A comparative study evaluating hospital demand forecasting has revealed that a simple time-varying linear model performs with accuracy comparable to an advanced long short-term memory neural network when predicting aggregate hourly patient arrivals. Accurate early prediction of patient flow is essential for effective hospital resource allocation and clinical decision-making, yet remains a pervasive operational challenge due to temporal uncertainty and the difficulty of anticipating future care trends.
The investigation analyzed historical operational records from Rambam Medical Center in Israel covering the period between 2004 and 2007. The analyzed dataset comprised aggregate hourly patient arrivals across hospital care pathways. Specifically, these recorded arrivals spanned emergency department attendances, direct ward admissions, and arrivals at the X-ray department.
To forecast these incoming patient volumes, the research team designed and applied a time-varying linear approach. This linear formulation predicted all patient arrivals to the hospital on an hourly basis over a three-day window and across a full seven-day horizon. The linear predictions incorporated specific calendar and temporal inputs, structuring future estimates on the day of the week as well as historical arrival patterns documented across the preceding seven days.
In parallel, the researchers implemented a machine learning architecture relying on a long short-term memory neural network model. This neural network method was leveraged to capture non-linear relationships between past data and future forecasting windows of up to seven days. By evaluating this deep learning approach alongside the time-varying linear formulation, the study examined the predictive capabilities of both proposed systems across identical operational timeframes.
The predictive performance of both proposed models was evaluated against three existing benchmark approaches. These comparison models included a seasonal naive approach, a reduced-rank vector autoregressive model, and the trigonometric seasonality, Box-Cox transformation, autoregressive moving average errors, trend, and seasonal components model, designated as TBATS. Model forecasts were issued every day across the 2007 test period and statistically compared using a paired block bootstrap.
Overall empirical results demonstrated that neither proposed model convincingly outperformed the other. At the seven-day horizon, the time-varying linear model achieved the lowest errors of all models, registering a mean absolute error of 4.24 with a 95 percent confidence interval between 4.11 and 4.38. This linear result demonstrated a borderline advantage over the seven-day long short-term memory network, showing an error difference of 0.08 with a confidence interval from minus 0.01 to 0.15 and a p-value of 0.08.
Evaluations at the three-day horizon revealed a similar absence of significant separation between the two architectures. At this three-day window, the long short-term memory model trained for that specific horizon attained the lowest mean absolute error of 4.20, with a confidence interval between 4.07 and 4.36. Nevertheless, its performance difference from the linear model measured 0.04 with a confidence interval from minus 0.05 to 0.13 and a p-value of 0.38, confirming no significant distinction.
Across the broader evaluation, predictive accuracy changed little between one and seven days ahead for either proposed system. Both the time-varying linear approach and the neural network significantly outperformed the seasonal naive and TBATS benchmarks. In addition, both methods demonstrated a clear ability to capture aggregate hourly variability, sustaining similar accuracy at both the three-day and seven-day forecasting horizons.
In light of these findings, the authors concluded that the time-varying linear model represents an attractive default tool for operational hospital planning. Although both analytical methods achieved comparable forecasting accuracy, the time-varying linear approach provides high interpretability, requires negligible training cost, accommodates online updating, and successfully serves every operational forecasting horizon with a single underlying model.
Subject of Research: Hospital patient arrival demand forecasting using machine learning and time-series models
Article Title: Enhancing uncertain demand prediction in hospitals using simple and advanced machine learning
Article References: Hu, F., Stockman, S., Zhi, B., Wood, R., Wu, X., & Chén, O. Y. (2026). Enhancing uncertain demand prediction in hospitals using simple and advanced machine learning. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03947-7
Image Credits: AI Generated
DOI: 10.1186/s12911-026-03947-7
Keywords: patient demand forecasting, hospital resource allocation, time-varying linear model, long short-term memory, LSTM neural network, Rambam Medical Center, emergency department attendances, ward admissions, X-ray department arrivals, hourly arrival prediction, seasonal naive benchmark, reduced-rank vector autoregressive, TBATS model, paired block bootstrap, mean absolute error, operational planning, temporal uncertainty, clinical decision-making
News Source: Cassandra Pierce. (November 18, 2026). Simple Linear Models Match Neural Networks for Predicting Hospital Patient Demand. Scienmag.



