Spatiotemporal Bayesian networks for malaria prediction

Area of Science:

• Epidemiology
• Computational Biology
• Environmental Health

Background:

• Effective malaria control requires targeted interventions and resource allocation, particularly in remote areas.
• Predictive modeling is crucial for informed decision-making in malaria control efforts.
• Existing malaria models have not utilized Bayesian networks, a technique adept at handling uncertainty and complex relationships.

Purpose of the Study:

• To explore the application of Bayesian networks for modeling malaria incidence and outbreaks.
• To develop and evaluate village-level, spatiotemporal Bayesian network models for malaria prediction.
• To compare the performance of Bayesian networks against traditional modeling approaches.

Main Methods:

• Learned Bayesian networks using historical malaria case data and environmental covariates at a village level.
• Developed three network types: numeric prediction, outbreak prediction, and spatial autocorrelation.
• Generated large-scale spatiotemporal models using probability logic rules for automated network construction.

Main Results:

• Numeric prediction Bayesian networks achieved mean absolute errors of 1.4 and 1.7 cases for 1-week and 6-week predictions, respectively.
• Outbreak prediction networks demonstrated an ROC AUC above 0.9 across all prediction horizons.
• Bayesian networks outperformed traditional models in longer-term predictions for high-incidence transmission areas, with spatial links improving accuracy in some villages.

Conclusions:

• Spatiotemporal Bayesian networks offer a powerful and promising alternative for predicting malaria and other vector-borne diseases.
• The developed models provide accurate and explainable predictions, supporting targeted interventions and resource allocation.
• Automated generation of complex spatiotemporal models is feasible and enhances predictive capabilities.