Machine Learning–Driven COVID-19 Hospitalization Forecasting: From Theory to Practice in a Major Northeastern Academic Medical Center - Scorecard - MDSpire
Advertisement
Machine Learning–Driven COVID-19 Hospitalization Forecasting: From Theory to Practice in a Major Northeastern Academic Medical Center
Clinical Scorecard: Utilizing Machine Learning for Predicting COVID-19 Hospitalizations: Implementation and Evaluation in a Leading Academic Medical Center in the Northeastern United States
At a Glance
Category
Detail
Condition
COVID-19 hospitalizations forecasting
Key Mechanisms
Machine learning model based on enhanced N-BEATS architecture integrating temporal convolutional network for exogenous variables and probabilistic predictions
Target Population
US population and local hospital patient populations
Care Setting
Academic medical centers and public health systems
Key Highlights
Developed an enhanced N-BEATS model (NBEATS-xd) incorporating exogenous variables via temporal convolutional network and probabilistic forecasting.
Model showed 34% improvement in mean absolute error over performance-weighted ensemble forecasts for US COVID-19 hospitalizations.
Successful real-world implementation in a large academic medical center enabled actionable forecasts for resource allocation and surge preparation.
Guideline-Based Recommendations
Diagnosis
Utilize aggregated incident hospitalization data and relevant exogenous variables (e.g., dew point temperature, test positivity rates) for forecasting.
Management
Apply machine learning forecasting models to anticipate hospitalization surges and optimize hospital resource planning.
Incorporate local data via transfer learning to improve hospital-level prediction accuracy.
Monitoring & Follow-up
Continuously update models with recent hospitalization and surveillance data to maintain forecast accuracy.
Use probabilistic forecasts to assess uncertainty and prepare for variable epidemic trajectories.
Risks
Model performance may vary with data quality and availability of exogenous variables.
Limited interpretability of neural network models may challenge clinical decision-making without proper validation.
Patient & Prescribing Data
Patients at risk of COVID-19 hospitalization within US regions and specific hospital catchment areas
Forecasting models aid hospital leadership in timely resource allocation and surge capacity planning but do not directly guide individual patient treatment.
Clinical Best Practices
Integrate multiple data sources including hospitalization counts, environmental factors, and test positivity for comprehensive forecasting.
Use machine learning models that incorporate exogenous covariates and provide uncertainty estimates to improve prediction reliability.
Implement transfer learning to adapt national models to local hospital data for enhanced decision support.
Leverage forecasts to inform public health interventions and hospital operational planning during respiratory virus outbreaks.
