Study design
In this retrospective cohort study, the sample included all episodes of inpatient care experienced by patients admitted from the ED to the inpatient setting at a 420-bed hospital in southern Sweden during 2011–2012 and discharged before 1 December 2012. The 30-day readmission rate for discharged patients at this hospital was previously estimated to be about 9 %. For the purposes of this study, an increase of 2 % was considered to be clinically relevant.
To limit bias, the study material was not subject to further restrictions. Post hoc power calculations were performed to determine the number of strata (see cut-offs in the Statistical Analysis section) of inpatient bed occupancy to use for group comparisons (α = 0.05, 1-β = 0.80) [17].
Data sources
Data on inpatient care episodes were retrieved from the hospital billing system PASiS®. Data on hospital occupancy per hour were retrieved from an occupancy database used by hospital management for purposes of quality assurance. Data on ED visits were retrieved from the ED information system Patientliggaren®. Data gathering and linking were performed by the hospital informatics unit using QlikView® software. The head of the division (KI) and the chair of the ED (FJ) granted access to data.
Setting
Helsingborg general Hospital is one of four hospitals that provide emergency care in the region of southern Sweden called Region Skåne. Its ED serves a population of roughly 250,000, which expands to more than 300,000 in summer due to tourism. The hospital is a teaching hospital that offers education for medical students as well as emergency medicine residents.
Its ED is separated into units by specialty, and in 2010, a complementary unit staffed by emergency physicians capable of handling all but psychiatric, otolaryngologic, ophthalmologic, and pediatric complaints was introduced that currently operates from 8:00–23:00 daily. There are separate EDs for children (<18 years of age) with medical conditions and for patients with obstetric, gynecologic, psychiatric, or ophthalmologic complaints. Patients admitted from these EDs were excluded from the study. Patients with suspected hip fractures or ST-elevation Myocardial Infarction (STEMI) diagnosed in an ambulance bypass the ED and were thus also excluded. Hand surgery, neurosurgery, and thoracic surgery are not available at the hospital, and the availability of endovascular surgery and PCI are limited after hours (17:00–08:00). Patients with such needs are referred to Skåne University Hospital (SUS) and were thus also excluded. At times of pronounced bed shortage, some patients are admitted from the ED to two other hospitals in the region. Patients admitted to these hospitals at index were likewise excluded.
Statistical analysis
Unplanned readmissions were defined as readmissions to the hospital through the ED within 30 days of discharge. We computed the readmission rate for inpatient bed occupancy rates of <95 %, 95–100 %, 100–105 %, and >105 % and compared proportions using Fisher’s exact test. Inpatient bed occupancy <85 % has traditionally been used for the reference level in the field, following Bagust et al [4]. Since the mean bed occupancy at the study site is around 95 %, <85 % is likely to reflect an artificial situation and hence <95 % was selected for reference. Post hoc power analysis revealed that the power to detect the pre-specified difference (2 %) was 84.2 % for the smallest category (>105 %). A binary logistic regression model was constructed in order to adjust for confounders and other factors liable to affect the outcome. Variables considered for inclusion in the model were sex, age group, IPLOS, the admitting specialty at index admission, day of the week of discharge, time of day of discharge (00:00-07:59, 08:00-15:59, 16:00-23:59), and inpatient bed occupancy at discharge. Age was grouped into intervals of 0–18 years, 18–40 years, 40–65 years, and ≥65 years. In Sweden, 18 years is the age of legal adulthood and 65 the age of retirement. For the binary logistic regression models, inpatient bed occupancy was categorized by the same intervals as in the crude analysis.
Predictors were tested for crude association with the outcome before entering the preliminary primary effects model. Associations weaker than p = 0.25 but of clinical importance were still included [18]. Multicollinearity testing was performed using Spearman correlation [19], and the selection of interaction terms screened for inclusion in the final models was governed by perceived clinical significance determined a priori. Variables were manually added to the models. Model fit was evaluated using Nagelkerke’s R2, the Hosmer & Lemeshow test and ROC-curves. The association between each predictor and the outcome was addressed by the -2LL and Wald statistic. Models were screened for influential cases by addressing standardized residuals and Cook’s distance. To prevent overfitting, the final model selected was that with the highest explanatory value relative to the number of predictors [19]. Statistical analyses were performed with IBM® SPSS® version 22. Data was anonymized before analysis. The regional ethical review board in Lund granted ethical approval for the study (dnr 2013/11).