Early-stage infectious disease outbreaks impose acute pressure on hospital supply systems, particularly for essential protective materials such as medical masks. Accurate short-term demand forecasting is critical for emergency preparedness; however, during the initial phase of a novel epidemic, historical data are extremely limited, rendering many conventional statistical and deep learning models impractical. This study aimed to evaluate the feasibility of using a shallow backpropagation (BP) neural network to forecast short-term hospital material demand under conditions of extreme data scarcity, and to compare its performance with commonly used statistical and deep learning baselines. Using a univariate time series of 24 consecutive days of hospital mask consumption from a tertiary hospital in China, a BP neural network was constructed with a four-day sliding window. Model training employed the Levenberg-Marquardt algorithm with early stopping. Forecasting performance was assessed using root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and symmetric MAPE, and compared against naïve persistence, ARIMA, and LSTM models under identical one-step-ahead forecasting settings. The BP neural network demonstrated stable convergence and competitive predictive accuracy despite the limited sample size. On denormalized data, the model achieved an RMSE of approximately 519 masks and an MAE of 415 masks, outperforming ARIMA and LSTM baselines in short-horizon forecasting. Regression analysis showed strong fitting in training and validation phases, with acceptable generalization to future time points. These findings suggest that shallow BP neural networks can provide useful short-term predictive signals for hospital material demand during early outbreak stages when data are scarce. While the results represent a proof-of-feasibility rather than a deployable forecasting system, this approach may serve as a rapid, data-efficient baseline for hospital emergency resource planning, warranting further validation using larger, multi-center datasets.
Wang et al. (Thu,) studied this question.