Using a dynamic stochastic simulation model, we have evaluated four intervention scenarios to control the spread of MRSA under comparable in silico conditions. Although universal screening at hospital admission leads to the fastest decline in both the hospital-wide and ICU prevalence of MRSA, it also requires the highest investment costs and the longest time till return of investment. In our analyses, screening all patients at ICU admission and those previously detected with MRSA (so-called flagged patients) or screening of flagged patients only were almost equally cost-saving in a 10 years period and were both associated with the fastest return of investment. These strategies should, therefore, be seriously considered by hospitals that aim to control the nosocomial spread of MRSA.
Our findings are complimentary to those of two other modelling studies on screening for carriage with antibiotic-resistant bacteria in hospitalized patients. In one study, Hubben and co-workers compared the effects of PCR-based and chromogenic screening tests 
. Determination of the optimal screening was not investigated in the current study, and we have, therefore, used a fixed time-to-result parameter. In the other modelling study, Robotham and co-workers investigated the effects of different screening tests in ICU patients, in combination with patient isolation and decolonisation 
. The latter study did not include the effects of ICU-screening on the non-ICU hospital population and did not include the possibility of patients being readmitted while still colonized.
Yet, this is an important aspect of the dynamics of nosocomial MRSA as it explains why control measures may have not only a direct, almost instantaneous, effect on the prevalence of the nosocomial MRSA in the hospital, but also an indirect effect due to interruption of the so-called feedback loop; when less patients acquire colonization during hospitalization, less patients will be colonized upon readmission to the hospital (see supplementary Figure S5
). This lower admission prevalence in time ensures that controlling spread of the nosocomial MRSA will become easier in time. Therefore, neglecting these feedback loop dynamics will underestimate the cost-savingness of interventions.
An important assumption of our model is that the pathogen spreads predominantly in health care settings. Interventions in health care settings will not be very effective in prevention of acquisitions in the community. With substantial spread in the community, a smaller fraction of the acquisitions can be prevented and also the fraction of the patients colonized on admission that are flagged will reduce. In the extreme case that transmission almost exclusively occurs outside health care settings, interventions in hospitals are ineffective and the cheapest strategy is the optimal one. For these reasons, our model is not applicable for community-associated MRSA, but is applicable for other pathogens with similar epidemiological characteristics as MRSA.
Although reductions in the occurrence of nosocomial MRSA infections have been reported 
, multi-resistant Gram-negative bacteria, such as those producing extended-spectrum β-lactamases (ESBL) or carbapenemases are emerging in health care settings worldwide 
. With no new antibiotics on the horizon to treat infections caused by these bacteria, effective transmission control strategies are needed. Yet, identifying the most effective control strategy for every possible setting through clinical trials seems impossible. Well-designed large clinical trials on rapid diagnostic testing of MRSA yielded highly variable results, varying from no effects on infection rates in surgical units 
to 69.6% reductions in hospital-wide infection rates 
. Moreover, the stochastic nature of ARB dynamics necessitates long study periods to avoid that conclusions are primarily based on chance events, rather than on true effects. We have, therefore, used mathematical modelling. Of note, mathematical models always are a simplification of real life complexities and cannot produce very precise predictions for a certain situation. For instance, we have assumed that all isolation measures were equally effective in all isolated patients and that all measures were executed with equal efficacy. One can easily think of scenarios in which these assumptions do not hold 
. Therefore, the main value of modelling is the comparison of different scenario analyses, while keeping other important parameters constant, rather than providing exact values.
In doing so, our analyses identified screening of flagged patients and ICU patients as a very powerful control strategy, even reducing prevalence levels in non-ICU wards. The central role of the ICU in our model follows from two assumptions. First, many patients discharged from ICU are transferred to other wards. Therefore, prevention of spread in ICUs will reduce the frequency at which MRSA is introduced in other wards. Second, the likelihood of cross-transmission is higher in ICUs than in non-ICU wards. This assumption is motivated by the more frequent (and possibly even more intense) contacts between patients and HCWs, allowing HCWs to act as transmission vectors of MRSA. Moreover, antibiotic selective pressure is higher in ICUs than in non-ICU wards, which may increase the likelihood that a HCW will pick up a pathogen during a physical patient contact and that another patient will be successfully colonized after being contacted by a temporarily contaminated HCW. Finally, the severity of disease of critically ill patients in ICU wards makes them more susceptible to acquire colonization with MRSA than patients in non-ICU wards. Several studies indeed support the potential effects of ICU-screening on hospital-wide resistance levels 
With regard to the costs of interventions, our analyses were most sensitive to the costs associated with an ICU-acquired infection caused by MRSA. Many studies have quantified the costs of ICU-acquired bacteremia and ventilator-associated pneumonia 
and these estimates were all in the range of the €30,000 that we used. However, these costs sensitively depend on the additional length of stay that can be ascribed to infections, which is difficult to determine, see e.g. 
. Another important aspect is the role of the ICU in the patient flow. We have used data on patient admissions to 13 ICUs in the Netherlands. Naturally, patient flow may be different in other hospitals.
One of the simplification of the model is that patients should be colonized with MRSA before they are at risk of getting an infection with MRSA, i.e., we did not explicitly incorporate that some patients may acquire MRSA infection directly without being colonized first, i.e., due to invasive medical procedures. A slight increase in the daily probability for colonized patients to acquire an infection would lead to the same ratio of colonized and infected patients. Therefore, our sensitivity analysis on the daily probability for colonized patients to acquire an infection can also be interpreted as a proxy for a sensitivity analysis to the parameter which determines how often patients acquire an infection without being colonized.
We also assumed that the rates of conventional microbiological cultures performed for clinical reasons are independent of screening on admission (0.03 and 0.3 per patient day in non-ICU and ICU wards). We have assumed that a clinical suspicion of infection is the main reason for obtaining clinical cultures, and that screening for MRSA-carriage on admission reduces the frequency of obtaining clinical cultures in case of a clinical suspicion of infection.
Our model contains many parameters and some parameter values are unknown, whereas others may differ between hospitals and countries. We have based our values on data from the literature and from our own hospital, where possible. To fully capture the effects of parameter uncertainty we would have considered to perform a probabilistic sensitivity analysis (PSA) for all parameters simultaneously, as was performed by Robotham et al. 
. However, due to the higher complexity of our simulation model, as compared to the model of Robotham et al., this was computationally unfeasible. We, therefore, had to restrict our sensitivity analysis primarily to univariate sensitivity analysis. As a result, there may be more uncertainty in the results as we have presented here.
We did not include decolonization of detected carriers as a measure to control MRSA. Naturally, adding this measure (if successful at low costs) would increase intervention effects and would make the duration till return of investments shorter. Although persistently colonized HCWs were included as potential sources for MRSA transmission, we did not include screening and decolonization of them as intervention measure. This intervention measure would - in most settings – only slightly enhance the control of MRSA transmission, at the cost of significant expenses due to the necessity to replace colonized HCWs.
The (cost)-efficacy of admission screening strategies critically depends on the effectiveness of the infection prevention measures taken when a carrier of MRSA is detected. If these measures are not very effective, it may not be wise to invest lots of efforts in detecting carriers. The effectiveness of barrier precautions has been sufficiently high in the Netherlands and the Scandinavian countries to prevent high prevalence levels of MRSA. However, it is still debated whether patient isolation prevents transmission at all 
, and a recent estimate indicated that the efficacy is in the order of 25% 
, We, therefore, advocate to perform more clinical studies to determine the efficacy of decolonization, isolation or cohorting measures in different settings.
In conclusion, our study demonstrates marked and robust differences in the costs and effects of different infection control measures for MRSA. Because of the central role of ICU wards in patient flow in hospitals, the vulnerability of ICU patients to infections caused by MRSA and the high costs associated with these infections targeted infection control measures in ICU wards are likely to be the most effective and cost-saving from a hospital perspective.