Signal screening methodology

Each exact disease, source, metric, normalized geography, and dimension identity appears once. Geography aliases are retained for audit but resolve to one canonical identity. Counts, rates, percentages, admissions, and positivity are never summed across incompatible definitions.

Countries and agencies publish on different schedules. For every source and cadence, GIDS derives a common analysis watermark from the latest periods available across eligible source identities. At least 80% of the active source cohort must reach the watermark, and a cadence-specific maturity interval must elapse before that period can enter the model. Newer provisional observations remain visible as latest received but are held out of analysis. Feeds beyond the configured delay warning are marked delayed and excluded from the current Benjamini–Hochberg family.

Common count metrics use robust quasi-Poisson regression with seasonal harmonics, optional long-term trend, over-dispersion, and historical-outlier down-weighting. Low-expected-count series use an exact Poisson or moment-matched negative-binomial upper tail. CUSUM is supporting persistence evidence only. A labelled seasonal empirical fallback can preserve review evidence but always remains private. Percentages and rates enter formal anomaly modeling only when numerator and denominator are available.

Within each run, one-sided p values are adjusted with Benjamini–Hochberg inside detector-tier, metric-type, and cadence families. Alert requires q ≤ 0.05; strong requires q ≤ 0.01. Common counts require at least 20 current cases, +10 absolute, and +25%; rare counts require +100% relative growth. Rate metrics require an explicit metric-specific gate.

Collection, modeling, FDR adjustment, evidence linking, deduplication, queueing, static builds, and deployment are automated. Statistical candidates remain private until independently verified. Rare-count and fallback candidates enter analyst review only when independently matched to an official event. Public-health risk appears only with attributable evidence and rationale.

Official updates for the same disease within 45 days and overlapping geographies resolve to a stable timeline before review state is applied. Weekly and monthly reports require every UTC daily member in the closed period to exist and pass its gate; missing or failed members block publication.

Pydantic is the contract source for JSON Schema, OpenAPI, and generated TypeScript. Every analysis run and review decision is immutable. A report must pass its quality gate and commit to the history store before the main publication pointer advances. Archive, contract, or build failure leaves the previous public version in place.