Fine-grained spatial data-driven ensemble modeling for predicting Sylvatic Yellow Fever environmental suitability in Brazil

This study presents a fine-grained, machine-learning ensemble model utilizing high-resolution environmental covariates to predict Sylvatic Yellow Fever suitability across Brazil, identifying Southern Brazil as the highest-risk region and highlighting Land use and cover as the primary influencing factor while noting significant data gaps in the North.

The Gist

Imagine Brazil as a giant, living puzzle. Some pieces of this puzzle are safe, while others are dangerous traps for a invisible enemy: the Sylvatic Yellow Fever virus. This virus doesn't just jump from person to person; it hides in the jungle, moving between monkeys and mosquitoes, waiting for the right moment to spill over into human cities.

The authors of this paper are like super-sleuths trying to map out exactly where these traps are hidden. Instead of guessing, they built a high-tech "Crystal Ball" using a massive amount of data and a special kind of computer brainpower.

Here is how they did it, broken down into simple concepts:

1. The Detective Work: Gathering Clues

To predict where the virus might strike, the team needed to know what the "perfect storm" looks like. They gathered clues from every corner of Brazil, looking at:

• The Weather: How much rain fell? Was it hot or cold? (Like checking if the jungle is too wet or too dry for mosquitoes).
• The Land: Is it a dense forest, a soybean farm, or a city? (Monkeys and mosquitoes have specific neighborhoods they love).
• The History: They looked at 545 confirmed cases of the virus from 2019 to 2024. These cases were their "crime scenes."

2. The "Zoom Lens" Strategy

One of the coolest parts of this study is the resolution. Imagine looking at a map of Brazil.

• Old maps were like looking at the country from an airplane: you could see states and big cities, but you couldn't see the trees.
• This new map is like looking through a microscope. They zoomed in to 30 meters (about the size of a small house) and looked at the data month-by-month.

They didn't just look at the exact spot where a monkey got sick. They looked at a 100-meter circle, a 500-meter circle, and a 1-kilometer circle around that spot. Why? Because mosquitoes and monkeys don't stay in one spot; they wander. This "multi-scale" approach ensures they catch the virus's habits, not just its address.

3. The "Crowd of Experts" (Ensemble Modeling)

Usually, scientists might ask one computer model to make a prediction. But what if that computer is having a bad day or is biased?

Instead, the authors created a team of 532 different computer models. Think of this like a jury of experts.

• Each expert looks at the clues slightly differently (some focus more on rain, others on forests, others on temperature).
• They all vote on whether a specific 1km square of land is "Safe" or "Dangerous."
• The final answer is the average of all their votes.

This "crowd wisdom" makes the prediction much harder to fool. If one expert is wrong, the other 531 correct them.

4. The Results: Where is the Danger?

When they ran the simulation over the entire country (checking over 7 million tiny squares), a clear picture emerged:

• The Hotspot (Southern Brazil): The southern states (like Rio Grande do Sul) are the most dangerous. The model says there is a 64% chance the environment there is perfect for the virus. This makes sense because that's where most of the recent monkey cases happened.
• The Middle Ground (Southeast & Central-West): These areas are also risky (around 44-46% suitability), especially near the edges of forests and cities.
• The Mystery Zone (The North/Amazon): This is the most interesting part. The Amazon is famous for Yellow Fever, but the model was very unsure (high uncertainty) about it. Why? Because there are very few reported cases there. It's not that the virus isn't there; it's that we haven't looked closely enough. The model is essentially saying, "I don't know enough about this area to give a good answer."
• The Low Risk (Northeast): This area showed the lowest risk, mostly because it's drier and has less of the specific forest the virus needs.

5. The "Why": What Drives the Virus?

The team used a special tool (called SHAP) to ask the computer: "What made you decide this spot is dangerous?"

The answer was surprisingly simple: The Land Use.

• High Risk: Areas with humid forests, river corridors, and mixed farming (where forests touch farms). The virus loves the "edge" where the wild jungle meets human activity.
• Low Risk: Huge, open fields of soybeans or dry savannas. The virus hates these places because the monkeys and mosquitoes can't survive there.

The Big Takeaway

This paper is like giving public health officials a high-definition weather forecast, but for disease outbreaks.

• Before: They had to guess where to send vaccines or mosquito control teams.
• Now: They have a map that says, "Send your teams to this specific patch of forest near this town, because the conditions are perfect for an outbreak."

It also highlights a critical gap: We need more data from the Amazon. The model is confident in the South but blind in the North, simply because we haven't been looking closely enough. By showing us where the data is missing, this study tells us exactly where to send the next team of researchers.

Technical Summary

1. Problem Statement

Sylvatic Yellow Fever (YF) is a high-lethality mosquito-borne disease affecting humans and non-human primates (NHPs) in tropical regions, particularly Brazil. Despite the existence of a vaccine, outbreaks persist due to logistical barriers, vaccine hesitancy, and changing environmental dynamics (deforestation, urbanization) that alter vector and host ranges.

• The Challenge: Traditional modeling approaches often struggle with:
  • Data Scarcity & Heterogeneity: Previous studies relied on scattered, low-quality, or aggregated data.
  • Scale Issues: Models often fail to capture non-linear relationships across varying spatial scales or operate at resolutions too coarse for precise intervention.
  • Presence-Only Data: In Brazil, confirmed YF cases are available, but confirmed "absence" data (truly unsuitable areas) is scarce, making standard supervised classification difficult.
  • Computational Load: Processing high-resolution environmental data for a continental-scale country like Brazil requires significant computational resources.

The goal is to develop a robust, high-resolution predictive tool to identify environmental suitability for YF, enabling efficient allocation of surveillance and vaccination resources.

2. Methodology

The authors propose a fine-grained, data-driven ensemble modeling framework utilizing machine learning on high-resolution georeferenced data.

A. Data Sources & Preprocessing

• Case Data: 568 confirmed YF cases (560 NHPs, 8 humans) from September 2019 to January 2025.
  • Split: 545 cases for training, 23 cases (from 2025) for validation.
  • Bias Note: 76% of cases are from Southern Brazil, creating a potential spatial bias.
• Environmental Covariates: Five data sources downscaled to 30m spatial resolution and aggregated to monthly temporal resolution:
  1. Rainfall: CHIRPS v3.
  2. Temperature: CHIRTS (min/max).
  3. Land Use/Cover: MapBiomas (Collection 9).
  4. Altitude: NASA SRTM.
  5. Climate Normals: WorldClim 2.1 (1970–2000).
• Grid: A prediction grid of 1×1 km covering Brazil (~7.1 million points).

B. Spatio-Temporal Characterization

• Multi-Scale Buffers: To account for vector/host movement and GPS error, zonal statistics were calculated within three circular buffer radii: 100m, 500m, and 1000m.
• Feature Engineering: 18 descriptive statistics (count, mean, min, max, percentiles, std dev) were computed for each buffer.
  • Total raw features: 51 land cover classes × 18 statistics = 918 features.

C. Feature Selection

To address the "curse of dimensionality" (545 samples vs. 918 features) and multicollinearity:

1. Low-Variability Filter: Removed features with a Coefficient of Variation (CV) < 0.05.
2. Multicollinearity Filter: Applied Variance Inflation Factor (VIF) with a threshold of 5.
   • Result: Reduced features to 140 unique variables (union of selected features across the three buffer sizes).

D. Ensemble Modeling (One-Class SVM)

Since the data is "presence-only," the authors used One-Class Support Vector Machines (OC-SVM), an unsupervised anomaly detection algorithm.

• Ensemble Strategy: Instead of a single model, they generated a massive ensemble via a Cartesian product of hyperparameters:
  • Buffer sizes (3) × PCA variance explained (9) × Kernel scaling (11) × PCA reduction (4) = 1,188 potential submodels.
• Filtering: Submodels with training accuracy < 80% were discarded, resulting in a final ensemble of 532 robust submodels.
• Prediction Aggregation: For each grid point, 532 predictions were aggregated into statistics (mean, median, percentiles, confidence levels).

E. Interpretability

• SHAP (SHapley Additive exPlanations): Used to quantify the contribution of individual environmental features to the model output, overcoming the "black box" nature of ensemble SVMs.

3. Key Results

A. Spatial Suitability Predictions

The model generated a national suitability map (0 to 1 scale, where 1 is highest suitability).

• Regional Rankings (Mean Suitability Level - MSL):
  1. Southern Brazil: 0.64 (Highest). Driven by high case density in training data (PR, SC, RS).
  2. Southeast: 0.46.
  3. Central-West: 0.44.
  4. North: 0.39.
  5. Northeast: 0.28 (Lowest).
• Validation: The model successfully predicted the location of 100% of the 23 unseen 2025 cases, with suitability scores ranging from 58% to 97% (mean 74%).
• Uncertainty: The North region showed high prediction uncertainty (low confidence) due to data scarcity, despite being historically endemic.

B. Feature Importance (SHAP Analysis)

• Dominant Factor: Land Use and Cover was the most influential variable group.
• Key Drivers of High Suitability (HMSL):
  • Forest Formation: +184% increase in high-suitability areas.
  • Urban Areas: +346% increase (suggesting peri-urban/fragmented forest risk).
  • Pasture: +36%.
• Key Drivers of Low Suitability (LMSL):
  • Soybean: -99% (almost absent in high-risk areas).
  • Savanna: -92%.
  • Open Agriculture: High suitability areas lack extensive, homogeneous open agriculture (soy, savanna) and favor humid forests, riparian corridors, and agroforestry mosaics.

C. Model Sensitivity

The ensemble approach allows for adjustable sensitivity. By using different percentiles (e.g., 10th vs. 90th), decision-makers can generate maps ranging from "highly conservative" (low false positives) to "highly sensitive" (prioritizing detection of all potential risks).

4. Key Contributions

1. Unprecedented Resolution: First fine-grained (30m input, 1km output) national-scale YF suitability model for Brazil, utilizing the most recent official georeferenced case data.
2. Robust Ensemble Framework: Development of a massive One-Class SVM ensemble (532 models) that handles presence-only data effectively and quantifies prediction uncertainty.
3. Multi-Scale Analysis: Integration of 100m, 500m, and 1000m buffers to capture landscape effects at different ecological scales.
4. Interpretability: Application of SHAP values to an ensemble of SVMs to identify specific land-use drivers (e.g., the critical role of forest fragmentation and peri-urban interfaces).
5. Operational Utility: Provision of multiple map products (confidence maps, municipality aggregates, sensitivity tiers) tailored for public health decision-making.

5. Significance

This study provides a critical tool for public health planning in Brazil. By identifying specific environmental signatures (humid forests, fragmented patches near urban areas) and quantifying regional risks, the model enables:

• Targeted Vaccination: Optimizing the allocation of limited vaccine doses to high-risk municipalities.
• Surveillance Prioritization: Directing vector control and NHP monitoring efforts to areas with high environmental suitability but low current reporting.
• Risk Communication: Offering a visual and statistical basis for communicating risk to stakeholders, particularly in data-scarce regions like the Amazon where the model highlights the need for improved surveillance.

The framework demonstrates that despite data imbalances and the complexity of sylvatic cycles, machine learning ensembles can effectively model disease suitability at continental scales, offering a blueprint for similar efforts in other tropical regions.