Short-term Forecasting of Ventilator and Oxygenation Device Demand During the COVID-19 Pandemic: A Comparison of Deep Learning and Statistical Models

Accurate forecasting of hospital-specific requirements during healthcare crises enables decision-makers to tackle key strategic challenges, such as determining the optimal allocation of beds, respiratory support machines, and other critical medical equipment. Identify departments that may need to suspend operations temporarily to efficiently reallocate their resources, and assess the feasibility of resource pooling, inter-hospital sharing, and patient transfers to optimize capacity during peak demand periods. This study evaluates the forecasting accuracy of LSTM, Bi-LSTM, ARIMA, Holt-Winters, and Exponential Smoothing models in predicting demand for ventilators and oxygenation devices over two weeks during the COVID-19 pandemic. Utilizing a local dataset with detailed records of respiratory support devices provides a more precise predictive analysis of critical hospital resource demand, thereby optimizing resource allocation and improving hospital preparedness during health crises. The proposed models were evaluated using MAPE, RMSE, MSE, MAE, and R-squared. While ARIMA outperformed LSTM in terms of accuracy and model fit, it struggled to capture sudden surges in resource demand, limiting its reliability during crisis peaks. LSTM exhibited lower accuracy, likely due to the lack of long-term dependencies and limitations of the dataset. Among all models, the Bi-LSTM showed the highest predictive accuracy, aligning closely with actual observations and achieving the lowest mean absolute error (MAE), making it the most suitable approach for forecasting hospital resources during critical periods.