A retrospective observational cohort study was performed using chart abstraction data linked to administrative datasets held at the Institute of Clinical Evaluative Sciences (ICES), Toronto, Ontario, Canada.
Data sources
We conducted a chart abstraction study in Ontario to capture all TAVR procedures performed from 2007 to 2013 [9, 10]. The database included detailed demographic, clinical and procedural data. Data from the clinical database were linked using unique encoded identifiers to administrative databases and analyzed at ICES to protect patient confidentiality. This comprehensive chart abstraction was performed by six trained nurse abstractors using standardized data definitions.
The Canadian Institutes for Health Information Discharge Abstract Database was used to identify hospital readmissions and comorbidities. Validated disease state definitions were used to establish comorbidities including diabetes, hypertension, and heart failure [11,12,13]. Frailty was identified based on John Hopkins Ambulatory Clinical Groups software which uses a proprietary set of diagnostic codes found in administrative data [14]. The Ontario Health Insurance Plan and the ICES Physician Database were used to identify physician visits and their specialty, respectively. The Home Care Database and the National Rehabilitation Database facilitated identification of home care services and use of in-patient rehabilitation. Finally, the Registered Persons Database was used to assess vital status.
Cohort
All patients who underwent TAVR in Ontario between April 1st, 2007 and March 31st, 2013 were included. Patients were excluded if they died prior to discharge during the index hospitalization. All patients had a minimum follow-up of 1 year.
Outcomes
The primary outcome was all-cause readmission within 1 year of discharge from the index TAVR hospitalization.
Transitional care factors
To evaluate the association between transitional care interventions and hospital readmission after TAVR, we assessed the following factors: follow-up with a family physician, follow-up with a cardiologist, use of home care services and in-patient rehabilitation. Home care services were considered to be present if any in-home service was delivered after discharge and includes: nurse visitation, physiotherapy, occupational therapy, personal support worker visits, as well as homemaking services. Due to the inability to accurately determine whether a physician visit occurred during hospitalization or as an out-patient, visits that occurred on the same day as hospital admission were excluded.
Statistical analysis
Patient characteristics were compared based on the presence or absence of readmission within 1 year. Standard descriptive statistics were used. Causes of readmission were characterized based on aggregate diagnoses using groupings of ICD-10 codes [9].
Cause-specific proportional hazards models were used for modeling the hazard of readmission, treating death prior to readmission as a competing risk (known to be 20% at 1 year) [3]. This approach models the instantaneous hazard of an of event of interest in subjects who are currently event-free (i.e, who are alive and have not yet been readmitted to hospital). This strategy is appropriate when studying ‘etiologic’ associations [15]. Variables for inclusion in the regression model were chosen ‘a priori’, based on the clinical knowledge of content experts and prior literature [2, 16,17,18]. The following variables were included: demographics (age, sex, living status), clinical characteristics [frailty, New York Heart Association (NYHA) functional class, left ventricular function], comorbidities (diabetes, dementia, heart failure, myocardial infarction, atrial fibrillation, lung disease, cerebrovascular disease, dialysis, peripheral vascular disease, liver disease, peptic ulcer disease, bleeding), previous cardiac interventions [percutaneous coronary intervention (PCI), coronary-aortic bypass surgery], lab values [hemoglobin, estimated glomerular filtration rate (eGFR)], prior health care utilization (hospitalization 30 days prior to TAVR), procedural factors (valve-in-valve, elective vs urgent, vascular access site, self-expandable vs balloon expandable valve), post-procedural complications (permanent pacemaker, bleeding/vascular complication/transfusion, stroke, delirium), post-TAVR echocardiographic findings (aortic regurgitation, mitral regurgitation), transitional care factors and year of the TAVR procedure. Collinearity was assessed using variance inflation factors. Time to first readmission was modelled in all analyses. Patients were censored at maximum follow-up or if death occurred before readmission. Transitional care factors were included in the regression model as time-varying covariates. Robust sandwich variance estimates were used to account for clustering of patients within hospitals. The incidence of readmission and death over time were estimated using cumulative incidence function (CIF) curves.
Two sensitivity analyses were performed. First, due to the concern that factors such as physician follow-up were in the causal pathway for readmission and to better understand the influence of transitional care factors, we performed a landmark analysis assessing outcomes occurring between day 31 and day 365. In this analysis, patients who were readmitted or died within the first 30 days were excluded. Transitional care factors were operationalized in a dichotomous fashion based on their occurrence within the first 30 days. Secondly, due to concern that early TAVR experience may not be representative of current practice, we selected a cohort of patients who underwent TAVR after 2010 for separate analysis. The primary analysis was repeated in this sub-cohort to ensure overall consistency in our results.
All analyses were performed using SAS statistical software, version 9.3 (SAS Institute, Cary, NC). A 2-tailed p value of < 0.05 was considered significant.
