Specialized AI Models for Predicting Dengue Disease

Saldanha, Raphael et al

Inria, LNCC

Building and Selecting Specialized AI Models for Predicting Dengue Disease

• SALDANHA, Raphael (Inria)
• AKBARINIA, Reza (Inria)
• PENA, Eduardo (UTFPR)
• RIBEIRO, Victor (LNCC)
• PORTO, Fabio (LNCC)

Context

• Raphael SALDANHA: Degrees on Geography, Statistics, Public Health and Health Information
• Postdoc call from International Relations Department of Inria. Duration of 24 months (2023-2024)
• Supervisors: Prof. Reza Akbarinia (Inria) and Prof. Fabio Porto (LNCC)

Arboviruses

• Dengue, Zika, Chikungunya, and other arboviruses imposes a significant burden over populations health
• Endemic vector borne disease (Aedes mosquitoes)
• Impacts all geographic regions in Brazil, with its continental extension
• Follow spatial and seasonal trends

Dengue spread

Guzman; Harris, 2015

Dengue transmission cycle

mosquitoalert.com

Dengue symptons

mosquitoalert.com

Aedes breeding sites

• No need of clean water

• Eggs sticks to container walls lige glue

• They can survive drying out for up to 8 months

Probable dengue cases in Brazil

per epidemiological week of symptoms onset

MS. Boletim epidemiológico vol. 54 n. 1 (2023)

Dengue on the news

Dengue outbreaks over time

Azevedo et al, 2020

Dengue and weather covariates

LSE Blogs

Dengue over time

and different regions

Brito et al. 2021

Territory diversity

• On rainy season, water accumulates on cans, pots and litter present at backyards, junkyards and on the streets.

_{Fortaleza, jan. 2016.}

Territory diversity

• On dry season or droughts, water is stored on open drums.
• The same predictor (ie. rain) may have different signal, in respect to other conditions.

Modeling dengue cases

• Dengue cases as target variable
• Predictors
  • Temperature, droughts, rainfall, floods, land use, deforestation
  • Living conditions, urban environment
  • Water supply, water rationing
  • Mosquito infestation? Only few and localized data available.

Modeling intuition

• General models tend to ignore the diversity of the territory
  • Different or even contradictory predictors relationship with the outcome
  • Dengue transmission follows different rules, affected by climate and culture
• Local models are basically restricted, not useful for different regions and dengue transmission regimes
• A single machine-learning model is not a good option due the Brazilian diversity

Approach

• Separate Brazilian municipalities into clusters
  • Maximizing their similarity regarding dengue transmissions and its covariate’s importance and signals
• Train several ML models of different strategies for each cluster
• Considering a query region \(r\), select models with the best accuracy from related clusters to \(r\), and ensemble the models into a single model, tailored for the specific region
• Predict cases using the ensemble model and compare its results with a baseline model trained with all data

Global and subsets models

_{Consider each square as a municipality}

Workflow

_{Short version}

flowchart LR
A[All data] --> G[Global Model]
A --> C(Clustering)
C --> K1[Model Cluster 1]
C --> K2[Model Cluster 1]
C --> Kn[Model Cluster n]
G --> P1[Predictions]
K1 --> P2[Predictions]
K2 --> P2
Kn --> P2
P1 --> AC[Accuracy comparison]
P2 --> AC

Key decisions

• How to cluster municipalities?

  • Time series

  • Outcome (dengue cases)

  • Predictors

• What are the best predictors for forecast?

  • Lags and rolling windows

Expected results

• More precise predictions for regions, with data “closer” to the training data.
• Parallelism in training models on different regions, less costs.
• Reduce the impact of concep-drift models of the affected regions.
• Provide health managers with tools, predictions and scenarios adequate to different scales of health surveillance, preparedness and field action
• Contribute to public health policies formulation and implementation