Feasibility of short-term hospital mask demand forecasting using a backpropagation neural network under data scarcity

Why mask planning matters in a crisis

When a new infectious disease hits, hospitals can run short of simple but vital supplies like medical masks. Ordering too few masks leaves doctors and nurses unprotected, while ordering too many ties up money and storage space. This article asks whether a small, data‑hungry computer model can still help hospitals make better short‑term guesses about tomorrow’s mask needs when only a few days of past data are available.

Watching hospital mask use day by day

The researchers studied mask use in a large hospital in China over just 24 days during an epidemic period. Each data point was the total number of masks used by the whole hospital on that day, without any information about individual patients. This very short record mirrors the early days of a new outbreak, when managers must act quickly but have little history to guide them. Demand over those days rose and fell sharply, reflecting changing patient loads, staff behavior, and control measures, which makes simple straight‑line forecasting unreliable.

Turning a short history into a learning problem

To turn these 24 daily counts into something a computer could learn from, the team used a sliding window approach. They grouped every four consecutive days of mask use as the “input” and asked the model to predict the mask use on the fifth day as the “output.” Moving this window along the time line produced 20 such examples. The earliest 17 days were used to train the models, and the last three days were held out to test how well the models could predict truly unseen future demand, mimicking how a hospital would use such a tool in real time.

A simple neural network versus familiar methods

The main tool tested was a shallow backpropagation neural network, a modest stack of connected nodes that can capture bends and twists in data patterns. The network had one hidden layer and used common training tricks such as scaling all values to the same range and stopping training early when performance on a small validation set stopped improving. Its job was to learn the link between mask use on the last four days and the next day. The study compared this network with three standard alternatives: a naive method that simply repeats yesterday’s value, a classical ARIMA time‑series model, and a long short‑term memory (LSTM) deep learning model.

How well the models followed real mask use

Despite the tiny dataset, the shallow neural network converged quickly and produced stable predictions. When converted back into real mask counts, its average error was about 415 masks per day, and its root‑mean‑square error was about 519 masks. In this experiment, that beat both the ARIMA and LSTM models, and clearly improved on the naive “tomorrow equals today” guess, especially during sudden drops and jumps in demand. Checks comparing predicted and actual values showed strong agreement on the training and validation data, and acceptable, though weaker, agreement on the test days, suggesting the model could track short‑term trends but still struggled with the most abrupt shifts.

Caution and next steps for real hospitals

The authors stress that their model is a proof of feasibility rather than a ready‑to‑deploy tool. With only 24 days of data from one hospital, there is a real risk that the network has partly learned quirks of this short series rather than general rules, and some signs of overfitting were present. Even so, the work shows that a simple neural network can provide more useful short‑term guidance than several common alternatives when data are scarce. For hospital planners, this suggests that lightweight models could serve as quick, stop‑gap aids for managing mask stocks during the early stages of an outbreak, to be replaced or refined once longer and richer datasets from multiple hospitals become available.

Citation: Wang, Y., Han, Y., Wang, S. et al. Feasibility of short-term hospital mask demand forecasting using a backpropagation neural network under data scarcity. Sci Rep 16, 14904 (2026). https://doi.org/10.1038/s41598-026-44754-4

Keywords: hospital supplies, mask demand, neural network, epidemic planning, time series forecasting