What question did this study set out to answer?

The study aims to create a hybrid forecasting framework to predict demand in the emergency care system and assess capacity thresholds.

March 17, 2026Open Access

Dynamic demand forecasting and capacity assessment for emergency care system: a hybrid time series and machine learning approach with evidence from Beijing

Key Points

The study aims to create a hybrid forecasting framework to predict demand in the emergency care system and assess capacity thresholds.
Analyzed data from 120 ambulance dispatches and hospital admissions from 2019 to 2024.
Categorized data into normal and peak periods for state-specific analysis.
Applied time series models (ARIMA, VAR, Prophet) and machine learning models (XGBoost, LightGBM, CatBoost).
Conducted comparison analyses to identify model performance in capturing dynamics.
Machine learning models outperform traditional time series methods in capturing nonlinear relationships.
Identified capacity expansion thresholds: 2,100 emergency beds and a bed usage-to-patient ratio of 6.5%–6.0%.
Ensemble average model provided robust forecasting accuracy.

Abstract

Accurately forecasting the demand of emergency care system (ECS) and assessing the system capacity are critical for effective resource allocation and crisis preparedness. This study aims to develop a robust hybrid forecasting framework to predict patient demand and identify critical capacity thresholds in Beijing’s ECS under varying operational conditions. Using comprehensive data on 120 ambulance dispatches and hospital emergency department admissions from 2019 to 2024, the data are explicitly categorized into normal and peak periods, enabling state-specific analyses that reflect the distinct operational dynamics of each mode. The analysis reveals that normal period data exhibit strong seasonality and nonlinearity, while peak periods follow a three-phase pattern: initial resource stress, capacity expansion, and system recovery. To capture the complex temporal dynamics of emergency care system, both time series models (ARIMA, VAR, Prophet) and machine learning models (XGBoost, LightGBM, CatBoost) were applied, the comparison analysis results show that machine learning models, particularly CatBoost and LightGBM, outperform traditional approaches in capturing nonlinear relationships, while Prophet excels in long-term trend and seasonality modeling. Consequently, an ensemble average model combining these three models could yield robust and accurate forecasts. Further scenario tests based on the ensemble model identified two key system capacity expansion thresholds: emergency bed use nearing 2,100 beds and a bed usage-to-patient ratio dropping to 6.5%–6.0%. The framework offers a practical, transferable approach for forecasting demand and guiding resource allocation in urban emergency systems.

Demander à l'IA

Bookmark

View Full Paper