Effect of Collaborative Care on Cost in an ICU
Effect of Collaborative Care on Cost in an ICU
This prospective study was performed in the 13-bed medical ICU of MetroHealth Medical Center, a 450-bed urban, county-owned, university-affiliated teaching hospital located in Cleveland, Ohio. The study used a database including all admissions to that ICU from July 14, 2002, to December 31, 2006, excluding March through May 2004, when data collection was suspended because of personnel limitations. This ICU is a closed-model unit in which care of all patients is provided by a rotating team comprising a board-certified intensivist, an ICU fellow, and a group of 5 house officers who take overnight call on a 1 in 5 rotation. Formal rounds on all patients are made twice daily by this team. The intensivist rotates for a period of 2 weeks. The ICU fellow and house officers rotate on a 4-week schedule. Neither the intensivist nor the ICU fellow routinely stay in the hospital overnight. The mean nurse to patient ratio is 1 to 2.
All physician members of the ICU team are licensed and can make independent decisions. Although all physician members of the ICU team can write orders, the orders are usually written by the more junior house officers. As usual in such teams, the team functions hierarchically, with the fellow providing guidance and direction to junior house staff, and the intensivist providing guidance and direction to all the others. Although physician orders can be written at any time, much of the major decision making and order writing occurs during formal rounds, which are led by the intensivist.
To evaluate the impact of the physicians on costs and adjust for potentially confounding variables, we used quantile regression to construct a multivariable model for the median value of discretionary costs on the initial day in the ICU (D1COST, indexed to 2012 US dollars). Discretionary ICU costs were defined as those generated during the patient's ICU stay from the pharmacy, radiology department (imaging and interventional radiology), laboratories, blood bank, and echocardiography. Costs were obtained from the hospital's cost analysis system (Trendstar, McKesson Corporation). Because of the structure of the hospital's cost accounting, the initial ICU day was taken as the calendar day, not the initial 24 hours in the ICU. The time spent in the ICU on ICU day 1 was measured in fractional hours.
The inclusion criterion was all ICU admissions during the study period. Because extreme outliers of cost are most likely quite different from costs for the remainder of patients, the outliers were excluded from analysis. The long, sparse tail of the distribution of D1COST is illustrated by noting that its 85th, 90th, 95th, 99th, and 100th percentiles were $3026, $3706, $5146, $15 261, and $34 670, respectively. Thus, admissions with D1COST in the top 1% of that distribution, that is, exceeding $15 261, were excluded. Also, in order to ensure that there was a reasonable time to incur costs, we excluded patients who were in the ICU for 2 hours or less on the initial ICU day.
The ICU database includes the identity of the responsible intensivist and fellow for each day. During the years under study, 9 intensivists and 12 ICU fellows rotated through the ICU. To assess the individual and joint influences of the intensivists and fellows on costs, a 2-stage modeling process was used. First, the independent influences of the 2 types of physicians were assessed by including them in the model as separate sets of indicator variables. That model was used to identify the rank order of the intensivists, and separately of the fellows, in generating ICU discretionary costs. Based on those ranks, the intensivists were divided into terciles according to the costs they generated; for example, the 3 intensivists associated with the lowest adjusted D1COST were grouped together into the lowest tercile of intensivist-associated costs. Similarly, the ICU fellows were grouped into terciles of fellow-associated costs. The rationale for grouping physicians into 3 groups was the intent to assess the conjoint effects of intensivists and fellows on costs by creating interaction terms in the second stage of modeling. Without such grouping, there would be 88 interaction terms; grouping them into terciles reduced this to 4.
The second stage of modeling incorporated the 2 indicator variables representing the terciles of intensivists, the 2 indicator variables representing the terciles of the ICU fellows, and the 4 interaction terms. Regression coefficients of the interaction terms indicate whether the combined influences on costs of intensivists and fellows are independent or whether discretionary costs are influenced by interactions between the 2 types of physicians.
Modeling for D1COST included adjustments for patients' demographics, comorbid conditions, the type and severity of acute illness, the source of ICU admission, existence of care limitation orders in ICU, whether ICU admission occurred during nights (8 PM-8 AM) or weekends, whether the ICU episode represented a readmission to the ICU (although not necessarily during the same hospitalization), whether the patient survived the stay in the ICU, and the number of hours spent in the ICU on ICU day 1 (D1LOS). Demographics were age, sex, and race dichotomized into white versus non-white. Comorbid illness was quantified as the presence of 31 specific preexisting conditions, as recorded in the hospital's administrative and billing database. The 31 conditions were subdivided into 2 groups. Group 1 comprised conditions that are predictive of poorer outcomes; group 2 comprised comorbid conditions that, because of coding bias present in administrative data, are associated with lower mortality. The 2 comorbid illness variables included in the models were the number of conditions present within each of these groups. The main diagnosis responsible for ICU admission was categorized into the organ system responsible for ICU admission (respiratory, cardiovascular, gastrointestinal, neurologic, miscellaneous medical conditions, and surgical conditions including trauma). Severity of acute illness was measured as: (1) the Glasgow Coma Scale (GCS) score, (2) the need for invasive mechanical ventilation during ICU day 1, and (3) the acute physiology score from the Acute Physiology and Chronic Health Evaluation (APACHE) II after its neurologic subcomponent is removed (APS-Neuro), because it was already explicitly included as the GCS score. The source of admission to the ICU was categorized as the emergency department, hospital general care area, another ICU, an outside hospital, or other sources. Care limitations were represented by 2 variables representing: (1) the presence of an order before ICU admission that put limits on the use of life-supporting therapy, and (2) the initiation of such an order on ICU day 1. The time spent in the ICU on ICU day 1 is important to account for differences in costs related to varying times spent in the ICU on that day.
Because of the expectation that a disproportionate fraction of the first day's discretionary costs are expended during the initial hours in the ICU, the relationship between D1COST and D1LOS was allowed to be nonlinear. Indeed, linearity of the relationship between each of the continuous independent variables and D1COST was assessed graphically, and by use of restricted cubic splines. Such variables were represented by cubic splines if linearity did not hold.
Model goodness-of-fit was assessed as the pseudo-R for quantile regression. Standard errors we calculated by bootstrapping, with 600 repetitions. Continuous data are presented as mean (SD) or median (interquartile range). Categorical data are presented as proportions. P values less than .05 are considered significant. Statistical analysis was performed by using Stata 10 (StataCorp LP). This study was approved by the hospital's institutional review board.
Methods
This prospective study was performed in the 13-bed medical ICU of MetroHealth Medical Center, a 450-bed urban, county-owned, university-affiliated teaching hospital located in Cleveland, Ohio. The study used a database including all admissions to that ICU from July 14, 2002, to December 31, 2006, excluding March through May 2004, when data collection was suspended because of personnel limitations. This ICU is a closed-model unit in which care of all patients is provided by a rotating team comprising a board-certified intensivist, an ICU fellow, and a group of 5 house officers who take overnight call on a 1 in 5 rotation. Formal rounds on all patients are made twice daily by this team. The intensivist rotates for a period of 2 weeks. The ICU fellow and house officers rotate on a 4-week schedule. Neither the intensivist nor the ICU fellow routinely stay in the hospital overnight. The mean nurse to patient ratio is 1 to 2.
All physician members of the ICU team are licensed and can make independent decisions. Although all physician members of the ICU team can write orders, the orders are usually written by the more junior house officers. As usual in such teams, the team functions hierarchically, with the fellow providing guidance and direction to junior house staff, and the intensivist providing guidance and direction to all the others. Although physician orders can be written at any time, much of the major decision making and order writing occurs during formal rounds, which are led by the intensivist.
To evaluate the impact of the physicians on costs and adjust for potentially confounding variables, we used quantile regression to construct a multivariable model for the median value of discretionary costs on the initial day in the ICU (D1COST, indexed to 2012 US dollars). Discretionary ICU costs were defined as those generated during the patient's ICU stay from the pharmacy, radiology department (imaging and interventional radiology), laboratories, blood bank, and echocardiography. Costs were obtained from the hospital's cost analysis system (Trendstar, McKesson Corporation). Because of the structure of the hospital's cost accounting, the initial ICU day was taken as the calendar day, not the initial 24 hours in the ICU. The time spent in the ICU on ICU day 1 was measured in fractional hours.
The inclusion criterion was all ICU admissions during the study period. Because extreme outliers of cost are most likely quite different from costs for the remainder of patients, the outliers were excluded from analysis. The long, sparse tail of the distribution of D1COST is illustrated by noting that its 85th, 90th, 95th, 99th, and 100th percentiles were $3026, $3706, $5146, $15 261, and $34 670, respectively. Thus, admissions with D1COST in the top 1% of that distribution, that is, exceeding $15 261, were excluded. Also, in order to ensure that there was a reasonable time to incur costs, we excluded patients who were in the ICU for 2 hours or less on the initial ICU day.
The ICU database includes the identity of the responsible intensivist and fellow for each day. During the years under study, 9 intensivists and 12 ICU fellows rotated through the ICU. To assess the individual and joint influences of the intensivists and fellows on costs, a 2-stage modeling process was used. First, the independent influences of the 2 types of physicians were assessed by including them in the model as separate sets of indicator variables. That model was used to identify the rank order of the intensivists, and separately of the fellows, in generating ICU discretionary costs. Based on those ranks, the intensivists were divided into terciles according to the costs they generated; for example, the 3 intensivists associated with the lowest adjusted D1COST were grouped together into the lowest tercile of intensivist-associated costs. Similarly, the ICU fellows were grouped into terciles of fellow-associated costs. The rationale for grouping physicians into 3 groups was the intent to assess the conjoint effects of intensivists and fellows on costs by creating interaction terms in the second stage of modeling. Without such grouping, there would be 88 interaction terms; grouping them into terciles reduced this to 4.
The second stage of modeling incorporated the 2 indicator variables representing the terciles of intensivists, the 2 indicator variables representing the terciles of the ICU fellows, and the 4 interaction terms. Regression coefficients of the interaction terms indicate whether the combined influences on costs of intensivists and fellows are independent or whether discretionary costs are influenced by interactions between the 2 types of physicians.
Modeling for D1COST included adjustments for patients' demographics, comorbid conditions, the type and severity of acute illness, the source of ICU admission, existence of care limitation orders in ICU, whether ICU admission occurred during nights (8 PM-8 AM) or weekends, whether the ICU episode represented a readmission to the ICU (although not necessarily during the same hospitalization), whether the patient survived the stay in the ICU, and the number of hours spent in the ICU on ICU day 1 (D1LOS). Demographics were age, sex, and race dichotomized into white versus non-white. Comorbid illness was quantified as the presence of 31 specific preexisting conditions, as recorded in the hospital's administrative and billing database. The 31 conditions were subdivided into 2 groups. Group 1 comprised conditions that are predictive of poorer outcomes; group 2 comprised comorbid conditions that, because of coding bias present in administrative data, are associated with lower mortality. The 2 comorbid illness variables included in the models were the number of conditions present within each of these groups. The main diagnosis responsible for ICU admission was categorized into the organ system responsible for ICU admission (respiratory, cardiovascular, gastrointestinal, neurologic, miscellaneous medical conditions, and surgical conditions including trauma). Severity of acute illness was measured as: (1) the Glasgow Coma Scale (GCS) score, (2) the need for invasive mechanical ventilation during ICU day 1, and (3) the acute physiology score from the Acute Physiology and Chronic Health Evaluation (APACHE) II after its neurologic subcomponent is removed (APS-Neuro), because it was already explicitly included as the GCS score. The source of admission to the ICU was categorized as the emergency department, hospital general care area, another ICU, an outside hospital, or other sources. Care limitations were represented by 2 variables representing: (1) the presence of an order before ICU admission that put limits on the use of life-supporting therapy, and (2) the initiation of such an order on ICU day 1. The time spent in the ICU on ICU day 1 is important to account for differences in costs related to varying times spent in the ICU on that day.
Because of the expectation that a disproportionate fraction of the first day's discretionary costs are expended during the initial hours in the ICU, the relationship between D1COST and D1LOS was allowed to be nonlinear. Indeed, linearity of the relationship between each of the continuous independent variables and D1COST was assessed graphically, and by use of restricted cubic splines. Such variables were represented by cubic splines if linearity did not hold.
Model goodness-of-fit was assessed as the pseudo-R for quantile regression. Standard errors we calculated by bootstrapping, with 600 repetitions. Continuous data are presented as mean (SD) or median (interquartile range). Categorical data are presented as proportions. P values less than .05 are considered significant. Statistical analysis was performed by using Stata 10 (StataCorp LP). This study was approved by the hospital's institutional review board.
Source...