Healthcare-associated infections

Patient data were collected by individual case report forms, and HAIs were defined as outlined in the ECDC protocol version 5.3. ^{Reference Zingg, Metsini and Balmelli3,9} Healthcare-associated infections were stratified into bloodstream infections (BSIs), lower respiratory tract infections (LRIs), surgical site infections (SSIs), urinary tract infections (UTIs), and other HAIs. In case of multiple HAIs, only the first was considered because only this is required for the incidence proportion and because we were interested in length of stay (LOS) with versus without acquiring an HAI. We distinguished between HAI present at admission, HAI occurring during hospitalization and up to the PPS (prevalent HAI), and HAI occurring during hospitalization but after the PPS (incident HAI). HAIs present at admission (ie, previously acquired at a different healthcare facility) were modeled separately from HAIs that occurred during hospital stay because these events represent distinct at-risk populations.

Point-prevalence survey data

The Swiss 2017 PPS database collected data from April to June from 12,931 patients in 96 acute-care hospitals, including information on age, sex, date of hospital admission, date of HAI, type of HAI, patient specialty, surgery since admission, and McCabe score. ^{Reference Zingg, Metsini and Balmelli3} Data at the hospital level included hospital size and provision type. Patients in long-term care, geriatrics, and rehabilitation were excluded, and length of stay was censored after 90 days to meet the definition of acute care.

Steady-state assumption

Our model makes use of the steady-state assumption, which, in broad terms, asserts that the number of patients entering and leaving a hospital is balanced at any period, and that the distribution of LOS does not change over time. The assumption should be confirmed, ^{Reference Doerken, Mandel, Zingg and Wolkewitz10} which can be done by comparing the distribution time from admission to PPS to the distribution time from PPS to discharge. ^{Reference Asgharian, Wolfson and Zhang11} Under a steady state, the distributions are the same. We confirmed that the steady-state assumption was met using the follow-up data (Supplementary Fig. 1 online).

Imputation

Follow-up information beyond the PPS was imputed using data for 1,714 patients from 3 hospitals that provided follow-up data sets of the patients beyond the date of the PPS (Bülach, Geneva, Zurich) (Fig. 1). Patients from Geneva and Zurich represented large Swiss hospitals, and patients from Bülach, a regional hospital, represented small to medium-sized hospitals (Table 1). Imputation was stratified by prevalent HAI type. For every patient with missing follow-up with a prevalent infection, we imputed the following:

1. 1. Time from PPS to discharge

Fig. 1.

Patient flow chart.

Table 1.

Descriptive Patient Summary Per Hospital

Note. HAI, healthcare-associated infection; ICU, intensive care unit.

^a Percentage of “ulimately fatal” or “rapidly fatal.”

For every patient with missing follow-up without a prevalent infection, we imputed the following:

1. 1. Time from PPS to discharge

2. 2. Time from PPS to incident infection

3. 3. Time from incident infection to discharge:

   • If discharge (1) occurred before incident infection (2), the patient was considered infection free.

   • If discharge (1) occurred after incident infection (2), the patient was considered undergoing an infection and to remain in hospital with additional time from incident infection to discharge (3).

Time to event T was modeled using an accelerated failure time model with a Weibull link function, $T\sim Weibull\left( {\lambda,\sigma } \right)$ with shape parameter λ and scale parameters $\sigma \left( x \right) = {\beta _0} + {\beta _1}{x_1} + \ldots + {\beta _p}{x_p}$ with covariate vector x. ^{Reference Kalbfleisch and Prentice12}

Time from PPS to infection and time from infection to discharge

To determine the time from PPS to incident infection and the time from incident infection to discharge, we fitted models to the follow-up data that were available using age, McCabe score, surgery since admission, and time from admission to PPS as covariates.

We drew from the distributions of our imputation models 5 times for each patient to generate 5 augmented data sets. ^{Reference Buuren and Groothuis-Oudshoorn13} Figure 2 is a depiction of the length of stay, showing both original and imputed data, of 1 augmented data set.

Fig. 2.

Visualization of point-prevalence survey (PPS) data. As part of the PPS only retrospective information was available. For a subset of patients from Bülach, Geneva, and Zurich, prospective information beyond PPS date on date of infection and discharge was made available. From this information, we developed a model to impute prospective information for the remaining patients. Note that this figure depicts only 1of 5 imputed data sets.

Length bias and weighting

After creating an augmented data set with follow-up information, length bias was corrected by weighting patients inversely to the LOS: overrepresented long LOSs were adjusted downward, and underrepresented short LOSs were adjusted upward. These weights were included as probability weights in the logistic and Cox regression models. ^{Reference Fluss, Mandel and Freedman14} Patients who were censored were given a weight of 0 because their LOS was not known; thus, they were effectively excluded from the model. ^{Reference De Uña-Álvarez15} For analyzing attributable LOS, weight adjustments were conducted using a data set in which observations were duplicated inversely proportionally to their LOS.

Incidence proportion and hazard rates

The incidence proportion of HAI was estimated using a mixed-effects multivariable logistic regression model, including hospital as a random effect. Hazard rates were estimated in a multistate multivariable Cox model. ^{Reference Wolkewitz, von Cube and Schumacher8} Patients were weighted to correct the length bias. ^{Reference Wang, Armitage and Colton16} Models were fitted to each of the 5 augmented data sets and the 5 data sets were pooled to obtain results. ^{Reference Buuren and Groothuis-Oudshoorn13} The confidence intervals of incidence proportions and hazard rates were calculated using robust estimates of the standard error.

Attributable length of stay

To evaluate the impact of HAIs on hospital stay, we compared the expected subsequent stay given infection status at day d ^{Reference Schulgen and Schumacher17} : patients who acquired an infection after d days were compared to infection-free patients with at least d days of hospitalization. To account for the time-dynamic at-risk sets, we compared the LOS using the transition hazards in a multistate model with transitions from hospital admission to infection and discharge. ^{Reference Allignol, Schumacher and Beyersmann18,Reference Allignol, Schumacher and Beyersmann19} Confidence intervals of attributable LOS were calculated by estimating the standard error by drawing 100 bootstrap samples in each of the 5 augmented data sets.

HAIs that were already present at admission had to be treated separately because the dates of onset were missing. To infer the attributable LOS for such patients, we matched patients with an HAI at admission to patients without an HAI at admission according to age, McCabe score, surgery, and time from admission to PPS.

We followed the STROBE guidelines for cross-sectional studies in reporting our findings. ^{Reference Von Elm, Altman and Egger20}

Simulations

Because the use of imputation for follow-up data and subsequent LOS-weighted regression are novel to the analysis of PPS data, we also applied our methods in an accompanying simulation study to confirm that they are appropriate (Supplementary Material online).

No institutional review board approval was deemed necessary, similar to the ECDC-PPS, given the quality-improvement nature of the survey. Only anonymous patient and ward data were collected and analyzed.