Real-time forecasting of COVID-19-related hospital strain in France using a non-Markovian mechanistic model

ABSTRACT Background Projects such as the European Covid-19 Forecast Hub publish forecasts on the national level for new deaths, new cases, and hospital admissions, but not direct measurements of hospital strain like critical care bed occupancy at the sub-national level, which is of particular interest to health professionals for planning purposes. Methods We present a sub-national French framework for forecasting hospital strain based on a non-Markovian compartmental model, its associated online visualisation tool and a retrospective evaluation of the real-time forecasts it provided from January to December 2021 by comparing to three standard statistical forecasting methods (auto-regression, exponential smoothing, and ARIMA). Results For anticipating risk of critical care unit overload, our model performed worse than pure statistical methods at the one- and two-week horizons, but had better point forecasts at the four-week horizon for 8 of the 13 regions considered. Our model also suffered from over-confidence with respect to its prediction intervals. Conclusions Online visualisation tools and consideration of how metrics can be affected by distortion from non-pharmaceutical government interventions are essential in the assessment of forecasting models for hospital strain. What is already known on this topic The US and European Covid-19 Forecast Hubs do not provide more direct measurements of hospital strain like critical care bed occupancy at the sub-national level, which was essential for the provisioning of healthcare resources during the COVID-19 pandemic. In France, statistical modelling approaches have been proposed to anticipate hospital stain at the sub-national level but are limited by a two-week forecast horizon. What this study adds We present a sub-national French modelling framework and online application for anticipating hospital strain at the four-week horizon that can account for abrupt changes in key epidemiological parameters. It was the only publicly available real-time non-Markovian mechanistic model for the French epidemic when implemented in January 2021 and, to our knowledge, it still was at the time it stopped in early 2022. How this study might affect research, practice or policy Further adaptations of this surveillance system can serve as an anticipation tool for hospital strain across sub-national localities to aid in the prevention of short-noticed ward closures and patient transfers.