Figure 1. Distribution of predicted mortality rates in medical patients admitted from the emergency department or the outpatient clinic. Density represents the likelihood of different values of patient severity among medical patients admitted from the emergency department or the outpatient clinic to the intensive care unit (ICU) (n = 31 555) or non-ICU (n = 257 755). The total area under each curve equals 1 when the x-axis is transformed to the log-odds scale.
Figure 2. Variation in hospital intensive care unit (ICU) admission rates for the typical patient. Each point-line combination represents a single hospital. The points correspond to the predicted probability of ICU admission for a typical patient (ie, 1.1% predicted 30-day mortality, diagnoses not separately modeled, mean occupancy, level 1 complexity). Lines represent 95% CIs.
Figure 3. Variation in hospital intensive care unit (ICU) admitting patterns for patients across the range of mortality risk. Rates are shown for patients with “other” diagnoses (ie, diagnoses that were not separately modeled), at the mean occupancy (57.8%), and at the highest ICU complexity (level 1). The band at the x-axis denotes the relative likelihood of predicted 30-day mortality values. Darker black indicates that the corresponding values were more likely, and lighter gray indicates that the corresponding values were less likely. The x- and y-axes are on the log-odds (or logit) scale.
Chen LM, Render M, Sales A, Kennedy EH, Wiitala W, Hofer TP. Intensive care unit admitting patterns in the Veterans Affairs health care system. Arch Intern Med.. Published online July 23, 2012. doi: 10.1001/archinternmed.2012.2606.
eTable 1. Inclusion Criteria for the Sample
eTable 2. Patient and Hospital Characteristics Associated with ICU Admission Rates, Adjusted
eAppendix. Mortality Analysis
This supplementary material has been provided by the authors to give readers additional information about their work.
Chen LM, Render M, Sales A, Kennedy EH, Wiitala W, Hofer TP. Intensive Care Unit Admitting Patterns in the Veterans Affairs Health Care System. Arch Intern Med. 2012;172(16):1220-1226. doi:10.1001/archinternmed.2012.2606
Author Affiliations: Veterans Affairs Health Services Research and Development Center of Excellence, Veterans Affairs Ann Arbor Healthcare System, Ann Arbor, Michigan (Drs Chen, Sales, Wiitala, and Hofer and Mr Kennedy); Division of General Medicine, Department of Internal Medicine, University of Michigan, Ann Arbor (Drs Chen and Hofer); Cincinnati Veterans Affairs Medical Center, Cincinnati, Ohio (Dr Render); Division of Pulmonary, Critical Care, and Sleep Medicine, University of Cincinnati, Cincinnati (Dr Render); and Veterans Affairs Inpatient Evaluation Center, Office of Informatics and Analytics, Cincinnati, Ohio (Dr Sales). Dr Sales is now with Veterans Affairs Health Services Research and Development Center of Excellence, Veterans Affairs Ann Arbor Healthcare System, and the School of Nursing, University of Michigan, Ann Arbor.
Background Critical care resource use accounts for almost 1% of US gross domestic product and varies widely among hospitals. However, we know little about the initial decision to admit a patient to the intensive care unit (ICU).
Methods To describe hospital ICU admitting patterns for medical patients after accounting for severity of illness on admission, we performed a retrospective cohort study of the first nonsurgical admission of 289 310 patients admitted from the emergency department or the outpatient clinic to 118 Veterans Affairs acute care hospitals between July 1, 2009, and June 30, 2010. Severity (30-day predicted mortality rate) was measured using a modified Veterans Affairs ICU score based on laboratory data and comorbidities around admission. The main outcome measure was direct admission to an ICU.
Results Of the 31 555 patients (10.9%) directly admitted to the ICU, 53.2% had 30-day predicted mortality at admission of 2% or less. The rate of ICU admission for this low-risk group varied from 1.2% to 38.9%. For high-risk patients (predicted mortality >30%), ICU admission rates also varied widely. For a 1-SD increase in predicted mortality, the adjusted odds of ICU admission varied substantially across hospitals (odds ratio = 0.85-2.22). As a result, 66.1% of hospitals were in different quartiles of ICU use for low- vs high-risk patients (weighted κ = 0.50).
Conclusions The proportion of low- and high-risk patients admitted to the ICU, variation in ICU admitting patterns among hospitals, and the sensitivity of hospital rankings to patient risk all likely reflect a lack of consensus about which patients most benefit from ICU admission.
Critical care resource use accounts for almost 1% of US gross domestic product and almost 15% of all hospital costs.1 Wide hospital variation in critical care resource use2- 5 has implications for the quality and cost of acute inpatient care. Underuse of critical care, as exemplified by high-acuity patients not admitted to the intensive care unit (ICU) in a timely manner, is associated with high mortality rates and greater resource use.6,7 On the other hand, unnecessary use of critical care wastes valuable resources. Better triage decisions on admission, that is, choices about the first hospital unit to which a patient is admitted, could aid in efforts to improve the quality and decrease the cost of acute inpatient care.
Previous research has explored the characteristics of patients referred to the ICU8- 11 and the severity of illness of patients once admitted to the ICU.12- 16 However, to our knowledge, no one has described what proportion of all the medical patients presenting for admission at a broad sample of hospitals in a national health care system are sent to the ICU or how this proportion varies across hospitals at all levels of patient severity. Given current efforts to increase health care efficiency, it is important to better understand variation in hospital admitting patterns (at the time of triage) and, ultimately, how to target use of the ICU to those who will benefit the most.
We used data from the Veterans Health Administration, the largest health care system in the United States and one that in 2009 provided care to almost 6 million US military veterans, to examine ICU admitting patterns for patients at their initial presentation to the hospital. With data from the Veterans Affairs (VA) Inpatient Evaluation Center,17 we addressed 3 questions: (1) What is the 30-day predicted mortality rate (severity) of medical patients admitted directly to the ICU from the emergency department (ED) or the outpatient clinic? (2) For patients with the same 30-day predicted mortality rate, how much does direct admission to the ICU vary between hospitals? (3) Are comparisons of hospital admitting patterns dependent on patient severity? In a secondary analysis, we examined the relationship between use of the ICU and 30- and 90-day mortality rates, although this analysis is complicated by the nonrandomized selection of patients into the ICU.
We performed a retrospective cohort study of all the adult nonsurgical admissions to any VA acute care hospital with on-site ICU admissions from the ED or the outpatient clinic between July 1, 2009, and June 30, 2010. The study was approved by the Veterans Affairs Ann Arbor Healthcare System institutional review board.
We used diagnosis, laboratory, and clinical data collected by the VA Inpatient Evaluation Center. We supplemented this information with data from the American Hospital Association and the Veterans Affairs Bed Control Database.
We sought to construct a cohort of patients for whom ED, outpatient clinic, and ICU physicians made triage decisions based largely on presenting patient severity. Therefore, we excluded admissions in which surgery was performed within 24 hours of presentation, as surgery acutely changes patient severity, and triage to the ICU for such patients may anticipate that change rather than reflect admission severity. Similarly, we excluded transfers because admitting decisions for these patients may be influenced by earlier severity assessment via telephone triage. We also excluded admissions missing patient identifiers (0.2% of remaining admissions). For each patient, we included only the first admission during the study period. We further excluded 2 sites missing ICU bed data and 1 site with an unadjusted ICU admission rate greater than 50%. The final sample included 289 310 patients, each with a single admission at 1 of 118 hospitals (eTable 1). In total, 92.6% of all patients (66 271 of 71 543) who went to the ICU when first arriving at the hospital came from the ED or the outpatient clinic.
The adjusted models included patient severity, diagnosis, and ICU occupancy (all at the time of admission), along with the level of specialty care offered at the hospital. We adjusted for these factors because they are not completely under hospitals' control yet may affect ICU admission rates.
Patient severity would seem to be the most obvious determinant of ICU use given that “[i]ntensive care units (ICU) are places in the hospital where the most seriously ill patients are cared for by specially trained staff.”18 We estimated patient severity (defined as the predicted 30-day mortality rate on admission conditional on not being admitted to the ICU)19 using a modified VA ICU severity score.20- 22 The score is derived from a model similar in principle to the APACHE (Acute Physiology, Age, Chronic Health Evaluation) ICU model23 and includes age on admission, 1 of 73 mutually exclusive diagnostic groups, comorbid disease burden adapted from the study by Elixhauser et al,24 admission source, and the worst values in the 24 hours surrounding hospital admission of 11 laboratory tests (sodium, serum urea nitrogen, glucose, albumin, and bilirubin levels, glomerular filtration rate, white blood cell count, hematocrit, pH, PaCO2, and PaO2).20- 22
We included indicators for the 6 most common diagnoses among ICU and non-ICU admissions using the Angus definition of sepsis25 and 5 other VA diagnostic cohorts defined by grouped International Classification of Diseases, Ninth Revision, Clinical Modification, codes.20 As common diagnoses for ICU and non-ICU admissions overlapped, we were left with 10 diagnostic categories (including the reference category of all the other diagnoses). We defined ICU occupancy for each patient as the proportion of ICU beds that were occupied by any VA patient (not only those in this sample) at the exact time of admission. Each VA hospital is assigned a type (1-4 included as dummy variables) representing a level of complexity based on the Society of Critical Care Medicine framework (eg, at level 1 hospitals, cardiology, neurosurgery, interventional cardiology, and radiology services are offered; at level 4 hospitals, more limited services are offered).26
We modeled the patient-level probability of admission to the ICU using multilevel models with random effects to account for the clustering of patients in hospitals.27 For the primary analysis, we fit 2 random intercept models. The first model was adjusted for severity of illness and case mix (admitting diagnosis). In the second model, we added ICU occupancy and level of complexity. We generated hospital-specific ICU admission rates using empirical Bayesian prediction, which accounts for hospital volume and accordingly shrinks less reliable estimates toward the overall mean.28
In a third model, we used a random slope to investigate whether the relationship between patient severity and triage to the ICU might differ across hospitals. We used the weighted κ of agreement29 with squared difference weights to describe the stability of hospital quartile rankings of ICU admission rates for patients with the lowest vs highest severity.
Ultimately, one of the most important questions is whether any variation in ICU use translates into differences in patient outcomes, the most compelling of which would be 30- and 90-day mortality rates. To attempt to control for the nonrandom selection of patients into the ICU, we used a generalized propensity approach,30 which ensures the balance of observed covariates and any highly correlated unobserved covariates when estimating a treatment effect of ICU hospitalization conditional on severity (eAppendix).
Although a protective effect of ICU hospitalization would be of interest, if ICU hospitalization is associated with increased mortality, we would have to infer that this might reflect increased unmeasured severity in individuals initially hospitalized in the ICU. We, thus, also examined an interaction of ICU hospitalization and severity as we hypothesized that we might detect a relative benefit of ICU use for patients of higher severity. We conducted all the analyses using a commercially available software program (STATA, version 11.0; StataCorp LP) and R 184.108.40.206
Of the 289 310 patients admitted from the ED or the outpatient clinic to 1 of 118 hospitals in 48 states, 31 555 (10.9%) were admitted directly to the ICU. Patients admitted to the ICU had higher mean predicted (and observed) 30-day mortality rates than did patients admitted to the non-ICU ward (7.5% [7.7%] vs 3.5% [3.5%]) (Table 1). Both populations had a relatively low predicted mortality rate at admission (median mortality of ICU and non-ICU patients of 1.7% and 1.0%, respectively). Compared with admissions to the ICU, the distribution of severity for non-ICU admissions was shifted toward lower severity (Figure 1).
Hospitals varied widely in the proportion of patients admitted to the ICU, even after adjusting for predicted mortality and diagnosis on admission (Table 2). Adjusted rates of admission to the ICU ranged from 1.6% to 29.5% across hospitals for patients with median predicted mortality. The median ICU admission rate was 6.9% (interquartile range, 4.7%-10.0%).
After further adjusting for occupancy and hospital type, rates of admission to the ICU still varied from 1.2% to 38.9% across hospitals. The median rate was 7.3% (interquartile range, 4.4%-10.7%) (Figure 2). Almost 10% of the total variation in probability of ICU admission was explained by patient severity and diagnosis, and an additional 0.4% was explained by ICU occupancy and the facility's complexity level (eTable 2).
Sicker patients were more likely to be admitted to the ICU (odds ratio [OR] for a 1-SD increase, 1.50; 95% CI, 1.44-1.55). However, for high-risk patients (predicted mortality >30%), ICU admission rates still varied widely (Figure 3). We found significant variation (SD of random slope, 0.18; 95% CI, 0.16-0.21) in the change in odds of ICU admission associated with a 1-SD increase in severity (ranging from a 15% decrease [OR = 0.85; 95% CI, 0.75-0.98] to a 122% increase [OR = 2.22; 95% CI, 1.93-2.55]) (represented by the varying slopes in Figure 3). From the median severity of 1.1%, a 1-SD increase in severity on the log-odds scale would correspond to an increase to 4.8% in severity. Three hospitals had negative slopes; sensitivity analyses with and without these hospitals did not change the results.
Given that hospitals' admitting patterns were not uniformly sensitive to changes in predicted mortality, 66.1% of hospitals were in different quartiles of ICU admission for patients with the lowest (30-day predicted mortality ≤2%) vs highest (30-day predicted mortality >30%) severity (weighted κ = 0.50) (Figure 3).
After adjusting for severity at admission, patients admitted to the ICU had a higher odds of death at 30 days (OR = 1.10; 95% CI, 1.03-1.17) but not at 90 days (OR = 0.96; 95% CI, 0.90-1.02; P = .16). However, when the treatment effect was estimated as a function of severity at admission, admission to the ICU was protective against death at 30 days for patients with expected mortality greater than 18.4% at admission (and protective against death at 90 days for patients with expected mortality >8.8%).
With growing focus on the nation's rising health care costs, understanding use of one of the most costly components of inpatient care—critical care—has become more important. In this context, this study had 4 key findings. Approximately half of all the patients admitted to the ICU had 30-day predicted mortality of 2% or less. In more than half the cases, patients with predicted mortality greater than 30% were not admitted to the ICU. At all levels of patient risk, hospitals varied widely in the proportion of patients admitted to the ICU. Comparisons of hospital admitting patterns were not stable for patients of differing risk.
Critical care units proliferated in the 1960s32 to offer life-sustaining therapy and closer monitoring for the sickest patients. However, the high costs associated with this level of care and finite ICU beds make the decision to send a patient to the ICU one that merits renewed attention. Ultimately, there are 3 possible outcomes of the triage decision: not using the ICU when it should be used, using it when it should not be used, and getting it right. The present findings of wide variation (1.2%-38.9%) in direct admission to the ICU for patients with median severity suggest that all of these outcomes may be occurring.
Many patients directly admitted to the ICU had a low chance of dying within 30 days. This, in part, reflects the fact that most patients presenting to the hospital were at low risk (ie, large denominator). The cohort was relatively healthy (first admission during the year and only direct admissions to the ICU). Considering this, the present results are consistent with those of earlier studies that found low median predicted mortalities (<9%) in ICU patients.12-15 We add to the existing literature by estimating the proportion of all patients of a given severity admitted to the ICU vs non-ICU and documenting wide variation in hospital admitting patterns at all levels of patient risk.
The admission of many low-severity patients to the ICU suggests that measureable severity of illness may represent only 1 of several factors contributing to the ICU triage decision. Other likely contributors include goals of care (comfort vs prolongation of life) and ability to solicit patient preferences, inherent risk of treatment (eg, insulin infusion and intravenous antihypertensive agents), and available alternative environments (eg, step-down units, telemetry units, and staffed procedure rooms). Admitting patterns may also reflect a belief that the ICU is better at preventing adverse outcomes in certain low-risk compared with high-risk patients (eg, a low-risk patient with diabetic ketoacidosis compared with a moderate-risk patient with congestive heart failure, chronic kidney disease, and pneumonia). However, given evidence that the value of ICU care may be limited for less severely ill patients,33- 37 it seems that reevaluation of the appropriate environment for low-risk patients would be timely and valuable.
Guidelines for the selection of patients to the ICU by the American College of Critical Care Medicine recommend that the lowest priority should be given to patients who are “too sick to benefit” or “too well to benefit” from ICU care.38 Although the present results suggest that we need to better understand who might be too well to benefit, similar conclusions might be made regarding ICU care for the sickest patients. Across all the hospitals in this sample, 69.6% of patients with predicted mortality of 30% or greater were not admitted to the ICU. We were unable to determine whether these patients were too sick to benefit or had changed their goals of care or whether the ICU might have provided added value for these patients.
The present findings on hospital rankings should be interpreted in the context of increasing efforts to improve hospital efficiency through the development of efficiency metrics.39,40 We found that hospitals most likely to use the ICU for low-risk patients frequently differed from hospitals most likely to use the ICU for high-risk patients. This finding suggests that ranking hospitals by their rates of ICU use would be misleading in the absence of stratification by severity.
In evaluating the mortality benefit of ICU treatment, we would not expect to see a constant protective effect of being treated in the ICU at all levels of severity. In fact, one would most reasonably expect that at low levels of severity, the ICU would confer little benefit, whereas it would add the most value for the severely ill. We found evidence supporting this general pattern because a clear benefit of ICU treatment was seen for the sickest patients. Although the likely selection effects of nonrandomized assignment to the ICU make it unwise to infer from these findings the precise threshold below which ICU admission confers more harm than benefit, the possibility that there might be such a threshold is provocative and provides further incentive to understand wide variation in ICU use.
The seemingly paradoxical finding that admission to the ICU confers a slightly higher risk of death at lower levels of severity may well reflect unmeasured factors used by physicians in selecting patients for ICU admission, although, again, propensity scores should ensure balance in the observed covariates and any unobserved covariates correlated with the severity measure. Better understanding of the factors used to select patients for ICU care or identification of instrumental variables that determine selection for treatment could help elucidate this apparent paradox.
This study has several limitations. Because we did not have information about patient preferences, patients who had declined ICU care or who were enrolled in hospice were included in the sample. In addition, although the measure of predicted mortality has been shown to have excellent predictive validity,20 we did not have vital sign data, and, thus, the mortality risk of patients who were evaluated with significant changes in vital signs but before organ decline might be underestimated. However, to undermine the present findings, these and any other unmeasured risk factors would have to be distributed across hospitals in very different ways than all the extensive comorbidity and physiology variables that we collected. The data set also lacked physician-level characteristics, so we could not make any conclusions about whether differing physician characteristics within and across hospitals might contribute to hospital admitting patterns. Finally, because we studied admitting patterns in a single health care system (albeit the largest one in the United States), caution should be used before generalizing these findings to other health systems.
In summary, we found that many high-risk patients were not directly admitted to the ICU but that approximately half of all medical patients directly admitted to the ICU were at low risk. Hospital admitting patterns varied widely for patients at every level of severity. These findings suggest that there may be a considerable lack of consensus about when to use the ICU. As a next step, it would be important to better understand (1) the value of ICU care for patients stratified by risk and (2) factors other than patient severity that are associated with hospital admitting patterns. This work could form the basis for validating standards for ICU admission that are tailored to different risk groups and that are sensitive to hospital-specific environments. Such standards could inform policymakers seeking to look beyond simple but incomplete hospital comparisons based only on aggregate rates of ICU use.
Correspondence: Lena M. Chen, MD, MS, Division of General Medicine, University of Michigan, 300 N Ingalls, Room 7E02, Ann Arbor, MI 48109 (firstname.lastname@example.org).
Accepted for Publication: April 26, 2012.
Published Online: July 23, 2012. doi:10.1001/archinternmed.2012.2606
Author Contributions: Dr Chen had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Study concept and design: Chen, Render, Sales, and Hofer. Acquisition of data: Chen, Render, Sales, Wiitala, and Hofer. Analysis and interpretation of data: Chen, Render, Sales, Kennedy, Wiitala, and Hofer. Drafting of the manuscript: Chen, Render, Sales, Kennedy, and Hofer. Critical revision of the manuscript for important intellectual content: Chen, Render, Sales, Kennedy, Wiitala, and Hofer. Statistical analysis: Chen, Render, Sales, Kennedy, Wiitala, and Hofer. Obtained funding: Render. Administrative, technical, or material support: Render and Wiitala. Study supervision: Render and Hofer.
Financial Disclosure: None reported.
Funding/Support: This material is the result of work supported with resources from the Veterans Affairs Health Services Research and Development Center of Excellence, Veterans Affairs Ann Arbor Healthcare System.
Disclaimer: The views expressed in this article are those of the authors and do not necessarily represent the views of the Department of Veterans Affairs or the University of Michigan.
Additional Contributions: Eve Kerr, MD, MPH, reviewed an earlier version of the manuscript.