Association between intrahospital transfer and hospital-acquired infection in the elderly: a retrospective case–control study in a UK hospital network

Background Intrahospital transfers have become more common as hospital staff balance patient needs with bed availability. However, this may leave patients more vulnerable to potential pathogen transmission routes via increased exposure to contaminated surfaces and contacts with individuals. Objective This study aimed to quantify the association between the number of intrahospital transfers undergone during a hospital spell and the development of a hospital-acquired infection (HAI). Methods A retrospective case–control study was conducted using data extracted from electronic health records and microbiology cultures of non-elective, medical admissions to a large urban hospital network which consists of three hospital sites between 2015 and 2018 (n=24 240). As elderly patients comprise a large proportion of hospital users and are a high-risk population for HAIs, the analysis focused on those aged 65 years or over. Logistic regression was conducted to obtain the OR for developing an HAI as a function of intrahospital transfers until onset of HAI for cases, or hospital discharge for controls, while controlling for age, gender, time at risk, Elixhauser comorbidities, hospital site of admission, specialty of the dominant healthcare professional providing care, intensive care admission, total number of procedures and discharge destination. Results Of the 24 240 spells, 2877 cases were included in the analysis. 72.2% of spells contained at least one intrahospital transfer. On multivariable analysis, each additional intrahospital transfer increased the odds of acquiring an HAI by 9% (OR=1.09; 95% CI 1.05 to 1.13). Conclusion Intrahospital transfers are associated with increased odds of developing an HAI. Strategies for minimising intrahospital transfers should be considered, and further research is needed to identify unnecessary transfers. Their reduction may diminish spread of contagious pathogens in the hospital environment.

: Flow chart depicting data linking, selection criteria and number of spells included in the analysis.  We concluded that while CDI cases compose a large proportion of the cases, they are not influential drivers of the effect between intrahospital transfers and HAI in these data.

Text 1: Correlation between bed-transfers and ward-transfers
Unlike the ward-transfer variable in our dataset, the bed-transfer variable was not time stamped. Once a case entered the ward on which their positive sample was collected, it was not possible to tell which bed transfers happened prior to the diagnosis and which happened after the diagnosis, and possibly due to it. With this limitation in mind, we conducted an exploratory sensitivity analysis in which we included any bed changes which occurred up to the ward entry time of the ward on which a case had their sample collected (therefore likely undercounting the number of bed-transfers for cases). Additionally, the bed-transfer variable captures instances where a patient is moved to a trolley while visiting a diagnostic area, and so includes some between, as well as within ward movements. Therefore, any changes of environment including a change of ward were counted in this sensitivity analysis. Results showed that bed/inter-ward transfers were highly correlated with ward-transfers (Pearson correlation = 0.998, Spearman correlation = 0.997). The residual intraclass correlation coefficient (ICC), which computes the proportion of variability explained by the presence of clusters was checked at the patient, treatment function code (TFC) and hospital level, and as clustering was found to be minimal a logistic regression was chosen as the final model. However, several other models were run for data exploration. Univariable and multivariable regressions were run using Stata"s estimation command with the vce(cluster clustvar) option to obtain a robust variance estimate that adjusts for within-cluster correlation at the patient level (Table 4A). In addition, univariable and multivariable logistic multilevel models (hospital spells within TFCs, and hospital spells within hospital site) were run with both random intercepts for TFCs and hospital sites, and with random slope for intrahospital transfers (Table 4B-D), using Stata"s melogit command.
The model for hospital site as a second-level cluster and intrahospital transfers as a random slope did not converge. All multivariable model results were adjusted for: age, gender, time at risk, Elixhauser comorbidities, hospital site of admission, dominant treatment function code, intensive care unit admission, number of procedures and discharge destination.

While Office of Population Censuses and Surveys Classification of Interventions and
Procedures (OPCS-4) intervention codes capture a wide range of procedures, they do not weigh majorly invasive procedures differently to minimally invasive procedures. They are broadly categorised into major, intermediate, minor and non-operative procedures, but the individual codes themselves are not classified. The national clinical coding standard does however provide some key words associated with each of these groups. Therefore, the OPCS-4 code descriptions were searched for the terms "Total removal", "Replacement", "Transplant", "Partial removal", "Destruction" and "Reconstruction Repair" which are associated with major and intermediate procedures. This returned 1,250 individual codes.
Spells which contained these codes at any point were removed from this analysis (n=1091) resulting in a final patient sample of 29,427.

Sensitivity analysis using only positive samples collected from sterile sites:
In order to address the potential for misclassification due to colonisations in all types of pathogens, we conducted a sensitivity analysis in which we restricted cases to positive samples collected from a sterile site. A sterile site included any pathogens identified from a blood or urine sample. Urine from catheters and nephrostomy bags were included, however samples associated with all other devices were excluded. Patients with an infection identified from any other source were assumed to be colonisations or contaminations, rather than true pathogenic infections, and allocated to the control group. This necessarily excluded any infections detected by targeted samples (such as CDI infections which are identified from stool samples) from the case count, which may have resulted in some controls being misclassified. The site of the sample collection in our data was a free text variable, giving over 2000 unique sites. We firstly filtered for blood and urine, and then excluded any sites that appeared to be associated with an indwelling device (e.g. blood line), or a wound/clot. We also note that it was assumed aseptic technique was performed during the sample collection, and a limitation of this analysis is that this cannot be confirmed. Results showed increase in the observed OR from 1.09 (95%CI 1.05 to 1.13) in the primary analysis to 1.11 (95% CI 1.06 to 1.16) ( Table 7). Although this estimate is less precise it suggests that some cases in our main analysis are misclassified colonisations, and therefore nondifferential misclassification of the outcome has occurred (as this misclassification does not depend on the main exposure) resulting in a lower OR.   Exploratory covariate for periods of higher admissions:

Univariable Results
Seasonal effects which correlate with higher numbers of hospital-acquired infections (HAIs) would ideally have been explored as a covariate, but we were unable account for seasonality in order to respect the terms set by the Data Protection Office, which stipulated that dates of patient admission remain de-identified.
An exploratory analysis was conducted in which admission dates covering the 36-months of admissions were allotted into four quarters of consecutive three-month intervals for each year, creating a categorical variable. We produced three such variables, the first of which began from Day 0 in the data, the second from day 30 and the third day 60, in order to account for the possibility that the data collection began mid-quarter. We then chose the variable with the largest difference in admissions between the smallest and largest quarter, which therefore best discriminated between 3-month intervals with higher and lower number of admissions (Table 10A). As influxes in admissions of any kind, not only the patients in our final cohort, would have contributed to resource strain we used everyone available in the dataset for this analysis (with the exclusion of some erroneous spells and those that belonged to the specialist hospitals). This variable was found to be a significant predictor of HAI development, but it did not impact the association between intrahospital transfers and HAI development (Table 10B).