MOSAIC: Calibrated Multi-Stream Outbreak Early Warning
Published:
MOSAIC fuses three asynchronous, differently-noised public surveillance streams (CDC wastewater, Nextstrain genomic lineages, and WHO/ProMED outbreak text) into a single unitless quantity: P(Rt > 1), the posterior probability that the effective reproduction number exceeds one. Then it does the thing most dashboards skip: it validates, on the real multi-year CDC NWSS record, that the probability it reports is calibrated.
Key work: Per-stream detectors, namely Bayesian Online Change-Point Detection on wastewater and text counts, Jensen-Shannon divergence on lineage frequencies, and an EpiEstim renewal posterior for Rt · Logistic / noisy-or fusion in a serverless lite tier that faithfully reproduces a Python NumPyro + NUTS hierarchical backend · Calibration validated on 1,334 day-ahead forecasts: ECE 0.086, AUROC 0.917, median early-warning lead ~68 days before a wave peak · 39 sewershed sites across 19 countries with real US data clearly separated from modeled demo sites · Claude assistant with tool use that both explains and drives the console
Stack: TypeScript, Next.js, Python, NumPyro/NUTS, Bayesian inference, Recharts