https://orcid.org/0000-0002-7206-8077
Abstract
The key epidemiological drivers of Clostridioides difficile transmission are not well understood. We estimated epidemiological parameters to characterize variation in C. difficile transmission, while accounting for the imperfect nature of surveillance tests.
We conducted a retrospective analysis of C. difficile surveillance tests for patients admitted to a bone marrow transplant (BMT) unit or a solid tumor unit (STU) in a 565-bed tertiary hospital. We constructed a transmission model for estimating key parameters, including admission prevalence, transmission rate, and duration of colonization to understand the potential variation in C. difficile dynamics between these 2 units.
A combined 2425 patients had 5491 admissions into 1 of the 2 units. A total of 3559 surveillance tests were collected from 1394 patients, with 11% of the surveillance tests being positive for C. difficile. We estimate that the transmission rate in the BMT unit was nearly 3-fold higher at 0.29 acquisitions per percentage colonized per 1000 days, compared to our estimate in the STU (0.10). Our model suggests that 20% of individuals admitted into either the STU or BMT unit were colonized with C. difficile at the time of admission. In contrast, the percentage of surveillance tests that were positive within 1 day of admission to either unit for C. difficile was 13.4%, with 15.4% in the STU and 11.6% in the BMT unit.
Although prevalence was similar between the units, there were important differences in the rates of transmission and clearance. Influential factors may include antimicrobial exposure or other patient-care factors.
Clostridioides difficile infection (CDI) contributes significantly to the overall burden of healthcare-associated infections [1, 2], causing approximately half a million infections and 30 000 deaths in the United States annually [3]. Despite substantial efforts to control and reduce the incidence of CDI, it remains one of the most common healthcare-associated infections worldwide. Approximately 50% of CDI cases develop during hospitalization [4], which results in significantly longer length of stay (LOS). Patients with CDI experience symptoms ranging from relatively mild and self-limiting diarrhea to intestinal perforation, sepsis, and death [5]. The Centers for Disease Control and Prevention (CDC) 2019 Antimicrobial Resistance Threat Report designated CDI as 1 of 5 urgent threats to public health and estimated 1 billion dollars in CDI-attributable healthcare costs in 2017 [6].
Current clinical practice guidelines to reduce incidence of CDI cover a broad range of domains including epidemiology, diagnosis, treatment, and infection control [7]. Guidelines on infection control are critical to employ once individuals are identified to prevent further spread, and include isolation and contact precautions, careful hand hygiene when caring for patients in isolation, and antibiotic stewardship. In addition, disinfection and cleaning of the patient room and reusable patient equipment are stressed to help prevent further contamination. However, there is no guidance on asymptomatic screening owing to limited data on which to base recommendations [8, 9]. Yet, asymptomatic colonization prevalence with C. difficile can be upwards of 20% for acute care hospitals [10], and possibly even larger in settings with higher-acuity patients. Colonization prevalence, therefore, may play an important role in patient care across a broad range of settings, especially because asymptomatic colonization may serve as a multiplier for future infections.
While a broad range of efforts have been undertaken to help control CDI, the epidemiology of C. difficile transmission in healthcare settings remains poorly understood. Dynamic transmission models have been used to estimate critical parameters underlying the transmission dynamics of several organisms, including methicillin-resistant Staphylococcus aureus (MRSA) [11]. Although promising, this method has not been applied to address the transmissibility of C. difficile. A better understanding of the transmission dynamics of C. difficile, including the proportion of asymptomatic carriers, would provide better insights into where to focus efforts and resources to more effectively promote infection control. The purpose of this study was to provide just such tools. Here, we estimate important parameters that underlie the dynamics of C. difficile transmission, to better understand C. difficile epidemiology and improve insights into disease control. Our approach accounts for the imperfect nature of surveillance tests and provides an estimate of importation, transmission, and clearance rates of C. difficile.
METHODS
Data
We analyzed data collected from 2 inpatient units at a 565-bed tertiary hospital in Milwaukee, Wisconsin, between 1 September 2016 and 30 April 2018. The data included times of admission and discharge times into the units, as well as C. difficile surveillance test times and results, which were collected upon admission and subsequently every 7 days from patients admitted to 5 units in the hospital. The surveillance was based on a nucleic acid amplification test, using a stool sample. Our analysis of the surveillance data focused on the bone marrow transplant (BMT) unit and solid tumor unit (STU) due to the relatively higher amount of longitudinal surveillance testing in those units.
Model
We used a previously developed Bayesian transmission model [12] that incorporates transmission dynamic parameters and mechanistic features including patient flow into and between units through admission, transfer, and discharge, as well as the acquisition and clearance of C. difficile (Figure 1). The model parameters were surveillance test sensitivity and specificity, importation probability, transmission rate parameter, and clearance rate.
Illustration of the relationships between the data, parameters, and underlying states in the Bayesian transmission model.
We combined the data from the STU and BMT unit into our Bayesian modeling framework and estimated parameters using Markov chain Monte Carlo (MCMC). Importantly, our model incorporates “hidden” states that are not directly observable, namely times of acquisition and clearance, although these times are informed by the results of the surveillance tests. Thus we have interval-censored data for these events and rely on a modeling framework that uses the acquisition and clearance times to completely specify the model likelihood function. We refer to the combined observed and unobserved data as the augmented data.
Patient Movement
To model patient flow and appropriately track spatial and temporal changes in the vulnerable and at-risk patient populations, we estimate unit-specific parameters for the transmission and clearance rate parameters. We also account for patient readmissions by allowing for dependence between consecutive hospitalizations.
Importation
At admission, patients were categorized as either colonized (importation) or uncolonized. Although our model allows for patients to have >1 admission, here we define the importation probability as the probability that an individual who is admitted for the first time is colonized at the time of admission. We also account for the dependence in patient colonization status between consecutive hospitalizations, which we call the readmission importation probability. Parameterization of importations in our model in this way implies that given a sufficiently long period between admissions, the more uncorrelated the colonization status between the 2 consecutive admissions will be.
We assume that patients with a recent hospitalization have a different likelihood of importation compared with those in the general community. To estimate the probability of importation at the time of readmission for a given patient, we used the times and results of the patient’s previous C. difficile surveillance tests. In particular, our model assumed that patients could acquire and lose colonization between consecutive admissions. Mathematically, this between-hospitalization model leads to a simple relationship between the importation probability and the acquisition and clearance rates between hospitalizations. This framework provides a simple formula for calculating the importation probability for readmissions based on the time since the previous discharge and colonization status at the time of the previous discharge. For an individual who was colonized at a previous hospitalization, as the time between hospitalizations increases, the probability of remaining colonized decreases as the chance of clearance increases. However, we assume that while the probability of remaining colonized decreases, it never becomes zero as there is always a risk of clearance followed by reacquisition in the community.
Transmission
We assume that transmission occurs within units, resulting in unit-specific transmission rates. In our transmission model, we assume that the force of infection is proportional to the prevalence of C. difficile on the unit, consistent with classical modeling assumptions about transmission of infectious diseases [13]. We estimate the transmission rate parameter, which for each susceptible patient is a measure of the rate of acquisitions per fraction of patients colonized per day, and which describes the intensity of the force of infection and serves as a measure of transmission that controls for prevalence. This provides us a metric with which to compare transmission intensity across units: Given 2 units with the same prevalence, a higher transmission rate parameter in 1 unit suggests an increased level of transmission compared to the other unit.
Clearance
We assume that colonized patients clear colonization during their hospitalization at a constant rate (colonized patients cleared per day). As a result of this assumption, we do not assume a particular mechanism of clearance, rather, a constant clearance rate reflects the variety of patient-care and facility-specific factors that may contribute to the loss of colonization with C. difficile. Once cleared, we assume patients are immediately at risk to reacquire C. difficile both during and between admissions, albeit at different respective rates.
Test Sensitivity and Specificity
To account for test sensitivity and specificity, we allow for both false positives and false negatives, which provides a more nuanced understanding of the transmission dynamics by allowing for imperfect classification of patient colonization status over time.
Estimation
We estimated the model parameters using an iterative MCMC algorithm to obtain posterior samples using both Metropolis-Hastings and Gibbs sampling [14–16]. To estimate the model parameters, in each iteration of the MCMC algorithm we generated a new sample of both the augmented data and the model parameters. Given the observed patient data (admission, discharge, and tests) and an initial set of model parameters, new augmented data (or patient histories) consistent with the observed data and the parameter values were generated through a process of iterative updating. This process resulted in a collection of parameter values with a distribution consistent with the likelihood, conditioned on all observed and unobserved data, known as the posterior distribution for the parameters. The posterior distribution of parameters, which were based on 5000 samples with a burn-in of 500 samples, formed the basis for our point estimates and credible intervals (CrIs). (Additional technical details related to the implementation of this model together with other model extensions are included in the Supplementary Materials.)
Analysis
We report both posterior means and 95% CrIs for the model parameters. We estimated importation probability and unit-specific transmission rate parameters and clearance rates. For our base model, we assumed a narrow prior distribution on surveillance test sensitivity having a mean of 75% and fixed surveillance test specificity at 100%. We then conducted a sensitivity analysis on the surveillance test parameters with the mean sensitivity varying across the range 75%, 80%, 85%, 90%, and specificity varying across the range 100%, 99.7%, 99.4%.
RESULTS
Patient Summaries
Our data show that the STU had more patients and admissions than the BMT unit (Table 1), and of the 2406 patients admitted, 450 (18.7%) spent time in both units. However, the STU had fewer surveillance tests than the BMT unit (Table 1). Less than 50% of patients admitted to the STU had a surveillance test (45.7%), and <50% of patients in either unit (16.9% in STU and 45.4% in BMT unit) had >1 surveillance test (Figure 2). The mean LOS for BMT patients was 8.8 days, whereas the mean LOS for STU patients was 4.9 days. Furthermore, the LOS distributions for BMT and STU patients were statistically distinct (Figure 2). The LOS distribution for the BMT unit is characterized by a longer and thicker tail than the LOS distribution for the STU, indicating a higher proportion of admissions to the BMT with particularly long LOS, comparatively.
Summaries of the Observed Data Used in the Bayesian Transmission Model
| Characteristic | Solid Tumor Unit | Bone Marrow Transplantation | Total |
|---|---|---|---|
| Patients, No. | 1760 | 1096 | 2406 |
| Admissions, No. | 2936 | 1725 | 4658 |
| Surveillance tests, No. (% positive) | 1346 (12.2%) | 2204 (10.0%) | 3550 (10.8%) |
| Tests per patient, No. | 0.76 | 2.01 | 1.48 |
| Tests per admission, No. | 0.46 | 1.28 | 0.76 |
| Characteristic | Solid Tumor Unit | Bone Marrow Transplantation | Total |
|---|---|---|---|
| Patients, No. | 1760 | 1096 | 2406 |
| Admissions, No. | 2936 | 1725 | 4658 |
| Surveillance tests, No. (% positive) | 1346 (12.2%) | 2204 (10.0%) | 3550 (10.8%) |
| Tests per patient, No. | 0.76 | 2.01 | 1.48 |
| Tests per admission, No. | 0.46 | 1.28 | 0.76 |
Summaries of the Observed Data Used in the Bayesian Transmission Model
| Characteristic | Solid Tumor Unit | Bone Marrow Transplantation | Total |
|---|---|---|---|
| Patients, No. | 1760 | 1096 | 2406 |
| Admissions, No. | 2936 | 1725 | 4658 |
| Surveillance tests, No. (% positive) | 1346 (12.2%) | 2204 (10.0%) | 3550 (10.8%) |
| Tests per patient, No. | 0.76 | 2.01 | 1.48 |
| Tests per admission, No. | 0.46 | 1.28 | 0.76 |
| Characteristic | Solid Tumor Unit | Bone Marrow Transplantation | Total |
|---|---|---|---|
| Patients, No. | 1760 | 1096 | 2406 |
| Admissions, No. | 2936 | 1725 | 4658 |
| Surveillance tests, No. (% positive) | 1346 (12.2%) | 2204 (10.0%) | 3550 (10.8%) |
| Tests per patient, No. | 0.76 | 2.01 | 1.48 |
| Tests per admission, No. | 0.46 | 1.28 | 0.76 |
Comparison of the length of stay distribution (left) and the number of tests per patient (right) in the solid tumor unit and bone marrow transplant unit. Abbreviations: 1Q, first quartile; 3Q, third quartile; BMT, bone marrow transplant unit; MED, median; STU, solid tumor unit.
Transmission Rate
Our estimated transmission rate parameter for the BMT unit was nearly 3-fold higher at 0.029 acquisitions per fraction colonized patients per day (95% CrI, .014–.048), compared to our estimate for the STU (0.010 [95% CrI, .001–.029]) (Figure 3).
Comparison of the posterior distributions for transmission rate (left) and the clearance rate (right) in the solid tumor unit (STU) and bone marrow transplant (BMT) unit.
Clearance Rate
Our estimated clearance rate for the BMT unit is nearly double that estimated in the STU, at 0.052 colonized patients cleared per day (95% CrI, .024–.074), compared with 0.025 (95% CrI, .001–.048), respectively (Figure 3). These rate estimates correspond to an estimated median time to clearance of 13 days in the BMT unit and 28 days in the STU.
Importation
Our model estimates suggest that 20% of individuals admitted into either the STU or BMT unit were colonized with C. difficile at the time of admission (95% CrI, 17.6%–22.6%). In contrast, the percentage of surveillance tests that were positive within 1 day of admission to either unit for C. difficile was 13.4%, with 15.4% in the STU and 11.6% in the BMT unit.
Sensitivity Analysis
The transmission and clearance rates increased as both the surveillance test sensitivity and specificity increased. Yet, there were differences in the sensitivity by unit, with the estimates in the BMT unit varying more with the surveillance test characteristics than the estimates in the STU (Figure 4). Nonetheless, across all combinations of sensitivity and specificity, the BMT unit had both a higher transmission rate and clearance rate compared with the STU.
Heat map illustrating how the estimates of the transmission rate (top) and the clearance rate (bottom) depend on our assumptions of the surveillance test sensitivity and specificity parameters. The left panels represent this relationship in the solid tumor unit (STU), and the bone marrow transplant (BMT) unit is illustrated in the right panels.
DISCUSSION
Our results indicate a higher importation of C. difficile into the STU compared with importations into the BMT unit. In contrast, we found that both the transmission rate and clearance rate were higher in the BMT unit than in the STU. This suggests potentially more transmission and shorter carriage durations for patients in the BMT unit. These differences in transmission dynamics between the STU and BMT unit suggest that there may be some different patient risk factors, or care factors that contribute to differential risk of transmission, clearance, or ability to detect C. difficile, such as antibiotic exposure. Our sensitivity analysis indicates that these findings of contrasting transmission dynamics between the STU and BMT unit were robust to surveillance test sensitivity and specificity parameters.
We found that the proportion of positive admission tests, as a measure of admission prevalence, substantially underestimates importation derived from the transmission model. This underestimate of admission prevalence is largely driven by false-negative test results, and suggests that the traditional approach for estimating admission prevalence based on the proportion of positive test results at admission leads to an underestimate of colonization burden. In a previous analysis, we identified factors that influence risk of colonization at admission [17], yet it would be useful to understand the variation even within high-risk settings.
The estimated clearance rates correspond to relatively short carriage durations. Clearance rate estimates in healthcare settings are rare, particularly for C. difficile. A study from almost 30 years ago on an elderly population in a long-term care facility found that 70% of residents who were found to be colonized with C. difficile were colonized for <2 months while 7% were asymptomatic for 3–6 months [18]. Clearance rates have been estimated for a variety of other antibiotic-resistant bacteria, with a broad range of estimates for time to clearance, which are similar to those we have obtained here [19, 20]. Similarly, in a separate study reevaluating the STAR*ICU cluster randomized trial [21], we estimated time to clearance for MRSA and vancomycin-resistant enterococci to be approximately 1 month in intensive care units across the United States [22]. An interesting observation in our current analysis is the contrast between the 2 units, with clearance being twice as fast in the BMT unit compared with that in the STU. Although we do not have direct data to inform this distinction, it is possible that other patient-care factors such as antibiotic exposure may influence this clearance differential. Direct estimates of the transmission rate for C. difficile are less common than those for clearance; however, we and others have indirectly estimated transmission through calibration of a dynamic transmission models to observed rates [23, 24]. These estimates, together with direct estimates of the transmission rate parameter for other antibiotic-resistant organisms, are in the same range [11, 22, 25].
Our sensitivity analysis indicates that the estimated parameters were influenced by model assumptions of surveillance test sensitivity and specificity. We opted for investigating such a small range of specificity because false positives tend to be rare enough to be ignored [25–29]. Nonetheless, there are reasons why false positives may arise in practice; for that reason, we explored their impact on the model across a plausible range. Although the rate of false positives is also assumed to be small, the number of patients at risk for a false positive can be large relative to acquisitions or importation, due to the fact that the patients at risk for a false positive (those who are uncolonized at the time of a test) represent most of the inpatients on the unit and acquisitions, which are imperfectly observed are rare. Thus, even a small false-positive probability can lead to a dramatically different interpretation of the underlying transmission dynamics. For example, increasing false positives lead to increased chances of observing a positive test followed by a negative test, thereby reducing the likelihood of true clearance, and a negative test followed by a positive test is less likely to represent a true acquisition.
This analysis represents an important initial step for better understanding transmission dynamics of C. difficile, which is important for understanding variation in overall burden of C. difficile. Because asymptomatic carriers can represent a high proportion of carriers in healthcare facilities, and can transmit C. difficile to other patients and contaminate environmental surfaces, and because acquisition is a necessary step in the development of CDI, understanding transmission dynamics is critical for refining infection control. Knowledge about which factors predict high burden or risk will enable the development of improved and more efficient practices for targeting infection control. For example, in this article, we show that the risk of transmission in the BMT unit is higher than in the STU, so targeting the spread of C. difficile in the BMT unit would be an appropriate strategy. Yet, if we understood why clearance occurs rapidly in the BMT unit, it would help to inform improved control for patients already colonized in the STU. We plan to extend this work by obtaining and incorporating patient and patient-care factors into our model, giving us a more nuanced interpretation of the transmission dynamics and a better understanding of infection control strategies that may work in these settings.
Our results should be interpreted in light of some limitations. Due to the relatively small amount of longitudinal data, we were not able to estimate surveillance test parameters but instead had to make assumptions about the test sensitivity and specificity parameters. Because our estimates of model parameters, including transmission and clearance, were based on longitudinal data, they were largely influenced by patients having long lengths of stay, although in reality long-stay patients are probably more likely to be colonized and drive transmission dynamics. Additionally, factors that impact transmission dynamics, such as antimicrobial exposure, were not explicitly incorporated into our analysis.
In summary, we found differences in the transmission dynamics that govern C. difficile transmission in 2 units in a large hospital. These differences are likely due to patient-level or unit-level factors. We also found that sensitivity and specificity are important parameters for understanding the dynamics of transmission. Therefore, estimates of surveillance test sensitivity and specificity using real data in healthcare settings would increase our understanding. Future studies should incorporate patient level data or care tasks into these models to expand our understanding of the transmission dynamics of C. difficile in inpatient units.
Supplementary Data
Supplementary materials are available at Clinical Infectious Diseases online. Consisting of data provided by the authors to benefit the reader, the posted materials are not copyedited and are the sole responsibility of the authors, so questions or comments should be addressed to the corresponding author.
Notes
Presented in part: European Congress of Clinical Microbiology and Infectious Diseases, Amsterdam, The Netherlands, 2019.
Acknowledgments. The authors acknowledge and thank the Medical College of Wisconsin (MCW) Collaborative for Healthcare Delivery Science and the MCW Cancer Center for their support in this work. Additionally, the authors thank Carrie Edlund, MS, for her many iterations reviewing and editing this manuscript, and to Jeanette Young for her time, attention, and expertise on the development of the model figure (Figure 1).
Disclaimer. The views expressed in this article are those of the authors and do not necessarily reflect the position or policy of the Department of Veterans Affairs or the United States government.
Financial support. This work was supported by the Department of Veterans Affairs, Veterans Health Administration, Health Services Research and Development Service (project REA 08-264); the Centers for Disease Control and Prevention (Prevention Epicenters award number U54CK000456 and MIND-Healthcare Program award number U01CK000538); the University of Utah Study Design and Biostatistics Center, with funding in part from the National Center for Research Resources and the National Center for Advancing Translational Sciences, National Institutes of Health (grant number 8UL1TR000105); the Center for Advancing Population Science at the Medical College of Wisconsin; and grants from Advancing a Healthier Wisconsin Endowment (to A. N.).
Supplement sponsorship. This supplement is sponsored by the Informatics, Decision-Enhancement and Analytic Sciences (IDEAS) Center.
Potential conflicts of interest. S. S. reports personal fees as a consultant for AstraZeneca. All other authors report no potential conflicts of interest.
All authors have submitted the ICMJE Form for Disclosure of Potential Conflicts of Interest. Conflicts that the editors consider relevant to the content of the manuscript have been disclosed.
