Social Networks and the Spread of Hospital-Acquired Infections
A primary focus of our research is on models of hospital-acquired infections. This work entails developing computational tools to collect, analyze, and visualize data in order to better understand the spread of infectious diseases in a healthcare setting.
Viewing human contact networks as special types of social networks (where social interactions are defined by physical proximity or contact) makes available a rich set of computational techniques for analysis. In fields where experiments are not possible, such models and algorithms are often used to support simulation studies; we have used these models to support our work on, for example, optimizing vaccination strategies.
Of course, to ensure meaningful simulation results, the structure of the underlying graph model must accurately match the structure of the human social interactions under study. Constructing accurate social network models for healthcare workers and patients in a hospital setting presents several problems, as interactions between patients and/or healthcare workers play out over a layer of subtle dependencies between physical infrastructure, administrative structure, hospital unit layout, staffing levels, patient loads, time-of-day constraints, and so on.
Because observational data are so expensive to collect, often lack location information, and are necessarily sparse, we have investigated the use of other existing data sources to provide finer-grained healthcare worker location information from which we could infer the underlying social network. One such source of data are de-identified electronic medical record (EMR) system logins: because many hospital employees use the EMR system multiple times over the course of a day, and because these events are logged by date, time, user and location, such keyboard data provide a rich context from which to infer contact and movement.
Contact networks can be used both to support simulation studies as well as to enable more analytical studies of their graph-theoretic properties. We have used both approaches can be used to understand how, e.g., different vaccination strategies might perform. If we accept the fact that the purpose of any vaccination program is to ‘‘shatter’’ the contact graph into many isolated components (by breaking the graph into many smaller compartments, infections that appear in one such compartment cannot find a path to infect individuals in other components — this is the same principle that underlies cohorting as an infection control strategy), we can evaluate the performance of a vaccination strategy by how well it shatters the graph.
We have been working with wireless sensor mote technology as a novel way to track healthcare worker movement, location and interactions, parameters that allow us to build near-real-time contact networks. Motes are active (i.e., battery powered, with some on-board memory and processing capability), inexpensive devices that can sense and record proximity with other motes, including time, mote identity, and perceived signal strength. Using an IEEE 802.15.4-compliant wireless radio, motes can be programmed to sense and record a number of parameters, such as temperature, humidity, proximity with other motes, and so on, depending on their hardware configuration and on-board software.
We currently have just over 100 first generation motes using commercial sensor mote boards and recycled pager cases and we have tested our mote system under controlled conditions in an temporarily unoccupied portion of UIHC. By logging messages between badge motes, we are able to record proximity. These data allow us to reconstruct the time and length of contacts between motes; with appropriate RSSI thresholds and transmission parameters, we can estimate the probability that one worker has been in contact range (i.e., direct touch) or within droplet range (i.e., three to six feet) of another. And by placing beacons at known locations in the hospital, we can use proximity to beacons to reconstruct the physical locations of these contacts when superimposed over a model of the hospital’s physical structure.
Improving Hand-Hygiene Practice in Healthcare Settings
A second focus of our work is to use the tools of social network analysis to improving hand-hygiene adherence. This work entails applying the social network measurement technology we have developed to measure hand-hygiene compliance and then designing interventions to improve adherence among healthcare workers.
According to the Centers for Disease Control and Prevention (CDC), healthcare-associated infections affect about 2 million patients in US hospitals each year and result in an estimated 99,000 deaths. Failure of healthcare workers to perform appropriate hand hygiene is one of the leading preventable causes of hospital-associated infections. Accurate measurement of hand-hygiene compliance is a component of every infection-control program. Measurements are required for benchmarking as well as to detect improvement; moreover, direct feedback of hand-hygiene adherence rates to healthcare workers is a critical element of many adherence campaigns. We have developed a wireless-mote-based data-collection system to capture healthcare worker hand-hygiene behavior along with healthcare worker interactions over time and space.
Measuring hand hygiene is an important component of all hospital infection control programs and is recommended by the Centers for Disease Control and Prevention and the World Health Organization. To address the shortcomings of existing paper-based or electronic recording tools, we have designed a hand-hygiene compliance recording application that is easy to use, requires minimal training, contains mechanisms to minimize data entry error, and decreases the amount of time required to process and feedback results. Our system is configurable, and flexible enough that it can be easily used by a wide range of institutions.
We are also actively working on computational epidemiology problems of broader geographic scope.
In the United States, sentinel surveillance for diseases such as seasonal influenza are conducted at the state level, usually by a network of volunteer facilities. Given a set of candidate volunteers, how do we choose optimally-located sentinal sites? Ideal locations are those that cover the most people for the least cost while also capturing disease incidence in a way that accurately reflects the greater whole.
The diffusion of information during the recent H1N1 or “swine” flu epidemic served to illustrate the kinds of difficulties public health officials are likely to encounter in distributing sound and reliable information to the population at large. Twitter is a free micro-blogging service that enables its millions of users to send and read each other’s messages, or “tweets.” Studying swine-flu-related tweets provides a picture of rapidly-evolving public sentiment specific that is easily tuned to a particular public health issue.