Forecasting regional COVID-19 hospitalisation in England using ordinal machine learning method.

Area of Science:

• Epidemiology
• Health Informatics
• Predictive Modeling

Background:

• The COVID-19 pandemic strained healthcare systems, necessitating effective short-term forecasts for respiratory infections.
• Existing quantitative forecasts treated hospital admissions as continuous, but discrete demand levels are preferred by health managers.
• Limited tools existed for precise sub-national forecasting of discrete healthcare demand.

Purpose of the Study:

• To develop and evaluate a method for precise sub-national ordinal forecasting of COVID-19 hospitalizations.
• To assess the impact of different data types (epidemiological, weather, mobility) and discretization methods on forecast accuracy.
• To provide health managers with a tool for better managing healthcare services during infectious disease waves.

Main Methods:

• COVID-19 hospitalizations in England were forecast using regional data (March 2020-December 2022).
• Hospital admission counts were transformed into ordinal variables using n-tile and n-uniform methods.
• An XGBoost model, adapted for ordinal data, incorporated epidemiological, weather, and mobility predictors.

Main Results:

• Mobility data significantly improved predictive performance compared to epidemiological data alone.
• Including weather data alongside epidemiological and mobility data yielded similar results to models using only epidemiological and mobility data.
• Forecast accuracy was robust across different numbers of ordinal levels.

Conclusions:

• Accurate ordinal forecasts for COVID-19 hospitalizations were achieved using XGBoost with mobility data.
• N-tile ordinal levels are recommended over uniform levels due to their richer information content.
• The developed method offers a valuable tool for public health and health system management.