COVID-19 ICU demand forecasting: A two-stage Prophet-LSTM approach

Area of Science:

• Healthcare Management
• Epidemiological Modeling
• Data Science

Background:

• Accurate forecasting of medical resources, particularly Intensive Care Unit (ICU) demand, is vital for hospital preparedness, especially during public health crises like the COVID-19 pandemic.
• Traditional demand forecasting models struggle to account for the dynamic and complex variables introduced during pandemics, such as disease transmission rates, vaccination progress, and public health interventions.

Purpose of the Study:

• To evaluate the hypothesis that a combined Prophet-LSTM forecasting model outperforms standalone Prophet and Long Short-Term Memory (LSTM) neural network models in predicting COVID-19 ICU demand.
• To compare the proposed hybrid model against established demand forecasting benchmarks using real-world data from a Brazilian municipality.

Main Methods:

• Developed and tested a hybrid Prophet-LSTM model incorporating time series components (trend, seasonality) alongside external variables.
• External variables included daily COVID-19 cases, vaccination rates, non-pharmaceutical interventions, social isolation index, and regional hospital bed occupancy.
• Compared the model's performance against standalone Prophet, LSTM, and other established forecasting benchmarks using Mean Average Error (MAE).

Main Results:

• The combined Prophet-LSTM model demonstrated superior forecasting accuracy for COVID-19 ICU demand.
• The proposed method achieved Mean Average Errors (MAE) 13% to 45% lower than standalone models and other established forecasting techniques.
• The inclusion of pandemic-specific variables significantly enhanced the predictive power of the forecasting model.

Conclusions:

• The Prophet-LSTM hybrid model offers a more accurate and reliable approach to forecasting critical healthcare demands during pandemics.
• Integrating diverse data sources, including epidemiological and socio-behavioral factors, is essential for robust demand prediction in turbulent scenarios.
• This methodology provides valuable insights for optimizing resource allocation and preparedness in healthcare systems facing infectious disease outbreaks.