All participating centers were members of the Academic Medical Center Consortium, which sponsored the study. All eight centers were large tertiary care academic medical centers with approximately 18 800–43 700 admissions annually 22. Patient enrollment occurred between January 1993 and April 1994.

Patients surveyed represented a stratified random sample of patients in ICUs and outside ICUs for whom blood cultures were either positive or negative, as previously described 22. Among non-ICU patients with blood cultures that were negative at 48 hours, sampling at all centers was fixed at 10%. Among ICU patients and patients with a blood culture positive for a pathogen within 48 hours, the fraction of patients sampled was established at each center and varied from 30% to 60%. In addition, all patients who died in an emergency department or an ICU and all patients who received a novel therapy for sepsis syndrome were surveyed. The ICUs involved included both open and closed units.

Although previous reports included 1342 episodes of sepsis among 1166 unique patients with 1190 hospital admissions 2223, the results presented here from the same database were based on the 1011 (87%) patients, representing 1028 (86%) admissions with 1173 (87%) episodes of sepsis syndrome, for whom billing information was available. When patients had more than one episode in an admission, only the first episode in that admission was evaluated.

Sepsis syndrome was defined as described by Sands and coworkers 22, using a modification of the criteria developed by Bone 24, and involved assessment of the presence of both screening criteria and confirmatory criteria for sepsis syndrome. Either of the following were required for a patient to satisfy the screening criteria: all four of temperature greater than 38.3°C or less than 35.6°C rectally, respirations greater than 20/min or mechanical ventilation, heart rate greater than 90 beats/min, and clinical evidence of infection; or one or more blood cultures positive for a pathogen at 48 hours following admission. Presence of any of the following seven features, without an alternative cause, was required for satisfaction of the confirmatory criteria: ratio of partial arterial oxygen tension to inspired fractional oxygen below 280 (intubated) or 40% face mask in use (nonintubated); arterial pH below 7.30; urine output below 30 ml/h; systolic blood pressure below 90 mmHg or a fall in systolic blood pressure by more than 40 mmHg sustained for 2 hours despite fluid challenge; systemic vascular resistance below 800 dynes·s/cm⁵; prothrombin time or partial thromboplastin time greater than normal or platelets below 100.0 × 10⁹/l or platelets decreased to less than 50% or most recent measurement before current day; and documentation of deterioration in mental status within 24 hours. In addition, all patients who received a form of novel therapy for sepsis syndrome were enrolled in the sepsis syndrome group. Because all patients with sepsis syndrome also meet the criteria for 'severe sepsis' (the term that is most popular now), below we use the term 'severe sepsis'.

Acute respiratory distress syndrome (ARDS) was defined as partial arterial oxygen tension below 50 mmHg despite fractional inspired oxygen in excess of 50%, and bilateral pulmonary infiltrates without signs of heart failure. Disseminated intravascular coagulation (DIC) was defined as a positive D-dimer or elevated fibrin degradation products plus a falling platelet count of 25% of less of baseline, and elevation of either prothrombin time or partial prothrombin time, or clinical evidence of bleeding. Central nervous system (CNS) dysfunction was defined as a Glasgow Coma Scale score below 15 in patients with normal results at baseline neurologic examination, or at least 1 point lower than baseline in patients with a previously abnormal examination result. Renal failure was defined as an increase in serum creatinine of at least 2.0 mg/dl during the sepsis episode if creatinine was below 1.5 mg/dl at baseline, or an absolute increase of 1.0 mg/dl or more if baseline creatinine was at least 1.5. Liver failure was defined as present if total bilirubin was 2.0 mg/dl or greater, and either the alkaline phosphatase or a transaminase level was greater than twice normal, in the absence of confounding disease. Shock (after onset of sepsis) was defined as either at least 1 hour of decreased systolic blood pressure by at least 40 mmHg, or systolic blood pressure of 90 mmHg or less after adequate volume replacement, and in the absence of antihypertensive agents, or continuous infusion of pressors. Uncontrolled hemorrhage was defined as active and live-threatening bleeding, requiring transfusions of more than 4 units within 24 hours of onset of sepsis. Moderate or severe liver disease was defined as cirrhosis with portal hypertension, or hepatic failure with coma or encephalopathy within the preceding 6 months.

Surveyors recorded the presence of all screening criteria for every day that a patient was a valid member of the surveillance group. If screening criteria were met, then the patient was enrolled as a case and detailed information, including demographic, historical, clinical and laboratory data, were abstracted from the medical record into a standardized data collection form for which all variables and criteria were defined. Additional data, including information on antibiotic use, treatment modalities (pulmonary artery catheter [PAC] use, ventilator support, pressor support, albumin infusion, het-astarch, and dextran), complications (ARDS, renal failure, liver failure, DIC, and shock), and associated outcomes, were abstracted 28 days after the first onset of severe sepsis or at the time of death or hospital discharge, whichever came first. Hospital length of stay (LOS) before and after the onset of severe sepsis and total hospital charge were obtained from the unified Academic Medical Center Consortium hospital billing database.

We calculated the patient's Charlson comorbidity score 25; a higher score reflects greater comorbidity. We also obtained information on discharge diagnosis-relative groups (DRGs) and DRG weight. The current generation of DRGs was originally based on ICD-9-CM and takes into consideration operations, complications, and comorbidities. Each DRG was preassigned an average LOS and reimbursement rate by the US Health Care Financing Administration. The higher the DRG weight, the higher the level of reimbursement.

Because the sampling fractions from different strata varied 22, individual cases carried different weight 23 and all calculations, including means and proportions, were performed using appropriate case weights.

We compared the proportion of sepsis patients receiving each intervention therapy and then the proportion with complications among eight academic centers, using weighted χ² statistics. We also calculated the ratio of the highest to lowest proportion for each intervention and complication. Multivariate comparisons were made using generalized linear modeling, with post-onset LOS and hospital charges as dependent variables and study center (coded as dummy variables) as the main covariate, adjusting for age, sex, Charlson comorbidity score, discharge DRG weight, organ dysfunction (ARDS, uncontrolled hemorrhage, moderate or severe liver disease, or acute renal failure), and service (surgical versus medical) at onset of sepsis. The same adjustment of variables was applied in logistic regression models to compare the likelihood that a sepsis patient at a particular center received an intervention and experienced a complication, arbitrarily choosing center #1 as the reference center. Because patients within the same center could share some common characteristics, generalized estimating equation regression models were used to perform the analyses on the comparisons of resource use between patients with and without use of modalities. The r² value was calculated for each model. All analyses were performed using the SAS statistical package (SAS Institute, Inc., Cary, NC, USA) 26.

There was significant variation in mean total charges among admissions with severe sepsis among the study centers, varying from $65 162 (median $42 802) to $244 293 (median $181 758; Table 2). The mean post-onset LOS varied from 16.1 days to 24.8 days (median range 9–16 days) among centers. Other measures of resource utilization including total LOS and LOS in the ICU were also significantly different across centers, with means of 22.7–36.9 days and 5.0–17.8 days, respectively. After adjusting for age, sex, Charlson comorbidity score, discharge DRG weight, organ dysfunction, and service at onset of sepsis, the variations in mean total charges and post-onset LOS remained significant, ranging from $69 429 ± $9562 to $237 898 ± $12 129, and from 15.9 ± 2.0 days to 24.2 ± 1.8 days, respectively (Table 3).

Interventions used frequently after the onset of severe sepsis included PACs (19.4%), ventilator support (21.8%), pressor support (45.8%), and albumin infusion (14.4%; Table 4). Although there were variations across medical centers for each therapeutic modality, variation profiles differed substantially among the modalities. Modalities with moderate variation profiles among the eight participating centers included PACs (varied 4.0-fold, 95% confidence interval [CI] of 2.6-fold to 6.4-fold; range 8.7% to 35.1%), ventilator support (varied 3.2-fold, 95% CI 2.1-fold to 4.9-fold; range 11.5% to 36.4%), and albumin infusion within 24 hours after sepsis onset (varied 4.9-fold, 95% CI 2.2-fold to 10.9-fold; range 6.0% to 29.1%). In contrast, use of pressor support varied only 1.9-fold (95% CI 1.4-fold to 2.6-fold) across centers, ranging from 31.4% to 60.4%, and antibiotics given within 24 hours after onset had the least variation, of 1.1-fold (95% CI 1.0-fold to 1.1-fold] and ranging from 87.0% to 98.5%. Use of hetastarch and use of dextran within 24 hours after onset in sepsis patients were the least common therapeutic modalities among the eight centers, with ranges of 0–6.8% and 0–6.2%, respectively. Of patients receiving pressors within 24 hours after sepsis onset, 28.3% received low-dose dopamine only (<5 μg/kg per min); the proportion varied 7-fold (95% CI 2.4-fold to 20.8-fold) across centers, ranging from 6.4% in center 5 to 44.8% in center 1 (data not shown).

The results of models adjusting for age, sex, Charlson comorbidity score, discharge DRG weight, organ dysfunction, and service at onset of sepsis are shown in Fig. 1. Centers that had rates significantly higher than the reference center were centers 3, 4, 5, 6, and 8 for PAC use; centers 2, 3, 4, 5, and 8 for ventilator support; center 7 for pressor support; and centers 2, 7, and 8 for albumin infusion. In contrast, significantly lower rates for antibiotic use within 24 hours after onset were identified in centers 3, 4, 5, 6, 7, and 8.

Among all patients with severe sepsis, frequent complications included shock (45.4%), renal failure (19.7%), CNS dysfunction (18.9%), liver failure (12.2%), DIC (10.3%), ARDS (9.1%), and death (30.0%) within 28 days after sepsis onset (Table 5). There was significant variation in each type of complication across the eight centers, with ranges of 23.9–70.0% for shock, 9.5–35.0% for renal failure, 0.3–37.5% for CNS dysfunction, 4.1–27.9% for liver failure, 5.5–15.0% for DIC, 2.8–21.8% for ARDS, and 19.3–49.1% for 28-day mortality rate.

Figure 2 shows the results of differences in rates of sepsis complications adjusted for age, sex, Charlson comorbidity score, discharge DRG weight, organ dysfunction, and service at onset of sepsis. Analysis of the 28-day in-hospital adjusted mortality rates revealed 3-fold variability among centers. Compared with the reference center, higher rates were found for ARDS among one center (C8), for renal failure among five centers (C3, C4, C5, C7, and C8), for liver failure among one center (C4), for shock among six centers (C3, C4, C5, C6, C7, and C8), and for 28-day mortality among four centers (C3, C6, C7, and C8).

Also, a significant inverse correlation was seen between antibiotic use within 24 hours after onset of episode and the 28-day mortality rates (r² = 0.72; P = 0.01; Fig. 3). This correlation remained significant even after excluding patients with 'do not resuscitate' orders (r² = 0.63; P = 0.02). Significant associations were not found between 28-day mortality and the use of other modalities across the eight centers (data not shown).

Comparisons of resource use among patients treated with specific modalities and those who were not showed that administration of antibiotics within 24 hours after sepsis onset had no impact on any of the LOS categories, but was associated with total and daily hospital charges (Table 6). Use of other modalities (including albumin infusion) was associated with increased resource use for daily hospital charges. Total hospital charges and LOS in the ICU were significantly higher among patients who were given pressor support and/or albumin infusion after onset of sepsis as compared with those who were not. Increased post-onset LOS and LOS in the ICU were associated with use of PACs. No associations were found between total LOS and the modalities studied.