Methicillin-resistant Staphylococcus aureus (MRSA) is increasingly implicated in potentially lethal hospital-acquired infections [1, 2]. Some of the MRSA strains that are found in patients in the community have been acquired during hospitalization or a visit to a hospital. However, MRSA strains likely originating outside of hospitals and different from those found in hospitals have been identified [3–5]. People who are infected by MRSA in the community have different risk factors than those infected in hospitals [1, 2], particularly as common risk factors in hospitals may be rare or absent in the community.

While hospital-acquired MRSA infections can be fatal [1, 2], cases of severe and life-threatening MRSA infections from the community have also been reported [6–8] and these case reports suggest that the prognosis of community-acquired MRSA infections may be poor [9]. However, hospital-based studies do not capture patients from the community whose infection does not require hospital care, and thus such studies or case reports cannot evaluate the impact of this infection on mortality. To date, no population-based study has evaluated the prognosis and mortality of MRSA patients diagnosed in the community.

We thus conducted a large cohort study to assess mortality associated with MRSA infections diagnosed in the community in comparison with a population-based sample of patients free of this infection.

We used the General Practice Research Database (GPRD) between 1 January 2001 and 31 December 2005 for this study. The GPRD contains the diagnostic, testing and prescribing records of approximately 3.2 million patients from over 400 general practices in the United Kingdom (UK) as recorded by their GPs. The GPRD is extensively used for medical research [10, 11] and has been used previously for prognostic work in cardiovascular disease [12], as well as infectious disease research [13, 14]. GPRD data is routinely audited for quality [10] and is considered to be of high validity for medical research.

The cohort included 1439 patients diagnosed with MRSA between 2001 and 2004, and 14,090 matching disease-free patients from the GPRD. Over the one-year period after cohort entry, 21.8% of patients diagnosed with MRSA and 5.0% of those without the diagnosis died. The Kaplan-Meier plot is shown in Figure 1. The loss to follow up owing to transfer out of practice was comparable in the patients with and without MRSA (4.2% and 3.0%, respectively). Patients with MRSA diagnosis were more likely to be male and to have co-morbid conditions predictive of mortality as compared with patients without a diagnosis of MRSA (Table 1).

The cumulative incidence of the composite outcome of death or hospitalization is depicted by the Kaplan-Meier plots in Figure 2. Table 2 shows that the adjusted HR of this outcome with MRSA infection is 3.31 (95% CI: 2.97–3.69).

The sensitivity analysis for informative censoring produced an adjusted HR of 4.91 (95% CI: 4.30–5.61) for worst-case imputation, and an adjusted HR of 2.69 (95% CI: 2.36–3.07) for best-case imputation. Prior antibiotic use had a small effect on mortality once we adjusted for other factors with an adjusted HR of 1.30 (95% CI: 1.14–1.49). Finally, among the subset of patients with no co-morbidities, the effect of MRSA infection on mortality was comparable to that obtained for all patients with an adjusted HR of 4.86 (95% CI: 4.10–5.77).

In this study, we demonstrated an increase in mortality within a year of a diagnosis with an MRSA infection in the community. This increased mortality cannot be completely explained by underlying co-morbid conditions as we adjusted for the major risk factors of death. The four-fold increase in mortality persisted across several sensitivity analyses. In addition, we found a risk increase of similar magnitude with the combined outcome of death and hospitalization.

Patients with nosocomial MRSA bacteraemia [18, 19] have been reported as having as great as a ten-fold increased risk of death [18]. However, these extremely severe infections are not representative of all nosocomial MRSA infections, and likely not representative of community-acquired MRSA infections, many of which do not require hospitalization [6]. As expected, our observed death rate is much lower than in these studies.

In this study we observe an average of 360 MRSA infections diagnosed in the community per year among the adult GPRD population during the study period. This is about 13 cases per 100,000 adults in the GPRD population, which is similar to the incidence found in other studies such as that by Fridkin et al [4]. This is especially true if we assume that some MRSA infections are diagnosed in hospital emergency rooms and thus not counted in our study. Fridkin et al also found a significant proportion of community-associated MRSA infections occurring in patients above 65 years of age [4].

In this study we considered a large cohort of patients diagnosed with MRSA infections in the community and a large group of matching disease-free patients all selected from a database that is representative of the UK population [10]. The death rate among the disease-free patients is broadly consistent for this age and sex distribution with that seen in UK vital statistics [20]. Using the GPRD allows us to study patients with all levels of disease severity encountered by GPs, including those not treated in hospitals and thus not part of any hospital-based investigation. The study design (frequency matched cohort study) ensured that the distribution of key covariates (general practice, age and calendar time) was balanced between patients with and without MRSA at baseline. By matching on general practice we indirectly matched on unmeasured factors such as socio-economic status, area of residence and nursing home care, which are usually common to the patient population of a given general practice. These demographic factors have been shown to be important in the epidemiology of MRSA [21]. By matching for practice we control for the potential confounding nature of these variables.

There are nevertheless some limitations to our study. In particular, we could not distinguish site or severity of individual MRSA infections. As we could not examine antibiotics given in hospital, it was not feasible to use post-diagnosis treatment either to separate infection site or infection source as different therapy is recommended for community-acquired MRSA [22]. However, our observed mortality of MRSA infections diagnosed by general practitioners in the community rules out that these are the extremely severe infections seen in hospitals [18]. On the other hand, these results imply that individuals with MRSA diagnoses in the GPRD are not mostly asymptomatic carriers but include clinically significant infections: a conclusion supported by other work in the GPRD on MRSA infections [13]. Moreover, we strongly believe that currently, general practitioners in the UK do not routinely screen and thus diagnose MRSA in asymptomatic individuals in the community.

We could not exclude the possibility that some MRSA infections were acquired in hospitals and not in the community, because we had no information on potential transmission from visits to or work in hospitals. We could not look at the spread of MRSA from close contacts [23] or clusters in workplaces or schools [24]. Therefore, we describe here the prognosis of MRSA infections detected by general practitioners, but we cannot demonstrate that these infections arose independently in the community. However, the high mortality rate underlines the importance of their control. To contribute to achieving this, we recently proposed the appropriate use of antibiotics [13].

The increased mortality observed for MRSA infections may be a result of a health state that predisposes to MRSA infections or residual confounding by this health state. Frailty and residing in a nursing home are examples of potential confounders that we were unable to adjust for in this study. If the mortality associated with MRSA is a manifestation of underlying susceptibility to infections, then past antibiotic use could serve as an estimate of the degree of bias in our study. Past antibiotic use should not be associated with distant outcomes but will show patients who have a history of infections. While we found a small increase in mortality predicted by antibiotic prescriptions given in the year prior to cohort entry, it was much too small to be a major confounder of the observed increase in mortality. Therefore, some degree of susceptibility to infection may explain a small part of the increased death rate. However, because of the small size of this effect we can infer that the risk of death in this study is not principally a marker of an underlying susceptibility to infections.

We did not report on the effect of antibiotic therapy on mortality after MRSA diagnosis for three reasons. One, the GPRD does not report drug exposures during hospitalization. This would result in the most severe cases of MRSA infection being misclassified as unexposed. This could lead to a falsely protective effect of antibiotic use if it was only recorded among the healthiest patients. Two, the use of antibiotics would be a dynamic treatment regime and more advanced statistical methods would be required to obtain a valid estimate [25]. Three, the GPRD recording system does not allow us to determine whether an antibiotic prescription is being given for the original infection or for a second infection that developed independently.

While previous work in this database has shown differences between classes of antibiotics and the development of MRSA diagnosed in the community [13], we did not see any strong link between antibiotic class and prognosis. Therefore, we considered antibiotics only as a single drug class for this analysis as the principal reason for including antibiotics as a covariate was to control for possible differences in susceptibility to infections between patients with and without MRSA.

The broad use of sensitivity analysis [26, 27] is a key strength of our study. The results of these analyses indicate that the increase in mortality after a diagnosis of MRSA is robust to changes in assumptions. The use of the READ code 4JP..00 for MRSA infection has been discussed previously [13] and patients with this code were more similar in characteristics to post-operative MRSA wound infections than to carriers. The very high mortality rate seen in patients with MRSA in the present study, independent of baseline health status, is also highly suggestive of infections rather than carrier status. However, carrier status may be associated with lower mortality and hospitalization rate in comparison to clinically relevant infections. Thus, including this code may have led to an underestimation of the impact of a diagnosis of MRSA in the community on hospitalization and mortality.

The largest remaining source of possible bias in this study is the possibility of unmeasured confounding or, more likely, residual confounding. While non-parametric approaches could be used to reduce residual confounding in this study, they are unlikely to explain effects of the size that we observe [28]. It is not likely that such large effects could be explained by anything other than many extremely strong confounders [29] based on estimates of the effect of confounding. We also avoided introducing bias from either the misallocation of person time [30] or from adjusting for variables potentially in the causal pathway between MRSA infection and mortality [31] by defining confounders using only the baseline characteristics of the subjects. Adjusting for factors (such as secondary infections) that arise after the diagnosis of MRSA infection could lead to overadjustment [32] and could bias the estimates towards the null.

In conclusion, our study suggests that MRSA can be a potentially serious infection in the community leading to increased mortality. Further research will allow a better understanding of MRSA diagnosed in the community and the prognosis of patients with this diagnosis. This is essential for effective prevention of MRSA arisen from the community, which will not only reduce hospital admissions with infections, spread into and within hospitals, and transmission in the community, but also likely will prevent deaths in the community.