Abstract
Early in the COVID-19 pandemic, localized evidence was critical for healthcare planning. This paper describes an open-source, age-structured SEIR (Susceptible-Exposed-Infectious-Recovered) mathematical model developed to estimate infection trajectories and hospital capacity requirements specifically for South West England (SW). By calibrating the model to regional data, the study provided the National Health Service (NHS) with estimates of the reproduction number ($R$), hospital bed occupancy, and the proportion of the population infected. As of May 11, 2020, the model estimated that 3.4% of the SW population had been infected and that lockdown measures had successfully reduced $R$ from 2.6 to 0.6.
Introduction
In the initial stages of the COVID-19 pandemic, national-level forecasting dominated the UK's public health strategy. However, these models were often heavily influenced by the epidemic's progression in London, which experienced an earlier and more intense surge than other regions. South West England presented a unique challenge: it has a distinct demographic profile—characterized by an older and more rural population—and its epidemic curve lagged behind the capital by approximately two weeks.
For local NHS authorities, "one size fits all" national projections were insufficient for managing hospital capacity and resource allocation. This research sought to provide a granular, region-specific modelling framework to support decision-making regarding acute and intensive care (IC) bed requirements and the timing of elective treatment resumption.
Methodology
Model Structure
The authors developed a deterministic, age-structured SEIR variant. To capture the nuances of disease transmission and severity, the population was divided into eight age cohorts (ranging from 0–4 to 70+). The model tracked individuals through several distinct states:
- Susceptible and Exposed: Individuals at risk and those infected but not yet infectious.
- Infectious States: Differentiated between asymptomatic and symptomatic individuals.
- Clinical States: Tracked patients requiring Acute care or Intensive Care (IC).
- Outcomes: Recovered or Deceased.
Calibration and Uncertainty
To address parameter uncertainty, the team employed Latin Hypercube Sampling (LHS) to generate 100,000 simulations across a range of plausible values. They then used Maximum Likelihood Estimation (MLE) to identify the 100 "best-fit" parameter sets that most closely aligned with observed cumulative cases and deaths in the South West.
Data Sources
The model integrated several key datasets:
- PHE COVID-19 Dashboard: Regional cumulative cases and death counts.
- Office for National Statistics (ONS): Weekly provisional death data and regional population estimates.
- Contact Matrices: Baseline contact patterns from the POLYMOD survey, adjusted for post-lockdown changes using data from the CoMix study.
Key Results
The Impact of Interventions
The model quantified the efficacy of public health measures in the South West. The initial reproduction number ($R$) was estimated at 2.6 (95% CrI 2.0–3.2). Following the introduction of social distancing on March 15, 2020, $R$ fell slightly to 2.3. However, the implementation of full lockdown and school closures on March 23 led to a dramatic reduction, bringing $R$ down to 0.6 (95% CrI 0.5–0.7).
Regional Status as of May 11, 2020
By the time lockdown measures began to ease in mid-May, the model provided the following estimates for the South West:
- Infection Prevalence: Approximately 5,793 individuals (0.10% of the population) were currently infectious.
- Cumulative Exposure: An estimated 189,048 people (3.4% of the population) had been infected and recovered.
- Healthcare Burden: Predicted hospital occupancy stood at 701 patients in acute beds and 110 in IC beds.
Discussion
The Importance of Granularity
By accounting for the South West’s older demographic, the model more accurately predicted the demand for hospital resources than national models might have. Furthermore, the two-week regional lag meant that national interventions were implemented at an earlier stage of the local outbreak, which likely played a significant role in preventing the regional healthcare system from being overwhelmed.
Estimating the Unobservable
A primary strength of this framework is its ability to estimate "unobservable" quantities. While testing was largely limited to hospital settings early in the pandemic, the model allowed researchers to estimate the total number of asymptomatic infections and the true extent of population immunity (recovered individuals).
Limitations and Future Work
The model operates as a closed system, assuming no migration in or out of the region—a reasonable but imperfect assumption during travel restrictions. Additionally, the current iteration does not account for nosocomial (in-hospital) transmission or deaths occurring in care homes and the wider community.
Future applications of this open-source tool include evaluating specific lockdown exit strategies and assessing how limited bed capacity might lead to excess deaths if demand exceeds supply. The framework remains portable and can be adapted for other geographies or more granular local authorities.
Summary updated