key: cord-0267719-r4xu6jve
authors: Madhobi, K. F.; Kalyanaraman, A.; Anderson, D. J.; Dodds-Ashley, E.; Moehring, R. W.; Lofgren, E. T.
title: Who Is Hospitalized With Whom? Inpatient Contact Networks and Mixing Patterns
date: 2022-02-24
journal: nan
DOI: 10.1101/2022.02.22.22271374
sha: 3267e6910e6bb8c49025c852ac305588d72e45da
doc_id: 267719
cord_uid: r4xu6jve

Importance: Person-to-person contact is important for the transmission of healthcare-associated pathogens. Quantifying these contact patterns is crucial for modeling disease transmission and understanding routes of potential transmission. Objective: Generate and analyze the mixing matrices of hospital patients based on their contacts within hospital units. Design, Setting, and Participants: The study was conducted in 24 hospitals in the Southeastern United States that were part of the Duke Antimicrobial Stewardship Outreach Network (DASON) between January 2015 and December 2017. There were a total of 1,569,413 patients and 299 hospital units. Main Outcome and Measures: The mixing matrices of patients for each hospital unit using age, Elixhauser Score, and a measure of antibiotic exposure. Results: Mixing matrices were calculated from a database of 24 hospitals, which included 2.9 million admission records for nearly 1.6 million patients. Some units had highly similar patterns across multiple hospitals although the number of patients might vary to a great extent. Within a period of 26 months (October 2015 and December 2017), the highest daily average is 765 patients in the ED of Hospital-12 and lowest daily average is only 2 patients in some of the smaller hospital units. For most of the adult inpatient units, frequent mixing was observed for older adult groups while outpatient units e.g. ED and Behavioral Health etc. units showed mixing between different age groups. From the mixing matrices by Elixhauser Score, we observed mixing between patients with relatively higher comorbidity index on the ICUs. Mixing matrices by Antibiotic Rank, a 4-point scale based on priority for antibiotic stewardship programs, resulted in six major distinct patterns due to the variation of the type of antibiotics used in different units. Conclusions and Relevance: The mixing patterns of patients both within and between hospitals followed broadly expected patterns, though with a considerable amount of heterogeneity. These patterns can be used to evaluate the appropriateness of policies and guidelines for smaller community hospitals, as well as improve the design of interventions that rely on altering patient contact patterns.

An individual's risk of acquiring an infectious disease is inherently a function of whom they contact, with currently infected individuals representing the exposure source for those individuals infected in the future, a phenomenon known as "dependent happenings" 1 . Therefore, understanding whom an individual contacts becomes critical for understanding their risk. One way of representing and studying this information is the development of a contact network represented by a population of individuals (nodes) and the contacts between them (edges), and studying the properties of this network (i.e. are particular types/classes of people more likely to come into contact with one another than others).

and other sexually transmitted diseases [2] [3] [4] [5] , and are being increasingly studied in infectious diseases more broadly [6] [7] [8] [9] . Hospitalized patients represent a particularly challenging population for contact network analyses due to the complexity of the hospital environment. Patients may or may not be contacting each other directly (depending on whether they are mobile and can interact with one another), but they may be exposed to pathogens through contamination on the hands and clothes of healthcare workers, on shared instruments, or persisting in the hospital environment as fomites. Several studies have collected hospital contact networks using a variety of methods [10] [11] [12] [13] . Many, however, were limited to a single hospital or a single study site.

Long-term and multi-site studies of these networks may be important for understanding how hospitals adjust to shifting demands for patient care (e.g. during a pandemic), the evolution of antibiotic stewardship programs, or other shifts to the flow of hospitalized patients, and how this in turn impacts infection control. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted February 24, 2022. ; https://doi.org/10.1101/2022.02. 22.22271374 doi: medRxiv preprint The focus of this study was to better describe the contact networks of hospitalized patients using a large, multi-hospital sample. By forming these contact networks, we aimed to visualize contact patterns of variables that affect susceptibility to hospital acquired infections and multidrug resistant organisms (MDRO): age, comorbidity, hospital unit type, and antibiotic exposure. We examine age because it is a known risk factor for infectious diseases such as COVID-19 14 as well as a number of healthcare-associated infections [15] [16] [17] . We also examine Elixhauser score 18 for comorbidity as a proxy for overall vulnerability to infection, and antibiotic usage as a measure for potential multidrug resistant organism (MDRO) colonization pressure from other patients within the unit 19 .

To estimate the patient contact networks, we used data from the Duke Antimicrobial Stewardship Outreach Network (DASON) 20 is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint 

Within the DASON data there are records for a patient's movement between units, as well as arrival and discharge times. Using this information, we estimated a colocation contact network -i.e., if two patients were recorded as being in the same hospital unit during a period of one day, they were counted as being in contact for that day. If there are k patients in a unit in a specific day, the number of contacts would be . It is important to note that the type of unit is based on its NHSN classification. This classification provides a useful, but ultimately imperfect, approximation of patient case mix, as patients may be placed in a unit for other reasons such as bed availability and hospital volume, a unit's definition may be shifting over time, etc.

Using the pairwise contact information, we computed three types of mixing matrices, based on patient age, Elixhauser score, and antibiotic agent exposure. These mixing matrices record the frequency of contacts between patients belonging to different classes of that category. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted February 24, 2022. ; https://doi.org/10.1101/2022.02.22.22271374 doi: medRxiv preprint j] corresponds to the frequency of contact between a patient having comorbidity score i with a patient having comorbidity score j.

For mixing matrices by antibiotic agents, we followed a ranking scheme proposed by RW Moehring et. al. 21 According to this scheme, antibiotic agents are categorized on a 4-point scale based on their spectrum of activity against bacterial pathogens and priority for antibiotic stewardship program which is as follows: Narrowspectrum (Rank 1), Broad-spectrum (Rank 2), Extended-spectrum (Rank 3) and

Protected (Rank 4). The resulting mixing matrices become 4 x 4 tables where cell [i, j] represents the number of contacts occurring between patient pairs exposed to agent ranks i and j respectively.

Note that the mixing matrices differed for different units within a hospital or for different hospitals. To help with comparisons between different hospital sizes, we also computed a normalized representation for the mixing matrices.

Data preparation and network extraction was performed using Python 3.6.9, including the Pandas library for data preparation, and the Bokeh 22 library for visualization and the creation of interactive plots. These interactive plots allow for comparing and contrasting the mixing matrices across different units, and across different hospitals, and are available at http://go.wsu.edu/hospitalmatrix. Further visualization was conducted in R 3.6.3 23 using the networkD3 library. The extracted patient contact networks, as well as the source code used for the analysis, are available on at https://github.com/epimodels/mixing_pattern. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint 

In Figure 1 , we added a snapshot of the patient contact networks that was generated using the records of one month (January 2017) for each of the hospitals included in this study. The nodes represent patients, and edges represent a pairwise contact between the corresponding two patients in the time interval considered. As expected, most networks were dense, particularly for hospitals with large unit capacities (e.g., Hospital-24, Hospital-14). However, networks were relatively sparse (e.g., Hospital-7, Hospital-16) in smaller hospitals. The largest network in this collection (for Hospital-24) has 9,778 nodes and average degree (the number of edges connecting to that node -in this case the number of co-located patients) of 329; while the smallest network (for Hospital-16) has 572 nodes and an average degree of 35.

There is considerable inter-hospital variability in the age-mixing patterns of patients as shown in Figure 2 , owing to the type of hospital, catchment population, etc.

There are primarily three major patterns visible: a) hospitals showing uniform mixing across all adult ages (e.g., Hospital-23); b) hospitals serving primarily younger age groups (e.g., Hospital-11); and c) hospitals, especially smaller ones, dominated by mixing between elderly patients (e.g., Hospital-5). These mixing patterns of hospitals were consistent with their respective age distributions (eFigure 1).

However, within a hospital, the patterns varied from unit to unit as different units cater to different types of patients. Figure is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted February 24, 2022. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint 

Thet Elixhauser Score in DASON data is ranged from 0 to 16, with a median and IQR of 0 due to most patients having lacked known comorbidities. Figure 4 

From the medication information of DASON data, we found that there were many (and often a large majority) of patients who were not on any form of antibiotic.

The ratio of contacts where neither patient in a connected pair were on antibiotics compared to pairs where one patient or both patients were on antibiotics varied widely by unit, shown in eFigure 2 for four selected units.

We observed six distinct patterns on the antibiotic mixing matrices based on the four-point ranking scheme as shown in Figure 5 . Across all hospitals, Gynecology, Labor and Delivery and Postpartum units predominantly involved patients on narrowspectrum agents ( Figure 5A ). This pattern also occurred in some, but not all, Operating is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted February 24, 2022. 

Mixing matrices by age mostly conform to the expected patterns one would expect heuristically with a few exceptions. Especially Emergency Departments and Behavioral Units showed areas with broader age-related mixing patterns. These units require special consideration when considering pathogens with markedly different agerelated risks, transmission potentials, or vaccination status. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint 

In comparison to general units, the dense area in the critical care units shifted up and to the right, indicating --not unexpectedly --an increase in patients with more prevalent comorbid conditions, though the peak for this distribution was more diffuse. 

Mixing matrices that showed dense mixing on one rank is an indicator of substantial use of a specific kind of antibiotic in a unit. On the other hand, the bimodal inter-spectrum mixing can arise from one of two possible mechanisms --two distinct groups of patients, one on one type of antibiotic, the other on another type of antibiotic, who happened to be co-located in the same unit. The second is that the same patients are prescribed drugs of two different ranks. To examine these two possibilities, we considered one ward with both a large number of patients as well as the distinctive Narrow-Extended spectrum pattern, a neonatal intensive care unit in a large academic medical center. The distribution of antibiotic exposures are shown in eTable 1. It is apparent that the vast majority of patients are exposed to both classes of antibiotics during their hospitalization, though the result was not statistically significant (Fisher's Exact Test p = 0.056). This pattern seems likely to arise from a commonly used combination of ampicillin and gentamicin for empiric coverage of neonatal sepsis 24 . is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint This, in turn, suggests that a patient's individual antibiotic exposure profile likely represents their primary exposure at any given time.

In contrast, the analysis of a unit (Emergency Department in Hospital-15) having a Broad-Extended pattern, reveals a different picture as can be seen from the distribution in eTable 2. This unit served a large elderly population with several Skilled Nursing Facilities, and among them 77.6% of patients were prescribed either broad-or extended-spectrum agents (33.2% and 44.4% respectively) (Chi-squared p > 0.001).

These findings once again highlight the importance of Emergency Departments as areas with far broader mixing patterns than the rest of the hospital environment, as well as the likelihood of the units discussed above being central points of empiric therapy within many hospitals.

The inter spectrum mixing of Narrow, Broad and Extended spectrum antibiotics was only observed in the Pediatric units of the large academic medical center. The academic medical center is notable for having both Cystic Fibrosis patients as well as a pediatric transplant program, which resulted in a markedly different patient profile to community hospital Pediatric units, and likely drove these different patterns.

This study presents several aspects of how hospitalized patients come into contact with each other. Understanding these contact patterns can provide vital information on infection transmission risk -for example, where patients with high potential susceptibility to acquiring multidrug resistant organisms might be in contact, directly or indirectly, with those at serious risk for adverse outcomes from infection.

While some of these patterns may be inferred heuristically, it is nevertheless beneficial to quantify these patterns. This enables their use in modeling studies of hospital-. CC-BY 4.0 International license It is made available under a perpetuity.

is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted February 24, 2022. ; https://doi.org/10.1101/2022.02.22.22271374 doi: medRxiv preprint acquired pathogens. Further, it could provide a means to quantitatively track shifts in antibiotic use or patient case mix patterns. For example, a unit shifted away from its NHSN unit-type designation to deal with surges of COVID-19 patients, and the population characteristics changed. As another example, the matrices might change as stewardship programs intervened on how antibiotics were used. Finally, quantifying the variability between hospitals can help assess the generalizability of effect estimates and the ensuring intervention and policy recommendations between the academic medical centers where these estimates are often obtained and rural and community hospitals, for any situation where patient-to-patient interaction is potentially at play.

There are important limitations to this study, arising from the use of an existing data source to reach multiple hospitals and a large number of patients. This study implicitly assumes that patients visiting a unit the same day had contact -primarily via indirect contact mediated by either healthcare workers or the environment. While age, underlying comorbidities and antibiotic exposure are certainly important risk factors for healthcare associated infections, they are far from an exhaustive list, and there are a number of risk factors that are beyond the reach of a single study of this type, or impractical to collect on an ongoing and continual basis in a broad network of hospitals of varying resource levels. Besides, patients that occupy the same unit for multiple days have repeated contacts, which we assume linearly add to the amount of mixing. Finally, we assume that an NHSN unit designation is an adequate proxy for the type of procedures and patients present in a given unit, which may result in a degree of misclassification.

Additionally, a number of contact patterns were either more diffuse (in the case of age and Elixhauser score) or unique (in the case of some antibiotic prescribing patterns in pediatric units) to the academic medical center present in the data set. This is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted February 24, 2022. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted February 24, 2022. ; is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint for the month January 2017. Each node is a patient, and each edge represents a contact between the two corresponding patients in that hospital during that month. One hospital is omitted due to a very sparse connectivity over the chosen month. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted February 24, 2022. ; https://doi.org/10.1101/2022.02.22.22271374 doi: medRxiv preprint Figure 5 : Mixing matrices by antibiotic rank measured by the number of pairwise contacts (shown by the color bar) for selected units in DASON network. Panel A represents a unit where the patients who came in contact were mostly exposed to one type of antibiotic namely Narrow spectrum antibiotics. Panels B and C show Broad and Extended spectrum heavy contact patterns. Panels D and E show highlighted interspectrum mixing patterns, e.g. Narrow-Extended and Broad-Extended. Panel F highlights a wide region encompassing Narrow-Broad-Extended spectrum antibiotics. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint

The copyright holder for this this version posted February 24, 2022. ; https://doi.org/10.1101/2022.02.22.22271374 doi: medRxiv preprint

An application of the theory of probabilities to the study of a priori pathometry.-Part I

Sexual mixing patterns of patients attending sexually transmitted diseases clinics

Modeling and analyzing HIV transmission: the effect of contact patterns

Concurrent partnerships and the spread of HIV

Sexual bridging socially and over time: a simulation model exploring the relative effects of mixing and concurrency on viral sexually transmitted infection transmission. Sexually transmitted diseases

Social contact patterns relevant to the spread of respiratory infectious diseases in Hong Kong

Social contacts and mixing patterns relevant to the spread of infectious diseases

Estimating contact patterns relevant to the spread of infectious diseases in Russia

The social life of infants in the context of infectious disease transmission; social contacts and mixing patterns of the very young

Estimating Time Physicians and Other Health Care Workers Spend with Patients in an Intensive Care Unit Using a Sensor Network

The Role of Healthcare Worker-Mediated Contact Networks in the Transmission of Vancomycin-Resistant Enterococci

Social network analyses of patienthealthcare worker interactions: implications for disease transmission

COVID-19) -United States

Additional risk factors for infection by multidrug-resistant pathogens in healthcareassociated infection: a large cohort study

Four Country Healthcare Associated Infection Prevalence Survey 2006: overview of the results

A risk factor analysis of healthcare-associated fungal infections in an intensive care unit: a retrospective cohort study

Comorbidity measures for use with administrative data

Receipt of Antibiotics in Hospitalized Patients and Risk for Clostridium difficile Infection in Subsequent Patients Who Occupy the Same Bed

Effects of a Collaborative, Community Hospital Network for Antimicrobial Stewardship Program Implementation

Development of an Electronic Definition for De-escalation of Antibiotics in Hospitalized Patients

R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing

The management of communityacquired pneumonia in infants and children older than 3 months of age: clinical