Utilizing a novel high-resolution malaria dataset for climate-informed predictions with a deep learning transformer...

Area of Science:

• Epidemiology
• Machine Learning
• Climate Science

Background:

• Climate significantly impacts malaria transmission by affecting the Anopheles vector and Plasmodium parasite.
• Accurate malaria incidence forecasting is crucial for early warning systems.
• Deep learning applications in epidemiological forecasting are emerging but underutilized in southern Africa.

Purpose of the Study:

• To compare the predictive accuracy of a deep learning Transformer model against statistical and XGBOOST models for malaria incidence in southern Africa.
• To evaluate the performance of a novel loss function for epidemiological data.
• To assess the potential of deep learning for daily malaria prediction using climate data.

Main Methods:

• Utilized a high-resolution, 23-year daily malaria incidence dataset (1998-2021).
• Developed and compared a deep learning Transformer model with statistical and XGBOOST models.
• Implemented a novel loss function, outperforming standard Mean Squared Error (MSE).

Main Results:

• The Transformer model achieved 98% overall accuracy, significantly outperforming statistical and XGBOOST models.
• Transformer model's AUROC was 20-40% higher than other models when predictions were converted to alert thresholds.
• A novel loss function improved performance by approximately +20% compared to standard MSE loss.

Conclusions:

• Deep learning, specifically the Transformer model, demonstrates superior performance in predicting daily malaria incidence in southern Africa.
• The model's accuracy increases with the inclusion of more climate variables, highlighting its potential for robust early warning systems.
• This prediction framework offers a valuable tool for public health interventions against malaria.