Patient Selection

Methodist University Hospital (MUH) in Memphis, Tennessee, is part of a 7‐hospital system with 697 licensed beds. MUH is a tertiary teaching hospital with centers of excellence in neuroscience and transplantation. Patients admitted to MUH diagnosed with VRE bacteremia between January 1, 2004, and July 31, 2009, were identified by the microbiology laboratory. All patients who were 18 years of age, had 1 documented positive VRE blood culture, and received linezolid or daptomycin for 5 days were eligible. Patients were excluded if they were treated simultaneously with more than 1 agent active against VRE. This study was approved by the MUH Institutional Review Board. Of note, use of linezolid or daptomycin at MUH is restricted to an infectious disease physician or pulmonologist. Currently, there are no protocols at our institution for treating VRE infections.

Data Collection and Definitions

Cerner Millennium was used to collect all pertinent patient information. Patient records were reviewed to determine demographic data, comorbid illnesses, laboratory data (from admission to discharge), medications, and discharge status (home, long‐term care facility, or death). Comorbid illnesses evaluated included: chronic obstructive pulmonary disease, diabetes mellitus, malignancy (solid or hematologic), transplant (liver or kidney), end‐stage renal disease (ESRD) (hemodialysis or nonhemodialysis), cirrhosis, and endocarditis. ESRD and endocarditis were defined per chart diagnosis. Laboratory data collected included serum creatinine, creatine phosphokinase, absolute neutrophil count (neutropenia defined as absolute neutrophil count <1000), and number and site (intravenous line or peripheral blood draw) of positive VRE blood cultures. Other data collected were (1) time elapsed to adequate antibiotic coverage, which was defined as microbiologic documentation of an infection that was being effectively treated at the time of its identification, and (2) time to appropriate antibiotic coverage, which was defined as antimicrobial treatment selected for efficacy based on presumptive identification of the causative pathogen, the antimicrobial agent's spectrum of activity, and local microbial resistance patterns.20 Doses of daptomycin and linezolid used in patients with VRE bacteremia were also documented.

Clinical cure was defined as a resolution of signs and/or symptoms of infection (white blood cell count <10,000/mm³, bands <5%, heart rate <90 beats per minute, respiratory rate <20 breaths per minute, and maximum oral temperature <38C) after gram‐positive therapy was discontinued. The definition of microbiologic cure was lack of positive blood cultures for VRE at least 14 days after cessation of gram‐positive therapy. Microbiologic failure was defined as positive VRE blood cultures obtained on gram‐positive therapy necessitating a change in treatment. Recurrence was defined as VRE bacteremia within 30 days after discontinuation of gram‐positive therapy. Reinfection was defined as VRE bacteremia that appeared 30 days after completion of primary gram‐positive therapy.

All isolates were tested for susceptibility to linezolid using the MicroScan system, whereas daptomycin susceptibility patterns were obtained by either the Etest or MicroScan system. Of importance, our laboratory did not routinely report isolate susceptibility for daptomycin until 2008. Clinical Laboratory Standards Institute breakpoint guidelines were used to delineate minimum inhibitory concentrations for linezolid and daptomycin.

Outcomes

The primary objective was to determine the cure rate, both clinical and microbiologic, of VRE bacteremia with the use of linezolid and daptomycin. Secondary outcomes were rates of recurrence and reinfection as well as 30‐day mortality. Clinical and microbiologic response rates for subsets of the patient population that were deemed immunocompromised or at an increased risk for VRE infections (neutropenic, transplant, malignancy, and ESRD on hemodialysis) were also evaluated.

Statistical Analysis

Data were analyzed using SAS version 9.2 (SAS Inc, Cary, NC). Patients with categorical characteristics were compared using a chi‐square test or Fisher's exact test. Continuous data were analyzed using a Student t test and are expressed as the mean standard deviation. The mean duration of initial antibiotics, time to appropriate antibiotics, time to adequate antibiotic therapy, and LOS were all calculated for the linezolid and daptomycin group with a Student t test used to compare the differences. Multivariate logistic regression was used for the following outcomes: clinical cure, microbiologic cure, mortality, reinfection, and recurrence. For the interval variable, LOS, stepwise multiple regression was used to choose significant independent variables. P < 0.05 was considered statistically significant.

Patient Characteristics

Of the 361 patients identified with a positive VRE blood culture, 201 were included in the study. The remaining 160 patients were excluded for one of the following reasons: <5 days of therapy (n = 87), no documented gram‐positive therapy (n = 49), simultaneous gram‐positive therapy (n = 10), or insufficient data to evaluate response rates (n = 14). For the treatment of VRE bacteremia, 138 patients received linezolid and 63 patients received daptomycin. Demographics, comorbid illnesses, and patient characteristics are shown in Table 1. There was a statistically significant difference in the average age, with the linezolid group consisting of older patients. The daptomycin group had more patients with hematologic malignancies than the linezolid group (33% vs 14%; P = 0.0021) and more patients who received liver transplants (13% vs 4%; P = 0.0264).

From the microbiology laboratory report of initial blood cultures, 78.6% of the isolates were noted as being E. faecium, with the remainder being E. faecalis (21.4%). One patient was classified as having linezolid‐resistant E. faecium (MIC >4 mg/L) upon repeat blood culture. Daptomycin MICs were obtained for 44 isolates using the Etest or MicroScan system; all isolates were susceptible with MICs ranging from 0.254 mg/L. As mentioned previously, our laboratory did not routinely report isolate susceptibility to daptomycin until 2008.

There were no statistically significant differences between the treatment groups with regard to time to appropriate or adequate antibiotic therapy (Table 1). However, there was a statistically significant difference in the mean duration of initial antibiotics between linezolid and daptomycin (11.1 days vs 14.1 days; P = 0.0401). Dosing strategies used in these patients were also evaluated. All linezolid patients received a dose of 600 mg every 12 hours by mouth or intravenously. The average dose of daptomycin was 6.1 mg/kg (range, 3.410.4 mg/kg; median, 6 mg/kg). The average LOS was 37 days for linezolid vs 40 days for daptomycin, which did not confer statistical significance. Overall mortality was 20%, occurring in 25 linezolid patients versus 15 daptomycin patients (P = 0.3481). The stepwise multiple regression analysis did not identify any statistically significant variables in patients treated with linezolid or daptomycin that affected any of the outcomes.

Outcomes and Analysis

As shown in Table 2, there were no statistically significant differences in clinical or microbiologic cure between the linezolid and daptomycin groups (74% vs 75% and 94% vs 94%, respectively). However, the linezolid group compared with the daptomycin group had fewer patients that developed a positive blood culture while on their initial antibiotic therapy (8% vs 22%; P = 0.0097). Follow‐up cultures were required to determine rates of recurrence and reinfection. Only 107/138 patients in the linezolid group and 51/63 patients in the daptomycin group had follow‐up cultures collected. Recurrence was documented in 3% of linezolid patients vs 12% of daptomycin patients (P = 0.0321). The odds ratio for developing a recurrent infection with daptomycin versus linezolid was 5.51 (95% confidence interval, 1.2524.28). Out of 6 patients that developed a recurrent VRE infection in the daptomycin group, 2 were prescribed doses <4 mg/kg with no reported MICs, and 2 patients received 6 mg/kg with reported MICs of 4 mg/L. No statistically significant difference existed for the rate of reinfection between linezolid and daptomycin (1% vs 6%; P = 0.0992).

Table 3 provides information on subsets of the patient population deemed high‐risk for VRE infection or immunocompromised. There was no statistically significant difference between the 2 antibiotic groups in clinical or microbiologic cure. In the subsets of immunocompromised patients, there was no difference in recurrence or reinfection between the linezolid and daptomycin patients. Furthermore, all groups had similar LOS regardless of the antibiotic used to treat the VRE BSI. Moreover, there were no statistically significant differences in 30‐day mortality in these subsets of the population with regard to initial antibiotic choice. No significant independent variables were found between linezolid or daptomycin that affected any of the outcomes listed in Table 3.

Patient Selection

Methodist University Hospital (MUH) in Memphis, Tennessee, is part of a 7‐hospital system with 697 licensed beds. MUH is a tertiary teaching hospital with centers of excellence in neuroscience and transplantation. Patients admitted to MUH diagnosed with VRE bacteremia between January 1, 2004, and July 31, 2009, were identified by the microbiology laboratory. All patients who were 18 years of age, had 1 documented positive VRE blood culture, and received linezolid or daptomycin for 5 days were eligible. Patients were excluded if they were treated simultaneously with more than 1 agent active against VRE. This study was approved by the MUH Institutional Review Board. Of note, use of linezolid or daptomycin at MUH is restricted to an infectious disease physician or pulmonologist. Currently, there are no protocols at our institution for treating VRE infections.

Data Collection and Definitions

Cerner Millennium was used to collect all pertinent patient information. Patient records were reviewed to determine demographic data, comorbid illnesses, laboratory data (from admission to discharge), medications, and discharge status (home, long‐term care facility, or death). Comorbid illnesses evaluated included: chronic obstructive pulmonary disease, diabetes mellitus, malignancy (solid or hematologic), transplant (liver or kidney), end‐stage renal disease (ESRD) (hemodialysis or nonhemodialysis), cirrhosis, and endocarditis. ESRD and endocarditis were defined per chart diagnosis. Laboratory data collected included serum creatinine, creatine phosphokinase, absolute neutrophil count (neutropenia defined as absolute neutrophil count <1000), and number and site (intravenous line or peripheral blood draw) of positive VRE blood cultures. Other data collected were (1) time elapsed to adequate antibiotic coverage, which was defined as microbiologic documentation of an infection that was being effectively treated at the time of its identification, and (2) time to appropriate antibiotic coverage, which was defined as antimicrobial treatment selected for efficacy based on presumptive identification of the causative pathogen, the antimicrobial agent's spectrum of activity, and local microbial resistance patterns.20 Doses of daptomycin and linezolid used in patients with VRE bacteremia were also documented.

Clinical cure was defined as a resolution of signs and/or symptoms of infection (white blood cell count <10,000/mm³, bands <5%, heart rate <90 beats per minute, respiratory rate <20 breaths per minute, and maximum oral temperature <38C) after gram‐positive therapy was discontinued. The definition of microbiologic cure was lack of positive blood cultures for VRE at least 14 days after cessation of gram‐positive therapy. Microbiologic failure was defined as positive VRE blood cultures obtained on gram‐positive therapy necessitating a change in treatment. Recurrence was defined as VRE bacteremia within 30 days after discontinuation of gram‐positive therapy. Reinfection was defined as VRE bacteremia that appeared 30 days after completion of primary gram‐positive therapy.

All isolates were tested for susceptibility to linezolid using the MicroScan system, whereas daptomycin susceptibility patterns were obtained by either the Etest or MicroScan system. Of importance, our laboratory did not routinely report isolate susceptibility for daptomycin until 2008. Clinical Laboratory Standards Institute breakpoint guidelines were used to delineate minimum inhibitory concentrations for linezolid and daptomycin.

Outcomes

The primary objective was to determine the cure rate, both clinical and microbiologic, of VRE bacteremia with the use of linezolid and daptomycin. Secondary outcomes were rates of recurrence and reinfection as well as 30‐day mortality. Clinical and microbiologic response rates for subsets of the patient population that were deemed immunocompromised or at an increased risk for VRE infections (neutropenic, transplant, malignancy, and ESRD on hemodialysis) were also evaluated.

Statistical Analysis

Data were analyzed using SAS version 9.2 (SAS Inc, Cary, NC). Patients with categorical characteristics were compared using a chi‐square test or Fisher's exact test. Continuous data were analyzed using a Student t test and are expressed as the mean standard deviation. The mean duration of initial antibiotics, time to appropriate antibiotics, time to adequate antibiotic therapy, and LOS were all calculated for the linezolid and daptomycin group with a Student t test used to compare the differences. Multivariate logistic regression was used for the following outcomes: clinical cure, microbiologic cure, mortality, reinfection, and recurrence. For the interval variable, LOS, stepwise multiple regression was used to choose significant independent variables. P < 0.05 was considered statistically significant.

Patient Characteristics

Of the 361 patients identified with a positive VRE blood culture, 201 were included in the study. The remaining 160 patients were excluded for one of the following reasons: <5 days of therapy (n = 87), no documented gram‐positive therapy (n = 49), simultaneous gram‐positive therapy (n = 10), or insufficient data to evaluate response rates (n = 14). For the treatment of VRE bacteremia, 138 patients received linezolid and 63 patients received daptomycin. Demographics, comorbid illnesses, and patient characteristics are shown in Table 1. There was a statistically significant difference in the average age, with the linezolid group consisting of older patients. The daptomycin group had more patients with hematologic malignancies than the linezolid group (33% vs 14%; P = 0.0021) and more patients who received liver transplants (13% vs 4%; P = 0.0264).

From the microbiology laboratory report of initial blood cultures, 78.6% of the isolates were noted as being E. faecium, with the remainder being E. faecalis (21.4%). One patient was classified as having linezolid‐resistant E. faecium (MIC >4 mg/L) upon repeat blood culture. Daptomycin MICs were obtained for 44 isolates using the Etest or MicroScan system; all isolates were susceptible with MICs ranging from 0.254 mg/L. As mentioned previously, our laboratory did not routinely report isolate susceptibility to daptomycin until 2008.

There were no statistically significant differences between the treatment groups with regard to time to appropriate or adequate antibiotic therapy (Table 1). However, there was a statistically significant difference in the mean duration of initial antibiotics between linezolid and daptomycin (11.1 days vs 14.1 days; P = 0.0401). Dosing strategies used in these patients were also evaluated. All linezolid patients received a dose of 600 mg every 12 hours by mouth or intravenously. The average dose of daptomycin was 6.1 mg/kg (range, 3.410.4 mg/kg; median, 6 mg/kg). The average LOS was 37 days for linezolid vs 40 days for daptomycin, which did not confer statistical significance. Overall mortality was 20%, occurring in 25 linezolid patients versus 15 daptomycin patients (P = 0.3481). The stepwise multiple regression analysis did not identify any statistically significant variables in patients treated with linezolid or daptomycin that affected any of the outcomes.

Outcomes and Analysis

As shown in Table 2, there were no statistically significant differences in clinical or microbiologic cure between the linezolid and daptomycin groups (74% vs 75% and 94% vs 94%, respectively). However, the linezolid group compared with the daptomycin group had fewer patients that developed a positive blood culture while on their initial antibiotic therapy (8% vs 22%; P = 0.0097). Follow‐up cultures were required to determine rates of recurrence and reinfection. Only 107/138 patients in the linezolid group and 51/63 patients in the daptomycin group had follow‐up cultures collected. Recurrence was documented in 3% of linezolid patients vs 12% of daptomycin patients (P = 0.0321). The odds ratio for developing a recurrent infection with daptomycin versus linezolid was 5.51 (95% confidence interval, 1.2524.28). Out of 6 patients that developed a recurrent VRE infection in the daptomycin group, 2 were prescribed doses <4 mg/kg with no reported MICs, and 2 patients received 6 mg/kg with reported MICs of 4 mg/L. No statistically significant difference existed for the rate of reinfection between linezolid and daptomycin (1% vs 6%; P = 0.0992).

Table 3 provides information on subsets of the patient population deemed high‐risk for VRE infection or immunocompromised. There was no statistically significant difference between the 2 antibiotic groups in clinical or microbiologic cure. In the subsets of immunocompromised patients, there was no difference in recurrence or reinfection between the linezolid and daptomycin patients. Furthermore, all groups had similar LOS regardless of the antibiotic used to treat the VRE BSI. Moreover, there were no statistically significant differences in 30‐day mortality in these subsets of the population with regard to initial antibiotic choice. No significant independent variables were found between linezolid or daptomycin that affected any of the outcomes listed in Table 3.

Patient Selection

Methodist University Hospital (MUH) in Memphis, Tennessee, is part of a 7‐hospital system with 697 licensed beds. MUH is a tertiary teaching hospital with centers of excellence in neuroscience and transplantation. Patients admitted to MUH diagnosed with VRE bacteremia between January 1, 2004, and July 31, 2009, were identified by the microbiology laboratory. All patients who were 18 years of age, had 1 documented positive VRE blood culture, and received linezolid or daptomycin for 5 days were eligible. Patients were excluded if they were treated simultaneously with more than 1 agent active against VRE. This study was approved by the MUH Institutional Review Board. Of note, use of linezolid or daptomycin at MUH is restricted to an infectious disease physician or pulmonologist. Currently, there are no protocols at our institution for treating VRE infections.

Data Collection and Definitions

Cerner Millennium was used to collect all pertinent patient information. Patient records were reviewed to determine demographic data, comorbid illnesses, laboratory data (from admission to discharge), medications, and discharge status (home, long‐term care facility, or death). Comorbid illnesses evaluated included: chronic obstructive pulmonary disease, diabetes mellitus, malignancy (solid or hematologic), transplant (liver or kidney), end‐stage renal disease (ESRD) (hemodialysis or nonhemodialysis), cirrhosis, and endocarditis. ESRD and endocarditis were defined per chart diagnosis. Laboratory data collected included serum creatinine, creatine phosphokinase, absolute neutrophil count (neutropenia defined as absolute neutrophil count <1000), and number and site (intravenous line or peripheral blood draw) of positive VRE blood cultures. Other data collected were (1) time elapsed to adequate antibiotic coverage, which was defined as microbiologic documentation of an infection that was being effectively treated at the time of its identification, and (2) time to appropriate antibiotic coverage, which was defined as antimicrobial treatment selected for efficacy based on presumptive identification of the causative pathogen, the antimicrobial agent's spectrum of activity, and local microbial resistance patterns.20 Doses of daptomycin and linezolid used in patients with VRE bacteremia were also documented.

Clinical cure was defined as a resolution of signs and/or symptoms of infection (white blood cell count <10,000/mm³, bands <5%, heart rate <90 beats per minute, respiratory rate <20 breaths per minute, and maximum oral temperature <38C) after gram‐positive therapy was discontinued. The definition of microbiologic cure was lack of positive blood cultures for VRE at least 14 days after cessation of gram‐positive therapy. Microbiologic failure was defined as positive VRE blood cultures obtained on gram‐positive therapy necessitating a change in treatment. Recurrence was defined as VRE bacteremia within 30 days after discontinuation of gram‐positive therapy. Reinfection was defined as VRE bacteremia that appeared 30 days after completion of primary gram‐positive therapy.

All isolates were tested for susceptibility to linezolid using the MicroScan system, whereas daptomycin susceptibility patterns were obtained by either the Etest or MicroScan system. Of importance, our laboratory did not routinely report isolate susceptibility for daptomycin until 2008. Clinical Laboratory Standards Institute breakpoint guidelines were used to delineate minimum inhibitory concentrations for linezolid and daptomycin.

Outcomes

The primary objective was to determine the cure rate, both clinical and microbiologic, of VRE bacteremia with the use of linezolid and daptomycin. Secondary outcomes were rates of recurrence and reinfection as well as 30‐day mortality. Clinical and microbiologic response rates for subsets of the patient population that were deemed immunocompromised or at an increased risk for VRE infections (neutropenic, transplant, malignancy, and ESRD on hemodialysis) were also evaluated.

Statistical Analysis

Data were analyzed using SAS version 9.2 (SAS Inc, Cary, NC). Patients with categorical characteristics were compared using a chi‐square test or Fisher's exact test. Continuous data were analyzed using a Student t test and are expressed as the mean standard deviation. The mean duration of initial antibiotics, time to appropriate antibiotics, time to adequate antibiotic therapy, and LOS were all calculated for the linezolid and daptomycin group with a Student t test used to compare the differences. Multivariate logistic regression was used for the following outcomes: clinical cure, microbiologic cure, mortality, reinfection, and recurrence. For the interval variable, LOS, stepwise multiple regression was used to choose significant independent variables. P < 0.05 was considered statistically significant.

Patient Characteristics

Of the 361 patients identified with a positive VRE blood culture, 201 were included in the study. The remaining 160 patients were excluded for one of the following reasons: <5 days of therapy (n = 87), no documented gram‐positive therapy (n = 49), simultaneous gram‐positive therapy (n = 10), or insufficient data to evaluate response rates (n = 14). For the treatment of VRE bacteremia, 138 patients received linezolid and 63 patients received daptomycin. Demographics, comorbid illnesses, and patient characteristics are shown in Table 1. There was a statistically significant difference in the average age, with the linezolid group consisting of older patients. The daptomycin group had more patients with hematologic malignancies than the linezolid group (33% vs 14%; P = 0.0021) and more patients who received liver transplants (13% vs 4%; P = 0.0264).

From the microbiology laboratory report of initial blood cultures, 78.6% of the isolates were noted as being E. faecium, with the remainder being E. faecalis (21.4%). One patient was classified as having linezolid‐resistant E. faecium (MIC >4 mg/L) upon repeat blood culture. Daptomycin MICs were obtained for 44 isolates using the Etest or MicroScan system; all isolates were susceptible with MICs ranging from 0.254 mg/L. As mentioned previously, our laboratory did not routinely report isolate susceptibility to daptomycin until 2008.

There were no statistically significant differences between the treatment groups with regard to time to appropriate or adequate antibiotic therapy (Table 1). However, there was a statistically significant difference in the mean duration of initial antibiotics between linezolid and daptomycin (11.1 days vs 14.1 days; P = 0.0401). Dosing strategies used in these patients were also evaluated. All linezolid patients received a dose of 600 mg every 12 hours by mouth or intravenously. The average dose of daptomycin was 6.1 mg/kg (range, 3.410.4 mg/kg; median, 6 mg/kg). The average LOS was 37 days for linezolid vs 40 days for daptomycin, which did not confer statistical significance. Overall mortality was 20%, occurring in 25 linezolid patients versus 15 daptomycin patients (P = 0.3481). The stepwise multiple regression analysis did not identify any statistically significant variables in patients treated with linezolid or daptomycin that affected any of the outcomes.

Outcomes and Analysis

As shown in Table 2, there were no statistically significant differences in clinical or microbiologic cure between the linezolid and daptomycin groups (74% vs 75% and 94% vs 94%, respectively). However, the linezolid group compared with the daptomycin group had fewer patients that developed a positive blood culture while on their initial antibiotic therapy (8% vs 22%; P = 0.0097). Follow‐up cultures were required to determine rates of recurrence and reinfection. Only 107/138 patients in the linezolid group and 51/63 patients in the daptomycin group had follow‐up cultures collected. Recurrence was documented in 3% of linezolid patients vs 12% of daptomycin patients (P = 0.0321). The odds ratio for developing a recurrent infection with daptomycin versus linezolid was 5.51 (95% confidence interval, 1.2524.28). Out of 6 patients that developed a recurrent VRE infection in the daptomycin group, 2 were prescribed doses <4 mg/kg with no reported MICs, and 2 patients received 6 mg/kg with reported MICs of 4 mg/L. No statistically significant difference existed for the rate of reinfection between linezolid and daptomycin (1% vs 6%; P = 0.0992).

Table 3 provides information on subsets of the patient population deemed high‐risk for VRE infection or immunocompromised. There was no statistically significant difference between the 2 antibiotic groups in clinical or microbiologic cure. In the subsets of immunocompromised patients, there was no difference in recurrence or reinfection between the linezolid and daptomycin patients. Furthermore, all groups had similar LOS regardless of the antibiotic used to treat the VRE BSI. Moreover, there were no statistically significant differences in 30‐day mortality in these subsets of the population with regard to initial antibiotic choice. No significant independent variables were found between linezolid or daptomycin that affected any of the outcomes listed in Table 3.

Design Overview

We used a prospective observational design to measure patient mortality and LOS within 2 medical ICU staffing paradigms. This was a collaborative study between the Division of Hospital Medicine and the Division of Pulmonary and Critical Care Medicine, with approval from Emory University's Institutional Review Board.

The hospitalist ICU model was staffed by a board certified internal medicine attending, with clinical responsibilities limited to the ICU. An intensivist‐led consult team (members distinct from the intensivist‐led ICU team) was staffed by a board certified pulmonary critical care attending and non‐physician providers. This consult team comanaged mechanically ventilated patients and was available for additional critical care consultation at the hospitalists' discretion. The intensivist‐led ICU model was staffed by a board certified pulmonary critical care attending, a pulmonary critical care fellow (postgraduate years 46) and 4 internal medicine residents (postgraduate years 23).

For their respective patients, the hospitalist and intensivist‐led teams participated in similar multidisciplinary ICU rounds with the charge nurse, respiratory therapist, and pharmacist. Both teams used the same evidence‐based ICU protocols and order sets. The hospitalists and intensivists were aware of the ongoing study.

Setting and Participants

Our study was conducted in an urban, community teaching hospital that is affiliated with a major regional academic university and has 400 medical‐surgical beds, including 56 ICU beds. All medical ICU patients receiving primary medical care from the hospitalist or intensivist‐led team were assessed for inclusion between October 2007 and September 2008. Predetermined exclusion criteria included surgery under general anesthesia, outside hospital transfers, pregnancy, and age under 18.

Selection of the admitting ICU team followed existing institutional referral patterns. For emergency department (ED) patients, the ED physicians made the decision to admit to the ICU and contacted an ICU team based on the respiratory support needs of the patient, not the admitting diagnosis. ED patients with respiratory failure who required invasive ventilatory support were admitted to the intensivist‐led ICU team. Those without invasive ventilatory support were admitted to the hospitalist ICU team, including ones with respiratory failure requiring noninvasive ventilation. Patients transferred from a hospital floor bed to the ICU by non‐hospitalist physicians were assigned to the intensivist‐led ICU team, while those transferred by hospitalist floor teams were assigned to the hospitalist ICU team, regardless of diagnosis or respiratory support needs. Patient assignments deviated from these patterns, however, based on ICU teams' census. The intensivist‐led ICU team had a strict limit of 20 patients, established by the residency program, and the hospitalist ICU team had a preferred limit of 12 patients.

Measurement of Outcomes and Follow‐Up

Study endpoints were in‐hospital and ICU mortality, as well as hospital and ICU LOS. Patient characteristics and outcome data were collected prospectively from medical records and hospital databases by 2 trained research nurses according to study protocol. For data collection training, 1 investigator (K.R.W.) reviewed sample data from 108 patients to ensure consistency and accuracy of data abstraction. Patients with several ICU admissions during 1 hospitalization had ICU data collected only from the first ICU entry, consistent with other trials' methodology.1, 4, 10, 1820 Additional ICU entries did not change ICU LOS derived only from the initial entry, but did contribute to hospital LOS. Data from patients with multiple ICU entries was analyzed with the original team assignment. All patients were followed until death or hospital discharge.

Statistical Analysis

Sample size was determined a priori using an expected inpatient mortality of 10% from historical data, power of 80%, and 2‐sided alpha of 0.05 to demonstrate no difference in outcomes, defined as a mortality difference of <5% between teams. This mortality difference used for the power calculation is consistent with other trial designs.2125 The required sample size was 1306 patients calculated using PASS software (version 2008, NCSS, Kaysville, UT), accounting for an expected 3:2 admission rate to the hospitalists. The statistician was blinded to team assignments.

Clinical characteristics of the groups were compared with the Student t test. The outcome and predictor variable distributions were examined with univariate analyses. Bivariate analyses were calculated for each predictor and endpoint. Multiple logistic and linear regression analyses were performed. Propensity scores were used and defined as the conditional probability of admission to the hospitalist versus intensivist‐led ICU team given a patient's covariates. It included all predictors in Table 1 and was calculated using logistic regression. Outcome measures were excluded from the regression.

A generalized linear model (GENMOD), using a binomial distribution and an identity link function,26 assessed the in‐hospital and ICU mortality rate differences between teams while controlling for major risk factors identified. GENMOD, however, does not accommodate several covariates, as it often fails for lack of convergence. Hence, logistic regression models with adjusted odds ratios (aOR) are reported as well.

The initial logistic regression model for in‐hospital and ICU mortality included all 20 independent variables from patient demographics, comorbidities, simplified acute physiology score (SAPS) II,27 respiratory support, central venous catheter (CVC) utilization, which included peripherally placed central catheters, and all terms for 2‐way interactions with team assignment. To determine the best model, a hierarchical backward elimination was executed while assessing for interactions, confounding, and estimate precision. Before removing a regression term, a likelihood ratio test was applied to each coefficient followed by Wald's chi square test.28 Collinearity diagnostic for nonlinear models was applied to look for multicollinearity. To exclude variables or regression terms, a condition index of 30 and variance decomposition proportion of 0.5 were used. The final model was evaluated for goodness‐of‐fit using the Hosmer and Lemeshow test.

LOS was analyzed with linear regression using the same covariates and backward elimination as the logistic model. Goodness‐of‐fit was evaluated using coefficient of determination (r²). A variance inflation factor of 10 was used to assess for collinearity. Two‐sided P values 0.05 were considered statistically significant. All analyses were performed with SAS software (version 9.1, SAS Institute, Cary, NC).

Patients

A total of 1747 patients received critical care from the hospitalist or intensivist‐led teams (Figure 1). Of the 1367 patients who met inclusion criteria, complete data was available for 1356 patients. The ED was the ICU admission source for 68.8% of hospitalist and 69.9% of intensivist patients. Baseline patient demographics were similar (Table 1). Among preexisting comorbidities, morbid obesity was more prevalent in hospitalist patients, whereas cancer and pulmonary and immunological diseases were more prevalent in intensivist patients (Table 1).

Figure 1

Hospitalist patients, compared to intensivist patients, had a lower mean SAPS II (37.4 vs 45.1, P < 0.001), less noninvasive (17.9% vs 25.8%, P < 0.001) and mechanical (11.0% vs 51.9%, P < 0.001) ventilation utilization, and fewer CVCs (29.1% vs 50.8%, P < 0.001) (Table 1). The intensivist‐led consult team comanaged 18.4% of the hospitalist patients. These 152 patients had a mean SAPS II of 41, and 19.7% required noninvasive ventilation while 44.1% required mechanical ventilation. For mechanically ventilated hospitalist patients, 80.2% were comanaged by the intensivist‐led consult team, 5.5% had palliative care for end‐of‐life management, 6.6% died imminently, and 7.7% received short‐term ventilation managed by the hospitalist. Hospitalist and intensivist‐led teams' ICU readmission rates were similar (7.0% vs 5.9%, P = 0.41).

Outcomes

Of the 1356 patients, there were 168 (12.4%) deaths and 135 (10.0%) occurred in the ICU. The overall mean ICU LOS was 4.0 days (SD 5.9), and mean hospital LOS was 9.1 days (SD 9.0). The mean hospital LOS for survivors was 9.0 days (SD 8.8).

Subgroup Analysis

Since each team's respiratory support utilization differed greatly and was a significant variable in the logistic and linear regression models, we performed subgroup analysis of mechanically ventilated patients (Table 4). Without mechanical ventilation, no significant outcome differences were detected between the intensivist and hospitalist groups when stratified by disease severity (Table 4).

Table 4.Subgroup Analysis: Stratified Mortality and Length of Stay of Patients With and Without Mechanical Ventilation

	Without Mechanical Ventilation			With Mechanical Ventilation
	Hospitalist % (No. Died)	Intensivist % (No. Died)	P Value	Hospitalist % (No. Died)	Intensivist % (No. Died)	P Value
	Hospitalist Days (Patients)	Intensivist Days (Patients)	P Value	Hospitalist Days (Patients)	Intensivist Days (Patients)	P Value
NOTE: All patients included in analyses (hospitalist ICU team patients: n = 828 with 91 requiring mechanical ventilation; intensivist‐led ICU team patients: n = 528 with 274 requiring mechanical ventilation). Abbreviations: LOS, length of stay; SAPS, simplified acute physiology score.
In‐hospital mortality
SAPS 33	1.4 (5)	2.2 (2)	0.63	0.0 (0)	8.9 (4)	0.22
SAPS 3451	5.2 (15)	4.3 (5)	0.72	27.5 (11)	17.4 (19)	0.18
SAPS 52	20.0 (20)	15.6 (7)	0.53	57.1 (20)	50.0 (60)	0.46
ICU mortality
SAPS 33	0.6 (2)	1.1 (1)	0.60	0.0 (0)	8.9 (4)	0.22
SAPS 3451	3.1 (9)	3.5 (4)	0.86	27.5 (11)	15.6 (17)	0.10
SAPS 52	11.0 (11)	11.1 (5)	0.98	54.3 (19)	43.3 (52)	0.26
Hospital LOS
SAPS 33	5.6 (347)	6.5 (93)	0.25	13.8 (16)	11.8 (45)	0.50
SAPS 3451	8.7 (290)	7.9 (116)	0.28	17.8 (40)	10.6 (109)	<0.001
SAPS 52	10.2 (100)	11.0 (45)	0.62	12.7 (35)	14.3 (120)	0.56
ICU LOS (days)
SAPS 33	1.9 (347)	2.2 (93)	0.30	8.0 (16)	7.0 (45)	0.67
SAPS 3451	2.6 (290)	2.8 (116)	0.52	10.6 (40)	7.2 (109)	0.02
SAPS 52	3.5 (100)	2.8 (45)	0.17	8.1 (35)	9.2 (120)	0.61

With mechanical ventilation, patients with intermediate illness severity had a significantly shorter hospital LOS (10.6 vs 17.8 days, P < 0.001) and ICU LOS (7.2 vs 10.6 days, P = 0.02) when managed by the intensivist‐led team (Table 4). When the calculations were repeated for only the patients who survived hospitalization, the shorter ICU LOS (6.5 vs 10.9 days, P = 0.01) remained significant but not the hospital LOS (10.3 vs 19.3 days, P = 0.10). The patients with intermediate acuity also showed a trend toward a decreased ICU mortality (15.6% vs 27.5%, P = 0.10) when managed by the intensivist‐led team (Table 4).

Adjusting for relevant risk factors, no statistically significant mortality rate difference was demonstrated between the hospitalist and intensivist‐led teams when evaluating all patients or patients without mechanical ventilation. The result, however, was inconclusive for patients with mechanical ventilation and did not allow refutation of the null hypothesis because the confidence interval for the mortality rate difference crossed the prespecified mortality difference threshold for clinical significance (Figure 2).

Figure 2

Study Limitations

Our study has several obvious limitations. It uses an observational design within a single hospital. However, this is seen in prior comparisons of intensivists to non‐intensivists.15, 810 Our study is unique with its prospective design and sample‐size calculation to demonstrate no difference in outcomes. Because our data is from a single center, it eliminates practice differences encountered when comparing multiple institutions, but it may also limit its generalizability.

Another major limitation in our comparison of an intensivist‐led ICU team to a hospitalist ICU team is their composition. Instead of 2 multidisciplinary teams, we compared a hospitalist's performance to that of a group of physicians at various levels of training. Similar comparisons have been seen in prior studies. For example, in the large study by Levy et al., half of the intensivists studied were in academic centers affiliated with teaching teams.18 Housestaff involvement, however, may have confounded the intensivist‐led team's patient outcomes. Tenner et al. demonstrated improved survival and decreased LOS in a pediatric ICU when hospitalists provided after‐hours coverage instead of residents.30 Furthermore, the patient census varied between the ICU teams, potentially impacting outcomes. While each service had only 1 attending, the hospitalist team had 1 clinician caring for patients whereas the intensivist‐led team had 5 to 6 clinicians. This study's implications may be more relevant to academic centers. A similar study of hospitalists and intensivists conducted in a nonteaching institution may yield different results.

Our 2 patient groups had substantial differences in illness severity and mechanical ventilation. Despite statistical techniques to address potential confounders in observational trials including stratification, multivariable adjustment, and propensity scores,29 residual confounders may still remain that influence the results and thus our conclusions. SAPS II is a validated method to objectively quantify disease severity and provide predictive mortality,27 however, it has known deficiencies. The use of propensity scores may not fully account for selection biases in team assignments introduced by the ED physicians. Biases may stem from the ICU teams' awareness of the ongoing study, and each team may have tried to maintain improved outcomes.

Additionally, the mortality outcomes represent in‐hospital mortality, not 30‐day mortality. This may be a less‐useful indicator of ICU performance because of post‐ICU transitions to extended care facilities and emphasis on end‐of‐life care. The majority of patients from both ICU models, however, did transfer to inpatient medical units under the care of non‐ICU hospitalist teams. Furthermore, this study did not capture important outcomes reported in other investigations, such as discharge disposition or quality of life after discharge.3132 Finally, the adjusted odds ratio for the intensivist‐led team's in‐hospital mortality (aOR 0.8, P = 0.23), referent to the hospitalist, does not eliminate the possibility that an intensivist‐led model may reduce mortality risk.

Our study suggests that intermediate‐ and high‐acuity, mechanically ventilated patients may benefit from care by intensivists rather than hospitalists. The results from this initial study could be used to design and estimate sample size for future studies of hospitalists and intensivists to elucidate risk reduction. Randomized and multicenter trials are needed to provide more robust data, because our subgroups were small and not accounted for in the sample size calculation. Considering the severe intensivist shortage, 1 strategy to provide effective and efficient coverage of the growing American ICU population may be to ask hospitalists to care independently for lower acuity ICU patientsespecially nonventilated patientswhile encouraging or requiring intensivist care for higher acuity patients, especially once mechanically ventilated.

Conclusion

We anticipate this initial study of hospitalist and intensivist‐led ICU teams will validate a hospitalist ICU staffing model for further investigation. We propose that hospitalists can provide quality care for lower acuity critical care patients. This may improve intensivist availability to higher acuity critically ill patients and allow for judicious utilization of the limited intensivist supply. Future studies may better delineate specific subgroups of critically ill patients who benefit most from intensivist primary involvement. Additional research may also help generate evidence‐based triage standards to appropriate critical care teams and foster guideline development. Hospitalists may be instrumental in the critical care staffing shortage, however, identification of their ideal role requires further study.

Design Overview

We used a prospective observational design to measure patient mortality and LOS within 2 medical ICU staffing paradigms. This was a collaborative study between the Division of Hospital Medicine and the Division of Pulmonary and Critical Care Medicine, with approval from Emory University's Institutional Review Board.

The hospitalist ICU model was staffed by a board certified internal medicine attending, with clinical responsibilities limited to the ICU. An intensivist‐led consult team (members distinct from the intensivist‐led ICU team) was staffed by a board certified pulmonary critical care attending and non‐physician providers. This consult team comanaged mechanically ventilated patients and was available for additional critical care consultation at the hospitalists' discretion. The intensivist‐led ICU model was staffed by a board certified pulmonary critical care attending, a pulmonary critical care fellow (postgraduate years 46) and 4 internal medicine residents (postgraduate years 23).

For their respective patients, the hospitalist and intensivist‐led teams participated in similar multidisciplinary ICU rounds with the charge nurse, respiratory therapist, and pharmacist. Both teams used the same evidence‐based ICU protocols and order sets. The hospitalists and intensivists were aware of the ongoing study.

Setting and Participants

Our study was conducted in an urban, community teaching hospital that is affiliated with a major regional academic university and has 400 medical‐surgical beds, including 56 ICU beds. All medical ICU patients receiving primary medical care from the hospitalist or intensivist‐led team were assessed for inclusion between October 2007 and September 2008. Predetermined exclusion criteria included surgery under general anesthesia, outside hospital transfers, pregnancy, and age under 18.

Selection of the admitting ICU team followed existing institutional referral patterns. For emergency department (ED) patients, the ED physicians made the decision to admit to the ICU and contacted an ICU team based on the respiratory support needs of the patient, not the admitting diagnosis. ED patients with respiratory failure who required invasive ventilatory support were admitted to the intensivist‐led ICU team. Those without invasive ventilatory support were admitted to the hospitalist ICU team, including ones with respiratory failure requiring noninvasive ventilation. Patients transferred from a hospital floor bed to the ICU by non‐hospitalist physicians were assigned to the intensivist‐led ICU team, while those transferred by hospitalist floor teams were assigned to the hospitalist ICU team, regardless of diagnosis or respiratory support needs. Patient assignments deviated from these patterns, however, based on ICU teams' census. The intensivist‐led ICU team had a strict limit of 20 patients, established by the residency program, and the hospitalist ICU team had a preferred limit of 12 patients.

Measurement of Outcomes and Follow‐Up

Study endpoints were in‐hospital and ICU mortality, as well as hospital and ICU LOS. Patient characteristics and outcome data were collected prospectively from medical records and hospital databases by 2 trained research nurses according to study protocol. For data collection training, 1 investigator (K.R.W.) reviewed sample data from 108 patients to ensure consistency and accuracy of data abstraction. Patients with several ICU admissions during 1 hospitalization had ICU data collected only from the first ICU entry, consistent with other trials' methodology.1, 4, 10, 1820 Additional ICU entries did not change ICU LOS derived only from the initial entry, but did contribute to hospital LOS. Data from patients with multiple ICU entries was analyzed with the original team assignment. All patients were followed until death or hospital discharge.

Statistical Analysis

Sample size was determined a priori using an expected inpatient mortality of 10% from historical data, power of 80%, and 2‐sided alpha of 0.05 to demonstrate no difference in outcomes, defined as a mortality difference of <5% between teams. This mortality difference used for the power calculation is consistent with other trial designs.2125 The required sample size was 1306 patients calculated using PASS software (version 2008, NCSS, Kaysville, UT), accounting for an expected 3:2 admission rate to the hospitalists. The statistician was blinded to team assignments.

Clinical characteristics of the groups were compared with the Student t test. The outcome and predictor variable distributions were examined with univariate analyses. Bivariate analyses were calculated for each predictor and endpoint. Multiple logistic and linear regression analyses were performed. Propensity scores were used and defined as the conditional probability of admission to the hospitalist versus intensivist‐led ICU team given a patient's covariates. It included all predictors in Table 1 and was calculated using logistic regression. Outcome measures were excluded from the regression.

A generalized linear model (GENMOD), using a binomial distribution and an identity link function,26 assessed the in‐hospital and ICU mortality rate differences between teams while controlling for major risk factors identified. GENMOD, however, does not accommodate several covariates, as it often fails for lack of convergence. Hence, logistic regression models with adjusted odds ratios (aOR) are reported as well.

The initial logistic regression model for in‐hospital and ICU mortality included all 20 independent variables from patient demographics, comorbidities, simplified acute physiology score (SAPS) II,27 respiratory support, central venous catheter (CVC) utilization, which included peripherally placed central catheters, and all terms for 2‐way interactions with team assignment. To determine the best model, a hierarchical backward elimination was executed while assessing for interactions, confounding, and estimate precision. Before removing a regression term, a likelihood ratio test was applied to each coefficient followed by Wald's chi square test.28 Collinearity diagnostic for nonlinear models was applied to look for multicollinearity. To exclude variables or regression terms, a condition index of 30 and variance decomposition proportion of 0.5 were used. The final model was evaluated for goodness‐of‐fit using the Hosmer and Lemeshow test.

LOS was analyzed with linear regression using the same covariates and backward elimination as the logistic model. Goodness‐of‐fit was evaluated using coefficient of determination (r²). A variance inflation factor of 10 was used to assess for collinearity. Two‐sided P values 0.05 were considered statistically significant. All analyses were performed with SAS software (version 9.1, SAS Institute, Cary, NC).

Patients

A total of 1747 patients received critical care from the hospitalist or intensivist‐led teams (Figure 1). Of the 1367 patients who met inclusion criteria, complete data was available for 1356 patients. The ED was the ICU admission source for 68.8% of hospitalist and 69.9% of intensivist patients. Baseline patient demographics were similar (Table 1). Among preexisting comorbidities, morbid obesity was more prevalent in hospitalist patients, whereas cancer and pulmonary and immunological diseases were more prevalent in intensivist patients (Table 1).

Figure 1

Hospitalist patients, compared to intensivist patients, had a lower mean SAPS II (37.4 vs 45.1, P < 0.001), less noninvasive (17.9% vs 25.8%, P < 0.001) and mechanical (11.0% vs 51.9%, P < 0.001) ventilation utilization, and fewer CVCs (29.1% vs 50.8%, P < 0.001) (Table 1). The intensivist‐led consult team comanaged 18.4% of the hospitalist patients. These 152 patients had a mean SAPS II of 41, and 19.7% required noninvasive ventilation while 44.1% required mechanical ventilation. For mechanically ventilated hospitalist patients, 80.2% were comanaged by the intensivist‐led consult team, 5.5% had palliative care for end‐of‐life management, 6.6% died imminently, and 7.7% received short‐term ventilation managed by the hospitalist. Hospitalist and intensivist‐led teams' ICU readmission rates were similar (7.0% vs 5.9%, P = 0.41).

Outcomes

Of the 1356 patients, there were 168 (12.4%) deaths and 135 (10.0%) occurred in the ICU. The overall mean ICU LOS was 4.0 days (SD 5.9), and mean hospital LOS was 9.1 days (SD 9.0). The mean hospital LOS for survivors was 9.0 days (SD 8.8).

Subgroup Analysis

Since each team's respiratory support utilization differed greatly and was a significant variable in the logistic and linear regression models, we performed subgroup analysis of mechanically ventilated patients (Table 4). Without mechanical ventilation, no significant outcome differences were detected between the intensivist and hospitalist groups when stratified by disease severity (Table 4).

Table 4.Subgroup Analysis: Stratified Mortality and Length of Stay of Patients With and Without Mechanical Ventilation

	Without Mechanical Ventilation			With Mechanical Ventilation
	Hospitalist % (No. Died)	Intensivist % (No. Died)	P Value	Hospitalist % (No. Died)	Intensivist % (No. Died)	P Value
	Hospitalist Days (Patients)	Intensivist Days (Patients)	P Value	Hospitalist Days (Patients)	Intensivist Days (Patients)	P Value
NOTE: All patients included in analyses (hospitalist ICU team patients: n = 828 with 91 requiring mechanical ventilation; intensivist‐led ICU team patients: n = 528 with 274 requiring mechanical ventilation). Abbreviations: LOS, length of stay; SAPS, simplified acute physiology score.
In‐hospital mortality
SAPS 33	1.4 (5)	2.2 (2)	0.63	0.0 (0)	8.9 (4)	0.22
SAPS 3451	5.2 (15)	4.3 (5)	0.72	27.5 (11)	17.4 (19)	0.18
SAPS 52	20.0 (20)	15.6 (7)	0.53	57.1 (20)	50.0 (60)	0.46
ICU mortality
SAPS 33	0.6 (2)	1.1 (1)	0.60	0.0 (0)	8.9 (4)	0.22
SAPS 3451	3.1 (9)	3.5 (4)	0.86	27.5 (11)	15.6 (17)	0.10
SAPS 52	11.0 (11)	11.1 (5)	0.98	54.3 (19)	43.3 (52)	0.26
Hospital LOS
SAPS 33	5.6 (347)	6.5 (93)	0.25	13.8 (16)	11.8 (45)	0.50
SAPS 3451	8.7 (290)	7.9 (116)	0.28	17.8 (40)	10.6 (109)	<0.001
SAPS 52	10.2 (100)	11.0 (45)	0.62	12.7 (35)	14.3 (120)	0.56
ICU LOS (days)
SAPS 33	1.9 (347)	2.2 (93)	0.30	8.0 (16)	7.0 (45)	0.67
SAPS 3451	2.6 (290)	2.8 (116)	0.52	10.6 (40)	7.2 (109)	0.02
SAPS 52	3.5 (100)	2.8 (45)	0.17	8.1 (35)	9.2 (120)	0.61

With mechanical ventilation, patients with intermediate illness severity had a significantly shorter hospital LOS (10.6 vs 17.8 days, P < 0.001) and ICU LOS (7.2 vs 10.6 days, P = 0.02) when managed by the intensivist‐led team (Table 4). When the calculations were repeated for only the patients who survived hospitalization, the shorter ICU LOS (6.5 vs 10.9 days, P = 0.01) remained significant but not the hospital LOS (10.3 vs 19.3 days, P = 0.10). The patients with intermediate acuity also showed a trend toward a decreased ICU mortality (15.6% vs 27.5%, P = 0.10) when managed by the intensivist‐led team (Table 4).

Adjusting for relevant risk factors, no statistically significant mortality rate difference was demonstrated between the hospitalist and intensivist‐led teams when evaluating all patients or patients without mechanical ventilation. The result, however, was inconclusive for patients with mechanical ventilation and did not allow refutation of the null hypothesis because the confidence interval for the mortality rate difference crossed the prespecified mortality difference threshold for clinical significance (Figure 2).

Figure 2

Study Limitations

Our study has several obvious limitations. It uses an observational design within a single hospital. However, this is seen in prior comparisons of intensivists to non‐intensivists.15, 810 Our study is unique with its prospective design and sample‐size calculation to demonstrate no difference in outcomes. Because our data is from a single center, it eliminates practice differences encountered when comparing multiple institutions, but it may also limit its generalizability.

Another major limitation in our comparison of an intensivist‐led ICU team to a hospitalist ICU team is their composition. Instead of 2 multidisciplinary teams, we compared a hospitalist's performance to that of a group of physicians at various levels of training. Similar comparisons have been seen in prior studies. For example, in the large study by Levy et al., half of the intensivists studied were in academic centers affiliated with teaching teams.18 Housestaff involvement, however, may have confounded the intensivist‐led team's patient outcomes. Tenner et al. demonstrated improved survival and decreased LOS in a pediatric ICU when hospitalists provided after‐hours coverage instead of residents.30 Furthermore, the patient census varied between the ICU teams, potentially impacting outcomes. While each service had only 1 attending, the hospitalist team had 1 clinician caring for patients whereas the intensivist‐led team had 5 to 6 clinicians. This study's implications may be more relevant to academic centers. A similar study of hospitalists and intensivists conducted in a nonteaching institution may yield different results.

Our 2 patient groups had substantial differences in illness severity and mechanical ventilation. Despite statistical techniques to address potential confounders in observational trials including stratification, multivariable adjustment, and propensity scores,29 residual confounders may still remain that influence the results and thus our conclusions. SAPS II is a validated method to objectively quantify disease severity and provide predictive mortality,27 however, it has known deficiencies. The use of propensity scores may not fully account for selection biases in team assignments introduced by the ED physicians. Biases may stem from the ICU teams' awareness of the ongoing study, and each team may have tried to maintain improved outcomes.

Additionally, the mortality outcomes represent in‐hospital mortality, not 30‐day mortality. This may be a less‐useful indicator of ICU performance because of post‐ICU transitions to extended care facilities and emphasis on end‐of‐life care. The majority of patients from both ICU models, however, did transfer to inpatient medical units under the care of non‐ICU hospitalist teams. Furthermore, this study did not capture important outcomes reported in other investigations, such as discharge disposition or quality of life after discharge.3132 Finally, the adjusted odds ratio for the intensivist‐led team's in‐hospital mortality (aOR 0.8, P = 0.23), referent to the hospitalist, does not eliminate the possibility that an intensivist‐led model may reduce mortality risk.

Our study suggests that intermediate‐ and high‐acuity, mechanically ventilated patients may benefit from care by intensivists rather than hospitalists. The results from this initial study could be used to design and estimate sample size for future studies of hospitalists and intensivists to elucidate risk reduction. Randomized and multicenter trials are needed to provide more robust data, because our subgroups were small and not accounted for in the sample size calculation. Considering the severe intensivist shortage, 1 strategy to provide effective and efficient coverage of the growing American ICU population may be to ask hospitalists to care independently for lower acuity ICU patientsespecially nonventilated patientswhile encouraging or requiring intensivist care for higher acuity patients, especially once mechanically ventilated.

Conclusion

We anticipate this initial study of hospitalist and intensivist‐led ICU teams will validate a hospitalist ICU staffing model for further investigation. We propose that hospitalists can provide quality care for lower acuity critical care patients. This may improve intensivist availability to higher acuity critically ill patients and allow for judicious utilization of the limited intensivist supply. Future studies may better delineate specific subgroups of critically ill patients who benefit most from intensivist primary involvement. Additional research may also help generate evidence‐based triage standards to appropriate critical care teams and foster guideline development. Hospitalists may be instrumental in the critical care staffing shortage, however, identification of their ideal role requires further study.

Design Overview

We used a prospective observational design to measure patient mortality and LOS within 2 medical ICU staffing paradigms. This was a collaborative study between the Division of Hospital Medicine and the Division of Pulmonary and Critical Care Medicine, with approval from Emory University's Institutional Review Board.

The hospitalist ICU model was staffed by a board certified internal medicine attending, with clinical responsibilities limited to the ICU. An intensivist‐led consult team (members distinct from the intensivist‐led ICU team) was staffed by a board certified pulmonary critical care attending and non‐physician providers. This consult team comanaged mechanically ventilated patients and was available for additional critical care consultation at the hospitalists' discretion. The intensivist‐led ICU model was staffed by a board certified pulmonary critical care attending, a pulmonary critical care fellow (postgraduate years 46) and 4 internal medicine residents (postgraduate years 23).

For their respective patients, the hospitalist and intensivist‐led teams participated in similar multidisciplinary ICU rounds with the charge nurse, respiratory therapist, and pharmacist. Both teams used the same evidence‐based ICU protocols and order sets. The hospitalists and intensivists were aware of the ongoing study.

Setting and Participants

Our study was conducted in an urban, community teaching hospital that is affiliated with a major regional academic university and has 400 medical‐surgical beds, including 56 ICU beds. All medical ICU patients receiving primary medical care from the hospitalist or intensivist‐led team were assessed for inclusion between October 2007 and September 2008. Predetermined exclusion criteria included surgery under general anesthesia, outside hospital transfers, pregnancy, and age under 18.

Selection of the admitting ICU team followed existing institutional referral patterns. For emergency department (ED) patients, the ED physicians made the decision to admit to the ICU and contacted an ICU team based on the respiratory support needs of the patient, not the admitting diagnosis. ED patients with respiratory failure who required invasive ventilatory support were admitted to the intensivist‐led ICU team. Those without invasive ventilatory support were admitted to the hospitalist ICU team, including ones with respiratory failure requiring noninvasive ventilation. Patients transferred from a hospital floor bed to the ICU by non‐hospitalist physicians were assigned to the intensivist‐led ICU team, while those transferred by hospitalist floor teams were assigned to the hospitalist ICU team, regardless of diagnosis or respiratory support needs. Patient assignments deviated from these patterns, however, based on ICU teams' census. The intensivist‐led ICU team had a strict limit of 20 patients, established by the residency program, and the hospitalist ICU team had a preferred limit of 12 patients.

Measurement of Outcomes and Follow‐Up

Study endpoints were in‐hospital and ICU mortality, as well as hospital and ICU LOS. Patient characteristics and outcome data were collected prospectively from medical records and hospital databases by 2 trained research nurses according to study protocol. For data collection training, 1 investigator (K.R.W.) reviewed sample data from 108 patients to ensure consistency and accuracy of data abstraction. Patients with several ICU admissions during 1 hospitalization had ICU data collected only from the first ICU entry, consistent with other trials' methodology.1, 4, 10, 1820 Additional ICU entries did not change ICU LOS derived only from the initial entry, but did contribute to hospital LOS. Data from patients with multiple ICU entries was analyzed with the original team assignment. All patients were followed until death or hospital discharge.

Statistical Analysis

Sample size was determined a priori using an expected inpatient mortality of 10% from historical data, power of 80%, and 2‐sided alpha of 0.05 to demonstrate no difference in outcomes, defined as a mortality difference of <5% between teams. This mortality difference used for the power calculation is consistent with other trial designs.2125 The required sample size was 1306 patients calculated using PASS software (version 2008, NCSS, Kaysville, UT), accounting for an expected 3:2 admission rate to the hospitalists. The statistician was blinded to team assignments.

Clinical characteristics of the groups were compared with the Student t test. The outcome and predictor variable distributions were examined with univariate analyses. Bivariate analyses were calculated for each predictor and endpoint. Multiple logistic and linear regression analyses were performed. Propensity scores were used and defined as the conditional probability of admission to the hospitalist versus intensivist‐led ICU team given a patient's covariates. It included all predictors in Table 1 and was calculated using logistic regression. Outcome measures were excluded from the regression.

A generalized linear model (GENMOD), using a binomial distribution and an identity link function,26 assessed the in‐hospital and ICU mortality rate differences between teams while controlling for major risk factors identified. GENMOD, however, does not accommodate several covariates, as it often fails for lack of convergence. Hence, logistic regression models with adjusted odds ratios (aOR) are reported as well.

The initial logistic regression model for in‐hospital and ICU mortality included all 20 independent variables from patient demographics, comorbidities, simplified acute physiology score (SAPS) II,27 respiratory support, central venous catheter (CVC) utilization, which included peripherally placed central catheters, and all terms for 2‐way interactions with team assignment. To determine the best model, a hierarchical backward elimination was executed while assessing for interactions, confounding, and estimate precision. Before removing a regression term, a likelihood ratio test was applied to each coefficient followed by Wald's chi square test.28 Collinearity diagnostic for nonlinear models was applied to look for multicollinearity. To exclude variables or regression terms, a condition index of 30 and variance decomposition proportion of 0.5 were used. The final model was evaluated for goodness‐of‐fit using the Hosmer and Lemeshow test.

LOS was analyzed with linear regression using the same covariates and backward elimination as the logistic model. Goodness‐of‐fit was evaluated using coefficient of determination (r²). A variance inflation factor of 10 was used to assess for collinearity. Two‐sided P values 0.05 were considered statistically significant. All analyses were performed with SAS software (version 9.1, SAS Institute, Cary, NC).

Patients

A total of 1747 patients received critical care from the hospitalist or intensivist‐led teams (Figure 1). Of the 1367 patients who met inclusion criteria, complete data was available for 1356 patients. The ED was the ICU admission source for 68.8% of hospitalist and 69.9% of intensivist patients. Baseline patient demographics were similar (Table 1). Among preexisting comorbidities, morbid obesity was more prevalent in hospitalist patients, whereas cancer and pulmonary and immunological diseases were more prevalent in intensivist patients (Table 1).

Figure 1

Hospitalist patients, compared to intensivist patients, had a lower mean SAPS II (37.4 vs 45.1, P < 0.001), less noninvasive (17.9% vs 25.8%, P < 0.001) and mechanical (11.0% vs 51.9%, P < 0.001) ventilation utilization, and fewer CVCs (29.1% vs 50.8%, P < 0.001) (Table 1). The intensivist‐led consult team comanaged 18.4% of the hospitalist patients. These 152 patients had a mean SAPS II of 41, and 19.7% required noninvasive ventilation while 44.1% required mechanical ventilation. For mechanically ventilated hospitalist patients, 80.2% were comanaged by the intensivist‐led consult team, 5.5% had palliative care for end‐of‐life management, 6.6% died imminently, and 7.7% received short‐term ventilation managed by the hospitalist. Hospitalist and intensivist‐led teams' ICU readmission rates were similar (7.0% vs 5.9%, P = 0.41).

Outcomes

Of the 1356 patients, there were 168 (12.4%) deaths and 135 (10.0%) occurred in the ICU. The overall mean ICU LOS was 4.0 days (SD 5.9), and mean hospital LOS was 9.1 days (SD 9.0). The mean hospital LOS for survivors was 9.0 days (SD 8.8).

Subgroup Analysis

Since each team's respiratory support utilization differed greatly and was a significant variable in the logistic and linear regression models, we performed subgroup analysis of mechanically ventilated patients (Table 4). Without mechanical ventilation, no significant outcome differences were detected between the intensivist and hospitalist groups when stratified by disease severity (Table 4).

Table 4.Subgroup Analysis: Stratified Mortality and Length of Stay of Patients With and Without Mechanical Ventilation

	Without Mechanical Ventilation			With Mechanical Ventilation
	Hospitalist % (No. Died)	Intensivist % (No. Died)	P Value	Hospitalist % (No. Died)	Intensivist % (No. Died)	P Value
	Hospitalist Days (Patients)	Intensivist Days (Patients)	P Value	Hospitalist Days (Patients)	Intensivist Days (Patients)	P Value
NOTE: All patients included in analyses (hospitalist ICU team patients: n = 828 with 91 requiring mechanical ventilation; intensivist‐led ICU team patients: n = 528 with 274 requiring mechanical ventilation). Abbreviations: LOS, length of stay; SAPS, simplified acute physiology score.
In‐hospital mortality
SAPS 33	1.4 (5)	2.2 (2)	0.63	0.0 (0)	8.9 (4)	0.22
SAPS 3451	5.2 (15)	4.3 (5)	0.72	27.5 (11)	17.4 (19)	0.18
SAPS 52	20.0 (20)	15.6 (7)	0.53	57.1 (20)	50.0 (60)	0.46
ICU mortality
SAPS 33	0.6 (2)	1.1 (1)	0.60	0.0 (0)	8.9 (4)	0.22
SAPS 3451	3.1 (9)	3.5 (4)	0.86	27.5 (11)	15.6 (17)	0.10
SAPS 52	11.0 (11)	11.1 (5)	0.98	54.3 (19)	43.3 (52)	0.26
Hospital LOS
SAPS 33	5.6 (347)	6.5 (93)	0.25	13.8 (16)	11.8 (45)	0.50
SAPS 3451	8.7 (290)	7.9 (116)	0.28	17.8 (40)	10.6 (109)	<0.001
SAPS 52	10.2 (100)	11.0 (45)	0.62	12.7 (35)	14.3 (120)	0.56
ICU LOS (days)
SAPS 33	1.9 (347)	2.2 (93)	0.30	8.0 (16)	7.0 (45)	0.67
SAPS 3451	2.6 (290)	2.8 (116)	0.52	10.6 (40)	7.2 (109)	0.02
SAPS 52	3.5 (100)	2.8 (45)	0.17	8.1 (35)	9.2 (120)	0.61

With mechanical ventilation, patients with intermediate illness severity had a significantly shorter hospital LOS (10.6 vs 17.8 days, P < 0.001) and ICU LOS (7.2 vs 10.6 days, P = 0.02) when managed by the intensivist‐led team (Table 4). When the calculations were repeated for only the patients who survived hospitalization, the shorter ICU LOS (6.5 vs 10.9 days, P = 0.01) remained significant but not the hospital LOS (10.3 vs 19.3 days, P = 0.10). The patients with intermediate acuity also showed a trend toward a decreased ICU mortality (15.6% vs 27.5%, P = 0.10) when managed by the intensivist‐led team (Table 4).

Adjusting for relevant risk factors, no statistically significant mortality rate difference was demonstrated between the hospitalist and intensivist‐led teams when evaluating all patients or patients without mechanical ventilation. The result, however, was inconclusive for patients with mechanical ventilation and did not allow refutation of the null hypothesis because the confidence interval for the mortality rate difference crossed the prespecified mortality difference threshold for clinical significance (Figure 2).

Figure 2

Study Limitations

Our study has several obvious limitations. It uses an observational design within a single hospital. However, this is seen in prior comparisons of intensivists to non‐intensivists.15, 810 Our study is unique with its prospective design and sample‐size calculation to demonstrate no difference in outcomes. Because our data is from a single center, it eliminates practice differences encountered when comparing multiple institutions, but it may also limit its generalizability.

Another major limitation in our comparison of an intensivist‐led ICU team to a hospitalist ICU team is their composition. Instead of 2 multidisciplinary teams, we compared a hospitalist's performance to that of a group of physicians at various levels of training. Similar comparisons have been seen in prior studies. For example, in the large study by Levy et al., half of the intensivists studied were in academic centers affiliated with teaching teams.18 Housestaff involvement, however, may have confounded the intensivist‐led team's patient outcomes. Tenner et al. demonstrated improved survival and decreased LOS in a pediatric ICU when hospitalists provided after‐hours coverage instead of residents.30 Furthermore, the patient census varied between the ICU teams, potentially impacting outcomes. While each service had only 1 attending, the hospitalist team had 1 clinician caring for patients whereas the intensivist‐led team had 5 to 6 clinicians. This study's implications may be more relevant to academic centers. A similar study of hospitalists and intensivists conducted in a nonteaching institution may yield different results.

Our 2 patient groups had substantial differences in illness severity and mechanical ventilation. Despite statistical techniques to address potential confounders in observational trials including stratification, multivariable adjustment, and propensity scores,29 residual confounders may still remain that influence the results and thus our conclusions. SAPS II is a validated method to objectively quantify disease severity and provide predictive mortality,27 however, it has known deficiencies. The use of propensity scores may not fully account for selection biases in team assignments introduced by the ED physicians. Biases may stem from the ICU teams' awareness of the ongoing study, and each team may have tried to maintain improved outcomes.

Additionally, the mortality outcomes represent in‐hospital mortality, not 30‐day mortality. This may be a less‐useful indicator of ICU performance because of post‐ICU transitions to extended care facilities and emphasis on end‐of‐life care. The majority of patients from both ICU models, however, did transfer to inpatient medical units under the care of non‐ICU hospitalist teams. Furthermore, this study did not capture important outcomes reported in other investigations, such as discharge disposition or quality of life after discharge.3132 Finally, the adjusted odds ratio for the intensivist‐led team's in‐hospital mortality (aOR 0.8, P = 0.23), referent to the hospitalist, does not eliminate the possibility that an intensivist‐led model may reduce mortality risk.

Our study suggests that intermediate‐ and high‐acuity, mechanically ventilated patients may benefit from care by intensivists rather than hospitalists. The results from this initial study could be used to design and estimate sample size for future studies of hospitalists and intensivists to elucidate risk reduction. Randomized and multicenter trials are needed to provide more robust data, because our subgroups were small and not accounted for in the sample size calculation. Considering the severe intensivist shortage, 1 strategy to provide effective and efficient coverage of the growing American ICU population may be to ask hospitalists to care independently for lower acuity ICU patientsespecially nonventilated patientswhile encouraging or requiring intensivist care for higher acuity patients, especially once mechanically ventilated.

Conclusion

We anticipate this initial study of hospitalist and intensivist‐led ICU teams will validate a hospitalist ICU staffing model for further investigation. We propose that hospitalists can provide quality care for lower acuity critical care patients. This may improve intensivist availability to higher acuity critically ill patients and allow for judicious utilization of the limited intensivist supply. Future studies may better delineate specific subgroups of critically ill patients who benefit most from intensivist primary involvement. Additional research may also help generate evidence‐based triage standards to appropriate critical care teams and foster guideline development. Hospitalists may be instrumental in the critical care staffing shortage, however, identification of their ideal role requires further study.