Gradient Biotech

Epidemic surveillance and forecasting

Pull versioned public-health surveillance signals, track epidemic curves and reproduction number across jurisdictions, forecast near-term trajectory, and screen for early-warning signals — through the area-level Epidemic Dynamics dashboard (/areas/infectious-disease/spread). This is distinct from a single study's Reproduction number analysis, which runs on one study's episode data rather than external surveillance feeds.

Research question

Is a monitored public-health signal rising or falling across geographies, what does a near-term forecast look like, and does a candidate leading indicator (such as wastewater) provide useful early warning relative to a clinical reference signal?

Who this is for

  • Public-health surveillance teams monitoring signals across multiple jurisdictions
  • Epidemiology researchers who need forecast model comparison with backtested performance, not a single point estimate
  • Teams evaluating a leading indicator (e.g. wastewater) against a clinical or reportable-disease signal

Data requirements

DataRequiredPurpose
A connected public surveillance sourceYesRaw signal to track, compare, or forecast
Geo-unit selectionYesScope for curves, comparisons, and maps
A second signal (e.g. wastewater vs. admissions)OptionalEarly-warning / leading-indicator comparison

Built-in connectors: CDC COVID-19 hospital admissions, CDC NNDSS reportable-disease counts, CDC NSSP emergency-department visits, CDC NWSS wastewater surveillance, CDC respiratory and state respiratory feeds, Census (denominators), and WHO GHO.

Workflow

Connect a source → Pull + immutable snapshot → Data quality check
  → Epidemic curve (+ Rt) → Compare across geographies or signals
  → Forecast (+ backtest evaluation) → Early warning → Map

Step 1 — Connect and snapshot a source

Pull data through a built-in connector. Each pull writes an immutable, versioned snapshot recording attribution, license terms, and a redistribution status (allowed / restricted / unknown) — snapshot downloads are blocked for sources not cleared for redistribution.

Step 2 — Check data quality

Review completeness and reporting cadence for a source, metric, and geo-unit before relying on a curve or forecast built from it.

Step 3 — Epidemic curve and Rt

The Epidemic curve tab charts the connected signal over time and can compute the reproduction number (Rt) and an Rt sensitivity sweep directly from the surveillance series, using the same Cori-style renewal-equation approach as the per-study Reproduction number pipeline, applied here to an external signal instead of study episodes.

Step 4 — Compare

Compare the same signal across geo-units, or compare two signals (for example hospital admissions against wastewater) side by side.

Step 5 — Forecast and evaluate

Generate forecasts from baseline models (no-change, seasonal average, trend) and a renewal-equation model, combine them into an ensemble, and evaluate performance with rolling-origin backtests scored by weighted interval score (WIS), mean absolute error, and interval coverage — gated by a promotion check before a model is treated as current.

Step 6 — Early warning

Screen a candidate leading indicator against a reference signal using lag-sweep alignment and trend comparison to flag early-warning signals.

Step 7 — Map

Render a choropleth map of a signal across geo-units, with small-count suppression applied for privacy.

Expected outputs

  • Versioned, immutable source snapshots with attribution and redistribution status
  • Data-quality summary (completeness, cadence) per source, metric, and geo-unit
  • Epidemic curve, Rt estimates, and Rt sensitivity from surveillance data
  • Cross-geography and cross-signal comparisons
  • Multi-model forecasts (baseline, renewal, ensemble) with backtested WIS/MAE/coverage scores and a promotion gate
  • Early-warning evaluation for a candidate leading indicator
  • Count-suppressed choropleth maps

Important caveats

These are retrospective and near-term research estimates, not real-time operational surveillance, nowcasting, or public-health decision support. Redistribution-restricted sources cannot be exported regardless of derived analysis.

Related guides