Firstly, we generated a database of mortality, population and environmental data, including temperature observations and weather forecasts. This database was used to fit epidemiological models, and the resulting associations were used to transform temperature observations and forecasts into temperature-related mortality estimates and predictions. For that purpose, weather forecasts were bias-corrected following standard techniques in the field of weather and climate forecasting. Next, we validated the weather forecasts stored in our database (i.e. daily time series of temperature) against climate observations, and the health predictions (i.e. daily time series of temperature related mortality) against the registered mortality counts. We found that temperature forecasts can be used to issue skilful predictions of heat and cold related mortality accounting for the real impacts of temperature on human health, although the window of predictability was differently reduced by season and location. However, in general, we found that skilful forecasts can be issued beyond 15 days in advance in winter, and beyond 10 days in summer. Consequently, we decided not to set bounds to the predictability window of the operational platform. We also found that the predictability of the early warnings is, to a very large extent, constrained by the original weather forecasts, and not by the epidemiological models. These findings have been included in a paper that is currently in revision in Science Advances.
Finally, we developed an operational, fit-for-purpose, early warning system representing the health impacts that environmental temperatures have on the exposed populations in 33 European countries. The health predictions that appear in the early warning system platform are automatically updated and uploaded every day by a series of protocols and scripts that we wrote, and that are synchronised in real time with the time of release of the daily weather forecasts in the ECMWF data servers.