Machine learning-based short-term forecasting of COVID-19 hospital admissions using routine hospital patient data.

During the COVID-19 pandemic, the field of infectious disease modeling advanced rapidly, with forecasting tools developed to track trends in transmission dynamics and anticipate potential shortages of critical resources such as hospital capacity. In this study, we compared short-term forecasting approaches for COVID-19 hospital admissions that generate forecasts one to five weeks ahead, using retrospective electronic health records. We extracted different features (e.g., daily emergency department visits) from an individual-level patient dataset covering six hospitals located in the region of Bern, Switzerland, from February 2020 to June 2023. We then applied five methods - last-observation carried forward (baseline), linear regression, XGBoost and two types of neural networks - to time series using a leave-future-out training scheme with multiple cutting points and optimized hyperparameters. Performance was evaluated using the root mean square error between forecasts and observations. Generally, we found that XGBoost outperformed the other methods in predicting future hospital admissions. Our results also show that adding features such as the number of hospital admissions with fever and augmenting hospital data with measurements of viral concentration in wastewater improves forecast accuracy. This study offers a thorough and systematic comparison of methods applicable to routine hospital data for real-time epidemic forecasting. With the increasing availability and volume of electronic health records, improved forecasting methods will contribute to more precise and timely information during epidemic waves of COVID-19 and other respiratory viruses, thereby strengthening evidence-based public health decision-making.