#15 · ML / AI methods

Structured models recover arboviral and SARS-CoV-2 dynamics beyond raw surveillance data

The common thread is inference under observational constraint: explicit model structure, mobility, or host variables extract dynamics that raw surveillance does not supply directly. TREA-Net adapts neural and foundation dengue forecasters to data-scarce surveillance through an SIR-informed residual correction. A hybrid SINDy-EnKF model learns chikungunya dynamics from noisy, partially observed data and reconstructs unobserved trajectories. A network renewal-equation framework defines multi-scale R(t) incorporating within-day mobility, applied to networks and mobile phone data. A simulation-validated framework estimates SARS-CoV-2 viral load trajectories from 322,218 sparse community PCR tests, linking variant, vaccination, and age to viral dynamics. The novelty lies in embedding structure, mobility, or host variables into inference; no shared mechanism across pathogens is implied.