Sites The Multicenter Hospitalist (MCH) study was a prospective multicenter observational study of the effect of hospitalist care on patients admitted to general medical services. Patient enrollment occurred between 1 July 2001 and 30 June 2003 at six geographically diverse academic medical centers: University of Chicago, University of Wisconsin, University of California at San Francisco, Brigham and Women’s Hospital, University of Iowa, and University of New Mexico. These sites were selected because hospitalist or non-hospitalist physicians (none of whom were cardiologists) cared for the majority of patients admitted to the general medical service at each center, and because they primarily used a system where patients were admitted to attending physicians essentially at random, according to day of the week. The MCH study was reviewed and approved by each centers’ institutional review board.
MCH Study Patients
Patients were eligible if they were admitted by a hospitalist or non-hospitalist physician, were 18 years of age or older, and were able to give consent themselves or had an appropriate proxy. Patients were admitted to a hospitalist or non-hospitalist based on a pre-determined call schedule. The MCH study excluded patients admitted specifically under the care of their primary care physician or specialty physician (e.g., oncologist), and those with Mini-Mental Status Examination score of 17 out of 22 or lower.15
Informed consent for chart abstraction and interviews was obtained from eligible patients.
Congestive Heart Failure Patients
Within the MCH-eligible patients, we retrospectively identified those with heart failure using International Classification of Diseases (ICD-9) diagnosis codes (Appendix 1
) assigned at discharge. Patients identified by ICD-9 codes were excluded if they had severe chronic obstructive pulmonary disease (defined as being oxygen dependent, on oral steroids, or having a forced expiratory volume in 1 s of <0.8 l/s), sepsis, fluid overload secondary to renal failure, constrictive pericardial disease, a cardiac surgery planned within 24 h of admission, or having had a thoracotomy in the preceding 2 weeks16
Data were obtained from administrative sources, patient interviews, chart abstractions, and the National Death Index (NDI) database. Administrative data were used to obtain dates of admission and discharge, diagnosis codes (used to identify patients with heart failure, as above), insurance type, age, race, and gender. Intake interviews collected socioeconomic information not available in administrative data (such as education), functional status, and comorbidity data. One month follow-up telephone interviews assessed whether or not the patient had any follow-up appointments or rehospitalizations.17
A NDI search was used to ascertain 30-day mortality from the date of hospital discharge.18
Congestive heart failure process and risk adjustment data (such as use of angiotensin-converting enzyme inhibitors at discharge or ejection fraction) were collected by chart abstraction. Principal investigators at each site were responsible for training and overseeing interviews and chart abstraction activities, with central oversight of data quality provided by the coordinating center.
Identification of Hospitalist and Non-Hospitalist Physicians
Hospitalists were defined as physicians whose primary focus is the care of general medical hospitalized patients, and whose activities include patient care, teaching, or research.19
Non-hospitalist physicians were most often outpatient general internal medicine faculty or non-cardiology subspecialists, who typically attended 1 month per year. Physicians were classified as hospitalists or non-hospitalists according to designations provided by each site. Physician designation was confirmed by site coordinators and linked to the attending physician at discharge using administrative data files.
We selected three process measures available in guidelines present at the time of the MCH, many of which are elements of current public-reporting initiatives.3,4,20
These processes included: (1) measurement of left ventricular ejection fraction (LVEF) anytime before or during hospital admission documented either by formal imaging study report, or within admission documentation, (2) prescription of an angiotensin-converting enzyme inhibitor (ACE-I) or angiotensin receptor blocker (ARB) at discharge for patients with left ventricular ejection fraction lower than 40%, highest creatinine <3.0 and highest potassium <5.5 in the 48 h prior to discharge, and (3) prescription of beta-blocker in patients with LVEF <40%. Patients were excluded from the two discharge medication measures if there was a documented allergy or adverse reactions. In addition, patients who expired, left against medical advice, or were discharged to hospice were also excluded.
We also examined two care processes highlighting coordination of care. First we assessed whether or not patients reported during the 30-day interview a physician visit within a month of discharge, a recommended element of the longitudinal care of heart failure patients.21
We limited the analysis to physician visits identified by patients as having been scheduled upon discharge. Next, we assessed whether formal inpatient cardiac consultation was obtained, based on literature suggesting improved outcomes with cardiology specialist consultation.22,23
Finally, as more stringent measures of quality, we determined the percentage of eligible patients who received all three care processes (LVEF assessment, ACE-I use, and beta-blocker use) and those who received all five measures (the first three plus care coordination measures).
Cardiac Test Utilization
Because guidelines recommend evaluation for cardiac ischemia in selected heart failure patients,21
we examined use of exercise stress testing (with or without scintigraphy or echocardiography) as well as the use of pharmacologic stress testing (with scintigraphy or echocardiography). In addition, we examined whether or not coronary artery catheterization, with or without percutaneous coronary interventions, was performed.
Cost, Length of Stay, and 30-Day Combined Readmissions and Mortality
Length of stay and cost data were obtained from administrative cost-accounting systems maintained at each site. Readmission within 30 days was defined using readmissions identified in administrative records combined with data collected at time of 30-day follow-up phone call. To guard against recall bias, self-report data were only included for non-site admissions.24
Additionally, patients who died within 30 days, or were discharged to hospice, were excluded from the readmission measure. The 30-day post-discharge mortality measure included all deaths identified by the NDI at discharge and up to 30-days, excluding patients discharged to hospice. For purposes of increasing statistical power, we combined both 30-day measures into a single outcome measure.
Statistical Analysis We first compared patient characteristics using chi-square tests for categorical variables and t- or Mann-Whitney rank sum tests for continuous variables. A p-value of ≤0.05 was considered statistically significant.
Next, we performed multivariable analyses to determine the independent association between hospitalist care and the odds of the patient receiving any process or outcome measures, after adjusting for confounding variables and accounting for clustering at the physician level in generalized estimating equation (GEE) regression models. Cost and length of stay regression models were also fitted using GEE, but the models were built on gamma distributions with log-link functions.
We tested socio-demographic, co-morbidities, and physiologic variables for inclusion in each model (see Appendix 2
). To minimize any potential bias or loss of power that might result from limiting the analysis to patients with complete data, we used the multivariate imputation by chained equations method of multiple imputation, as implemented in STATA, to create ten imputed datasets.25
Imputation of missing values was restricted to confounding variables. Standard methods were then used to combine results over the ten imputed datasets and to compute summary standard errors, confidence intervals, and p-values taking account of the additional uncertainty due to imputation. We also performed a complete-case analysis as a check on the sensitivity of our results to the missing data.
Covariates were considered for inclusion in multivariable models if they were associated with the primary predictor or outcome of interest at a statistical significance of p
0.10 in unadjusted analysis. Backwards deletion was then used to select final adjusted models, retaining variables reaching statistical significance at p
0.05. To further guard against confounding, we included site of care, the strongest correlate of the primary predictor, as well as cardiac consultation (when not treated as an outcome itself) in models with a minimum of 50 events. All analyses were performed using STATA 9.0 (StataCorp, College Station, TX).