Enhancing uncertain demand prediction in hospitals using simple and advanced machine learning

Background Early and timely prediction of patient care demand not only affects effective resource allocation but also influences clinical decision-making as well as patient experience. Accurately predicting patient care demand, however, is a ubiquitous challenge for hospitals across the world due, in part, to the demand’s temporal uncertainty, and to the difficulty in modelling future trends. Methods We used data from the Rambam Medical Center in Israel between the years of 2004 and 2007. The dataset comprised aggregate hourly patient arrivals to the hospital, spanning emergency department attendances, direct ward admissions, and X-ray department arrivals. The time-varying linear approach predicted all patient arrivals to the hospital hourly over a week and over 3 days based on factors including day of the week and previous 7-day arrival patterns. The neural network method leveraged a long short-term memory (LSTM) model, capturing non-linear relationships between past data and forecasting windows of up to 7 days. We evaluated the predictive capabilities of the two proposed approaches compared to three existing approaches-a seasonal naive approach, a reduced-rank vector autoregressive (RVAR) model and the trigonometric seasonality, Box-Cox transformation, autoregressive moving average errors, trend, and seasonal components (TBATS) model. Forecasts were issued every day of the 2007 test period and compared with a paired block bootstrap. Results Neither proposed model convincingly outperformed the other. At the 7-day horizon, the time-varying linear model achieved the lowest errors of all models (MAE 4.24 [95% CI 4.11,4.38]), with a borderline advantage over the 7-day LSTM (difference 0.08 [ $$-$$ 0.01,0.15], p = 0.08). At the 3-day horizon, the LSTM trained for that horizon had the lowest MAE (4.20 [4.07,4.36]), but its difference from the linear model was not significant (0.04 [ $$-$$ 0.05,0.13], p = 0.38). Accuracy changed little between one and seven days ahead for either model. Both proposed models significantly outperformed the seasonal naive and TBATS benchmarks. Conclusion Both methods demonstrated an ability to capture hourly variability, with similar accuracy at both horizons. Given its interpretability, negligible training cost, online updating, and ability to serve every horizon with one model, the time-varying linear model is an attractive default for operational planning.

Authors

Institutions

Publication Details

Journal
BMC Medical Informatics and Decision Making
Published
2026-11-18
DOI
https://doi.org/10.1186/s12911-026-03947-7
Citations
2
Primary Topic
Healthcare Operations and Scheduling Optimization
Type
article
Field-Weighted Citation Impact
2.30
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Enhancing uncertain demand prediction in hospitals using simple and advanced machine learning

Samuel Stockman, Richard M. Wood, Bangdong Zhi, Oliver Y. Chén et al.
2 citations
BMC Medical Informatics and Decision Making
Healthcare Operations and Scheduling Optimization
2.30
article

Enhancing uncertain demand prediction in hospitals using simple and advanced machine learning

Samuel Stockman, Richard M. Wood, Bangdong Zhi, Oliver Y. Chén, Xun Wu
article en
2 citations

Abstract

Background Early and timely prediction of patient care demand not only affects effective resource allocation but also influences clinical decision-making as well as patient experience. Accurately predicting patient care demand, however, is a ubiquitous challenge for hospitals across the world due, in part, to the demand’s temporal uncertainty, and to the difficulty in modelling future trends. Methods We used data from the Rambam Medical Center in Israel between the years of 2004 and 2007. The dataset comprised aggregate hourly patient arrivals to the hospital, spanning emergency department attendances, direct ward admissions, and X-ray department arrivals. The time-varying linear approach predicted all patient arrivals to the hospital hourly over a week and over 3 days based on factors including day of the week and previous 7-day arrival patterns. The neural network method leveraged a long short-term memory (LSTM) model, capturing non-linear relationships between past data and forecasting windows of up to 7 days. We evaluated the predictive capabilities of the two proposed approaches compared to three existing approaches-a seasonal naive approach, a reduced-rank vector autoregressive (RVAR) model and the trigonometric seasonality, Box-Cox transformation, autoregressive moving average errors, trend, and seasonal components (TBATS) model. Forecasts were issued every day of the 2007 test period and compared with a paired block bootstrap. Results Neither proposed model convincingly outperformed the other. At the 7-day horizon, the time-varying linear model achieved the lowest errors of all models (MAE 4.24 [95% CI 4.11,4.38]), with a borderline advantage over the 7-day LSTM (difference 0.08 [ $$-$$ 0.01,0.15], p = 0.08). At the 3-day horizon, the LSTM trained for that horizon had the lowest MAE (4.20 [4.07,4.36]), but its difference from the linear model was not significant (0.04 [ $$-$$ 0.05,0.13], p = 0.38). Accuracy changed little between one and seven days ahead for either model. Both proposed models significantly outperformed the seasonal naive and TBATS benchmarks. Conclusion Both methods demonstrated an ability to capture hourly variability, with similar accuracy at both horizons. Given its interpretability, negligible training cost, online updating, and ability to serve every horizon with one model, the time-varying linear model is an attractive default for operational planning.

BMC Medical Informatics and Decision Making
National Health Service (GB), Centre Hospitalier Universitaire Vaudois (CH), University of Bristol (GB), NHS Bristol, North Somerset and South Gloucestershire Integrated Care Board (GB), University of Bath (GB), University of Lausanne (CH)
Good health and well-being
Openalex Percentile: Top 18%
Healthcare Operations and Scheduling Optimization
2.30
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon