📄 health informatics

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

This study demonstrates that integrating real-time physician digital search data with climate indicators significantly improves the accuracy of short-term dengue hospitalization forecasts in Brazil, particularly when overcoming the limitations of delayed official reporting systems.

Original authors: QUINTANILHA, D. D. O. Q., Motta, M., Moura, E., Xavier, D., Caseri, A., Schittine, G., Gismondi, R.

Published 2026-01-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: QUINTANILHA, D. D. O. Q., Motta, M., Moura, E., Xavier, D., Caseri, A., Schittine, G., Gismondi, R.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine trying to predict a sudden flood in a city. Usually, the city officials wait until they see the water actually rising in the streets (hospitalizations) to sound the alarm. But by the time they see the water, it's often too late to move people to safety. The problem is that the official reports on how many people are in the water take weeks or even months to arrive.

This paper is like a team of meteorologists and local guides trying to solve that problem in Brazil, where a mosquito-borne disease called Dengue causes massive "floods" of sick people.

Here is the simple breakdown of what they did and what they found:

The Problem: The "Late News" Effect

The researchers noticed that the official data on how many people were hospitalized for Dengue was always "late news." By the time the government knew how many people were sick, the outbreak might have already peaked. Relying only on this old data is like trying to drive a car while only looking in the rearview mirror.

They also knew that weather (rain, heat, humidity) plays a huge role in Dengue, kind of like how rain makes puddles where mosquitoes breed. But weather alone isn't a perfect crystal ball, especially when the official hospital numbers are delayed.

The New Idea: Listening to the Doctors' "Whispers"

The team had a clever idea. They looked at a popular app used by doctors in Brazil called Whitebook. This app is like a digital encyclopedia where doctors look up how to treat diseases.

The researchers asked: "If a lot of doctors are suddenly searching for 'Dengue' on their phones, does that mean an outbreak is starting before the hospitals are even full?"

Think of it this way:

Official Hospital Data: The loud siren that goes off after the fire has already started.
Weather Data: The dark clouds gathering before the storm.
Doctor Search Data: The sound of people grabbing their raincoats and umbrellas just as the first few drops start falling.

How They Tested It

They built a "time-traveling" computer brain (a type of AI called LSTM) to predict Dengue hospitalizations for the next 8 weeks. They tested this brain in two different worlds:

The "Perfect World" Scenario: They pretended the hospital data arrived instantly. In this world, the Weather was the best predictor. It makes sense; if you know the rain and heat patterns, you can guess where the mosquitoes will be.
The "Real World" Scenario: They made the computer wait for the hospital data to arrive late (just like in real life, with a delay of up to two months). In this world, the Weather alone wasn't good enough anymore because the "news" was too old.

The Big Discovery

When the hospital data was delayed, the Integrated Model (Weather + Doctor Searches) won every time.

Why? The doctors' searches acted as a "real-time" signal. When doctors started searching for Dengue treatments, it meant they were seeing patients with symptoms right now. This signal arrived much faster than the official hospital paperwork.
The Result: By combining the "dark clouds" (Weather) with the "people grabbing umbrellas" (Doctor Searches), the computer could predict the flood much better than by looking at the old maps (Weather) or waiting for the siren (Hospital Data).

What They Found in the Data

The Signal: In almost every region they studied, the spike in doctors searching for Dengue happened before or at the same time as the spike in hospitalizations. It was a reliable early warning.
The Mix: Different regions needed different "ingredients." Some places were driven by humidity, others by temperature, but the doctor searches were a helpful constant that filled in the gaps when the official data was missing or late.

The Bottom Line

The paper concludes that doctors' digital behavior is a valuable, real-time tool. When the official reports are stuck in traffic (delayed), listening to what doctors are looking up on their phones helps public health officials see the outbreak coming sooner.

It's not about replacing the hospital data or the weather reports; it's about adding a third, faster voice to the conversation so everyone can react before the "flood" gets too high.

Technical Summary: Dengue Hospitalizations in Brazil: Forecasting with Climatic and Physicians' Digital Search Data under Real-World Reporting Delays

Problem Statement
Timely forecasting of dengue hospitalizations is critical for public health preparedness in Brazil, yet it is frequently hindered by significant reporting delays in official hospitalization systems (SIH/SUS), where data often becomes available weeks or months after patient admission. While climatic conditions are known drivers of dengue transmission, models relying solely on historical hospitalization data or climate variables struggle to provide actionable early warnings when official data is lagged. The study addresses the need for more timely signals of emerging disease activity by exploring the integration of digital clinical behavior—specifically, physicians' search queries on a clinical decision-support platform—with traditional meteorological and epidemiological data.

Methodology
The study developed a predictive framework using Long Short-Term Memory (LSTM) neural networks to forecast weekly dengue hospitalizations across 27 Immediate Geographic Regions (IGRs) in Brazil from 2021 to 2024. The analytical pipeline involved:

Data Integration: The model combined three primary data sources:
1. Hospitalization Data: Weekly counts from the SIH/SUS database.
2. Clinical Search Data: Anonymized, real-time search counts for dengue-related terms by verified licensed physicians on the Afya Whitebook® platform.
3. Climate Data: Daily meteorological observations (precipitation, temperature, humidity) from INMET stations, aggregated into weekly series and transformed into 42 potential predictors (including lags, extremes, and interactions).
Feature Selection: To manage dimensionality, the model selected one representative variable from each of four climatic categories (humidity, temperature, precipitation, seasonality) per IGR based on the lowest Root Mean Squared Error (RMSE) during training.
Modeling Scenarios: Two distinct scenarios were evaluated to assess performance under different data availability conditions:
1. Ideal-Data Scenario: Assumes immediate availability of hospitalization data (no reporting delay).
2. Real-World Scenario: Incorporates typical reporting delays by shifting hospitalization data by up to eight weeks to simulate real-world latency.
Comparative Models: Five model structures were compared: Hospitalization-only (LSTM and SARIMAX), Clinical-search only (Hospitalization + Search), Climate-only (Hospitalization + Climate), and an Integrated model (Hospitalization + Search + Climate).
Evaluation: Performance was assessed using RMSE, Mean Squared Error (MSE), Mean Absolute Error (MAE), and the coefficient of determination ( $R^2$ ). Feature importance was analyzed using SHAP (SHapley Additive Explanations) values.

Key Results

Correlation and Temporal Alignment: Analysis revealed that physician search activity on Afya Whitebook® demonstrated the strongest and most consistent positive correlation with dengue hospitalizations across IGRs. Time-series analysis indicated that peaks in search activity frequently preceded or coincided with increases in hospital admissions, suggesting search behavior acts as an early clinical indicator.
Ideal-Data Performance: In the absence of reporting delays, models relying primarily on climate variables achieved the highest predictive accuracy (lowest RMSE). The integrated model performed well but showed only modest gains over climate-only models in this scenario.
Real-World Performance: Under realistic reporting delays, the performance dynamics shifted significantly. The Integrated model (combining hospitalization history, climate, and physician search data) consistently outperformed all other approaches, achieving the lowest prediction errors across most regions.
- Models relying solely on climate or historical hospitalization data suffered reduced accuracy due to the latency of official records.
- The inclusion of physician search data compensated for the loss of signal from delayed hospitalization counts, maintaining forecast robustness.
Feature Importance (SHAP): SHAP analysis confirmed that under ideal conditions, recent hospitalization history and climate variables were the dominant predictors. However, under real-world delays, the relative importance of hospitalization history dropped sharply, while the contribution of the clinical-search indicator increased, highlighting its utility as a real-time proxy for disease activity.

Significance and Claims
The paper claims that integrating digital clinical behavior with climate data enhances forecasting performance specifically under real-world reporting constraints. The authors posit that physician search behavior constitutes a valuable, real-time indicator of dengue activity that reflects immediate clinical engagement with cases.

The study concludes that while climate-based models are effective when data is timely, they are insufficient for early response when official reporting is delayed. By incorporating digital signals from physicians, the proposed framework strengthens early-warning systems and supports more timely public health decision-making. The authors emphasize that this approach provides a "more rapid and clinically grounded signal" to support outbreak detection, particularly in settings where reporting delays are structural. They acknowledge limitations regarding ecological inference, potential biases in the physician user base, and the opacity of LSTM models, but assert that the integration of these data domains offers a practical improvement for operational forecasting in Brazil.