Modelling the impact of an influenza A/H1N1 pandemic on critical care demand from early pathogenicity data: the case for sentinel reporting

Authors


Correspondence to: Dr Ari Ercole
E-mail: ae105@cam.ac.uk

Summary

Projected critical care demand for pandemic influenza H1N1 in England was estimated in this study. The effect of varying hospital admission rates under statistical uncertainty was examined. Early in a pandemic, uncertainty in epidemiological parameters leads to a wide range of credible scenarios, with projected demand ranging from insignificant to overwhelming. However, even small changes to input assumptions make the major incident scenario increasingly likely. Before any cases are admitted to hospital, 95% confidence limit on admission rates led to a range in predicted peak critical care bed occupancy of between 0% and 37% of total critical care bed capacity, half of these cases requiring ventilatory support. For hospital admission rates above 0.25%, critical care bed availability would be exceeded. Further, only 10% of critical care beds in England are in specialist paediatric units, but best estimates suggest that 30% of patients requiring critical care will be children. Paediatric intensive care facilities are likely to be quickly exhausted and suggest that older children should be managed in adult critical care units to allow resource optimisation. Crucially this study highlights the need for sentinel reporting and real-time modelling to guide rational decision making.

A pandemic due to the novel strain of swine-origin influenza A/H1N1 [1] has now been declared. At the time of writing, cases have been largely confined to the community and there has consequently been little discussion regarding the potential impact on critical care provision. However, this outbreak appears to still be in the early stages and the relatively small numbers of confirmed cases together with a lag between presentation and more severe illness means that there is statistical uncertainty regarding the true pathogenic potential of this virus. Since only the most severe cases would be hospitalised, it seems likely that critical care services would be needed to treat a significant proportion of any future hospital cases and thus capacity projections are essential. Using a model based on reasonable historical assumptions, it has been shown that a flu epidemic could potentially overwhelm critical care bed and ventilator capacity in England [2]. Early experience of the present strain suggests that the attack rate is particularly high in the young [3] and that this virus may have the potential to elicit an immunologically severe host response [4]. Such considerations serve to strengthen concerns that critical care provision may prove to be inadequate. Whilst additional capacity may be created to cope with the worst-case scenario, this would have important implications for ongoing acute and elective service provision. Contingency planning is vital if proportionate resources are to be made available at relatively short notice. Accurate forecasting of potential demand is clearly central to such planning.

The complex nature of the population-pathogen interaction may lead to considerable regional and temporal variability in clinical impact. The use of historical pathogenicity data may thus be inappropriate for real-time planning during an actual outbreak. Instead, it is preferable to use contemporary and, preferably, local estimates for the behaviour of the pathogen in question. In the initial stages of an outbreak when the number of cases in a locality is still relatively small, however, estimates of the attack rate, hospitalisation rate and case-fatality ratio are subject to inevitable statistical uncertainty. Furthermore, published information regarding the current state of the outbreak is currently limited to the total number of suspected and confirmed cases as well as fatalities. In particular, current hospitalisation figures are not published. Crucially, if real-time information regarding current hospitalisation and critical care admission is not available, then accurate prediction of likely resource requirements in a particular locality will be difficult.

This paper attempts to quantify the potential critical care burden of influenza A/H1N1 using the FLUSURGE 2.0 model [5] provided by the Centers for Disease Control and Prevention in the USA [6]. Using data on disease severity from the USA [1] and Mexico [3], estimates are presented for the likely peak bed occupancy and ventilator usage in England (for which unified critical care bed data are available) under the assumption that, in the early stages of this pandemic, few or no patients have yet been admitted to hospital. The effect of different assumed hospital admission rates on our predictions is also investigated. We have elected to undertake this exercise at the current time, as the change in recent public health strategy [7] means that community testing of potential cases will no longer be undertaken. Consequently, accurate updated estimates for community incidence of the illness will no longer be available to serve as a denominator in analyses of Intensive Care Unit (ICU) resource utilisation. The current statistics therefore provide the best estimate we will have to calculate the burden of critical care bed usage in relation to community disease levels.

Methods

Age-stratified data for the English population were obtained from the Office for National Statistics [8]. Total critical care capacity was taken as the sum of total adult level 3 beds (2030) [9] and total paediatric intensive care beds (265) [10]. The total number of available ventilators was assumed to be equal to the number of critical care beds. The current total number of confirmed cases of influenza A/H1N1 in England at the time of writing was 6162 [11].

The influenza A/H1N1 virus in Mexico and the United States appears to be genetically homogeneous [12]. It seems reasonable to believe therefore that, although socio-demographic factors may vary, the inherent pathogenicity of this virus is also homogeneous across the world. From the published data for Mexico, an attack rate of 61% for age < 15 and 29% for age > 15 [3] is assumed. By applying these estimates to the current population demographics [8] we calculated the total expected number of cases. It may be that many of these cases are mild. However if the disease is sufficiently severe as to require hospitalisation then we assume that 36% will require critical care admission and 18% will require ventilatory support [1].

Data on total hospital admissions are not published for England but are low at the time of writing and subject to statistical uncertainty. We calculated the 95% confidence intervals for different assumed hospital admission rates. Corresponding total hospital admissions were computed for these confidence intervals given our estimates of total cases above. These data were used in place of the default virulence assumptions in the FLUSURGE model, which are based on historical assumptions. The model was then used to calculate peak critical care bed occupancy and ventilator usage for these confidence intervals. A pandemic duration of 12 weeks was assumed (optimistic from the point of view of peak resource demand). Other parameters in the FLUSURGE model were left unaltered from their default values. Of particular note were death rate and acute hospital bed number which did not affect calculations of critical care bed occupancy and could therefore be set to arbitrary values.

Results

If none of the first 6162 confirmed cases had been admitted to hospital, 95% confidence intervals for the hospital admission rate are calculated as 0–0.06%. This is significantly different to the United States admission rate of 9% (95% CI: 7–12%) [1]. With these values, the modelled peak required critical care requirements for influenza cases alone ranges between 0% and 37% of capacity, with peak ventilator usage ranging between 0% and 19% of total capacity. As increasing numbers of patients are hospitalised the lower bound on peak predicted impact rises rapidly (Fig. 1). Best case critical care bed capacity is exceeded for hospitalisation rates above 0.25% (approximately 15 out of the current 6162 confirmed cases).

Figure 1.

 Peak demand for critical care beds (left axis) and ventilator requirements (right axis) during a modelled 12-week pandemic as a function of recorded case hospital admission rate (bsl00001). The top and bottom solid lines represent upper and lower bounds for peak demand based on 95% confidence intervals when hospital admission rates are estimated from 6162 total cases (the situation at time of writing). The dashed lines are the equivalent ranges for the same admission rate estimates but based on 10 times the number of cases.

As the outbreak progresses and more cases appear, estimates of parameters such as hospital admission rates improve in accuracy. Also shown in Fig. 1 are best and worst credible predictions when hospital admission rates are calculated based on 61 620 cases (i.e. 10 times that of the time of writing). It is evident that the statistical uncertainty is greatly reduced. However maximum capacity is still rapidly exceeded as the hospital admission rate is increased.

These estimates are based on critical care admission rates of 36% of hospitalised patients and a requirement for mechanical ventilation of half of this (18%), based on the best available data. Early results from England suggest that these figures are broadly correct, but numbers are still very small. Of the 7904 confirmed cases in the United Kingdom (UK) at the time of writing, there have been 123 hospital admissions (1.7%), of which 20 patients have required ICU admission [∼17% of hospital admissions (95% CI 10–25%)]. Changes in the critical care admission rate would impact substantially on resource utilisation, and since the estimates of hospital and ICU admission rates are based on small samples, we modelled the impact of changing hospital and critical care admission rates on critical care bed occupancy, so as to provide a model that could inform service planning. This is shown in Fig. 2, which maps peak predicted critical care bed occupancy for different hospital and critical care admission rates. This graph provides a means of estimating resource requirement early in the course of a pandemic, with increasing confidence in estimates as the number of cases increase.

Figure 2.

 Graph of predicted peak critical care bed occupancy as a percentage of total capacity for different assumed rates of hospital and critical care admission. Approximately half of critical care patients would be expected to need mechanical ventilation.

The critical care availability data and demographic data for the 10 strategic health authorities have been disaggregated to examine regional differences in potential impact. The regional data for the current upper bound on hospital admission rate above (0.40%) are shown in Table 1. The projected critical care burden varies inversely with critical care provision per capita.

Table 1.   Predicted peak critical care bed occupancy and ventilator utilisation for the 10 strategic health authority regions. A hospital admission rate of 0.25% is assumed (upper 95% confidence boundary for no observed admissions out of 6162 total cases at the time of writing).
RegionTotal critical care beds/100 000 populationPredicted peak critical care bed occupancy (percent total capacity)Predicted peak ventilator utilisation (percent total capacity)
England total4.516078
London7.59447
North East5.912059
North West5.014069
Yorkshire and the Humber4.416080
South Central4.216082
West Midlands4.217084
East of England3.2220110
East Midlands3.2220110
South West3.1220110
South East Coast3.0230120

As of 6th July 2009 [13], the combined international mortality was 0.45% (429 of 94 512 cases). Mortality figures for countries with very small numbers of cases are likely to be unreliable. Thirteen countries have reported over 1000 cases, and this group shows a mean mortality of 0.42%, with a range of 0–2.4%. The country with the largest number of cases, and hence (arguably) the best single national estimate is the USA, with 33902 cases and 170 deaths (0.5% mortality) [14]. The mortality in the UK is substantially lower than the overall average, at 0.04%. It is likely that a large proportion of the variation in mortality rates may be attributed to different rates of community testing for H1N1 infection, with resulting large differences in the denominator used for calculating these figures.

Discussion

Pandemic modelling and forecasting depends critically on epidemiological data. Unfortunately there is considerable statistical uncertainty in these parameters in the early stages of any outbreak, when the case numbers are small. Furthermore the available data are inevitably biased due to under-reporting and the lag between disease presentation, diagnostic confirmation and clinical progression making real-time prediction difficult. In the most uncertain scenario, before any patients are reported to have been hospitalised, our model suggests that the predicted effect on critical care resources ranges from no impact to a significant proportion of beds (37%) and ventilators (19%) being utilised. Best case predictions suggest that all critical care beds will be filled with influenza A/H1N1 patients if the hospital admission rate is greater than 0.25%. Worryingly, even small increases to the observed admission rate make the overwhelming scenario increasingly statistically credible.

We have used the FLUSURGE model to estimate peak critical care bed occupancy. Whilst our estimates have employed contemporary virulence data, the model makes assumptions regarding the detailed kinetics of disease spread which cannot be prospectively verified. This is an unavoidable limitation of our study. We have assumed an arbitrary pandemic duration of 12 weeks. However, our results scale inversely with duration and are thus easily generalisable.

An observed attack rate in Mexico of 61% in the under-15 age group has been assumed. This is much higher than that of seasonal flu and clearly socio-economic factors may be important as well as local availability of antiviral therapy. In support of generality of this assumption it is interesting to note that the younger age group were similarly over-represented in the case mix from the United States [1]. If this is indeed the case, extrapolation to the demographic profile for England suggests that approximately one-third of cases would be less than 15 years of age. Since paediatric intensive care beds represent only just over 10% of the total number of critical care beds in the UK, it is clear that paediatric resources may be particularly badly affected with many children having to be looked after in adult units. Even if we apply the much lower adult attack rate of 29% to the population as a whole, this only reduces our required critical care capacity estimates by less that 20%. In this case, approximately one-fifth of cases would be under 15 years of age; still significantly greater than paediatric critical care bed availability. Despite these prepandemic estimates, it is important to report from the US experience that, while a large proportion of hospitalised patients are in the paediatric age range, very few of these have required critical care [15].

It is assumed that the total critical care bed capacity for England comprises 2030 adult level 3 beds [9] and 265 paediatric intensive care beds [10]. From these published data an additional 1607 adult plus 43 paediatric high dependency beds are also available, many of which may potentially be used to ventilate patients. Nevertheless the calculations still show that even this number could be far too small to cope with demand. Additionally, since many intensive care units (and acute hospital beds) run at high occupancy, much of this capacity would not be available in the event of a pandemic. Creation of additional ventilated beds would require, for example, cessation of normal elective surgical activities but even this capacity could not be realised instantaneously. This is a crucial consideration for planning and illustrates the need for accurate real-time forecasting. Whilst regional variations in critical care provision exist, the data suggest that these are small and so interhospital transfer is unlikely to provide a solution in an overwhelming pandemic. The projected impact on the London region is less due to the high per capita critical care capacity. However this is an area of high population density, which may conceivably increase the attack rate locally.

Critical care patients will represent the most severe cases in a pandemic and contingency planning specific to this group is of paramount importance. The potential range of impacts that an influenza A/H1N1 pandemic might have on critical care service provision for different hospital admission rates has been estimated. The data presented support the conclusion that accurate real-time forecasting in the early stages of this rapidly evolving outbreak is critically dependent on up-to-date epidemiological estimates. Clearly other parameters such as the attack rate, the proportion of cases developing clinically severe disease, and the length of ICU stay are also subject to uncertainty. In particular, it is a cause for concern that at least some patients with H1N1 infection severe enough to need ventilatory support have required prolonged ICU stay, in excess of the assumptions in our model. Finally, the limiting resource may not be critical care beds, or even ventilators, since early reports suggest the common need for sophisticated ventilatory modes such as high frequency oscillation and a common requirement for renal replacement therapy [16].

These considerations serve to underline the need for local data acquisition and strengthen the case for sentinel reporting and rapid dissemination of clinical information. Rational policy making and contingency planning necessitate such continuous refinement of input data along with mathematical modelling capable of incorporating the effects of inevitable statistical uncertainty. There are current plans underway to set up such sentinel centres in the UK, and data collection from these centres will provide useful information regarding hospital admission rates early in the course of a pandemic. However, the critical care admission rate may represent a small proportion of hospital admissions, and the derivation of confident estimates to guide clinical practice may require reporting from a wider base than the approximately six sentinel centres that are currently envisaged. Fortunately, the UK critical care community has existing arrangements for reporting such data, with approximately 80% of critical care units in England and Wales contributing to the Case Mix Program of the Intensive Care National Audit and Research Centre [17] and all Scottish critical care units contributing to an ongoing multicentre audit process of the Scottish Intensive Care Society Audit Group [18]. This infrastructure is now being used to provide a mechanism for real time reporting of critical care utilisation.

Indeed, we need to go beyond national boundaries to make best use of the data that emerges from the early phases of a pandemic. In particular, the likely impact of the illness on southern hemisphere countries during their current winter may provide important lessons for managing the pandemic in the coming northern hemisphere winter. The best use of such data demands close international collaboration, commonality of data collection, and rapid organisation and analysis of such data. An emerging collaboration between critical care clinicians in Canada, Australia and New Zealand, France, and the UK and Ireland may provide the foundation for such a process. Until such data are available to enable calculation of likely bed usage, our nomograms provide a means for ready reckoning of resource utilisation early in the pandemic, and will aid rational allocation of critical care resources, particularly in subsequent (and probably inevitable) phases of a pandemic.

Funding

This study was entirely funded from departmental resources. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Acknowledgements

DKM is funded as a Senior Investigator of the National Institute for Health Research, UK; and supported by the BOC Professorship of the Royal College of Anaesthetists.

Ancillary