Machine Learning Model Enhances COVID-19 Hospitalization Forecasting and Implementation

Overview

A novel machine learning model based on an enhanced N-BEATS architecture significantly improved COVID-19 hospitalization forecasts in the US, outperforming ensemble models by over 30% in mean absolute error. The model was successfully adapted and implemented in a large academic medical center, providing actionable predictions for hospital resource planning.

Background

Accurate forecasting of respiratory virus hospitalizations is critical for public health response and resource allocation. Traditional models include mechanistic and statistical approaches, but neural network-based machine learning models have shown promise for improved prediction. However, these models often struggle to incorporate exogenous variables important for transmission dynamics. This study developed an enhanced N-BEATS model integrating exogenous data and probabilistic forecasting to improve COVID-19 hospitalization predictions nationally and at the hospital level.

Data Highlights

Key Findings

• The enhanced N-BEATS model incorporated exogenous variables via a temporal convolutional network, enabling improved forecasting accuracy.
• The model outperformed COVID-19 Forecast Hub ensemble models by 34% to 37% in mean absolute error for US hospitalization predictions.
• Probabilistic forecasting was enabled through additional residual blocks, allowing uncertainty quantification.
• Transfer learning allowed adaptation of the national model to local hospital data for site-specific predictions.
• Implementation in a large academic medical center provided actionable forecasts to optimize resource allocation and surge preparation.
• The model effectively anticipated hospitalization peaks and resurgences, enhancing pandemic response capabilities.

Clinical Implications

This advanced machine learning model offers clinicians and hospital administrators more accurate and timely forecasts of COVID-19 hospitalizations, facilitating improved resource planning and surge management. Incorporating exogenous variables and uncertainty estimates enhances confidence in predictions, supporting proactive decision-making during respiratory disease outbreaks.

Conclusion

The enhanced N-BEATS machine learning model significantly improves COVID-19 hospitalization forecasting accuracy and has demonstrated practical utility in a real-world hospital setting, underscoring its potential to aid public health and clinical decision-making during pandemics.

References

1. COVID-19 Forecast Hub -- Data Source and Ensemble Models
2. N-BEATS Model Architecture -- Oreshkin et al. 2020
3. NOAA Dew Point Data -- National Oceanic and Atmospheric Administration
4. CDC National Respiratory and Enteric Virus Surveillance System