Respiratory Illness Forecasts

Incident hospitalizations due to respiratory viruses in the United States, forecast using neural network time series models

Predicting seasonal and emerging waves of respiratory viruses is crucial for effective public health responses. We developed and trained neural network models based on the Neural Basis Expansion Analysis for Time Series Forecasting (N-BEATS; Oreshkin et al. 2000) and Temporal Convolutional Network (TCN; Bai et al. 2018) architectures. Past hospitalization data, along with associated predictor variables, are encoded into a latent representation using a TCN, which is then decoded into a forecast of future hospitalizations using N-BEATS. The latter was extended with additional residual blocks to generate probabilistic predictions. Our code and data are available here.

• Influenza

For influenza forecasts, three models are provided: NHSN/FluSight, FluSurv, and FluSurv original. All models were trained on: historical hospitalization data reported by FluSurv-NET for 12 participating hospital systems across the United States; weather data at each location obtained from the National Centers for Environmental Information; and laboratory surveillance and outpatient illness data collected through the U.S. World Health Organization (WHO), National Respiratory and Enteric Virus Surveillance System (NREVSS), and U.S. Outpatient Influenza-like Illness Surveillance Network (ILINet). The NHSN/FluSight model was further trained on incident hospitalizations reported for US states and territories by the National Healthcare Safety Network (NHSN). The historical data (dark line) shown for this model reflects the NHSN data, whereas for the FluSurv models it reflects the FluSurv-NET data. The first two models were tuned for improved accuracy in forecasting annual peaks; FluSurv original was not, but may be more accurate at timepoints away from the peak.