Methods

Study design and population

The present study was a post-hoc analysis of the MedBridge trial [11,12]. The trial was conducted from February 2017 until October 2018 at eight wards within four hospitals in Sweden: Uppsala University Hospital and the hospitals in Enköping, Gävle and Västerås. The wards differed in terms of medical specialty: internal medicine (three wards), stroke and neurology (two wards), acute internal medicine, diabetes and nephrology, and geriatrics. The trial population (n=2,637; median age 81 years; median number of medications 9) was used to identify risk factors for drug-related readmissions. To assess preventability, Microsoft Excel was used to randomly select a sample of 400 patients from among all trial participants, stratified by county (hospital): 200 from Uppsala County (Uppsala and Enköping), 100 from Gävleborg County (Gävle) and 100 from Västmanland County (Västerås). We aimed for a representative sample, but no formal sample size calculation was performed.

Outcomes, data collection and assessment

Baseline (index admission) and outcome data were extracted from the patients’ electronic health records (EHRs) and the counties’ healthcare registries. The primary outcome for risk factor analysis was experiencing a possibly drug-related readmission within 12 months after hospital discharge from the index admission. In the MedBridge trial, all participants’ unplanned hospital readmissions were assessed by a pair of final-year pharmacy students with a validated tool to identify readmissions that were possible drug-related or unlikely to be drug-related (AT-HARM10, [13]). In case of doubt, an experienced clinical pharmacist was available to cast a deciding vote. In a validation study, the tool’s inter-rater reliability was moderate to substantial (Cohen’s kappa values within pairs were between 0.45 and 0.75 and Fleiss’ kappa values between pairs were between 0.46 and 0.58 [13]). Sensitivity, specificity and positive and negative predictive values were between 70% and 86%. In the present study, all possibly drug-related readmissions were used as the primary outcome. Secondary outcomes were all-cause unplanned hospital readmissions and all-cause ED visits.
The assessment of preventability of drug-related revisits followed a stepwise approach:

Statistical data analysis

Categorical variables were summarised as frequencies and percentages. Numerical variables were summarised as mean, median, standard deviation and quartile. To investigate differences in baseline characteristics (potential risk factors) for each primary (drug-related readmission) and secondary outcome (all-cause readmission and all-cause ED visit), categorical baseline variables were compared using the χ² test and continuous variables using the Wilcoxon non-parametric test. Baseline characteristics included were sociodemographics, unplanned hospital visits within 12 months prior to admission, diagnoses in medical history, medication use, estimated glomerular filtration rate (eGFR) upon admission, length of hospital stay and discharge diagnosis (full list of variables in Supporting Information S2). In order to a test for multicollinearity, the Cramer’s V correlation and Point-Biserial correlation were calculated. Highly correlated variables were not used in the same model.
A multivariate Cox proportional hazards model was developed for each primary and secondary outcome, with adjustment for MedBridge trial treatment group. All baseline characteristics that were significant in the univariate test were initially included. All non-significant variables were then removed from the multivariate model in a stepwise way, starting with the least significant, until only significant characteristics remained. Hazard ratios (HRs) and 95% confidence intervals (CIs) were calculated. The underlying proportional hazards assumptions of the Cox proportional hazards models were verified by visual inspection of Schoenfeld residuals. Significance was specified as p < 0.05. All statistical analyses were performed with R version 3.5.1 (R Foundation for Statistical Computing, Vienna, Austria). For preventability of drug-related visits, descriptive statistics were analysed with Microsoft Excel.