Generative embedding of sparse data with a tabular foundation model for dengue anticipatory action: a machine...

Area of Science:

• Epidemiology and Public Health
• Machine Learning and Artificial Intelligence
• Environmental Health

Background:

• Traditional early outbreak detection models are data-intensive and less effective in low-resource settings.
• State-of-the-art tabular foundation models require extensive data for fine-tuning disease transmission dynamics.

Purpose of the Study:

• To develop a domain-mechanistic generative embedding from sparse case and rainfall data for early epidemic onset detection.
• To improve the applicability of surveillance models in data-limited environments.

Main Methods:

• Constructed a generative, domain-mechanistic embedding generating 132 features from limited case and rainfall data.
• Evaluated a tabular foundation model using leave-one-year-out validation across 17 Philippine regions and eight dengue-endemic countries.
• Benchmarked performance against raw data columns and catch22 features.

Main Results:

• The generative embedding significantly improved dengue outbreak detection accuracy (AUROC 0.77 across countries, 0.89 across regions) compared to raw data (0.56-0.70).
• Calibration error was lower at the regional scale (0.067) than the country scale (0.149).
• Predictive accuracy was highest in strongly seasonal regions (e.g., Philippine Type I region mean 0.87) and countries (e.g., Mexico 0.94, Brazil 0.93).

Conclusions:

• The generative embedding effectively translates climate and epidemiological variables into early-warning signals by capturing transmission mechanisms.
• This approach offers a viable pathway for extending prospective outbreak surveillance in data-limited settings.
• Mechanism-grounded embeddings can calibrate transmission-acceleration models for improved predictions at aggregated scales.