Modelling study to estimate the health burden of foodborne diseases: cases, general practice consultations and hospitalisations in the UK, 2009

Objective To generate estimates of the burden of UK-acquired foodborne disease accounting for uncertainty. Design A modelling study combining data from national public health surveillance systems for laboratory-confirmed infectious intestinal disease (IID) and outbreaks of foodborne disease and 2 prospective, population-based studies of IID in the community. The underlying data sets covered the time period 1993–2008. We used Monte Carlo simulation and a Bayesian approach, using a systematic review to generate Bayesian priors. We calculated point estimates with 95% credible intervals (CrI). Setting UK, 2009. Outcome measures Pathogen-specific estimates of the number of cases, general practice (GP) consultations and hospitalisations for foodborne disease in the UK in 2009. Results Bayesian approaches gave slightly more conservative estimates of overall health burden (∼511 000 cases vs 566 000 cases). Campylobacter is the most common foodborne pathogen, causing 280 400 (95% CrI 182 503–435 693) food-related cases and 38 860 (95% CrI 27 160–55 610) GP consultations annually. Despite this, there are only around 562 (95% CrI 189–1330) food-related hospital admissions due to Campylobacter, reflecting relatively low disease severity. Salmonella causes the largest number of hospitalisations, an estimated 2490 admissions (95% CrI 607–9631), closely followed by Escherichia coli O157 with 2233 admissions (95% CrI 170–32 159). Other common causes of foodborne disease include Clostridium perfringens, with an estimated 79 570 cases annually (95% CrI 30 700–211 298) and norovirus with 74 100 cases (95% CrI 61 150–89 660). Other viruses and protozoa ranked much lower as causes of foodborne disease. Conclusions The 3 models yielded similar estimates of the burden of foodborne illness in the UK and show that continued reductions in Campylobacter, Salmonella, E. coli O157, C. perfringens and norovirus are needed to mitigate the impact of foodborne disease.


Bootstrapping of outbreak data to estimate the proportion of cases hospitalised
Data on hospitalisation in outbreaks were only available from England and Wales. For each reported outbreak in the England and Wales dataset (excluding outbreaks that occurred in hospitals and residential institutions), we calculated the proportion of outbreak cases that was hospitalised and plotted the resulting distribution for the proportion of cases hospitalised. We calculated this by causative organism and separately for all outbreaks and for foodborne outbreaks only. There was no major difference in hospitalisation between all outbreaks and foodborne outbreaks, so we based estimates of hospitalisation on data from all outbreaks. To account for uncertainty in hospitalisation parameters, we used a two-step approach. For each pathogen, we first obtained an empirical distribution for the proportion of cases hospitalised by bootstrapping 4,999 replicate samples of the outbreak data. For example, if there were 50 reported outbreaks for a given pathogen, we sampled 50 outbreaks with replacement from this set and calculated the mean proportion of cases hospitalised across the outbreak sample, weighted by outbreak size. This was repeated 4,999 times for each pathogen. The hospitalisation proportion was weighted by outbreak size because many reported outbreaks involve few cases and are therefore unlikely to involve hospitalised cases. The small number of larger outbreaks, on the other hand, is potentially more informative for estimating hospitalisation. We then fitted a Beta distribution to the bootstrapped data and estimated the corresponding a and b parameters using maximum likelihood. The mean hospitalisation proportions for each pathogen and Beta parameters used in Model 1 are given in Table A1. The fits of the Beta distributions to the outbreak data are show graphically in Figure A1.
For Listeria, all reported outbreaks occurred in hospitals, so it was not possible to estimate the hospitalisation rate from outbreaks. For adenovirus and sapovirus, no outbreaks were reported. For these two pathogens, parameters based on analysis of rotavirus and norovirus outbreaks respectively were used. Bootstrap estimates with fitted Beta distributions for the remaining 10 pathogens are shown below.

Deriving priors for the proportion hospitalised (p) from the IID1 and IID2 Studies
We pooled data from the IID1 and IID2 Studies and calculated, by pathogen, the proportion of cases presenting to the GP that were hospitalised. Applying this proportion to the rate of GP consultation gave an estimate of the hospitalisation rate. For each pathogen, we used the ratio of this rate to the rate of community IID to obtain an estimate of the proportion of cases hospitalised. This approach implicitly assumes that hospitalised cases always consult a GP. This is reasonable in the UK, as hospitalisation is likely to occur through a GP referral, but potentially disregards a fraction of more severe cases (e.g. cases admitted as a result of an emergency hospital visit). However, it was not possible to estimate hospitalisation directly from the IID1 and IID2 community cohort study components, as hospitalisation is very uncommon and the two cohort studies were not designed to measure the rate of hospitalisation.
To account for uncertainty in the hospitalised proportion, we took 100,000 random samples from the distributions of the overall IID rate, c p , and the proportion of GP cases hospitalised, and fitted a Beta function to the resulting distribution for the hospitalised proportion using maximum likelihood methods. The estimated parameters from this Beta distribution were used to inform the prior values for  p in the Bayesian approach (Table A2). For VTEC O157, for which hospitalisation information was not available from IID1 and IID2, we used a non-informative prior defined by the distribution Beta(1,1). For pathogens for which no hospitalisations were observed in the IID1 and IID2 studies, we specified limits to the fitted Beta distributions by assuming that the next case observed would have been hospitalised. Thus, for Shigella, with 11 cases and no hospitalisations, we obtained Beta parameters for a distribution with a mean equivalent to 1/12=0.087. Empirical bootstrap distributions with fitted Beta functions are shown below.

Monte Carlo approach (Model 1)
We obtained estimates of F p , G p and H p using Monte Carlo simulation, each time drawing at random from each parameter distribution in the model. We carried out 100,000 simulations, discarding the first 10% and retaining the model outputs for every 10 th simulation. We checked model convergence graphically by plotting parameter values over time to verify adequate mixing, plotting autocorrelograms and comparing density plots for outcome variables by tertile of the simulation chain. The model and associated parameter distributions are described below: From the ensuing distributions of F p , G p and H p , we used the median and central 95% of the distributions as the point estimates and 95% credible intervals respectively. Parameter values for each pathogen are given in table A1 below.

Bayesian approach (Models 2 and 3)
In the Bayesian approach, we included parameters for the prior distributions of  p and  p . These priors were used, together with the outbreak data to obtain posterior distributions for these parameters, which were then used in the model as described below: consultations as observed in IID1 and IID2. The prior values for parameters  p and  p are defined by uniform and Beta distributions respectively as described above. In Model 2, the uniform distributions for  p were informed by data from published multi-pathogen food attribution studies. We used a further model, Model 3, with the same structure as Model 2, but with parameters for the prior distribution of  p being derived from case-control and food attribution studies from the literature review. A full description of parameters for models 2 and 3 is given in the technical appendix.
For each model, we carried out 100,000 simulations to obtain posterior distributions for F p , G p and H p , discarding the first 10% and retaining the model outputs for every 10 th simulation. We checked for model convergence as described for the Monte Carlo approach above. Parameter values for each pathogen are given in tables A2 and A3 below.