Social Network Analyses of Patient-Healthcare Worker Interactions: Implications for Disease Transmission

Abstract

Patients and healthcare workers (HCW) in healthcare settings represent a unique social network in which the risk of transmission of an infection is considered to be higher for both HCW and patients. Using data from existing clinical informatics resources, we constructed social networks of patient-HCW interactions in the emergency department of a tertiary care pediatric hospital. The structural properties of these networks were analyzed and compared to other well known networks. Patient-HCW networks do not demonstrate the classical power-law distribution of scale-free networks, thus indicating that they are different from social networks of individuals in a community. The clustering coefficient is larger as compared to a random network, indicating small world properties. The eigenvector centrality, used to identify the most important nodes, reveals HCW to be more connected than patients. These properties imply differences that must be taken into account when analyzing patient-HCW networks and planning interventions and mitigation strategies to prevent the spread of infectious diseases in healthcare settings.

Introduction

Interactions between patients and healthcare workers (HCW) are an integral component of our healthcare system. While all interactions are to benefit the patient, there are unintended consequences that result in the transmission of infectious diseases between patients and HCW. Outbreaks of respiratory diseases such as TB, SARS and pertussis and the fear of pandemic influenza have highlighted the occupational risk to HCW [1] The risk is increased in pediatric emergency settings where there are frequent interactions of HCW with many patients who are infectious. Patients are also at risk and HCW are considered vectors in the nosocomial transmission of diseases such as Methicillin Resistant Staphylococcus aureus (MRSA) and other multi--drug resistant organisms [2].

Social network analyses and network epidemiology have been used to study the spread of infectious diseases [3–7]. These methods offer benefits over models that assumed homogeneous mixing among individuals. A key concept is that contact patterns determine the spread of diseases, be it casual contact in communities, sexual contact among individuals or close contact as that seen in families and healthcare. Most studies have focused on infections in communities, while transmissions in healthcare settings have been studied to a lesser extent. Social networks in healthcare pose unique challenges in that they are usually constrained (patient to HCW, then HCW to other patient, HCW to HCW) and the probability of transmission is higher given close interactions between infectious patients and HCW during routine medical care. A better understanding of these networks and contact patterns is needed to prevent the spread of infections and protect HCW and patients alike.

One limitation in further developing social networks for healthcare has been the lack of detailed data on interactions between patients and HCW [4, 7]. Investigators have used manual observations and electronic data such as medical record log-ins to infer movement and interaction of patients with HCW and HCW with other HCW [8, 9].

Our working hypothesis is that an enhanced comprehension of social networks of patient-HCW interactions will improve our understanding of the potential transmission of infectious diseases in healthcare. This pilot project addresses the following research questions: (1) Can existing clinical informatics data be used to develop social networks to study patient-HCW interactions in healthcare? (2) How do these social networks differ from others with respect to their structural properties? (3) What are the implications of these unique structural properties for preventing the spread of infectious diseases in healthcare settings?

Setting

This study was carried out at the emergency department (ED) of Primary Children’s Medical Center (PCMC) in Salt Lake City, Utah. This Intermountain Healthcare (IH) facility is a level I trauma and tertiary care center that provides care for patients from Utah and 5 surrounding states. The hospital averages 40,000 ED visits per year, along with 160,000 outpatients and 13,000 in-patients annually. IH is a leader in clinical informatics and PCMC has advanced informatics systems for tracking the movement of patients and HCW in the emergency department.

This study received approval from the Institutional Review Board of the University of Utah and the Privacy Board of Intermountain Healthcare.

Methods

Existing Clinical Informatics Data Sources

The Logicare® Patient Flow System: This is a proprietary patient flow management system in use at the PCMC ED for the past 5 years. The systems tracks the time of registration, triage, entry into the examination room, evaluation by providers and final disposition from the ED, using automatic and manually entered time-stamps. Data from the system are archived and accessible for patient care and research.

The Hill-Rom® ComLinx NCM Locator system: This has been in operation at the PCMC ED since 1998 and upgraded in 2006. The purpose of the locator system is to account for HCW location and movement for risk management. Physicians, mid-level practitioners, nurses and technicians who work in the ED are provided with individual badges that emit unique infra-red signals every 2 to 6 seconds from two light emitting diodes. Infrared receivers placed in various locations in the ED including each patient room, hallways and common areas such as triage, nurses station etc. detect the signals allowing the system to identify and timestamp the staff member to that location. Data of location and time stamps of HCW are stored in a stand-alone system and reports can be generated to view staff activity by location and by name. The reports can be exported in Microsoft Excel for linking to other data and analyses.

Data from the two systems were downloaded and merged using time-stamps and room locations. The merged data were used to create a data set that contained interactions (frequency and duration) between patients and their HCW, along with their spatio-temporal co-localization. Twelve calendar months of data were available and data from a randomly selected month and day were analyzed for this pilot project.

Developing and Analyzing Social Networks of Patient-HCW Interactions

Social networks were constructed using UCINET software (Version 6.216, Feb 23, 2009, Analytic Technologies, Lexingon, KY). Data were analyzed using SAS (Version 9.1.3,SAS Institute Inc.).

The network was characterized for a month using simple descriptive statistics and for a day using network properties in terms of:

1. Power-law distribution: It has been found that a diverse range of phenomena in life and nature such as populations of cities, the frequency of use of words in any human language, the number of hits on a website, etc. follow a power law distribution and are described as scale-free [10]. This refers to networks where a few nodes are connected to many other nodes and the majority of the nodes have few connections. The power law distribution is plotted on a log-log scale using the formula ln p(x)= −a ln x+c where x is the random variable, c is a constant. The exponent a is the slope of a straight line when plotting the distribution on a log-log scale and c is the intercept of this line. A modification offered by Newman [10] provides a superior method of plotting the power law distribution on a log-log scale using a reverse cumulative distribution using the formula $P(X\ge x)=\frac{C}{a-1}{x}^{-(a-1)}$. The function P(x) is the degree distribution and should be a straight line if the data follow a power law distribution. A regression line was fit and the residuals were used to assess the fit of the regression line in order to decide whether or not the data follow a power law distribution.

2. Small world phenomenon: Small world network is another type of widely discussed network. In a small world network, nodes are highly clustered in the neighborhood; at the same time, the path length between any two nodes is short, compared to the size of the network [11]. The clustering coefficient is simply the average of the densities of the neighborhoods of all of the actors. Average path length (or geodesic distances) is the average number of relations in the shortest possible walk from one node to the other node in the network. The average path length and clustering coefficient were calculated and compared to a similar random network [12].

3. Centrality: In social network analysis, centrality measures are used to identify the most important nodes in the network. Several studies have demonstrated the validity of using such measures to identify the most connected individuals/nodes for surveillance in the transmission of infectious diseases [13, 14]. One important centrality measure is the eigenvector centrality as shown by Bonacich [15]. Conceptually, eigenvector centrality weights an actor’s degree centrality proportional to that of its neighborhood. Thus, nodes that are strongly tied to other central nodes are proportionally more central than those that are tied to less central nodes. These values are then normalized by dividing each by the maximum eigenvector centrality, and expressing this ratio as a percentage. Eigenvector centrality is only valid in the context of symmetric data such as that seen in our network.

Results

For a randomly chosen month, there were 1261 patients and 87 HCW. The patients had contact with an average of three unique HCW (standard deviation = 2). The HCW saw an average of six patients per 12-hour shift (standard deviation = 8). These simple statistics describe only patient to HCW interactions, not HCW to HCW interactions, but are consistent with the one-day average links described by the network analysis below. These simple statistics however can capture neither the connectedness of the group of patients and HCW as a whole nor any neighborhood or clustering properties.

Existing clinical informatics data sources can be harnessed to develop detailed social networks of patients and HCW

For the one day of data analyzed, the network of patient-HCW interactions and patient interaction contains 61 nodes (21 HCW and 40 patients). The average contact time between any two nodes is 20.16±131.74 seconds; and on average, each node is connected to 4.36±15.48 other nodes. Moreover, for this network, there is no interaction between patients.

The Patient-HCW network is not scale free

In looking at the reverse cumulative distribution of the network interactions (Figure 1), this network does not follow a power law distribution. This was concluded as the residuals from the linear regression fit to the log-log data were significantly non-normal.

Figure 1.

Cumulative distribution of degree on log-log scale. Note that the distribution is not a straight line, indicating the data likely do not follow a power-law distribution.

The network shows strong small world properties

The average path length was 2.577 and the clustering coefficient was 0.541. A random network generated using a homogenous Bernoulli distribution with the same network size and mean connections had an average path length of 2.881 and a clustering coefficient of 0.047. These results indicate that the network has about the same path length as that of a random network, whereas it is 11 times more clustered in the neighborhood. In sum, the patient- HCW network exhibits strong small world property. A comparison of these metrics is made to known networks from the literature [12] as shown in the table below. The patient-HCW network is more similar to the film actors network than it is to the power grid network, which does not change over time or include human interaction.

Network	Path Length	Path Length of random network	Clustering coefficient	Clustering coefficient of random network
Patient-HCW	2.577	2.881	0.541	0.047
Film actors	3.65	2.99	0.79	0.00027
Power Grid	18.7	12.4	0.080	0.005

The HCW are more strongly connected as compared to patients (Figure 2)

Figure 2.

The patient-HCW interaction network, scaled by eigenvector centrality. Squares presents patients; circles represent healthcare workers (HCW). The size of nodes is scaled to each node’s eigenvector centrality, and the width of ties is scaled to the length of the interaction time between nodes. Please note that the HCW are connected to both patients and other HCW. Patients do not have connections to other patients as this network represents patients in individual rooms. In color: red lines are patient-HCW connections and blue lines are HCW-HCW connections.

The mean of normalized eigenvector centrality for the health worker group is 14.677 (±22.615), and that for the patient group is 6.9 (±9.257). The Kruskal-Wallis test shows that statistically, there are no significant differences among the different groups of HCW. This result follows the nature of the network that HCW contact different patients and other HCW, whereas patients are not mobile and do not contact other patients as they are all in single rooms.

Limitations

The limitations of this pilot study include the short focus of the network analysis and some properties of the data collection process. The network properties may vary somewhat over time and the standard deviations reported are for a single day, not across the year of data collected. There are limitations of the data collection system due to line-of-sight issues for the Hill-Rom system sensors and the requirement that HCW wear the badge at all times. The network analyses are missing the patient to patient interactions that takes place in the waiting room. These limitations should not unduly affect the results or conclusions and are recognized openings for future work.

Conclusions

This is the first detailed description of a social network of patients and HCW in an emergency department setting. The network was developed using existing clinical informatics systems that track patients and HCW. These types of systems exist in several forms and are available in many large hospitals; thus this has the potential to be developed as a practical application for use by clinicians and researchers alike.

The patient-HCW network has unique system properties

Unlike many other networks, this network is not scale-free. One possible reason is that in other networks such as the internet, new nodes are added based on preferential attachment or affinity for an existing mode. In contrast, in the emergency department and other healthcare settings, the patients are assigned and seen based on who is on duty that day. The patients and HCW have little or no control over this. Also, we may be capturing only a small subset of the entire network and these have been described to be scale-free [16]. The neighborhood is strongly clustered, although its average path length is as long as that of a random network. HCW, compared to patients, are more important nodes with numerous connections.

Implications of these properties for disease transmission and interventions

At one level, a highly clustered network such as the one described here would imply rapid spread of an infection with a decrease in the threshold required to cause an epidemic [6, 17]. As the neighborhood is densely connected, disease may spread fast within each neighborhood, however, it will spread to other neighborhoods in the entire network only as fast as in a random network, because the average path length is nearly the same as that of a random network. This would be true for an emerging infectious disease with a short incubation period such as SARS (severe acute respiratory syndrome) to which all individuals are equally and uniformly susceptible. However, for other diseases to which individuals are already exposed or may be resistant from either natural immunity or vaccination or prophylactic medications, a large clustering coefficient would slow down the epidemic as the susceptibles are depleted rapidly. New patients and new HCW entering the ED would change this dynamic. The concept of individuals with many connections would also partially explain “super spreaders” for certain diseases such as SARS. Finally, detailed analyses of the structural properties of these networks would be needed to improve current infection control strategies for mitigating disease and preventing spread in healthcare settings. These strategies could potentially target those HCW with the most connections for screening, prophylaxis and vaccination.

Further work is needed to study these networks in more detail and use these models to study transmission of disease, contact tracing and mitigation strategies. Possible extensions of these analyses are to improve the efficiency in the use of hospital resources, restructuring emergency department workflow and design and implementation of new clinical information systems.