Model structure

We developed a hierarchal metapopulation model ^{Reference Watts, Muhamad, Medina and Dodds12} of Maryland to capture the patient population hospital movement network. The model is comprised of a set of ordinary differential equations at 2 scales: (1) within a healthcare facility or community, and (2) movement between facilities and communities. Within this model, patients are assigned to 1 of 4 health states: susceptible (S), infected (I), or colonized with CRE (C), or susceptible but at increased risk of CRE colonization due to antibiotic use (X). ^{Reference Marchaim, Chopra and Bhargava13–Reference Falagas, Rafailidis and Kofteridis15} An additional state for each patient population was included to track patients with an electronic registry, which identified whether a patient has been previously identified as colonized or infected with CRE (Fig. 1 and Supplementary Fig. S1 online). All parameter values that correspond with biological values are listed in Supplementary Table S4 (online).

Fig. 1. Generalized schematic of the hierarchal metapopulation model. The compartmental state transition for each population is shown in more detail in the supplement. The diagram assumes there is an M number of long-term care facilities (LTCFs), N acute-care hospitals (ACHs), and P communities. The right-middle component in the diagram shows the regional flows of patients between the LTCFs, ACHs, and communities. The compartments for each population shown in the top, left-middle, and bottom components. There are 4 primary compartments in our model susceptible (S), higher susceptible (X), colonized (C), and infected (I). For patients that are identified with CRE, they are indicated with a hat, i.e., Ŝ, $\hat X$ , and Ĉ.

Data sources

To model patient flow, we used 2012–2016 patient-mix data from the Maryland Health Services Cost Review Commission (HSCRC), which included patient admissions and transfers for all ACHs in Maryland. Maryland hospitals must report detailed patient visit data to the Health Services Cost Review Commission to obtain reimbursement for care, including demographics, patient home ZIP code, source of admission, discharge location, and whether the patient spent time in an ICU. Data were anonymized; however, a unique patient identifier was assigned to track returning patients to be tracked over time and between hospitals. From these data, we developed a patient movement network that captured relative travel behavior between 51 ACHs (see Supplementary Table S1 online). We further extended this based on admission and discharge data to include movement between ACHs and long-term care facilities (LTCFs) and skilled nursing facilities (SNFs), communities, and patients from out of state. Based on data from Centers for Medicare and Medicaid Services (CMS) and HSCRC, we included 222 LTCFs and SNFs and determined their population size using the number of certified beds. ¹⁶ Communities were defined using the 464 ZIP codes in the state of Maryland, and population sizes were estimated based on data from the US Census Bureau. ¹⁷ Parameters were based on data from the literature as well as hospital-level data (Tables S1 and S2 online).

Interventions

Our primary aim was to explore the effect of implementing an electronic registry to facilitate targeted IPC implementation at ACH admission for all patients likely to be colonized (Table 1). We assumed patients that were under IPC measures had a 90% reduction in transmission compared to undetected CRE carriers. ^{Reference Lin, Lolans and Blom18} We compared the institution of an electronic registry with increased rates of culture-based screening for CRE. To our knowledge, only 1 ACH in Maryland has an established CRE screening program for patients entering select intensive care units (surgical and medical ICUs) and oncology units. Thus, we assumed in our simulated interventions that every other hospital would need to add a program to screen all patients upon admittance to the ICU and weekly thereafter. We considered 2 possible regional screening programs: (1) selective, in which an active CRE surveillance policy was implemented at the 5 ACHs with the highest connectivity to other hospitals, and (2) complete, in which active surveillance was implemented at all 46 ACHs with ICUs in Maryland. We assumed that IPC interventions were implemented for patients with positive results. As an alternative, we considered the implementation of a predictive screening program that assumed that machine-learning algorithms can identify patients at high-risk of CRE colonization at hospital admission with 80% positive predictive value. ^{Reference Goodman, Simner and Klein19}

Table 1. Summary of Scenarios and Interventions ^a

Note. ACH, acute-care hospital; ICU, intensive care unit; CRE, carbapenem-resistant Enterobacteriaceae; IPC, infection prevention and control. “Yes” in the center grid indicates whether the intervention is implemented for that specific scenario.

^a All scenarios (top) in our simulation with corresponding interventions are shown (left).

Sensitivity analysis

Given the structural assumptions of the model, we analyzed the main factors driving variation in the rates of CRE colonization and incidence using Latin Hypercube Sampling (LHS). ^{Reference McKay, Beckman and Conover20,Reference McKay21} Studies have shown that LHS is significantly more efficient than simple random and fractional stratified sampling designs (see Blower and Dowlatabadi ^{Reference Blower and Dowlatabadi22} ) and is a method commonly used to evaluate models within the field of epidemiology. ^{Reference Blower and Dowlatabadi22–Reference Marino, Hogue, Ray and Kirschner26} For each scenario, we sampled the parameter space over 300 simulation runs. We sampled outcomes between scenarios using the Welch 2-sample t test, and P values were reported. Additionally, we conducted a sensitivity analysis regarding the effectiveness of the IPC bundle to reduce transmission (Appendix G online).

Cost-effective analysis

The average cost for a single CRE surveillance screening was estimated to be $8.65, which is calculated from total costs of swabbing, culturing, conducting organism identification test, antimicrobial susceptibility testing, phenotypic testing, and molecular analysis (see Appendix F online for detailed cost breakdown).

For patients identified as positive, either through culture screening or by the registry, we assumed that the hospital would implement a bundled IPC intervention, which was estimated to cost $639.48 per CRE patient (Appendix F online). Costs were based on a review of the literature and included the cost of placing a patient on contact precautions and the cost of implementing daily chlorhexidine (CHG) bathing for decolonization. Contact precautions were defined as the use of personal protective equipment, including disposable gloves and gowns, which require 1 minute to don and doff per contact. All costs included both material and staffing costs (see Appendix F online). Because any surveillance method trades off between sensitivity and specificity, we assumed that with any method, some false positives resulted in erroneous utilization of the bundled IPC intervention on uncolonized patients, which increased costs with no additional benefit. For the implementation of the electronic registry, costs of modifications to each hospital were based on internal hospital estimates ($10,000 per hospital) so real-time connections to a state server could be made. No additional costs were assumed for the implementation of a predictive screening algorithm.

To determine the cost-effectiveness of each scenario, we calculated the net cost to deliver each intervention scenario, adjusted for the cost savings due to averted infections. The average cost per CRE infection was estimated to be $30,484. ^{Reference Almario, May and Shaheen27–Reference Cosgrove, Kaye, Eliopoulous and Carmeli29} The incremental cost-effectiveness ratio (ICER) was calculated by dividing the average difference in net costs between 2 intervention scenarios by the average difference in their effect (eg, number of infections averted by each scenario). Costs and benefits were measured per simulation, and averages and 95% uncertainty Intervals were calculated from the resulting outcomes.