Medical Centre

Effective management of hospital resources was a critical component of the response to the COVID-19 pandemic. With fluctuating waves of infection and emerging virus variants, accurately predicting the demand for hospital beds has proven to be a complex but essential task. Our recent study, led by Derryn Lovett and published in BMJ Health Care Informatics, evaluates a pragmatic approach to forecasting COVID-19-positive hospital bed occupancy using simple, accessible methods.

Why Predicting Bed Occupancy Matters

During the COVID-19 pandemic, healthcare systems around the world faced unprecedented challenges, with surges in demand for acute care beds due to severe cases of the virus. The ability to predict future bed occupancy is vital for several reasons:

1. Resource allocation: Effective forecasting helps healthcare leaders plan staffing, equipment needs, and additional capacity.
2. Crisis management: Accurate predictions enable health systems to anticipate surges and manage overflow by opening additional care facilities when needed.
3. Cost efficiency: Reducing over- or under-preparation minimizes waste of resources and ensures that patients receive timely care.

However, many prediction models require complex statistical knowledge, making them difficult to deploy at local or regional levels. This study sought to evaluate a simpler, more pragmatic model suitable for use by typical health system teams.

The Study: A Pragmatic Approach

The research focused on North West London (NWL) during two major COVID-19 waves, driven by the Delta variant in summer 2021 and the Omicron variant in winter 2021-2022. The model used observational data from community testing, vaccination records, and hospital admissions, with linear regression as the primary tool for prediction.

Key Data Sources:

• COVID-19-positive test results from NWL’s Whole Systems Integrated Care (WSIC) dataset
• Vaccination status by age group
• Daily hospital bed occupancy reports

Model Design:

The team developed two models:

1. Simple linear regression model: This model used the number of COVID-19 cases among unvaccinated individuals as the main predictor.
2. Multivariable model: This model incorporated additional variables, such as age bands, recognizing that older populations are more likely to be hospitalized.

Both models accounted for a lag period between positive cases in the community and hospital admissions, allowing for predictions of bed occupancy several days in advance.

Results and Performance

The models were evaluated using mean absolute percentage error (MAPE), a measure of prediction accuracy.

Key Findings:

• Accuracy before the Omicron wave: The multivariable model performed well, with a MAPE of 10.8% during the Delta-driven wave from July to October 2021.
• Decline in accuracy during the Omicron wave: The accuracy of predictions deteriorated significantly during the Omicron wave, with MAPE rising to over 110%.
• Age band considerations: While the multivariable model generally outperformed the simple model, it also faced challenges such as multicollinearity—an issue where variables are highly correlated, leading to unstable predictions.

The rapid spread and distinct characteristics of the Omicron variant, including its lower severity and higher transmissibility compared to Delta, likely contributed to the reduced model performance. The study highlights the importance of continually monitoring prediction errors and adapting models as needed.

Practical Applications

The study demonstrated that even relatively simple models can provide useful predictions during stable periods of a pandemic. Importantly, the predictions generated by the model were shared with healthcare leaders twice a week and used in planning discussions to manage resources effectively.

Key Lessons for Future Pandemics:

1. Monitoring and adaptation: Prediction models require ongoing monitoring to detect shifts in accuracy and adapt to changing epidemiological conditions.
2. Collaborative decision-making: The integration of model outputs into strategic meetings allowed for proactive responses to surges in demand.
3. Scalability: The simplicity of this approach makes it scalable and deployable in other settings, particularly when more sophisticated models are not feasible.

Limitations and Future Directions

While the study’s pragmatic approach offers many advantages, several limitations should be addressed in future research:

• Accounting for prior infections: The current model did not include the protective effect of previous COVID-19 infections, which could improve accuracy.
• Vaccine efficacy variability: Incorporating dynamic estimates of vaccine efficacy based on variant-specific data and individual factors could enhance predictions.
• Geographic granularity: Future models could explore more localized predictions by accounting for differences in prevalence and hospital capacity across regions.

Additionally, as COVID-19 testing availability changes over time, alternative data sources, such as primary care or emergency department data, may be required to maintain reliable predictions.

Conclusion: Balancing Simplicity and Accuracy

Our study highlights the value of pragmatic, data-driven models in supporting healthcare system resilience during pandemics. While more complex models may offer greater accuracy, the simplicity and accessibility of this approach make it a valuable tool for rapid response scenarios. The findings underscore the importance of collaboration between research teams and healthcare providers to develop, implement, and refine predictive models tailored to real-world needs.

As we continue to navigate the challenges of COVID-19 and prepare for future health crises, pragmatic prediction models will remain an essential component of effective healthcare planning and delivery.

Reference: Lovett D, Woodcock T, Naude J, et al. Evaluation of a pragmatic approach to predicting COVID-19-positive hospital bed occupancy. BMJ Health Care Inform 2025;32:e101055.