Forecasting Acute Respiratory Infections (ARI) and Influenza Like Illness (ILI) in Europe

Background

The disposal of accurate forecasts of ILI and ARI incidence can support preparedness and decision making by anticipating respiratory disease activity and pressure on healthcare systems. This is the focus of the ECDC-led RespiCast initiative.

Methods

ARI and ILI incidences in several European countries are forecasted. Models are trained with historical data (years 2014 to 2022) and then tested for year 2023. Forecasting horizons range from 1 to 4 weeks. Accuracy metrics considered are MAE, RMSE and MASE. Univariate baselines, Seasonal ARIMA models, local Random Forests (RF), pooled multi-output RF across countries and syndromes, a global Gated Recurrent Unit (GRU) Deep Learning model with country and disease embeddings and a bivariate VARIMA for the countries with paired ARI and ILI series were implemented and compared.

Results

Best models achieved MASE<1 (indicating better performance than the benchmark) for both syndromes at all horizons. In the full-coverage evaluation, the pooled RF was the best for ARI at horizons 1 to 4 weeks ahead (MASE ranging from 0.553 to 0.865). For ILI, the seasonal ARIMA was the best for one-week-ahead forecasts and the global GRU when the forecasting horizon ranges from 2 to 4 weeks (MASE varies between 0.709 and 0.864). In the common (for both ILI and ARI) six countries subset, the global GRU performed better for horizons 2 to 4 weeks ahead, both for ARI and ILI. Bivariate VARIMA was feasible, but did not win any disease-horizon comparison.

Conclusions

Pooling across countries and syndromes can improve forecasts, but performance depends on syndrome and horizon. Pooled RF were strongest for ARI in the full evaluation, whereas the global GRU was more consistent for multi-week forecasts and paired-country analyses. The ranking of forecasting methods obtained (by disease and horizon) helps improving epidemiological situation awareness, supporting earlier public health actions and anticipating pressure on healthcare systems.

Key messages

• Accurate 1 to 4 week-ahead ARI and ILI forecasts can support preparedness, risk assessment and healthcare planning.

• Best forecasting methods varies by syndrome and horizon, with pooled RF strongest and global GRU for multi-week ARI and ILI, respectively.