Forecasting new diseases in low-data settings using transfer learning.

Area of Science:

• Epidemiology
• Machine Learning
• Computational Biology

Background:

• Accurate forecasting of novel infectious diseases is challenging due to limited early-stage data.
• Traditional epidemiological models and machine learning require substantial data, often unavailable during new outbreaks.
• Knowledge of related diseases can potentially aid predictions in data-scarce environments.

Purpose of the Study:

• To investigate the effectiveness of transfer learning for predicting novel infectious diseases in data-scarce settings.
• To compare knowledge transfer between related diseases using both empirical and synthetic data.
• To assess the impact of source disease selection on prediction accuracy.

Main Methods:

• Empirical analysis using dengue/Zika and influenza/COVID-19 case data from Brazil.
• Synthetic data generation using an SIR model with varying transmission and recovery rates.
• Implementation and comparison of different machine learning transfer learning methods.

Main Results:

• Transfer learning demonstrated potential to enhance infectious disease predictions, sometimes outperforming models trained solely on target disease data.
• The effectiveness of transfer learning was dependent on the careful selection of the source disease.
• Both empirical and synthetic analyses confirmed the utility of transfer learning in data-limited scenarios.

Conclusions:

• Transfer learning offers a promising approach to improve early-stage infectious disease forecasting.
• Strategic selection of related diseases for knowledge transfer is critical for maximizing predictive accuracy.
• These models provide valuable supplementary tools for pandemic response decision-making.