The FluSight Challenge returns in Metaculus’s third annual tournament contributing forecasts to the Centers for Disease Control and Prevention (CDC) FluSight forecasting efforts. After two years of unusual influenza activity during the COVID pandemic, last year’s US flu season saw an early and intense peak, an estimated range of 300,000 to 650,000 flu hospitalizations, and 19,000 to 58,000 deaths. By forecasting the timing and intensity of this year’s flu season, the Metaculus community can aid public health officials and healthcare providers preparing to allocate resources for flu mitigation and care.
CDC: Percentage of Outpatient Visits for Influenza-like Illness (ILI)
This year, the CDC and the Council of State and Territorial Epidemiologists (CSTE) only funded the winning forecasting teams from last year’s FluSight competition. Metaculus is honored to belong to this group and to extend our partnership with Dr. Thomas McAndrew, who leads Lehigh University’s Computational Uncertainty Lab. Dr. McAndrew's team also includes Dr. Maimuna (Maia) Majumder, an Assistant Professor in the Computational Health Informatics Program (CHIP) at Boston Children’s Hospital and Harvard Medical School, as well as Dr. Shaun Truelove, an Assistant Scientist in the International Health and Epidemiology Departments in the Johns Hopkins Bloomberg School of Public Health.
Tournament Structure
The tournament features two question categories: flu intensity and flu timing. Flu intensity question groups ask for the weekly count of influenza hospitalizations in a given state on:
- December 9th
- March 2nd
- The seasonal peak, or week of the highest number of hospitalizations
The flu timing question group asks when hospitalizations will peak for a given state.
Participants forecast the intensity and timing for 10 states:
-
Massachusetts
-
New York
-
Pennsylvania
-
Florida
-
Illinois
-
Texas
-
Missouri
-
Colorado
-
California
-
Washington
Forecasts will serve as inputs to a computational model that projects flu counts for every state and territory in the US. See the Model Description for further details.
Community predictions will be hidden for the first 14 days a question is open, and 50% of forecasters' scores will be determined by participation during the hidden period. Learn more about tournament scoring here.
Two Metaculus Pro Forecasters will participate and share their predictions and rationales with the community.
Flu Hospitalizations Data
The below plot displays the number of confirmed influenza hospitalizations across the US and will be updated weekly. Find weekly data for all states here.
Model Description
How can forecasts for only 10 states be used to predict the flu countrywide?
A mechanistic model integrates the tournament forecasts with surveillance data on flu hospitalizations from all 50 states to generate a combined (chimeric) forecast for every state.
Interested in the details? This model is a ‘metapopulation model’ (SEIR dynamics) with Erlang-distributed exposed and infectious periods. A state-to-state commuter matrix from the American Community Survey is used to estimate interactions of infected individuals and susceptible individuals between states. The model generates a trajectory of incident hospitalizations by assuming a specific proportion of infected individuals are admitted to the hospital. The number of incident hospitalizations is considered latent. It is assumed that the number of observed hospitalizations is normally distributed, where the expected value is the latent number of hospitalizations and the standard deviation is a parameter to fit. To bias estimates toward human judgment forecasts, human judgment ensemble predictions are included as log-likelihood terms. The model is Bayesian and fit with a combination of NumPyro (MCMC platform) and JAX.
