Short-term forecasting of mortality rates based on operational data using machine learning methods

ABSTRACT

The study examined the possibility of creating and comparing short-term predictive mortality models of a region's population in the pre-COVID (2019) and COVID period (2020) using machine learning methods (CatBoost). We used operational data on the number of deaths from the Federal State Statistics Service, as well as additional reference materials of the constituent entities of the Russian Federation (demographic and general geographic data, information about healthcare facilities, health system indicators, medical monitoring, risk indicators, etc.).
For the 2019 data, the model error decreased as the learning period increased, from 13% to 0.5%. In the 2020 data, this decrease was not observed, and the error varied between 8 and 16%. It was not possible to improve the accuracy of forecasts by adding regional characteristics. The aggregated data had the features of a random process and no predictors that had a significant impact on the causes of death or were significantly associated with them were observed