To combat the spread of AMR, hospitals are advised to screen high-risk patients for carriage of antibiotic-resistant bacteria on admission to limit opportunities for onward transmission. One commonly used risk factor is previous admission to a hospital with a known resistance problem, such as a current outbreak, or high endemic prevalence of resistant organisms. Screening patients from such ‘high-risk’ hospitals will identify the largest proportion of colonised patients per screened patient because of the greater relative prevalence in ‘high-risk’ hospitals. However, in terms of the risk of introduction and subsequent transmission to new patients, it is the absolute number of colonised patients received by a hospital which determines the chance of other patients becoming colonised. For example, if 50% of two patients received from a high-risk hospital are colonised, this one patient poses a lower onward transmission risk if unidentified than 1% colonised patients out of 1000 patients previously admitted to the same hospital (10 patients). The structure of the patient referral network dictates that not all hospitals pose an equal risk to others because patient exchanges are not randomly distributed. As larger outbreaks are more likely to draw more attention, even if they occur further away, this can result in a discrepancy between the perceived and actual risk of receiving colonised patients.
To illustrate this, we considered, as an exemplar, CPE introductions into NHS hospital trusts in different referral regions in England. Substantial increases in CPE over the period 2008–2014 reflect what would be expected for a new AMR threat, particularly one initially predominantly confined to healthcare settings. Overall, CPE has made alarming gains in England over the last decade, with many cases forming part of larger clusters as opposed to individual introductions. Although many of the single CPE isolates found in various hospitals may still be the result of introductions from abroad [15, 16], the spread of CPE within English hospitals is gaining importance in the overall epidemiology. Our results should therefore be broadly generalisable to future AMR threats.
In contrast with the common practice of import screening based on assigning a few hospitals as high risk, we found that transfer of patients between hospitals with small AMR problems in the same region often pose a larger threat in terms of absolute numbers of colonised patients admitted to a hospital than transfer of patients from hospitals in other regions, even if the prevalence of AMR in the other regions is orders of magnitude higher. This means that, with the current pattern of hospital referrals, the closeness of other hospitals is a much greater driver for the spread of antibiotic-resistant bacteria than their AMR prevalence. As a rule of thumb, for any given referral region, the prevalence of AMR in another referral region needs to be at least 100-times higher to contribute substantially to the admission of patients colonised with antibiotic-resistant bacteria; equivalently, without knowledge of the regional structure, hospitals which share most patients pose a greater threat than those with the greatest prevalence, unless there are marked disparities in prevalence.
An outbreak of KPC-positive Enterobacteriaceae limited to hospitals in the Manchester referral region demonstrated the multi-institutional component of CPE dispersal, largely driven through patient movements. The observation that hospitals in the Liverpool referral region were less affected by KPC-positive bacteria, despite their geographical proximity to Manchester, illustrated the lower risk of cross-regional introduction. Screening of relatively few patients could therefore potentially be used to mitigate inter-regional spread at the start of an outbreak, although our results suggest this would become less effective with time. Such compartmentalisation only prolongs the time to a successful introduction  because the chance remains that the admission of one undetected colonised patient may seed a new outbreak. Any inter-regional screening efforts should therefore always be accompanied by intra-regional or hospital-specific control efforts as our results show that this is often the most likely initial source of resistant micro-organisms.
Several studies have highlighted the necessity of coordinating IPC activities within regions , often building on the structure of the patient referral network [19–21]. This can, for example, be done by sharing information about the colonisation status of patients between healthcare institutions, through centralised registers  or by patient held cards , alerting hospitals to take appropriate action to prevent onward transmission and saving money from unnecessary repeated screening. Regional coordination may also aid effective contact tracing in outbreaks that span multiple institutions. Our study shows that it is essential for hospitals to get an up-to-date overview of the true prevalence of CPE, or AMR in general, in their surrounding hospitals to estimate the risks they pose through shared patients. The improved sharing of prevalence estimates, for instance obtained through periodic point-prevalence surveys for AMR, between all healthcare trusts, and in particular those referring many patients, should therefore be promoted.
We did not find evidence that the increase in CPE over the last decade was due to large-scale breakdowns in IPC standards, again suggesting our results should be broadly applicable to new AMR threats arising on a background of ongoing IPC efforts. The single isolates or small outbreaks of CPE affecting hospitals throughout the country represented many carbapenemase types, but hospitals with longer outbreaks usually only ‘suffered’ from a single dominant carbapenemase and did not have parallel outbreaks involving multiple carbapenem resistance mechanisms despite what appears to be continuous new exposures. However, it is possible that temporary lapses in IPC give any AMR threat the opportunity to spread unseen after introductions, giving rise to larger outbreaks and highlighting the importance of continued vigilance. Moreover, the length of some of these outbreaks indicate that they may be hard to eradicate once established. This could be caused by colonised patients returning to hospital, thus reintroducing the bacteria repeatedly, or the establishment of environmental AMR reservoirs within the hospital.
The main study limitation is that we were only able to approximate the prevalence of our exemplar, CPE, since data came from a voluntary system of submitting isolates to a reference laboratory and reporting rates may differ considerably between hospitals and over time as screening efforts change depending on the perceived problem. However, firstly, the overall increase in CPE nationally is unlikely to be solely the result of changes in screening practice, and regional differences in the occurrence of the different resistance mechanisms should not be affected. Secondly, it is likely that hospitals with a known CPE problem, such as those in the Manchester referral region, engaged in a more active search for possible colonised patients, apparent from the higher proportion of isolates originating from rectal swabs, and therefore submitted relatively more isolates. If anything, this would have caused the differences between regions and the effect of reported problems further away to be over-estimated, which means, in turn, that the contributions from hospitals within the referral region are probably even greater than we estimate. Finally, the goal of this study was to investigate the interplay between hospital networks and recommendations for dealing with new AMR threats that have generally focussed on screening patients from ‘high-risk’ hospitals because of their greater relative prevalence. We investigated CPE only as an exemplar of such a new infection threat – our conclusions do not depend on the specific organism, only on variation in prevalence across regions and changes in its distribution being broadly what one would anticipate from a new AMR threat, and the main current method of transmission being hospital based.
Our calculations were based on the structure of the inter-hospital patient referral pathways. This implicitly assumes that, for the organism in question, community spread of AMR occurs rarely, and leaves out other healthcare facilities, such as long-term care facilities or nursing/residential homes. If community acquisition becomes an important part of the dispersal mechanisms, the regional referral networks may become less meaningful, since community acquisition would impose an additional relatively uniform probability of introduction on each hospital, diluting the effect of import screening from other hospitals, irrespective of how far away they are in the network. However, it is likely that contact patterns between patients in other healthcare institutions, or other social structures or community interactions, reflect the structure of the patient referral pathways. Furthermore, the observed regional differences in CPE resistance mechanisms do suggest that, at present, the dispersal of this AMR threat is primarily taking place within the healthcare regions.
We were unable to estimate the introductions by patients previously admitted to foreign hospitals due to the lack of reliable admission numbers. However, it is unlikely that the number of admissions from abroad would surpass the number of admissions from other regions in England. We would therefore tentatively suggest that the same conclusion applies to patients from abroad, namely that the number of introductions from abroad only contributes considerably if the prevalence within the region is orders of magnitude lower than the prevalence in the other country.
While we treated all exchanged patients as being equally likely to be colonised, the risk of colonisation will differ depending the ward/unit they were admitted to and/or their underlying health status. Although this may alter the exact risk posed by exchanged patients, the risk posed by far away hospitals relative to neighbouring hospitals is unlikely to increase dramatically if patient-specific risks are taken into account given the large difference in patient flows within and between regions. Adjusting the screening policy to differentiate patients based on the ward of current admission might be feasible , however, it would be more difficult to rely on information about the visited ward during the previous hospital visit.