Materials and Methods
Data Source
Data for this study were obtained from the Healthcare Cost and Utilization Project (HCUP) State Inpatient Database (SID) maintained by the Agency for Healthcare Research and Quality (AHRQ). We focused analyses on data from California, as it is a large, ethnically diverse state with readmissions data reported annually in the SID. Discharge and visit records from the SID include de-identified information on patient demographics, expected payer, ZIP code of residence, diagnosis and procedure codes, admission source, disposition after hospitalization, length of stay, total charge for hospitalization, and hospital identification code. Cost of each hospitalization was calculated from total charges using hospital-specific cost-to-charge ratios provided by HCUP. The hospital-level cost-to-charge ratios were merged with the SID database by the hospital identification code, and the total costs were estimated by multiplying the total charges with the cost-to-charge ratio for each calendar year. The costs were adjusted to 2011 using the medical component of the Consumer Price Index. Hospital characteristics were obtained by linking the SID database with California Office of Statewide Health Planning and Development (OSHPD) Patient Discharge Pivot Database by hospital identification codes ("DHOSPID" in the SID to "oshpd_id" in the OSHPD Patient Discharge Pivot Database) for each calendar year.
Study Design
This was a retrospective cohort study of adult patients (age > 18) who were hospitalized in California for sepsis, CHF, and AMI during 2009–2011. Hospitalizations for sepsis were identified based on the presence of a compatible International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM) code as the principal diagnosis for hospital admission. The ICD-9-CM codes that were used for sepsis were based on the Martin implementation: 038 (septicemia), 020.0 (septicemic), 790.7 (bacteremia), 117.9 (disseminated fungal infection), 112.5 (disseminated Candida infection), and 112.81 (disseminated fungal endocarditis). These codes have been previously used in population-based studies of sepsis, and the 038 code was found to have a positive predictive value of 97.7% and a negative predictive value of 80.0% using the clinical definition of sepsis as the gold standard. In our study population, 98.5% of the sepsis hospitalizations were identified based on the 038 code. For sensitivity analysis, we also examined the frequency, cost, and risk factors of 30-day readmissions in 2011 using the Angus implementation to identify hospitalizations for severe sepsis. The ICD-9-CM codes used to define the cohorts with CHF (402.01, 402.11, 402.91, 404.01, 404.03, 404.11, 404.13, 404.91, 404.93, 428.XX) and AMI (410.XX) were consistent with the definitions used in the CMS reporting program and in previous studies.
Patients were excluded from the analysis if they were hospitalized at a non–acute care facility, younger than 18 years, out-of-state residents, or missing information on length of stay or the unique encrypted patient identifier used in the HCUP database (VisitLink). Hospitals that reported less than 10 annual hospitalizations for sepsis, CHF, or AMI were excluded for analyses of each condition. For calculating index admissions for 30-day readmissions, hospitalizations occurring in December of each year were excluded to allow a 30-day window for rehospitalization. Similarly, for calculation of 90-day readmission rates, hospitalizations occurring in October, November, and December were excluded. The study was approved as an exempt protocol by the institutional review board at Los Angeles Biomedical Research Institute.
Statistical Analysis
Outcomes. The primary outcome was the all-cause 30-day readmission rate following hospitalization for sepsis. The unit of analysis was individual hospitalizations. The 30-day readmission rate for sepsis was calculated by dividing the number of hospitalizations with at least one subsequent hospital stay within 30 days of an index hospitalization by the total number of index hospitalizations for sepsis. Index hospitalizations were defined as all hospital admissions with a primary diagnosis of sepsis that did not result in transfer to another acute care hospital or patient death. Hospitalizations resulting in a transfer to another hospital were excluded because in the SID database if a patient is admitted to one hospital and transferred to another facility, two hospitalization records are created (one for each hospital). We felt that despite being hospitalized at two separate facilities, this should be analyzed as one hospitalization with 30-day readmissions being ascribed to the discharging hospital. With the exception of the first hospitalization in each calendar year, hospitalizations were eligible to be both an index hospitalization for future readmissions and a readmission for a prior hospitalization. Similar approaches were used to calculate the 30-day readmission rates for CHF and AMI. We chose to compare sepsis with CHF and AMI because these, along with CAP, are among the high-risk conditions that are tracked in the CMS Readmissions Reduction Program. We did not compare sepsis with CAP in our study because these conditions are not mutually exclusive, and coding patterns for sepsis and CAP are likely to vary between hospitals and over time. Length of stay, hospital mortality, and costs of hospitalization were calculated for all hospitalizations, index hospitalizations, and 30-day readmissions for sepsis, CHF, and AMI. Because length of stay and costs of hospitalization were not normally distributed, both medians and means are presented.
Causes of Readmissions. The most common diagnoses on readmission following sepsis were identified based on the primary Clinical Classification Software (CCS) code for each rehospitalization, which collapses ICD-9-CM codes into clinically meaningful categories. The cumulative frequencies of the 20 most common CCS primary diagnosis codes for 7, 14, 30, and 90-day readmissions were examined for longitudinal changes in causes of readmission. We further consolidated the CCS diagnoses into seven clinical categories (infection, pulmonary, complications of care, cardiovascular, renal/genitourinary, gastrointestinal) to facilitate more meaningful data interpretation. The CCS diagnoses included in each clinical category are provided in Supplemental Table 1 (Supplemental Digital Content 1, http://links.lww.com/CCM/B362).
Risk Factors for 30-Day Readmissions. To examine the relationship between patient- and hospital-level factors and 30-day readmissions following sepsis, we generated hierarchical mixed-effects logistic regression models using presence of a 30-day readmission compared with no 30-day readmission as the dependent variable. The independent variables in the model were hospital-associated factors (number of beds, teaching status, proportion of minority patients, and hospital type), patient factors (age, ethnicity/race, gender, median income, rural or urban residence, and comorbid conditions), and healthcare systems and hospitalization-related factors (payer status, disposition after hospitalization, and length of stay). Comorbid conditions were identified using discharge data codes to calculate a Charlson comorbidity index. The proportion of racial and ethnic minority patients seen by each hospital was estimated from the number of Black, Hispanic, Asian, and Native American patients admitted compared with total number of hospital admissions. Individual hospitals were used as random effects in the models to adjust for clustering of hospitalizations. We generated separate models for sepsis (model 1), CHF (model 2), and AMI (model 3) using the same dependent and independent variables to examine whether the patient- and hospital-level associations identified in the models are specific to each disease or more broadly applicable to odds of 30-day readmissions. Performance of each model was assessed by the c-statistic, calculated as the area under the receiver operating characteristic curve based on fitted probabilities from the model and the true values. The data are presented as adjusted odds ratios (ORs) with 95% CIs. Adjusted ORs with a 95% CI excluding 1.00 were considered statistically significant.
Hospital-Level Variation in 30-Day Readmissions. To assess variation in readmission rates across hospitals, we created hierarchical logistic regression models to estimate the 30-day readmission rates for sepsis, CHF, and AMI in each hospital. We adjusted for case-mix by adding age and medical comorbidities as fixed effects and used individual hospitals as the random effect. These models were different than the models used to examine the association between patient- and hospital-level risk factors to the odds of 30-day readmissions as they only contained age and medical comorbidities as independent variables. We did not include patient race/ethnicity, gender, insurance status, or hospital characteristics in the models because such an adjustment may obscure factors that inappropriately influence the quality of care across hospitals. Using empirical Bayesian posterior estimates from the logistic regression model, we determined the predicted 30-day readmission rate (risk and reliability adjusted) and 95% CIs at the hospital level for an average hospitalization and displayed the ranked order of adjusted rates across hospitals in a caterpillar plot. We used the variance component for the hospital random effect to evaluate the statistical significance of the variation in 30-day readmission rates across hospitals (p < 0.05 for statistical significance). All data analyses were performed using JMP version 11.0 (SAS Institute, Cary, NC) and SAS for Windows version 9.4 (SAS Institute, Cary, NC).