina. Observatório Demografia Médica – Painel de Demografia Médica [Internet]. 2024 [cited 2025 Dec 2]. Available from: https://observatorio.cfm.org.br/demografia/dashboard/ 20. AFYA. Dados de Acesso e Engajamento Afya Whitebook. Plataforma interna da empresa; 2025. [ Google Scholar ] 21. Instituto Nacional de Meteorologia (INMET). BDMEP — Banco de Dados Meteorológicos para Ensino e Pesquisa [Internet]. 2025 [cited 2025 Mar 3]. Available from: https://bdmep.inmet.gov.br/ 22. Python Software Foundation. Python: A Programming Language [Internet]. 2021 [cited 2025 Jul 18]. 23. Statsmodels Developers. statsmodels.tsa.statespace.sarimax.SARIMAX [Internet]. 2025 [cited 2025 Dec 2]. Available from: https://www.statsmodels.org/ 24. Gers FA, Schmidhuber J, Cummins F. Learning to forget: continual prediction with LSTM. Neural Comput. 2000;12(10):2451–71. doi: 10.1162/089976600300015015 [ DOI ] [ PubMed ] [ Google Scholar ] 25. Marques-Toledo CDA, Degener CM, Vinhal L, Coelho G, Meira W, Codeço CT, et al. Dengue prediction by the web: Tweets are a useful tool for estimating and forecasting Dengue at country and city level. PLoS Negl Trop Dis. 2017;11(7):e0005729. doi: 10.1371/journal.pntd.0005729 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Romeo-Aznar V, Picinini Freitas L, Gonçalves Cruz O, King AA, Pascual M. Fine-scale heterogeneity in population density predicts wave dynamics in dengue epidemics. Nat Commun. 2022;13(1):996. doi: 10.1038/s41467-022-28231-w [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Codeco C, Coelho F, Cruz O, Oliveira S, Castro T, Bastos L. Infodengue: A nowcasting system for the surveillance of arboviruses in Brazil. Revue d’Épidémiologie et de Santé Publique. 2018;66:S386. doi: 10.1016/j.respe.2018.05.408 [ DOI ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials S1 Table. Number of physicians with access to the Afya Whitebook platform per IGR and year 2021–2024. (DOCX) pdig.0001206.s001.docx (40.9KB, docx) S2 Table. Climatic and clinical variables used in the forecasting models. Description of all climatic, seasonal, and physician clinical-search variables, including units, derivation, and lagged indicators. (DOCX) pdig.0001206.s002.docx (41.9KB, docx) S3 Table. Predictors selected across Immediate Geographic Regions. Climatic and digital predictors selected by the machine learning model across all Immediate Geographic Regions. (DOCX) pdig.0001206.s003.docx (40.2KB, docx) S4 Table. Model performance under ideal-data conditions. Root mean squared error values across regions for all predictive model configurations. (DOCX) pdig.0001206.s004.docx (41.4KB, docx) S5 Table. Statistical comparison of models under ideal-data conditions. Pairwise statistical tests comparing predictive performance. (DOCX) pdig.0001206.s005.docx (41.6KB, docx) S6 Table. Model performance under real-world reporting conditions. Root mean squared error values across regions accounting for reporting delays. (DOCX) pdig.0001206.s006.docx (41.9KB, docx) S7 Table. Statistical comparison of models under real-world reporting conditions. Pairwise statistical tests comparing predictive performance under reporting delays. (DOCX) pdig.0001206.s007.docx (43.5KB, docx) S8 Table. Model diagnostics under ideal-data conditions. (CSV) pdig.0001206.s008.csv (25.9KB, csv) S9 Table. Model diagnostics under Real-world conditions (eight-week reporting delay). (CSV) pdig.0001206.s009.csv (26KB, csv) S1 Fig. Distribution of root mean squared error (RMSE) across triplicate model runs — complementary set of Immediate Geographic Regions (IGRs). (TIF) pdig.0001206.s010.tif (9.1MB, tif) S2 Fig. SHAP-based feature importance across all 27 IGRs under the Ideal-data scenarios. (TIF) pdig.0001206.s011.tif (12.9MB, tif) S3 Fig. SHAP-based feature importance across all 27 IGRs under the Real-world scenarios. (TIF) pdig.0001206.s012.tif (12.9MB, tif) Data Availability Statement All data and code used in this study are publicly available. Hospitalization data from the Brazilian Hospital Information System (SIH/SUS), meteorological data from the National Institute of Meteorology (INMET), and weekly aggregated clinical search data from the Afya Whitebook platform, together with all code used for data processing, model development, and analysis, are deposited at https://github.com/marcelamottadecastro/dengue-hospitalization-brazil-forecasting . The clinical search dataset comprises de-identified, weekly aggregated counts of dengue-related queries per IGR, with no information on individual physicians, patients, or query content. The full analytical protocol is also available at https://www.protocols.io/view/lstm-based-forecasting-of-dengue-hospitalizations-q26g7ojk8vwz/v1 . Articles from PLOS Digital Health are provided here courtesy of PLOS ACTIONS View on publisher site PDF (2.6 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Assessing the impact of flood duration on dengue burden across 108 countries.#223210
• Assessing the impact of flood duration on dengue burden across 108 countries#217758
• Assessment of the healthcare burden of dengue disease in England: a retrospective analysis (2010-2023).#1147335
• Understanding the epidemiological trends and economic impact of dengue in Indonesia: A systematic literature review#5213
• Complications, Deaths, and Disability Burden in the 2 Years Following Dengue Infection#1120033