Materials and Methods
Clinical Trial Selection and Data Abstraction
We used a sensitive strategy (Supplemental Methods, Supplemental Digital Content 1, http://links.lww.com/CCM/A757) to search MEDLINE for randomized trials enrolling patients with severe sepsis. Based on a review of abstracts, two independent investigators (A.R.R., G.T.R.) selected prospective studies enrolling patients with sepsis that reported a mortality outcome. The full text of these studies was then reviewed in detail by two independent investigators (A.J.W., E.K.S.) to identify multicenter randomized controlled trials that enrolled patients using a modified 1991 American College of Chest Physicians/Society of Critical Care Medicine Consensus definition of severe sepsis and included patients with suspected infection and acute organ dysfunction. Single-center studies were excluded out of concern that the reported mortality rates may be center specific and not representative of more widespread trends. Observational studies were excluded because very few multicenter observational studies of prospectively identified patients with severe sepsis (that were not secondary analyses of trial data) met our criteria of providing study start dates, hospital or 28-day mortality, and baseline severity-of-illness scores. Within the eligible observational studies, three of the four eras in our analysis were represented by only one study limited to patients from one country, confounding our ability to separate differences over time from national differences in severe sepsis mortality. Characteristics of the eight multicenter observational studies that met our criteria are shown in Supplemental Table 1 (Supplemental Digital Content 1, http://links.lww.com/CCM/A757 In addition, we excluded clinical trials that were not published in English and that did not specify study start dates. Because small trials may be more likely to have biased estimates. we performed a sensitivity analysis excluding trials in which the control group N was less than or equal to 20.
Using a standardized data abstraction form, two independent investigators (A.J.W., E.K.S.) recorded the following data from each trial: enrollment start date and end date, average age, sex distribution, severity-of-illness score (APACHE II, SAPS II, and LODS, enrollment of patients with severe sepsis versus septic shock, number of patients enrolled in the usual care group, and hospital and 28-day mortality of usual care group patients. Hospital mortality data were available from only five of 36 trials (14%), and thus, we were unable to analyze trends in hospital mortality among trials. Because we used administrative data from the United States to compare with the multicenter trials, we performed a sensitivity analysis stratifying by whether trials enrolled subjects from the United States or did not include any U.S. centers.
Administrative Data Analysis
We examined hospitalizations from adults (age ≥ 18 yr) using year 1993–2009 discharge data from the Nationwide Inpatient Sample (NIS), Healthcare Cost and Utilization Project, and Agency for Healthcare Research and Quality. Based on a large increase in population coverage starting in 1993, use of NIS data prior to 1993 for trend analyses is not recommended. The NIS is an approximate 20% stratified probability sample of all U.S. non-Federal acute care hospitals and contains deidentified clinical and resource use information from approximately 5–8 million hospital discharges yearly. NIS sampling strata are based on five hospital characteristics: ownership/control, teaching status, urban/rural location, geographic region, and bed size. The 1993 NIS contained data from about 900 hospitals in 17 states and the 2009 NIS included data from approximately 1,000 hospitals in 44 states. NIS elements include demographics, admission and discharge status, length of stay, up to 15 ICD-9-CM diagnosis and procedure codes (increased to 25 diagnosis codes in 2009), and hospital characteristics.
We identified cases of severe sepsis in the NIS based on two previously published and validated algorithms. The "Angus" algorithm selected severe sepsis cases based on the presence of an ICD-9-CM code for infection and acute organ dysfunction, whereas the "Martin" algorithm identified severe sepsis cases based on the presence of ICD-9-CM codes for septicemia, bacteremia, or fungemia, as well as an acute organ dysfunction code. Both algorithms included explicit severe sepsis (995.92) and septic shock (785.52) codes introduced in 2002 and 2003, respectively.
Statistical Analyses
We used standardized mortality ratios (SMRs) to adjust for baseline case-mix differences between trials. SMRs were calculated from the ratio of observed mortality to predicted mortality in the usual care arm. Predicted mortality for usual care arm trial participants was calculated from previously published regression equations based on the baseline severity-of-illness score available from each trial (APACHE II, SAPS II, or LODS score). In years during which multiple trials contributed data, SMRs were calculated from the pooled number of observed and predicted deaths from each trial.
We analyzed trends in severe sepsis mortality in clinical trial data using two methods. First, we pooled trials by the year of first patient enrollment. We used Joinpoint Regression Program version 4.0.0 (Statistical Research and Applications Branch, National Cancer Institute, Bethesda, MD) to evaluate trends in the observed, predicted, and standardized 28-day severe sepsis mortality rates from trials that began enrolling patients from 1991 to 2009. Joinpoint models were constructed using a heteroscedastic errors weighted least squares regression in which the SE of the mortality estimate from each year was input to the model. We identified trends in mortality using the annual percent change (APC) in yearly pooled 28-day mortality rates. Second, we pooled clinical trials by the start date of patient enrollment into four time frames (1991–1995, 1996–2000, 2001–2005, and 2006–2009).
We used SAS version 9.3 (SAS, Cary, NC) to identify the survey-weighted hospital mortality of patients identified from the NIS with severe sepsis and septic shock using Angus and Martin administrative data algorithms. We compared trends in severe sepsis 28-day mortality identified with clinical trial data with trends in hospital mortality identified using administrative data during the years 1993–2009 using the Joinpoint Regression Program test of parallelism. Because of large differences in sample size between the clinical trials and the administrative data, resulting in different power to detect inflection points of trend changes, we did not evaluate for joinpoints in the models. Instead, we analyzed APC over the entire 1993–2009 date range to capture trends. Because individual patient-level data were unavailable from the trials, we did not adjust for potential confounders of the comparison between trial and administrative data mortality trends. A Cochran-Armitage test for interaction was used to test for differences in mortality trend based on whether trials were conducted among U.S. centers or outside the United States. A two-sided α level of 0.05 as selected for statistical significance. All study procedures were approved by the Boston University Medical Campus Institutional Review Board.