Respiratory Forecast Prototype

County-level 1–4 week forecasts of RSV, influenza and COVID-19 emergency-department visits, from a graph neural network that has never seen the disease it is forecasting. Data through ….

Disease
Weeks ahead
Map shows

County map

Each dot is a county (contiguous US shown; Alaska, Hawaii and territories are in the table). Hover or click a dot.

Hollow ring: county has no county-level data; CDC’s state estimate is used

United States

Population-weighted average of county values, % of ED visits. Last 52 weeks and the next 4.

ObservedForecast

Select a county

Click a dot on the map or a row in the table.

ObservedForecastApprox. 80% band

Counties

Sorted by forecast level at the chosen horizon. Search by name or state.

CountyLatestForecastChange

How has this model actually done?

Weeks aheadSquared error, full historySquared error, recentAbsolute error, full history

Each cell is how much better (positive) or worse (negative) the model is than simply repeating the latest week’s value. Squared error (the paper’s metric) punishes large misses, so it rewards the model for catching big rises and falls. Absolute error treats every miss alike, and by that measure repeating last week is usually closer, so the model is worse in the typical county and week. The recent column covers the low-activity off-season, where the model has tended to overshoot.