This study used observational data collected as part of systematic routine surveillance procedures in a university hospital ward, with an endemic level of multidrug-resistant bacteria. This surveillance protocol followed the official recommendations of the French Ministry of Health and the French Society for Hygiene (http://sante.gouv.fr/les-infections-nosocomiales-recommandations-aux-etablissements-de-soins.html) and was approved by the Nosocomial Infections Fighting Committee. All the patients’ parents received general information about the hospital infection control strategy at admission. No more information than those collected by routine procedures was used; in particular, no additional individual data, biological collection or sample was required. Therefore, an ethics committee approval was not required for this study.

Necker Enfants–Malades is a 650-bed tertiary-care teaching hospital that handles 55 000 admissions per year. The pediatrics department includes a 21-bed unit that admits 300–350 children annually. Approximately half of the children are referred from other hospitals, 20% from other units in the hospital, and the remaining 30% from the emergency department. In May 2009, a whole-ward screening revealed an unusually high ESBL-E prevalence among patients, which prompted the subsequent interventions, beginning in June 2009 and described below. All children admitted from 1 June 2009 through 15 July 2010 were included in our study. All episodes of ESBL-E colonization or infection diagnosed during the stay or up to 2 days after discharge from the pediatric ward were included. During the study period, a rectal swab specimen was obtained at admission and once a week throughout each child’s stay. Swabs were plated on a selective chromogenic medium for ESBL screening (chromID ESBL Agar, bioMérieux, Marcy-l’Etoile, France). Enterobacteriaceae were isolated and identified according to the recommendations of the Comité de l’Antibiogramme de la Société Française de Microbiologie. ESBL production was evaluated with the double-disc synergy test and the Etest (AB Biodisk, Solna, Sweden) for ceftazidime and ceftazidime–clavulanate. Patients with ESBL-E isolated within 48 hours following admission from screening samples or from an infected site were considered imported. Other cases were considered acquired.

From June 2009 through February 2010 (first period, P₁), the following baseline infection-control practices were in place in the pediatrics ward. All children with ESBL-E–positivity in clinical or screening specimens were placed in single rooms and in contact isolation, within 24 h following test results. Briefly, these measures consisted of flagging microbiological reports, charts and doors of rooms of ESBL-E–positive patients with a warning symbol; wearing gowns and gloves when caring for these patients; emphasizing hand hygiene before and after patient contact; and notifying ESBL-E carriage when patients were transferred to another unit. Contact isolation was maintained throughout the entire duration of hospitalization. During the second period (P₂), from February 2010 to July 2010, a cohorting protocol was added to the aforementioned measures. All known ESBL-E carriers were moved and grouped at one end of the pediatrics ward in an 8-bed unit, and cared for by a dedicated nursing team. The isolation policy was similar to P₁, except that previously known carriers (i.e. patients with an episode of ESBL-E carriage in the past 6 months) were placed in isolation immediately upon admission.

We assessed the effectiveness of isolation measures by means of a stochastic, population-based transmission model, building on previously described models 17 18 . Patients can be in four mutually exclusive states, depending on their colonization and isolation status (Figure 1). Patients can be susceptible (S) or colonized, with the latter subdivided as isolated (I) or unisolated, distinguished as imported cases (patients colonized at admission C ₁ ) or acquired cases (patients acquiring ESBL-E in the unit, C ₂ ). Susceptible patients are at risk of acquiring ESBL-E at a rate λ = β ₀ + β ₁(C ₁ + C ₂) + β ₂ I, where β ₁ and β ₂ are the transmission rates from unisolated and isolated ESBL-E carriers, respectively, and β ₀ represents the non-cross–transmission acquisition rate. Although it is common practice to include both cross-transmission and non-cross–transmission terms in transmission models, their exact interpretation is subtle. In our model, these quantities are just two components of the attack rate: β ₁ and β ₂ scale that component proportional to the number of carriers (unisolated and isolated, respectively), while β ₀ quantifies the component independent of the number of carriers. With respect to ESBL-E dynamics, the cross-transmission characteristic is not whether it is transmitted through the contaminated hands of healthcare workers, environmental contamination, fomites etc., but whether its rate depends on previous numbers of carriers in the ward (the so-called colonization pressure 19 ). In contrast, sporadic acquisitions are decoupled from the previous ESBL-E dynamics in the ward. Within hospital settings, these acquisitions can include endogenous selection of a previously undetected colonizing strain after antibiotic therapy, which can reasonably be considered to be independent of ESBL-E prevalence in the ward 20 21 . Below, we refer to such acquisitions as sporadic, quantified by the sporadic acquisition rate β ₀. Cross-transmission is quantified by the transmission rates β ₁ and β ₂, and, because isolated ESBL-E–positive patients are managed under special contact precautions, we expect β ₁>β ₂.

Imported cases are isolated at rate δ ₁, and δ ₂ for acquired cases, with δ ₁>δ ₂. When the isolation ward reaches full capacity, the possibility of isolating a patient no longer exists, until an isolated patient is discharged, i.e. the isolation rate is Δ(C _*,I)=δ _* C _* if I < N _I, 0 otherwise. Because the duration of ESBL-E carriage (estimated at 6 months 22 ) typically exceeds hospital lengths of stay, we did not account for possible carriage clearance during the stay. The preemptive isolation of previously known carriers in P₂ is modeled as an additional parameter ∈ , representing the proportion of colonized inpatients immediately placed in isolation at admission.

The parameters used in the model are defined in Table 1. Some of them could be computed directly from the dataset. Discharge rates were computed from the patients’ observed lengths of stay recorded in the unit, assuming that all unisolated patients had the same risk of being discharged. Isolation rates were computed as the reciprocal of the mean time to isolation. For imported cases, this time was 2 days during P₁, and 0 days during P₂, i.e. we assumed that all colonized inpatients were put in preemptive isolation for this period. For simplicity, we also assumed that all patients put in preemptive isolation during P₂ were indeed carriers. For acquired cases, because the exact time from acquisition to isolation was unknown, it was assumed to be 4 days. Compared to single-bed isolation rooms, an isolation ward has a clear face validity, meaning that a well-implemented isolation ward with designated staff will effectively prevent all transmission from cohorted patients 23 . However, that assumption does not hold for weaker types of isolation, including single-room isolation, whose effectiveness is a priori unclear 18 . Therefore, we set β ₂ = 0 for P₂, but estimated the P₁ value. Other acquisition parameters, namely β ₀ and β ₁, and the admission prevalence σ were estimated from data for both periods.

Parameters in boldface were estimated from data using the iterated filtering algorithm, as explained in the main text. For these parameters, the maximum likelihood estimates (95% confidence intervals) are indicated. Rates are day^–1.

The model was implemented in the pomp package 24 , operating in the R environment 25 . Stochastic simulations of the model were performed using Gillespie’s exact algorithm 26 . Parameters were estimated with the iterated filtering algorithm, described elsewhere 27 . Profile likelihoods were used to derive 95% confidence intervals (CI) 28 . Technical details about model implementation and estimation are given in the electronic supplementary material (Additional file 1); the estimation procedure was also verified by using simulated data (Additional file 2).

Model fitting was assessed by visual inspection for both weekly point prevalence (i.e. point prevalence on Monday) and weekly incidence data. 10 000 model simulations were used to derive mean values and 95% prediction intervals, and were compared to observed P₁ and P₂ data.

P₁ and P₂, respectively, consisted of 4165 and 2305 patient-days (pt-d). The incidence of ESBL-E acquisition was 7.0 per 1 000 pt-d during P₁ and decreased to 4.3 per 1 000 pt-d during P₂. A similar trend was observed for ESBL-E prevalence, from 26% during P₁ to 21% during P₂ (Table 2). Figure 2 reports the time-series for point prevalence and incidence data. Although estimates for P₁ were consistent with higher transmission from unisolated patients (β ₁ > β ₂, Table 1), the difference was not statistically significant compared to isolated patients (95% CI for β ₁ − β ₂, -0.005–0.018 per day). Moreover, only the sporadic acquisition rate was statistically significant, while transmission rates were not (likelihood ratio tests, null hypotheses β ₁ = 0 and β ₂ = 0, p-values 0.16 and 0.35, respectively). Admission prevalence was estimated at 0.18. Model assessment suggested very good fit to point prevalence and incidence data (Figure 3). The fitted model predicted a total of 30 (20–42) acquisitions, i.e. a 7.2 (4.8–10.1)/1 000 pt-d incidence, compared with 29 observed acquisitions, i.e. 7.0/1 000 pt-d incidence, corresponding to a predicted 3.4/1 000 pt-d incidence due to sporadic sources, 2.6/1 000 pt-d to transmission from unisolated carriers and 1.2/1 000 pt-d from isolated carriers. Therefore, sporadic acquisitions accounted for nearly 50% of new cases in the ward during P₁.

The prevalence was computed as the ratio of the total number of colonized patient-days to the total number of patient-days for each period. The incidence was computed as the ratio of the number of acquisitions to the total number of patient-days for each period. The staff-to-patient ratio was computed as the daily ratio of the number of staff members (nurses and ward assistants) to the number of patients. pt-d: patient-days.

During P₂, the isolation ward had little impact on acquisition rates: the sporadic acquisition rate remained the only statistically significant source of acquisitions in the ward, while transmission from unisolated patients remained unchanged and insignificant (likelihood ratio test, null hypothesis β ₁ = 0, p-value 0.2). Admission prevalence was estimated at 0.15 for this period. Again, model fitting was adequate, with a total of 10 (4–18) predicted acquisitions, a 4.2 (1.7–7.6)/1 000 pt-d incidence, compared to 10 observed acquisitions (Figure 3), corresponding to a predicted 3.4/1 000 pt-d incidence resulting from sporadic sources and 0.8/1 000 pt-d a consequence of transmission from unisolated carriers, while the incidence due to cohorted patients was hypothetically set at 0. Crucially, the incidence resulting from sporadic acquisitions was unaffected by the isolation-ward implementation.

Although we estimated parameters separately for P₁ and P₂, estimations of acquisition parameters for the entire period are relevant. Doing so, β ₁ was estimated at 0.006 (0–0.015) per day and β ₀ at 0.008 (0.002–0.015) per day. Thus, transmission from unisolated carriers remained statistically insignificant but a trend was observed (likelihood ratio test, null hypothesis β ₁ = 0, p-value 0.08).

Because the preintervention period was not observed, a baseline incidence could not be computed to compare the relative efficacies of the two interventions. Furthermore, the isolation-policy change for imported cases during P₂ (preemptive isolation of previously known carriers) might obfuscate the precise contribution of the isolation ward to the observed incidence decline. Therefore, in addition to the observed interventions during P₁ and P₂, we used model simulations to predict the impact of two hypothetical situations: no isolation whatsoever, and an isolation ward without preemptive isolation, as summarized in Table 3. In the absence of isolation measures (β ₂ = β ₁), the predicted incidence was 14.6/1 000 pt-d, a two-fold increase compared to P₁. Had the isolation ward been implemented without preemptive isolation of known carriers, the predicted incidence was 5.9/1 000 pt-d, 16% lower than the observed incidence during P₁. Overall, these simulations suggested that single-room isolation effectively lowered ESBL-E incidence and that the isolation ward would have had little benefit compared to single-room isolation, had the isolation policy been similar. However, broadly overlapping prediction intervals, reflecting parameter-estimate uncertainty (particularly the insignificance of transmission parameters), preclude a definitive conclusion regarding interventions from these model simulations.

Predicted intervals were based on 1000 model simulations. Dashes indicate hypothetical interventions that were only simulated, and, therefore, not observed. pt-d: patient-days.

We attempted to investigate the expected impact of contact isolation (aimed at reducing patient-to-patient transmission) in combination with measures targeting other acquisition sources (e.g. antimicrobial stewardship program to reduce endogenous acquisition). We used incidence of colonization per 1 000 pt-d as the outcome criterion and simulated the expected impact of contact isolation for various levels of isolation effectiveness (0–100% effective) and different values for sporadic acquisition (Figure 4). When sporadic acquisition sources are high, the model predicted that even substantial efforts to interrupt transmission from isolated patients would have limited impact on lowering the ESBL-E incidence. For example, for β ₀ = 0.02 acquisitions per susceptible patient per day, no contact isolation resulted in an 18/1 000 pt-d incidence, whereas completely effective contact isolation yielded a 10/1 000 pt-d incidence, i.e. a 44% incidence reduction. Conversely, when acting simultaneously on sporadic sources of acquisition and cross-transmission, a better control would be expected. Assuming that sporadic sources are fully contained, totally effective contact isolation would be expected to reach a 2/1 000 pt-d incidence, compared to 11/1 000 pt-d without any contact isolation, an 82% decline.

We investigated the impact of screening and isolating ESBL-E-positive patients at admission while varying the admission prevalence from 0 to 20%, and assuming that cross-transmission occurs in the ward (Figure 5). When the admission prevalence was low, screening patients at admission had little or no benefit, regardless of the isolation rate. For rising admission prevalences, a higher degree of control could be achieved by screening and isolating colonized inpatients more-and-more rapidly. For example, for an admission prevalence of 20%, the predicted incidence was 11/1 000 pt-d without any screening and could be lowered to 6–7/1 000 pt-d when colonized inpatients were detected and isolated within 2 days.

For acquired cases, isolation was assumed to occur 4 days after acquisition. However, because the exact acquisition date was unknown, this value was uncertain. Likewise, the value of direct-isolation proportion during P₂ was uncertain, because previously unknown carriers colonized at admission could have been missed. Therefore, we conducted sensitivity analyses for these two parameters and found that even large variations of their values had little impact on our estimates (Additional file 1).