Abstract
Predicting bed occupancy for hospitalised patients with COVID-19 requires a nuanced understanding of length of stay (LoS) across different bed types. LoS often varies depending on a patient’s "bed pathway"—the specific sequence of transfers between general wards and critical care (CC) units. This study characterizes these pathways and evaluates their impact on the accuracy of hospital bed occupancy forecasts. Using data from University College Hospital (UCH) and the ISARIC4C COVID-19 Clinical Information Network (CO-CIN), we developed a discrete-time model to compare pathway-specific predictions against models using only average LoS. Our results identified five distinct bed pathways, with 82% of patients remaining solely in general wards. We found that incorporating bed pathways significantly improved local prediction accuracy at UCH. However, national LoS averages failed to replicate regional occupancy data across England, highlighting substantial geographical heterogeneity. These findings suggest that local data, rather than national averages, are essential for reliable hospital resource planning.
Introduction
During the COVID-19 pandemic, hospital bed capacity became a critical bottleneck for healthcare systems. To manage these limited resources, planners relied on predictive models to project bed demand based on admission rates. However, many early models operated on two simplifying assumptions: that patients occupy only one bed type during their stay, and that the length of stay (LoS) can be represented by a single fixed average.
In practice, patient journeys are more complex. A patient might start in a general ward, deteriorate and move to critical care, and eventually return to a ward for recovery. This study addresses the gap between these complex "bed pathways" and simplified modeling assumptions, investigating whether accounting for these transitions improves the accuracy of bed occupancy forecasts.
Methods and Data Sources
The researchers developed a flexible discrete-time model to simulate bed occupancy. Unlike standard models, this approach maps specific sequences of care and utilizes probabilistic sampling. For each simulated patient, the model assigns a pathway based on observed probabilities and samples a LoS for each stage from a Weibull distribution.
Three primary datasets informed the study:
- UCH Dataset: Patient-level data from 168 COVID-19 inpatients at University College Hospital (London) during the initial wave (March–April 2020).
- CO-CIN (ISARIC4C): A national dataset covering 42,980 patients across 208 acute care hospitals in the UK, representing approximately 80% of all COVID-19 hospitalizations in England during the study period.
- UK Government Dashboard: Publicly available time-series data used to validate model predictions against real-world occupancy.
Characterizing Bed Pathways
The study identified five distinct pathways present in both local and national data:
- Ward only: The most common pathway, accounting for 82% of patients.
- Ward $\rightarrow$ Critical Care (CC) $\rightarrow$ Ward
- Ward $\rightarrow$ CC
- CC only
- CC $\rightarrow$ Ward
Mean LoS varied significantly by bed type, pathway, and dataset, ranging from 1.78 to 13.53 days. Notably, the national CO-CIN dataset showed a much higher mean ward LoS (8.36 days) compared to the local UCH dataset (3.98 days), likely reflecting differences in clinical stages and regional pressures.
Key Findings
Pathway-Specific Accuracy
At the hospital level (UCH), using pathway-specific data improved peak bed demand predictions by 17–18% compared to models using simple averages. Models relying on simple averages consistently underestimated the true occupancy, failing to account for the "tail" of longer stays associated with complex pathways.
The Failure of National Averages
When the model was applied to national data, a significant discrepancy emerged. Models parameterized with national CO-CIN averages failed to accurately predict occupancy in specific NHS regions. To match real-world data, the researchers had to "down-adjust" the LoS for almost every region. For example, while the national average ward LoS was 8.78 days, the weighted regional average required to match observed data was only 7.62 days.
Regional Heterogeneity
The study revealed significant geographical variation in how long patients remained hospitalized. Best-fit ward LoS varied by nearly two days between regions, while critical care LoS varied by up to four days. This heterogeneity may be driven by local differences in patient demographics, clinical strategies, or resource availability (e.g., discharging patients faster to accommodate new admissions).
Discussion and Limitations
While the inclusion of bed pathways improves model fidelity, several limitations remain. The model assumes that LoS in one stage (e.g., CC) is independent of the next (e.g., recovery ward), which may not reflect clinical reality. Furthermore, the model assumes constant parameters over time, though clinical "best practices" and hospital strain evolved throughout the pandemic. Finally, inconsistencies in how "critical care" is defined across hospitals may introduce noise into national datasets.
Conclusion
This research demonstrates that bed pathways are a vital component of accurate hospital capacity forecasting. The substantial variation in LoS across different pathways and regions suggests that national averages are often inappropriate for local planning. For future pandemic preparedness, models should prioritize real-time local data and potentially expand to include other resource constraints, such as staffing and specialized equipment like ventilators.
Summary updated