Associations Between Short or Long Length of Stay and 30-Day Readmission and Mortality in Hospitalized Patients With Heart Failure

Data sources

Patients were identified using the Canadian Institute for Health Information Discharge Abstract Database (CIHI-DAD), which contains information on all hospital admissions in Ontario, Canada. Admissions with a primary or most responsible diagnosis of HF (International Classification of Diseases-Tenth Revision Canadian Enhancement code I50) were identified. Code I50 has demonstrated a high degree of accuracy compared with the clinical Framingham criteria (11,12). Using the patients’ unique encoded health card number, we linked the National Ambulatory Care Reporting System for emergency department (ED) visits, the Registered Persons Database for mortality, and the Ontario Registrar General Vital Statistics Database for cardiovascular deaths. The Ontario Drug Benefit Database was used to identify prescription medications, and the Ontario Diabetes Database was used to identify patients with diabetes. Diagnostic and interventional cardiac procedures were identified using the CIHI-DAD, the CIHI Same Day Surgery Database, and the Ontario Health Insurance Plan physician claims database. The accuracy of these databases has been described previously (13–15).

Study population

We conducted a population-based analysis between April 1, 2003 and March 31, 2012 of adult patients ≥65 years of age and residing in Ontario, Canada who were hospitalized with a primary diagnosis of HF (n = 178,905). To capture acute presentations of HF, hospitalizations in which patients were not admitted directly from the ED were excluded (n = 37,268). In the event of multiple HF admissions for the same patient, only the first admission was used (n = 48,786). We studied patients discharged home to independent living, excluding those who died before discharge from our primary cohort. Patients were also excluded if they were not discharged home (i.e., to another facility) from the hospital (n = 16,431), or were transferred from or to another acute care or long-term care facility (n = 1,305). Patients with missing demographic data (n = 228), left ventricular assist device or cardiac transplant recipients (n = 21), and patients with an admission date ≥1 days after the ED presentation date (n = 75) were also excluded. Because an increased risk was observed previously for very prolonged hospitalizations (i.e., >14 days) (16), we excluded these patients in the primary analysis. The primary study cohort consisted of 58,230 patients with a hospital LOS of 1 to 14 days.

Definitions and outcomes

We identified patients’ comorbidities within 3 years before the index HF-related hospitalization by using the CIHI-DAD (17–19). Socioeconomic status was determined by quintile of median neighborhood income, and the admitting unit was defined as the intensive or coronary care unit if the first day of admission in the CIHI-DAD was coded as a special care unit. Hospital type was defined by the classification system of the Ontario Hospital Association (20).

The primary outcomes were cause-specific 30-day hospital readmissions for cardiovascular disease and noncardiovascular causes, as previously defined (21,22). Secondary outcomes included readmissions for HF and all causes. We also analyzed 30-day all-cause mortality data from hospital discharge, subdivided into cardiovascular and noncardiovascular mortality as described previously (23).

Statistical analysis

Continuous variables were summarized as medians with interquartile ranges (interquartile range [IQR]: 25th, 75th percentiles) and compared between exposure categories using the Kruskal-Wallis test. Categorical variables were compared using the chi-square test. Hospital admission rates and 95% confidence intervals (CIs) were determined using the γ-distribution (24).

We used a marginal cause-specific hazard regression model with a robust (sandwich-type) variance estimator to account for clustering of patients within hospitals. We modeled the effect of LOS on the hazard of hospital readmission for a specific cause and treated mortality as a competing event (25). We then modeled the effect of LOS on the hazard of cardiovascular mortality and treated noncardiovascular mortality as a competing risk (and vice versa). Time was measured from the date of index hospital discharge.

Initially LOS was treated as a continuous predictor in the cause-specific Cox regression model. To test the hypothesis that the relationship between LOS and each outcome was nonlinear, we modeled LOS with restricted cubic splines with 3 knots at the 10th, 50th, and 90th percentiles corresponding to 2, 5, and 11 days, respectively. When reporting hazard ratios (HRs) with LOS as a continuous variable, the mean LOS (5.7 days) was used as the reference value. We performed a Wald test to test the null hypothesis that the relationship between LOS and the hazard of each outcome was linear. Subsequently, LOS was modeled using the following categories: 1 to 2 days, 3 to 4 days, 5 to 6 days, 7 to 8 days, and 9 to 14 days. We chose 5 to 6 days as the reference category because it reflected the mean and median LOS of our cohort and in previous reports (26).

Multivariable models were adjusted for variables identified in published risk models, including demographics (age, sex), cardiovascular comorbidities (myocardial infarction, angina, unstable angina, chronic atherosclerosis, coronary revascularization, implantable cardiac defibrillator, permanent pacemaker, hypertension, cerebrovascular disease, peripheral vascular disease, valvular and rheumatic heart disease, arrhythmias, and cardiopulmonary-respiratory failure and shock), and noncardiovascular comorbidities (diabetes, cancer, pneumonia, trauma, major psychiatric disorders, decubitus skin ulcers, chronic obstructive pulmonary disease, rheumatologic disease, renal disease, gastrointestinal disorders) (17–19). We adjusted for HF history according to whether there was a history of HF without previous hospitalization, 1 HF hospitalization, ≥2 HF-related hospitalizations, or no diagnosis of HF in either ambulatory or hospital-based settings within the previous 3 years. Additionally we adjusted for the following: day and year of admission; socioeconomic status; admitting unit (ward or intensive care unit); specialty of the responsible physician; and pre-admission medications (beta-adrenergic receptor blockers, angiotensin-converting enzyme inhibitors or angiotensin receptor blockers, digoxin, furosemide, metolazone, spironolactone, antiplatelet agents, warfarin, 3-hydroxy-3-methyl-glutaryl-coenzyme A reductase inhibitors) (Online Appendix).

To explore characteristics of patients with short or long LOS, we used multiple logistic regression models, with LOS category as the outcome. We developed parsimonious models for short LOS (1 to 2 days) versus nonshort LOS (3 to 14 days) (27). Then we developed a second model to predict long LOS (9 to 14 days) versus shorter LOS (1 to 8 days). Candidate predictor variables were considered for entry into the model if the univariate p value was <0.25, and they were retained in the final model if the multivariable p value was <0.05.

We performed sensitivity analyses to determine the robustness of our findings. We modeled the effects of LOS on each outcome, as described earlier, including the following: 1) patients with an LOS up to 35 days (n = 64,165); 2) patients ≥18 years of age (n = 71,586 without outpatient medication data); and 3) all index and repeat HF-related hospitalizations during the study period (n = 85,885). This last analysis of repeat 30-day hospitalizations was performed for each HF-related hospitalization occurring over the duration of follow-up, by resetting time to zero with each new hospital admission.

A 2-sided p value < 0.05 was considered statistically significant. All analyses were performed using SAS software version 9.3 (SAS Institute, Inc., Cary, North Carolina). Research ethics board approval was obtained from Sunnybrook Health Sciences Centre in Toronto.