UKHSA's Machine-Learning Pilot for Real-Time Winter Health-Care Demand Forecasting

During the 2022–23 winter respiratory season, the UK Health Security Agency (UKHSA) piloted a new automated forecasting pipeline built on its existing daily syndromic surveillance data for England. The pilot targeted two indicators sensitive to respiratory syncytial virus (RSV) in children under five: NHS 111 telehealth calls for "cough" and emergency department attendances for "acute bronchiolitis," aiming to forecast the timing and intensity of seasonal peaks up to 28 days ahead so NHS service managers could plan capacity.
UKHSA epidemiologists tested 252 model specifications combining seven supervised-learning methods (including random forest, support vector machines, elastic-net regression and gradient boosting) with different seasonality and trend terms, publishing the full comparison in a peer-reviewed evaluation. Random-forest models with Fourier seasonal terms performed best, achieving peak-timing and intensity forecast errors of 0.0195 (NHS 111 calls) and 0.0267 (ED attendances). In one real-time test, the model correctly forecast the November 29, 2022 bronchiolitis peak date and estimated its intensity within roughly 9% of the actual figure.
The same evaluation documents where forecasting broke down: an unprecedented December 2022 surge in NHS 111 cough calls — 39.3% above the previous record — was driven partly by public anxiety over invasive group A streptococcal infections rather than RSV alone, degrading the cough indicator's specificity and confounding forecasts trained on historical patterns. The authors conclude that automated forecasts fall outside their training range in genuinely novel seasons and "should always be accompanied by expert interpretation" rather than used as a standalone decision tool — an unusually candid, peer-reviewed account of both the method's value and its limits.

Read the full analysis: https://www.medrxiv.org/content/10.1101/2024.09.20.23296441v1.full