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

Comprehensive Summary

Wohlfender et al. evaluated machine learning models to predict COVID-19 hospital admissions one to five weeks in advance using electronic health record data. The researchers extracted features from electronic health records, including emergency department visit volumes and admissions with fever, and compared five forecasting approaches: a baseline model, linear regression, XGBoost, and two neural network architectures. Models were evaluated using root mean square error between predicted and observed admissions. XGBoost demonstrated superior performance, providing the most accurate forecasts of future COVID-19 hospitalizations. Incorporating fever-related admission trends as a predictive feature significantly enhanced forecast accuracy. The authors conclude that XGBoost-based forecasting can be integrated into hospital operations to support resource allocation and pandemic preparedness.

Outcomes and Implications

Incorporating fever-related admission trends as a predictive feature significantly enhanced forecast accuracy. The authors conclude that XGBoost-based forecasting can be integrated into hospital operations to support resource allocation and pandemic preparedness.This study demonstrates that machine learning forecasting can provide hospitals with actionable one-to-five week admission projections using readily available EHR data. Such forecasting tools could enhance patient outcomes by enabling more effective bed management, optimizing staffing levels, and improving triage decisions during periods of high demand. Proactive capacity planning may reduce operational costs and clinician burnout by preventing crisis-level resource shortages. Beyond individual hospitals, regional aggregation of these forecasts could support public health agencies in coordinating pandemic responses and allocating resources across healthcare systems.