Web-Based Interactive Tool to Identify Facilities at Risk of Receiving Patients With Multidrug-Resistant Organisms

Rany Octaria; Allison Chan; Hannah Wolford; Rose Devasia; Troy D. Moon; Yuwei Zhu; Rachel B. Slayton; Marion A. Kainer


Emerging Infectious Diseases. 2020;26(9):2046-2053. 

In This Article


Patient Matching

We constructed interfacility patient-sharing networks from the Tennessee Hospital Discharge Data System (HDDS) inpatient admissions, and Centers for Medicare and Medicaid Services (CMS) claims and Minimum Data Set (MDS; https://www.cms.gov). The HDDS dataset included all inpatient admissions to Tennessee acute-care hospitals (ACHs) licensed by TDH; admission to LTCFs and Department of Veterans Affairs hospitals were not captured in this dataset. We used HDDS data to summarize patient-sharing data among Tennessee facilities, including ACHs critical access hospitals (CAH), long-term acute-care hospitals (LTACH), and inpatient rehabilitation facilities (IRFs) from January 2014–December 2017.

We linked admissions of each patient in the HDDS dataset with a multilevel matching process by using patient identifiers. First, we linked consecutive admissions for ≤365 days by matching the combination of date of birth, sex, and Social Security number (SSN). In this step, we considered admissions of the same person to be those that matched for date of birth, sex, and first and last name, even if the SSN was missing or had a 1-digit difference. Subsequently, we linked admissions that did not generate matches in the first step by matching the combination of date of birth, sex, and full name, even with ≥2-digit differences in the SSN. To protect patient privacy, patient-level admission data used for matching were saved in secured hard-drives that were connected to the computer only when generating facility-level data.

The CMS dataset included claims data and data from the MDS, which captured all inpatient admissions of CMS fee-for-service beneficiaries to Tennessee hospitals and SNFs. We used MDS admission and discharge assessments to identify all visits of Medicare beneficiaries to SNFs, a type of LTCF that is not as intensive as hospital but offers more intensive medical and nursing services, such as subacute care.[11] We combined MDS visits with CMS claims data that included admissions in all types of hospitals in the HDDS to create a more complete dataset of visits for Medicare beneficiaries. We linked admissions in MDS to patients by matching the CMS beneficiary identification number. We aggregated facilities by using the facilities' CMS certification number, which is different than facility aggregation in the HDDS dataset. The CDC modeling unit conducted aggregation by using the secure environment of the CMS Virtual Research Data Center before sharing the facility-level aggregate data to TDH.

Network Construction

Because of differences in aggregation, we constructed the CMS and HDDS networks separately. The CMS dataset aggregated facilities based on their CMS certification number, but HDDS aggregated based on the assigned Tennessee state licensing registration.

From each data source, we constructed 2 types of networks that connected healthcare facilities through uninterrupted patient sharing (UPS) and total patient sharing (TPS).[12] UPS, or direct transfers, connect a pair of facilities when a patient is discharged from 1 facility and admitted directly to another facility within 1 day. We accounted for patients who spent time in the community between healthcare admissions through the TPS network, which connects a pair of facilities through direct and indirect transfers. An indirect transfer occurs when a patient is discharged from 1 facility and readmitted to another facility within 2–365 days. The number of days between consecutive admissions was calculated by subtracting the next admission date and the current discharge date. We constructed subnetworks from the overall TPS network for 30 and 365 days in the community.

In our visualizations, each healthcare facility is represented by a node. A pair of nodes is connected by a line, also known as an edge in network analysis, weighted by the number of 1-way patient shares between pairs of facilities. For example, a patient discharged from a facility on September 30, 2015 and admitted to another on September 29, 2016, represents 1 indirect transfer. A patient can be represented by multiple edges in the same network. For example, if hospital A discharged Mr. X on January 30 and hospital B admitted Mr. X 2 weeks later, the TPS network graph would represent this connection as an edge with a weight of 1 going from node A to node B. If Mr. X is then admitted to SNF C 2 months later, this indirect transfer will be represented only as another edge from node B to C, but not A to C.

Network Analysis

We used an ego network design for the tool; this type of social network consists of a focal node (ego) and the nodes to which it is connected, directly or indirectly. In our tool, the facility of interest in each web session acts as the ego facility. We defined the following centrality measures for each ego facility by calendar year and by type of network: in-degree, out-degree, weighted in-degree, and weighted out-degree. We defined in-degree as the total number of facilities that sent transfers to a given facility and out-degree as the total number of facilities that received transfers from a given facility.[12] We defined weighted in-degree as the total number of patient transfers sent to a given facility and weighted out-degree as the total number of patient transfers sent from a given facility.[8] We used the Fruchterman-Reingold force-directed graph drawing algorithm to assign the relative positions of each facility in the network graph.[13] We accounted for several characteristics of the healthcare facility, including the type of facility and the Emergency Medical Services (EMS) region in which the facility is located (Figure 1). Tennessee EMS regions represent the referral patterns of the EMS services and hospitals, and the coordinating areas for emergency preparedness activities, which TDH uses to aggregate MDRO surveillance data.

Figure 1.

Emergency Medical Services (EMS) regions in Tennessee, USA: 1) Northeast; 2) East; 3) Southeast; 4) Upper Cumberland; 5) Mid-Cumberland; 6) South Central; 7) West; and 8) Memphis-Delta. The 8 EMS regions represent the referral patterns for EMS services and hospitals and for coordination for emergency preparedness activities. The Tennessee Department of Health uses EMS regions to aggregate multidrug-resistant organisms surveillance data. Stars indicate metropolitan areas within EMS regions.

At-risk facilities targeted in public health containment efforts can vary based on the circumstances of each outbreak. For our purposes, we defined at-risk facilities as downstream facilities that historically were identified to have received patients from the ego facility. At-risk facilities also were classified as the facilities receiving the most historical transfers from the ego facility if there were >10 downstream facilities. To evaluate the long-term stability of these identified at-risk facilities in the HDDS network, we evaluated the top downstream facilities of 5 randomly selected ego facilities across different EMS regions from 2014–2017. For each ego facility, we compared the 5 downstream facilities receiving the most transfers between pairs of consecutive years to quantify the aggregate percent change in the top 5 downstream facilities.

Web-based Application

We developed a password-protected web-based application using Shiny (R Studio Inc., https://www.rstudio.com) to enable public health personnel to access network visualizations and transfer statistics easily. Approved usernames and passwords are managed internally by TDH Healthcare Associated Infections and Antimicrobial Resistance (HAI/AR) program. Authorized users can access the Shiny web application to visualize the network of a facility of interest (ego) through a user-friendly interface at the website, https://tnhealthhai.shinyapps.io/patientsharing (Figure 2). The ego facility is the facility of interest that serves as the center of the visualized network for the current online session. Users can select from menus to tailor the displayed plot based on the data source, HDDS or MDS; year; length of interim time in the community; and the ego facility.

Figure 2.

Screenshot of the initial user interface and network graph visualization tab of the web-based application developed to identify healthcare facilities at risk of receiving patients with multidrug-resistant organisms. The application was designed using Shiny (R Studio Inc., https://www.rstudio.com). The network graph is visualized by using a force-directed layout. Black node in the center indicates the facility of interest (ego facility). Tennessee EMS regions are represented by the node color for connected facilities and is represented by the color of the node border for the ego facility. Users can change visualizations interactively during real-time use. EMS, Emergency Medical Services.

In the network plot, the node color represents Tennessee EMS regions, node size represents number of beds, and node shape represents facility type. The thickness of the edge is weighted on the number of 1-way transfers, including multiple transfers of 1 patient, between a facility pair. When users place the cursor over a node, the tooltip function displays the facility name, facility type, and number of beds. A slider widget enables users to set the lower threshold of 1-way transfers between each pair of facilities displayed for the session (Figure 2).

The Shiny application has 2 display tabs, plot and transfer statistics. The plot tab displays a visualization of the ego-network and all facilities that shared patients with the ego facility (Figure 3). When users hover the cursor over an edge, the application displays the number of 1-way transfers. Users can interact by applying filters for region or facility of interest, and by dragging the position of different nodes.

Figure 3.

Varying user-tailored ego network visualizations in the web-based interactive tool to identify facilities at risk of receiving patients with multidrug-resistant organisms. Panels demonstrate options for visualizations for a large academic hospital from the HDDS and Centers for Medicare and Medicaid claims and MDS. Real-time use of the application enables users to tailor visualizations by facility, patient transfer threshold, and type of network. Black node in the center indicates the facility of interest (ego facility). The EMS region is represented by the node color for connected facilities and is represented by the color of the node border for the ego facility. Displays shown use the HDDS dataset (A–C) and MDS dataset (D–F). Panels A and D demonstrate a total patient sharing network; B and E demonstrate an uninterrupted patient sharing network; C and F are examples of alterations in patient threshold transfers and displays facilities that have >50 patient transfers to or with the ego facility. EMS, Emergency Medical Services; MDS, minimum dataset; HDDS, Tennessee Hospital Discharge Data System.

The transfer statistics tab displays facility-level characteristics and facilities most at risk to receive transfers from the ego facility (Figure 4). It also lists the ego facility's type, city, EMS region, number of licensed beds, and centrality measures and displays a table of facilities at risk to receive transfers from the ego facility. The list defaults to a descending order of facilities by the number transfers from the ego facility. Users can filter or sort the table display based on facility name, facility type, number of beds, county, and EMS region. A download button allows users to import the table as comma-separated values, or as Microsoft Excel (https://www.microsoft.com) or portable document format (PDF) files.

Figure 4.

Screenshot of the transfer statistics tab of the web-based interactive tool to identify facilities at risk of receiving patients with multidrug-resistant organisms. This function displays facility characteristics and downstream facilities that are most likely to receive transfers from the ego facility. The second tab of the application's user interface includes 2 tables. The top table displays detailed social network and facility characteristics for the ego facility. The bottom table displays the facilities at highest risk to receive patients from the ego facility, which are downstream facilities. The table defaults to sort the number of patient transfers in descending order. Users can interactively choose a column to sort and filter this table, which can be used to identify facilities at risk during outbreaks or regional detection of a novel organism. Hospital names have been de-identified for privacy.


Data cleaning and person-matching were completed in SAS 9.4 (SAS Institute, Cary, NC, USA). We conducted network analyses by using the Statnet and network visualization by using visNetwork packages, both in Rstudio version 3.5.2 (R Studio Inc.).[14,15] As described previously, we developed the interactive web-based network visualization application by using Shiny. We uploaded de-identified facility-level datasets to the shinyapps.io server hosted on Amazon Web Services (Amazon, https://aws.amazon.com) infrastructure in the United States. These datasets had facility-level patient transfer statistics and characteristics, including licensed facility names, number of beds, facility type, and city and county of address.

Ethics Considerations

The patient-sharing network project was exempted from the institutional review boards (IRBs) at CDC (IRB no. 032416JO), TDH (IRB no. 923990–1), and Vanderbilt University (IRB no. 161676). This work was conducted under a data use agreement between CDC and CMS. CDC's Human Research Protection Office determined this project was exempt from regulations governing the protection of human subjects in research under 45 CFR 46.101(b).