Summary

In the infectious disease group here at Predictive Science, we run simulations to predict influenza burden on a week-to-week basis in the United States. Here the overall burden of infuenza is measured using influenza hospital admissions. Thus our model seeks to forecast new hospital admissions for future weeks.

Model

We use a Suseptible-Infectious-Recovered/Hospitalized (SIRH) model with a time-dependent basic reproduction number R₀(t). The model is available as an open-source R package PROF. More information about the model and code examples are available at https://predsci.github.io/PROF/

Optimization

Fitting the model to daily data is done via Markov Chain Monte Carlo (MCMC) process. This procedure maps the model-data likelihood distribution over parameter-space, revealing not only the best fit, but also the most likely model variations.

Forecasting

The posterior distribution of the fit is used to make state-specific probabilistic forecasts of the daily number of confirmed influenza hospital admissions. The national forecast is constructed using our 'error-correlation' procedure for combining the state level forecasts. States with low hospitalization rates are not fitted and are forecasted with a simple statistical model. We also use a statistical model when we believe that there is large uncertainty in future trends. The daily data and forecasts are aggregated to the weekly scale.