Social Network Analysis of Patient Movement Across Health Care Entities in Orange County, Florida

Abstract

Objective

Multidrug-resistant organisms (MDROs) are continually emerging and threatening health care systems. Little attention has been paid to the effect of patient transfers on MDRO dissemination among health care entities in health care systems. In this study, the Florida Department of Health in Orange County (DOH-Orange) developed a baseline social network analysis of patient movement across health care entities in Orange County, Florida, and regionally, within 6 surrounding counties in Central Florida.

Materials and Methods

DOH-Orange constructed 2 directed network sociograms—graphic visualizations that show the direction of relationships (ie, county and regional)—by using 2016 health insurance data from the Centers for Medicare & Medicaid Services, which include metrics that could be useful for local public health interventions, such as MDRO outbreaks.

Results

We found that both our county and regional networks were sparse and centralized. The county-level network showed that acute-care hospitals had the highest influence on controlling the flow of patients between health care entities that would otherwise not be connected. The regional-level network showed that post–acute-care hospitals and other facilities (behavioral hospitals and mental health/substance abuse facilities) served as the primary controls for flow of patients between health care entities. The most prominent health care entities in both networks were the same 2 acute-care hospitals.

Practice Implications

Social network analysis can help local public health officials respond to MDRO outbreak investigations by determining which health care facilities are the main contributors of dissemination of MDROs or are at high risk of receiving patients with MDROs. This information can help epidemiologists prioritize prevention efforts and develop county- or regional-specific interventions to control and halt MDRO transmission across a health care network.

The field of public health is continuously evolving with new technologies, changing health care systems, globalization, and broader societal transformations.^1,2 In response, public health officials are using numerous tools to analyze and visualize complex data to answer important health questions (eg, about disease propagation) to protect the health of the community.^3,4 Social network analysis is an emerging analytical tool in public health that describes the connections of persons, partnerships, disease transmission, and interorganizational structure of health systems.⁴ Social network analyses visually describe the existence and strengths of relationships of participants in a network. These visualizations are depicted through sociograms (ie, drawings of points connected by lines).⁵ The points in the network (nodes) represent the persons or organizations of interest. The lines that connect nodes (edges) represent the relationship between nodes. In a directed sociogram, arrows are also used to connect nodes and show direction from a source node to a receiving node.

In applied public health, social network analyses have been used to understand disease transmission for HIV/AIDS, sexually transmitted infections, tuberculosis, and severe acute respiratory syndrome.⁴ Other approaches in applied public health have examined health communication, health behaviors, partnerships, and community collaboration.^4,6-10 Social network analyses allow public health officials to assess the relationships of interest through strengths of connections rather than common variables (eg, type, location, size, population served). However, social network analyses have not traditionally been used by public health officials at local health departments. Explanations for the scarcity of application are not fully understood, but on the basis of our experience, we posit that training and resources for local public health officials are lacking.

Multidrug-resistant organisms (MDROs) are continually emerging and threatening health care systems by increasing health care–associated morbidity, mortality, and costs.^11,12 Efforts to control transmission of MDROs are a top priority for public health officials and are directed toward enhancing infection control practices within a health care entity. Although reinforcing infection control practices within a health care entity has been successful in halting the transmission of MDROs, little attention has been paid to understanding the effect of patient transfers on MDRO dissemination in a county’s health care infrastructure.¹²

Understanding that patient transfers play an integral role in the health care infrastructure, we saw a meaningful application of social network analyses to local health department operations.¹³ Health care entities in the same region or county are often connected through the patients they share.¹⁰ Patients are transferred among health care entities for various reasons, such as health insurance providers, specialty medical procedures, referrals from physicians, or improved social support.^10,13-15 In this study, the Florida Department of Health in Orange County developed a baseline social network analysis of patient movement across health care entities countywide and regionally to learn how each health care entity contributes to the health care infrastructure. Central Florida’s health care system is complex because it services an estimated 1.35 million persons and approximately 72 million tourists annually.¹⁶ Understanding how each health care entity contributes to the health care infrastructure can assist us in guiding targeted MDRO intervention strategies and identifying facilities at risk for receiving or discharging a patient with an MDRO.

Materials and Methods

Data Collection

We received data on health care transfers for Orange County from the 2016 100% Medicare claims and Minimum Data Set files, provided as part of a collaboration with the Centers for Disease Control and Prevention (CDC) Division of Healthcare Quality Promotion. This work was conducted under a data-use agreement between CDC and Centers for Medicare & Medicaid Services (CMS) and was determined by CDC’s Human Research Protection Office to be exempt from the regulations governing the protection of human subjects in research.

The data set includes health care entity attribute variables, which distinguish the location of each facility (eg, entity CMS identification number, entity state location, entity county location, entity name, national provider identifier [NPI], CMS provider identification, and postal ZIP codes of service). In addition, the data set identifies patient transfers between health care facilities through the source and destination provider identification, NPI (ie, source and destination), postal ZIP code (ie, source and destination), and number of Medicare fee-for-service beneficiaries (patients) shared between the discharge facility and receiving facility. Adhering to the CDC Division of Healthcare Quality Promotion’s data-use agreement, the data set is coded with “–9” if the facility transferred or discharged ≤10 patients.

We used 2 data categories, provider ID and postal ZIP code, to fully identify the health care entities. Once we identified the entities, we were able to obtain data on additional descriptive statistics and understand the magnitude of the social connections/network. To maintain the anonymity of the entities, we labeled each with the letter “X” and a corresponding number as they appeared in the data set (ie, entity 1 is labeled X1). The health care entities represented in the data set comprised acute-care hospitals, post–acute-care hospitals (ie, long-term acute-care hospitals and rehabilitation hospitals), skilled nursing facilities, and other facilities (ie, mental health and behavioral hospitals, substance abuse rehabilitation facilities, and mental health facilities).

Network Analysis

We managed data by using R-dplyr version 0.7.5.¹⁷ Network visualizations and statistics were conducted using RStudio package R-statnet version 2016.9.^18,19 In 2018, we constructed 2 directed networks to understand county and regional influences in Orange County’s health care system. The county sociogram includes all CMS-certified health care entities (ie, acute-care hospitals, post–acute-care hospitals, skilled nursing facilities, and other) located in Orange County, Florida. To understand the role of surrounding county entities on our county network, we constructed a regional sociogram. The regional network includes all the CMS-certified health care entities located in Orange County plus entities that are geographically located in the 6 surrounding counties (Brevard, Osceola, Lake, Polk, Seminole, and Volusia). Patients shared in the networks were transferred from 1 to 365 days of the health insurance claim.

We created network visualizations by using the Fruchterman-Reingold 1991 layout (ie, force-directed graph drawing algorithm).²⁰ Each health care entity is represented by a node in both networks. The health care entities (nodes) are categorized by facility type (ie, acute-care hospitals, post–acute-care hospitals, skilled nursing facilities, and other) and have weighted edges depicting the volume of patients transferred. To control for fluctuation of patient volume across entities that transferred ≤10 patients, we valued the weighted edges as the midpoint (weight = 5). Intrafacility transfers were represented by a weighted edge from the node back into itself (self-loop).

In this study, we primarily focused on metrics that may be most useful for targeted MDRO reduction interventions or active case finding during an MDRO outbreak. Information flow depends on individual and group network properties. Thus, we calculated cohesion²¹ (ie, density), magnitude of influence²¹ (ie, centralization), and central importance of each entity²¹ (ie, centrality) for each network and the individual node. Measures of centralization include the extent to which 1 or a few health care entities influence the network by serving as a connector between 2 health care entities²¹ (ie, betweenness) and the extent to which 1 or a few health care entities influence the network by having the most connections²¹ (ie, degree). Measures of centrality include degree (total number of connections,²¹ indegree (ie, number of connections received²¹/outdegree (ie, number of connections sent/discharged,²¹ and flow betweenness (ie, betweenness that incorporates all possible pathways weighted with patients). In addition, we calculated node-based measures for the edge metrics to understand the effect each entity contributed toward the health care infrastructure. These descriptive statistics included degree, betweenness, thickness (ie, weighted), and direction (ie, indegree/outdegree). Nodes are described by the facility type, the total number of beds in the health care entity, and the total number of transfers. Density and centrality network measures have been identified as potentially the most informative network measures in the examination of public health systems.^22,23 These measures provide us with the baseline understanding of the influences of each health care entity in our system.

Results

The Orange County network consisted of 46 health care entities (Figure), and the regional network consisted of 82 health care entities (Figure). Of the 46 health care entities in the Orange County network, 8 were acute-care hospitals, 2 were post–acute-care hospitals, 33 were skilled nursing facilities, and 3 were other facilities. The Orange County network was sparse, with a density of 0.28 and a reciprocity of r = 74 (edgewise reciprocity; Table 1), indicating that there is a high proportion of edges reciprocated (ie, mutual sharing of patients). Centralization by indegree, outdegree, and betweenness revealed that the Orange County network was centralized (indegree = 0.78, outdegree = 0.78, betweenness = 0.86). The regional network consisted of 27 acute-care hospitals, 3 post–acute-care hospitals, 49 skilled nursing facilities, and 3 other facilities. Similar to the Orange County network, the regional network was sparse (density = 0.11), with a high proportion of edges reciprocated (edgewise reciprocity r = 74%; Table 1). The regional network was centralized (indegree = 0.79, outdegree = 0.88, betweenness = 0.79) by 1 health care entity. Neither network displayed fragmentation, because no isolates or components (health care entities not connected to the network) were identified.

Figure.

Directed social networks of patient transfers through (A) County (Orange County, Florida) and (B) Regional (central Florida) Centers for Medicare & Medicaid Services–accepting health care entities, 2016. Each node (circle) depicts a health care entity; the size of the node depicts the total degree, defined as the overall volume of connections that exist for each health care entity (indegree + outdegree)^0.3, edges (lines) are weighted with a log reduction (divided by 50) of the volume of patients shared between facilities, arrows depict the direction of patients being shared, curved lines depict loops (an entity shared patients with itself), and length of lines depict the speed of patient movement. Administrative data from the 2016 100% Medicare claims and Minimum Data Set files, provided as part of a collaboration with the Centers for Disease Control and Prevention Division of Healthcare Quality Promotion.

Table 1.

Directed sociogram attributes of patient transfers through CMS^a–accepting health care facilities, stratified by county level (Orange County, Florida) and regional level (Central Florida), November 2016

Attribute	County	Regional
Network size,b no.	46	82
Total edges,c no.	600	736
Dyad count,d no.	2116	6724
Densitye	0.28	0.11
Reciprocity “weighted edges,”f %	74	74
Indegree centralizationg	0.78	0.79
Outdegreeh	0.78	0.88
Betweennessi	0.86	0.79

Abbreviation: CMS, Centers for Medicare & Medicaid Services.

^aAdministrative data from the 2016 100% Medicare claims and Minimum Data Set files, provided as part of a collaboration with the Centers for Disease Control and Prevention Division of Healthcare Quality Promotion.

^bNetwork size is the total number of nodes (circles) in the sociogram(s) (network visualization).

^cTotal edges are the number of lines that connect nodes to each other in the sociogram(s).

^dDyads are the total number of node pairs that exist in the network.

^eDensity is the ratio of observable edges (physical lines that are observed connecting nodes in the network) to potential edges (total number of connections that could have connected nodes together) in a network.¹⁹

^fReciprocity is the proportion of dyad(s) (pairs of health care entities) that are mutually linked.¹⁹

^gIndegree centralization is the extent to which 1 or a few health care entities influence the network by receiving a substantially higher amount of connections from other health care entities within the network.¹⁹

^hOutdegree centralization is the extent to which 1 or a few health care entities influence the network by sending a substantially higher amount of connections to other health care entities within the network.¹⁹

ⁱBetweenness is the extent to which 1 or a few health care entities influence the network by creating a bridge between health care entities within the network.¹⁹

Node-level descriptive statistics for Orange County health care entities are described with a labeling scheme (Table 2). The number of beds at acute-care hospitals in the Orange County network ranged from 100 to 1366 (median = 266), and the number of patients shared across the network ranged from 15 (entity no. X26) to 6538 (entity no. X13) (median = 2025). Post–acute-care hospitals had a range of 35-40 beds (median = 38) and shared a range of patients from 188 (entity no. X38) to 240 (entity no. X39) (median = 214) across the network. The number of beds at skilled nursing facilities ranged from 39 to 420 (median = 120), and the total number of patients shared across the network ranged from 46 (entity no. X22) to 886 (entity no. X20) (median = 219). Other facilities had a range of 90-174 beds (median = 112) and shared a range of patients from 106 (entity no. X44) to 309 (entity no. X3) (median = 259) across the network.

Table 2.

Characteristics of facilities in a county-level network of patient transfers through CMS^a–accepting health care facilities, Orange County, Florida, November 2016

Facility Typeb	No. of Beds	Total No. of Patients Shared	Indegree Centralityc	Outdegree Centralityd	Flow Betweenness Centralitye
Acute-care hospital
X11	120	329	11	14	299
X12	295	2549	31	37	1241
X13	1366	6538	48	48	1980
X14	320	2588	37	36	1440
X16	211	1500	27	31	1022
X26	100	15	2	2	0
X29	237	1311	27	37	1038
X31	866	2900	42	45	1728
Other
X3	90	309	6	9	140
X5	174	259	11	10	257
X44	112	106	8	6	149
Post–acute-care hospital
X38	35	188	8	19	335
X39	40	240	6	20	290
Skilled nursing facility
X1	103	152	9	8	108
X2	120	411	12	13	261
X4	118	152	8	6	51
X6	180	139	10	7	142
X7	120	719	13	14	229
X8	120	113	7	5	104
X9	60	194	12	11	245
X10	120	499	10	13	200
X15	120	140	10	8	142
X17	228	226	9	7	136
X18	116	219	11	7	157
X19	120	509	11	11	209
X20	120	886	11	12	218
X21	138	235	8	11	248
X22	40	46	4	4	90
X23	60	310	12	8	202
X24	120	211	13	9	199
X25	120	110	8	7	89
X27	120	179	9	10	117
X28	420	261	12	10	152
X30	168	399	14	10	199
X32	120	330	13	7	149
X33	120	81	9	6	72
X34	120	367	12	10	209
X35	180	237	9	10	174
X36	120	133	9	10	138
X37	39	98	5	4	50
X40	120	300	10	8	157
X41	115	86	6	6	45
X42	120	532	25	17	291
X43	180	170	9	6	58
X45	120	146	7	4	39
X46	80	247	9	7	78

Abbreviation: CMS, Centers for Medicare & Medicaid Services.

^aAdministrative data from the 2016 100% Medicare claims and Minimum Data Set files, provided as part of a collaboration with the Centers for Disease Control and Prevention Division of Healthcare Quality Promotion.

^bThe health care entities represented in the data set comprised acute-care hospitals, post–acute-care hospitals (ie, long-term acute-care hospitals and rehabilitation hospitals), skilled nursing facilities, and other facilities (ie, mental health and behavioral hospitals, substance abuse rehabilitation facilities, and mental health facilities).

^cIndegree centrality is the number of connections a health care entity receives from other health care entities across the network.

^dOutdegree centrality is the number of connections a health care entity sends to other health care entities across the network.

^eFlow betweenness centrality is the frequency with which a health care entity serves as a connector within the network that incorporates all possible pathways weighted with patients.

Centrality measures further identify the heterogeneity across health care entities (Table 3). In Orange County, at acute-care hospitals, the weighted indegree (ie, patients being admitted) ranged from 10 (entity no. X26) to 4403 (entity no. X13) (median = 1202), and the weighted outdegree (ie, patients being discharged) ranged from 10 (entity no. X26) to 4947 (entity no. X13) (median = 1384). The median number of patients discharged from post–acute-care hospitals (median weighted outdegree = 121) was higher than the median number of patients admitted into post–acute-care hospitals (median weighted indegree = 96), whereas skilled nursing facilities in the Orange County network admitted more patients (median weighted indegree = 123) than they discharged (median weighted outdegree = 83). Other health care entities in the Orange County network had a median weighted indegree and outdegree of 173 (Table 3). In addition, acute-care hospitals played an integral role in controlling the flow of patients between health care entities that would not otherwise be connected (flow betweenness centrality = 0 [entity no. X26] – 1980 [entity no. X13]).

Table 3.

County-level (Orange County, Florida) and regional (Central Florida) sociogram of CMS^a–accepting facility-level centrality measures, 2016

Facility Typeb	Centrality Measures
	Median Indegree Centrality (Range)	Median Weighted Indegreec (Range)	Median Outdegree Centrality (Range)	Median Weighted Outdegreed (Range)	Median Flow Betweenness Centralitye (Range)
County (n = 46)
Acute-care hospital (n = 8)	29 (2-48)	1202 (10-4403)	37 (2-48)	1384 (10-4947)	1140 (0-1980)
Post–acute-care hospital (n = 2)	7 (6-8)	96 (71-121)	20 (19-20)	121 (117-124)	313 (290-335)
Skilled nursing facility (n = 33)	10 (4-25)	123 (26-550)	8 (4-17)	83 (20-348)	149 (39-291)
Other (n = 3)	8 (6-11)	173 (66-206)	9 (6-10)	173 (55-236)	149 (140-257)
Regional (n = 82)
Acute-care hospital (n = 27)	2 (0-73)	73 (0-5608)	3 (1-80)	115 (10-6259)	30 (0-5393)
Post–acute-care hospital (n = 3)	8 (4-10)	146 (111-311)	19 (4-20)	124 (117-246)	396 (24-413)
Skilled nursing facility (n = 49)	9 (1-25)	95 (11-574)	7 (0-17)	73 (0-359)	103 (0-466)
Other (n =3)	9 (7-12)	188 (77-224)	10 (6-11)	189 (55-252)	238 (195-363)

Abbreviation: CMS, Centers for Medicare & Medicaid Services.

^aAdministrative data from the 2016 100% Medicare claims and Minimum Data Set files, provided as part of a collaboration with the Centers for Disease Control and Prevention Division of Healthcare Quality Promotion.

^bThe health care entities represented in the data set comprised acute-care hospitals, post–acute-care hospitals (ie, long-term acute-care hospitals and rehabilitation hospitals), skilled nursing facilities, and other facilities (ie, mental health and behavioral hospitals, substance abuse rehabilitation facilities, and mental health facilities).

^cWeighted indegree is the volume of patients a given health care entity received from other health care entities in the network.

^dWeighted outdegree is the volume of patients a given health care entity sent to other health care entities in the network.

^eFlow betweenness centrality is the frequency with which a health care entity serves as a connector within the network that incorporates all possible pathways weighted with patients.

In the regional network, of the types of entities, acute-care hospitals received and transferred the highest number of patients (weighted indegree range = 0-5608; weighted outdegree range = 10-6259; Table 3). However, the other facilities group (median weighted indegree = 188; median weighted outdegree = 189) and post–acute-care hospitals (median weighted indegree = 146; median weighted outdegree = 124) consistently shared patients across the network. Also, in the regional network, post–acute-care hospitals served as the primary controls in the health care system (flow betweenness centrality median = 396).

Discussion

Social network analyses provide a flexible framework for analyzing the association of health care entities in a community. The network visualizations allow public health officials to identify critical stakeholders and facilitate discussions to implement targeted interventions.^10,23 We developed and analyzed 2 directed networks (ie, county and regional) to understand the actors and influences in our health care infrastructure. Our analysis demonstrated a centralized network (indegree = 0.78, outdegree = 0.78, betweenness = 0.86) in Orange County, with 6 acute-care hospitals (entities no. X12-X14, X16, X29, X31) serving as the most prominent facilities for controlling the flow of patients. When we expanded the analysis regionally, the centralization measures were constant (indegree = 0.79, outdegree = 0.88, betweenness = 0.79), but the influential facilities that shared a consistent number of patients shifted from acute-care hospitals to post–acute-care hospitals and the other facilities group. Centralization by degree and betweenness measures are important because they indicate which health care entities in a network may have the greatest influence on changes within the network.

The network of Orange County was not tightly knit (density = 0.28), and comparatively, the regional network was more sparse (density = 0.11). This finding was not unexpected because larger networks tend to have lower densities than smaller networks.²⁴ Although this sparseness could be viewed as a positive, because it could translate into lower rates of disease transmission, it could also be viewed as a negative, because it could result in slower rates of information dispersal and intervention implementation. The node-level analysis revealed that with 46 facilities, we could reach 8231 beds in our network. The most influential entities in the county network, excluding intrafacility transfers, were X13 and X31. Both facilities admitted the most patients (weighted indegree: X13 = 1581; X31 = 815) and discharged the most patients (weighted outdegree: X13 = 2125; X31 = 1068) in the county. Although X13 and X31 were the most active health care entities in the county, X13, X31, X14, and X12 all had a similar amount of control (flow betweenness centrality: X13 = 1980, X31 = 1728, X14 = 1440, X12 = 1241) based on the total volume of patients shared between health care entities. When we included intrafacility transfers, both X13 (weighted indegree = 4403, weighted outdegree = 4947) and X31 (weighted indegree = 1832, weighted outdegree = 2085) were the most active and controlling health care entities in the county. Although, overall, acute-care hospitals were the most prominent health care entities in the county network, skilled nursing facilities and other facilities also provided contributions, albeit moderate contributions, to the network. In particular, X42 (skilled nursing facility) and X7 (other facility) were the most active health care entities among their facility types (weighted indegree: X42 = 341, X7 = 437; weighted outdegree: X42 = 211, X7 = 299) by admitting and discharging the highest volume of patients. The information gained from the baseline county network provide us with reason to apply our network when we are in need of communicating information or applying countywide infection control interventions.

Post–acute-care hospitals admit patients with chronic critical illnesses that require specialized care (eg, mechanical ventilation), have multiple comorbidities, and have high rates of antibiotic use.²⁵ These complex treatment plans and prolonged hospitalization courses place patients at post–acute-care hospitals at risk of being colonized with MDROs.²⁵ Our baseline regional network identified that our post–acute-care hospitals may be largely affected through admissions or may directly influence the dispersal of patients with MDROs across the network. Thus, this analysis allowed us to readily identify important facilities in MDRO outbreaks and direct our efforts toward targeted interventions in preventing and containing MDROs.

Practice Implications

State and local health departments have a unique and eminent position in assessing emerging trends or gaps in disease prevention and are a consistent facilitator when health care communication between facilities shifts.^26,27 As MDROs increasingly emerge and are transferred across our health care system, we need to be vigilant and strategic to contain and halt transmission. A lack of coordination and communication between health care facilities may increase the risk of MDRO acquisition.²⁸ More recently, public health authorities have served as the lead in coordinating and alerting facilities of MDROs and outbreaks in a community.²⁸ Public health officials can use social network analyses to identify and address gaps in communication across the health care network. Moreover, social network analyses can help determine the direction of transmission between facilities and indicate which health care entities are at an increased risk for receiving a patient with an MDRO.¹² Once identified, local public health officials can prioritize facilities at high risk for outbreaks and develop county- or regional-specific interventions. Access to analysis platforms to perform social network analyses is a potential limitation for public health officials. However, the R software environment that we used for statistical computing and graphics is free, which allows for low-cost reproducibility for any health department.

Limitations

Our study had several limitations. First, the data set provided was collected from the CMS billing system, which is limited to patients insured by CMS and facilities that accept CMS health insurance. Thus, proportional selection bias of the total patient volume shared across the facilities may have been introduced, leading to an underestimation of the effect of patient transfers. However, although the volume of patients may have been underestimated, we captured 94% of the acute-care hospitals, post–acute-care hospitals, and skilled nursing facilities located in Orange County.²⁹ Finally, closeness centrality is a metric we would have found most useful at the county level, because it is associated in communication sociograms with the speed of information dissemination. Closeness centrality allows for the shortest paths (geodesic distance) between entities to be determined on a close-knit network. However, our networks were not tightly knit, and the metric would have been ill-defined.²⁰

Conclusion

To better understand the associated connectiveness of health care entities as it relates to their shared patients, we created this baseline social network analysis of patient movement. In our county, acute-care hospitals served as the main entities, but post–acute-care hospitals and other facilities (mental health/substance abuse facilities) were most influential in patient flow in our regional network. Patients are transferred across health care entities for several reasons, and their infections and conditions travel with them. This knowledge and understanding can have a substantial effect on our response to MDRO outbreak investigations, developing targeted infection control interventions, preventing the spread of MDRO infections, and dispersing information.

Future social network analyses should include eigenvector centrality to capture the importance of entities in the network that can have a substantial effect on propagation of information, interventions, or spread of pathogens. Analyzing the transitivity of a community’s health care infrastructure can assist health departments in outbreak situations to identify clusters. As part of this analysis, we created a social network analysis toolkit to provide local and state health departments with guidance in the construction of social networks.²¹

ORCID iDs

Danielle A. Rankin https://orcid.org/0000-0003-3018-3373

Sarah D. Matthews https://orcid.org/0000-0001-8372-1185