Each node is based on available data from registry patient hospitalizations
for each predictive variable presented. BUN indicates blood urea nitrogen.
To convert BUN to mmol/L, multiply by 0.357; creatinine to μmol/L, multiply
Log odds of mortality was calculated for all records in the derivation
cohort and risk group cut points established at percentile rankings equivalent
to those of the classification and regression tree model (65th, 78th, 94th,
and 98th percentiles).
Fonarow GC, Adams KF, Abraham WT, Yancy CW, Boscardin WJ, ADHERE Scientific Advisory Committee,
Study Group, and Investigators FT. Risk Stratification for In-Hospital Mortality in Acutely Decompensated Heart FailureClassification and Regression Tree Analysis. JAMA. 2005;293(5):572-580. doi:10.1001/jama.293.5.572
Author Affiliations: Ahmanson–UCLA Cardiomyopathy
Center, University of California, Los Angeles Medical Center, Los Angeles
(Dr Fonarow); Department of Biostatistics, University of California, Los Angeles
(Dr Boscardin); Division of Cardiology, University of North Carolina, Chapel
Hill (Dr Adams); Department of Cardiology, Ohio State University Medical Center,
Columbus (Dr Abraham); and Division of Cardiology, University of Texas Southwestern
Medical Center, Dallas (Dr Yancy).
Context Estimation of mortality risk in patients hospitalized with acute decompensated
heart failure (ADHF) may help clinicians guide care.
Objective To develop a practical user-friendly bedside tool for risk stratification
for patients hospitalized with ADHF.
Design, Setting, and Patients The Acute Decompensated Heart Failure National Registry (ADHERE) of
patients hospitalized with a primary diagnosis of ADHF in 263 hospitals in
the United States was queried with analysis of patient data to develop a risk
stratification model. The first 33 046 hospitalizations (derivation cohort;
October 2001-February 2003) were analyzed to develop the model and then the
validity of the model was prospectively tested using data from 32 229 subsequent
hospitalizations (validation cohort; March-July 2003). Patients had a mean
age of 72.5 years and 52% were female.
Main Outcome Measure Variables predicting mortality in ADHF.
Results When the derivation and validation cohorts are combined, 37 772 (58%)
of 65 275 patient-records had coronary artery disease. Of a combined cohort
consisting of 52 164 patient-records, 23 910 (46%) had preserved left ventricular
systolic function. In-hospital mortality was similar in the derivation (4.2%)
and validation (4.0%) cohorts. Recursive partitioning of the derivation cohort
for 39 variables indicated that the best single predictor for mortality was
high admission levels of blood urea nitrogen (≥43 mg/dL [15.35 mmol/L])
followed by low admission systolic blood pressure (<115 mm Hg) and then
by high levels of serum creatinine (≥2.75 mg/dL [243.1 μmol/L]). A simple
risk tree identified patient groups with mortality ranging from 2.1% to 21.9%.
The odds ratio for mortality between patients identified as high and low risk
was 12.9 (95% confidence interval, 10.4-15.9) and similar results were seen
when this risk stratification was applied prospectively to the validation
Conclusions These results suggest that ADHF patients at low, intermediate, and high
risk for in-hospital mortality can be easily identified using vital sign and
laboratory data obtained on hospital admission. The ADHERE risk tree provides
clinicians with a validated, practical bedside tool for mortality risk stratification.
Heart failure causes considerable morbidity and mortality and is responsible
for a tremendous burden on the health care system in the United States.1 It accounted for approximately 1 million hospital
discharges in 2001, an increase of 164% since 1979, and is associated with
an overall annual cost of nearly $29 billion.1 Reported
in-hospital mortality ranges from as low as 2.3% among patients enrolled in
clinical trials to 19% in referral hospital series.2,3
Despite this dramatic increase in the public health burden of hospitalization
for heart failure, models for the risk stratification of patients during admission
for acute decompensated heart failure (ADHF) are not well established. Clinical
risk prediction tools may be helpful in guiding medical decision making. Patients
estimated to be at a lower risk may be managed with less intensive monitoring
and therapies available on a telemetry unit or hospital ward, whereas a patient
estimated to be at a higher risk may require more intensive management in
an intensive or coronary care unit. Previous studies, generally based on outpatients
with chronic heart failure, have identified a number of variables that are
associated with increased mortality, including etiology,4 patient
age,5 clinical assessment at the time of hospitalization,4 cardiothoracic ratio,5,6 peak
oxygen consumption,6 left ventricular ejection
fraction,6 serum sodium concentration,5 serum creatinine concentration,4,5 and
B-type natriuretic peptide concentration.7
In contrast, several factors have limited the development of similar
models in patients with ADHF. Lack of a consistent definition of ADHF, different
nomenclature to describe its clinical features, incomplete data available
in administrative data sets, and varying statistical methods have hindered
the development of risk stratification tools.8- 10 Consequently,
unlike acute coronary syndromes, in which several systems have been developed
for risk stratification,11- 17 no
clinically practical method of risk stratification exists for patients hospitalized
The objective of the present analysis was to develop and validate a
practical and user-friendly method of risk stratification for in-hospital
mortality among patients admitted with ADHF that could be applicable to routine
clinical practice. Data used to model risk were taken from the Acute Decompensated
Heart Failure National Registry (ADHERE).18,19 This
registry collects detailed hospitalization data from initial presentation
in the hospital or emergency department until discharge, transfer, or in-hospital
death.19 As an observational database, these
data reflect current real-world treatment patterns and in-hospital clinical
outcomes for patients hospitalized with ADHF.
The ADHERE registry contains data on patients hospitalized with ADHF
in 263 community, tertiary, and academic centers from all regions of the United
States.20 For the purpose of the registry,
ADHF is defined as new-onset decompensated heart failure or decompensation
of chronic, established heart failure with symptoms sufficient to warrant
hospitalization. The design, methods, and patient characteristics in the ADHERE
registry have been described previously.20 Briefly,
medical records are reviewed by trained abstractors at participating study
sites and data from consecutive eligible male and female patients aged 18
years or older at the time of hospitalization are entered into the registry
using an electronic case report form incorporating real-time validity checking.19,20 These data include demographic information,
medical history, baseline clinical characteristics, initial evaluation, treatment
received, procedures performed, hospital course, and patient disposition.18- 20 Standardized definitions
are used for all patient-related variables, clinical diagnoses, and hospital
Institutional review board approval is required for all participating
centers; however, informed consent of individuals is not required for registry
entry.19,20 To preserve patient
confidentiality, direct patient identifiers are not collected. Data are reported
only in aggregate format. Therefore, registry entries reflect individual hospitalization
episodes, not individual patients, and multiple hospitalizations of the same
patient may be entered into the registry as separate records.
The current analysis is based on the initial 65 275 patient-records
entered into the registry. For this analysis, predictors of in-hospital mortality
were determined from an initial derivation cohort consisting of data from
October 2001 to February 2003 (33 046 hospitalizations). These data were subjected
to classification and regression tree (CART) analysis to identify the best
predictors of in-hospital mortality and develop the risk stratification model.
The validity of this model was then independently assessed using data from
the second validation cohort, consisting of the subsequent 32 229 hospitalization
episodes (March 2003 to July 2003).
Admission and/or medical staff recorded race/ethnicity, usually as the
patient was registered, using hospital-defined race/ethnicity. Patients were
assigned to only 1 race/ethnicity category. Prior studies in patients hospitalized
with heart failure have suggested different mortality risk based on race/ethnicity.
Race/ethnicity was also a significant univariate predictor of in-hospital
mortality in the ADHERE derivation cohort. Race/ethnicity was thus included
as one of the 39 variables for CART and logistic regression analysis.
The CART method is an empirical, statistical technique based on recursive
partitioning analysis.22- 24 Unlike
multivariable logistic regression, it is well suited to the generation of
clinical decision rules.23,24 Furthermore,
because it does not require parametric assumptions, it can handle numerical
data that are highly skewed or multimodal and categorical predictors with
either an ordinal or nonordinal structure.23,24 The
CART method involves the segregation of different values of classification
variables through a decision tree composed of progressive binary splits. Every
value of each predictor variable is considered as a potential split, and the
optimal split is selected based on impurity criterion (the reduction in the
residual sum of squares due to a binary split of the data at that tree node).
When missing values are encountered in considering a split, they are ignored
and the probability and impurity measures are calculated from the nonmissing
values of that variable. Each parent node in the decision tree produces 2
child nodes, which in turn can become parent nodes producing additional child
nodes. This process continues with both tree building and pruning until statistical
analysis indicates that the tree fits without overfitting the information
contained in the data set.23 As a result, CART
analysis produces decision trees that are simple to interpret and may be applied
at the bedside.
An open-source adaptation of the CART algorithm (tree library in S-PLUS,
version 6.0, Insightful Corp, Seattle, Wash) was used to analyze 39 potential
clinical variables of interest in the derivation cohort (Table 1). These variables were selected from 80 variables collected
in the ADHERE registry by virtue of predicting in-hospital mortality on univariate
analysis or having been identified in previous published studies as risk factors
for mortality. Nodes in the CART tree were constrained to have a minimum size
of 800 records in parent nodes and 400 records in final child nodes. A 10-fold
cross-validation was used to assess the predictive ability of the tree model.
Mortality was calculated for each of the terminal nodes in the CART tree and
used to generate the risk stratification model. The predictive value of this
model was then assessed by determination of mortality odds ratios (ORs) and
95% confidence intervals (CIs) between risk groups.
The ability of the derived risk tree to identify ADHF patients at low,
intermediate, and high risk for in-hospital mortality was tested. The patients
from the validation cohort were classified into risk groups based on the CART
tree. Mortality for these risk groups and the mortality ORs and 95% CIs between
risk groups were determined and these data were compared with those of the
Finally, a multivariate logistic regression model was constructed from
the derivation cohort (logistic procedure in SAS version 8.2, SAS Institute
Inc, Cary, NC), tested in the validation cohort, and the accuracy of the CART
and logistic regression models was compared using area under receiver operating
characteristic curves in the derivation and validation cohorts.
Baseline characteristics and main outcomes of the 33 046 hospitalization
episodes used to develop the model (derivation cohort) and the 32 229
hospitalization episodes used to test the model (validation cohort) are shown
in Table 2 and Table 3. The derivation and validation cohorts were similar with
respect to age at admission; sex distribution; heart failure history; medical
history; and clinical symptoms, vital signs, and laboratory values (Table
2). Clinical outcomes were also similar between the 2 cohorts (Table 3).
Of the 39 variables evaluated, the CART method identified blood urea
nitrogen (BUN) level of 43 mg/dL or higher (≥15.35 mmol/L) at admission
as the best single discriminator between hospital survivors and nonsurvivors.
The next best predictor of in-hospital mortality in both the higher and lower
BUN nodes was systolic blood pressure (SBP) at a discrimination level of less
than 115 mm Hg. For the node with patients having a BUN level of 43 mg/dL
or higher (≥15.35 mmol/L) and SBP of less than 115 mm Hg, a serum creatinine
level of 2.75 mg/dL or higher (≥243.1 μmol/L) provided additional prognostic
value. Figure 1 depicts the final tree
generated by the CART analysis along with the mortality data for each child
node of this tree. These branch points permit patient stratification into
5 risk groups: high risk (BUN level ≥43 mg/dL [≥15.35 mmol/L], SBP <115
mm Hg, and creatinine level ≥2.75 mg/dL [≥243.1 μmol/L]), intermediate
risk 1 (BUN level ≥43 mg/dL [≥15.35 mmol/L], SBP <115 mm Hg, and
creatinine level <2.75 mg/dL [<243.1 μmol/L]), intermediate risk
2 (BUN level ≥43 mg/dL [≥15.35 mmol/L] and SBP ≥115 mm Hg), intermediate
risk 3 (BUN level <43 mg/dL [<15.35 mmol/L] and SBP <115 mm Hg),
and low risk (BUN level <43 mg/dL [<15.35 mmol/L] and SBP ≥115 mm
Hg). Table 4 summarizes the clinical
characteristics of patients in these 5 risk groups. The mortality OR between
the high- and low-risk groups was 12.9 (95% CI, 10.4-15.9), with statistically
significant differences in mortality risk detected between all risk groups
except intermediate risk groups 2 and 3 (Table
5). Although additional risk nodes involving additional variables
could be generated, they offered little incremental risk discrimination.
The decision tree generated by CART analysis of the derivation cohort
was tested for its ability to risk stratify patients in the validation cohort.
This risk tree was able to stratify patients into high, intermediate, and
low risk (Figure 2). The mortality OR
between the high- and low-risk groups was 10.4 (95% CI, 8.4-13.0), with statistically
significant differences detected between all risk groups except intermediate
risk groups 2 and 3 (Table 4). These absolute mortality rates, as well as
the clinical characteristics and mortality ORs between risk groups, were similar
to those of the derivation cohort (Table 4 and Table 5)and comparable risk
stratification occurred when the analysis was limited to the subset of validation
patients with new onset heart failure (in-hospital mortality: 23.6% in the
high-risk group; 20.0%, intermediate risk group 1; 5.0%, intermediate risk
group 2; 5.1%, intermediate risk group 3; and 1.8%, low-risk group).
Multivariate logistic regression identified BUN level, SBP, heart rate,
and age as the most significant mortality risk predictors:log
odds of mortality =0.0212 × BUN − 0.0192 ×
SBP+ 0.0131 × heart rate + 0.0288×
age − 4.72.
The addition of 24 predictors did not meaningfully increase the accuracy
of this model. Figure 3 compares in-hospital
mortality rates in the derivation and validation cohorts based on risk groups
determined by logistic regression. Based on the area under the receiver operating
characteristic curves, the accuracy of the CART model (derivation cohort:
68.7%; validation cohort: 66.8%) was modestly less than that of the more complicated
logistic regression model (derivation cohort: 75.9%; validation cohort: 75.7%).
This analysis of more than 65 000 recent ADHF hospitalizations
in patients demographically and clinically similar to those seen in other
large community- or Medicare-based evaluations25- 27 demonstrates
that the risk of in-hospital mortality can be reliably estimated using routinely
available vital signs and laboratory data obtained on hospital admission.
Overall, in-hospital mortality was 4.1%, but this mortality risk varied more
than 10-fold (from 2.1% to 21.9%) based on the patient’s initial SBP
and levels of BUN and creatinine.
Multiple evaluations of patients hospitalized for heart failure have
demonstrated an association between clinical outcomes and indices of renal
function and blood pressure.28- 32 In
the Enhanced Feedback for Effective Cardiac Treatment study, increasing BUN
levels and decreasing SBP were significant and independent predictors of both
30-day and 1-year mortality.28 In the Outcomes
of a Prospective Trial of Intravenous Milrinone for Exacerbations of Chronic
Heart Failure study, these same parameters were significant and independent
predictors of mortality or rehospitalization.31 Similarly,
in a retrospective review of 1004 consecutive patients hospitalized for heart
failure at 11 geographically diverse hospitals, worsening renal function was
associated with a 7.5-fold increase (95% CI, 2.9- to 19.3-fold increase) in
the adjusted risk of in-hospital mortality.32 In
a prospective analysis of 412 patients hospitalized at a single institution,
the 6-month mortality risk increased with decreasing renal function, which
was determined by the change in creatinine levels relative to baseline.30 Renal dysfunction causes further congestion and neurohormonal
activation, which are factors associated with adverse outcomes in patients
with heart failure.33
In addition to these parameters, other parameters that have been correlated
with clinical outcomes in patients hospitalized with heart failure include
sex34 ; heart failure etiology34,35 ;
history of previous heart failure hospitalizations31 ;
comorbid conditions such as cerebrovascular disease, dementia, chronic obstructive
pulmonary disease, hepatic cirrhosis, and cancer28,34 ;
respiratory rate28 ; anemia31,36,37 ;
serum sodium concentration28 ; B-type natriuretic
peptide levels38,39 ; left ventricular
ejection fraction29 ; and heart failure therapy.3 Because multiple risk factors can exist in the same
patient, risk factor analysis (to be meaningful) must consider factors in
combination rather than isolation. Because previous evaluations tended to
treat these factors as isolated entities, they have not produced aclinically
practical way of integrating various factors to stratify risk in heart failure
Unlike ADHF, several risk stratification schemes already exist for patients
with acute coronary syndromes.11- 17 These
schemes are typically based on multivariable analysis using logistic regression
or Cox proportional hazards models. Although schemes using as few as 313,17 to more than 20 variables are available,12 an acute coronary syndromes risk scoring scheme generally
involves 7 to 10 variables.11,14- 16 Such
clinical prediction models have been interpreted to be helpful for risk stratification
and management of acute coronary syndrome patients and have been integrated
into national guidelines.11
Although no in-hospital mortality risk stratification scheme is available
for patients hospitalized with heart failure, a heart failure survival score
has been developed and independently validated for ambulatory outpatients
with heart failure.40,41 This
score, based on 7 variables—heart failure etiology, heart rate, blood
pressure, serum sodium concentration, intraventricular conduction time, left
ventricular ejection fraction, and peak oxygen consumption—stratifies
patients into low (16%), medium (39%), and high (50%) mortality risk categories.40 In addition, hospitalization data have been used
to develop a risk score for heart failure readmission.42 This
risk score, which is based on 16 variables, was moderately predictive in a
derivation cohort but it has not been independently validated in a second
A significant disadvantage of multivariable-generated risk schemes is
their complexity. The number of variables and mathematical functions involved
frequently require access to a computer or an electronic calculator to generate
a score and to determine risk, making them impractical for bedside assessment
unless such tools are readily available. Even when converted to point scores,
the tools derived from a multivariate model still require a nomogram reference
to convert a point score to risk. Similar to multivariate regression techniques,
the CART method can detect interactions between variables.24 Moreover,
it yields a decision tree that is relatively easy to apply at the bedside,
leading to its potential use in a wide variety of clinical conditions, including
infections43 and neurological,44 oncological,45,46 and cardiac47 disorders.
In the current evaluation, the CART method identified 3 of 39 potential
variables as significant predictors of in-hospital mortality risk. In a simple
2- to 3-step process, these variables permit identification of patients with
low, intermediate, or high risk for in-hospital mortality. Furthermore, the
accuracy of this model, which can be easily applied at the bedside, is close
to that of the more complicated model derived from logistic regression. Alternately,
if computer access or a pocket digital assistant is available at the bedside,
the model derived from logistic regression may have advantages.
These validated models should aid medical decision making in patients
hospitalized with ADHF. Patients judged to be at higher risk may receive higher-level
monitoring and earlier, more intensive treatment for ADHF, while patients
estimated to be at lower risk may be reassured and managed less intensively.
In addition, these models may prove to be valuable in designing clinical trials
to evaluate heart failure therapies because they permit risk to be balanced
across treatment groups45 and allow for selective
inclusion criteria to enroll only patients at high risk for in-hospital mortality.
Potential limitations of the current analysis must be acknowledged.
Real-world practice information can be both an advantage and a disadvantage
of analyses based on registry data. Study results can be influenced by differences
in disease assessment, treatment, and documentation patterns at participating
institutions. The ADHERE registry reflects patients cared for by thousands
of clinicians at hundreds of hospitals across the country and thus has an
excellent chance to adjust for this variation and create a risk prediction
model that is robust for most situations. However, this model may not apply
to patients who are cared for in settings that deviate substantially from
those in ADHERE. In addition, each patient’s actual risk may be influenced
by many factors not measured or considered in this model. The CART method
favors variables available for analysis in the greatest proportion of patients.
Some potential risk factors, such as B-type natriuretic peptide, were obtained
in less than 25% of patients. Consequently, there may be additional variables
that either were not considered or were considered and rejected because of
limited data that could ultimately improve the risk discrimination if assessed
in a sufficient number of patients.
Therefore, this model enhances, not replaces, physician assessment.
Moreover, because the ADHERE registry does not contain specific patient identifiers,
information regarding patient status following hospital discharge is not available.
Thus the effects of any of the variables evaluated in this analysis on intermediate-
and long-term mortality risks cannot be determined. Similarly, because of
the lack of patient identifiers, the analyzed cohorts may contain multiple
admissions for the same patient. However, this should not have influenced
the study results because the outcome parameter, in-hospital mortality, and
the identified risk factors (admission SBP and admission levels of BUN and
creatinine) are specific to individual hospitalization episodes. Lastly, the
derivation and validation cohorts come from periods that differ both temporally
and in duration. Nonetheless, these cohorts are similar in size, baseline
characteristics, and clinical outcome. Despite these potential limitations,
the current CART-based analysis of the ADHERE registry has created a simple
robust tool to predict in-hospital mortality that is easy to use and has good
In patients hospitalized with ADHF, the risk of in-hospital mortality
can be quickly and accurately determined using admission clinical and laboratory
variables. Of 39 variables, BUN level of 43 mg/dL or higher (≥15.35 mmol/L),
serum creatinine level of 2.75 mg/dL or higher (≥243.1 μmol/L), and
SBP of less than 115 mm Hg were independent predictors of high risk for in-hospital
mortality in the current CART analysis. On the basis of these 3 variables,
ADHF patients can be readily stratified into groups at low, intermediate,
and high risk for in-hospital mortality, with mortality risks ranging from
2.1% to 21.9%. The finding that indices of renal status are 2 of the 3 predictors
providing the best mortality risk discrimination underscores the importance
of renal function in ADHF patients. The continued high mortality for patients
hospitalized with ADHF provides a compelling indication to apply tools, such
as the risk tree derived in this study, to improve the evaluation and, potentially,
management and outcomes of these patients.
Corresponding Author: Gregg C. Fonarow,
MD, Division of Cardiology, University of California, 10833 Le Conte Ave,
Los Angeles, CA 90095 (firstname.lastname@example.org).
Author Contributions: Dr Fonarow had full access
to all of 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: Fonarow, Adams, Abraham,
Acquisition of data: Fonarow, Adams, Abraham,
Analysis and interpretation of data: Fonarow,
Adams, Abraham, Yancy, Boscardin.
Drafting of the manuscript: Fonarow.
Critical revision of the manuscript for important
intellectual content: Adams, Abraham, Yancy, Boscardin.
Statistical analysis: Boscardin.
Administrative, technical, or material support:
Study supervision: Fonarow, Adams, Abraham,
Financial Disclosures: Drs Fonarow, Adams,
Abraham, and Yancy have received research grant funding and honoraria and
have served as consultants for Scios Inc. Dr Boscardin reported no financial
ADHERE Scientific Advisory Committee: William
T. Abraham, MD, Division of Cardiology, Ohio State University Medical Center,
Columbus; Kirkwood F. Adams, Jr, MD, Division of Cardiology, University of
North Carolina, Chapel Hill; Robert L. Berkowitz, MD, Heart Failure Program,
Hackensack University Hospital, Hackensack, NJ; Maria Rosa Costanzo, MD, Midwest
Heart Specialists, Edward Hospital, Naperville, Ill; Teresa DeMarco, MD, Division
of Cardiology, University of California, San Francisco; Charles L. Emerman,
MD, Department of Emergency Medicine, Cleveland Clinic Foundation and MetroHealth,
Cleveland, Ohio; Gregg C. Fonarow, MD, Ahmanson–UCLA Cardiomyopathy
Center, University of California, Los Angeles; Marie Galvao, CANP, Congestive
Heart Failure Program, Montefiore Medical Center, Bronx, NY; J. Thomas Heywood,
MD, Cardiomyopathy Program, Adult Cardiac Transplant, Loma Linda University
Medical Center, Loma Linda, Calif; Thierry H. LeJemtel, MD, Cardiology Division,
Albert Einstein Hospital, Bronx, NY; Lynne Warner Stevenson, MD, Cardiovascular
Division, Brigham and Women’s Hospital, Boston, Mass; Clyde W. Yancy,
MD, University of Texas Southwestern Medical Center, Cardiovascular Division,
St Paul University Hospital, Dallas.
ADHERE Study Group: Departments of Biostatistics
(Yu Ping Li, Janet Wynne, Mei Cheng, David Arnold) and Clinical Registries
(Jeannie M. Fiber), Scios Inc, Fremont, Calif.
The list of hospitals participating in the ADHERE Registry can be found
Funding/Support: The ADHERE registry and this
study were funded by Scios Inc.
Role of Sponsor: The sponsor used a committee
of academic advisors who were intimately involved in the design, conduct,
and interpretation of the ADHERE registry and this analysis of registry data.
The sponsor provided financial and material support for the registry and performed
the classification and regression tree analyses for this study based on the
input of the advisory committee. Dr Boscardin, who has no financial relationship
with the sponsor, recalculated and confirmed the classification and regression
tree analyses and also performed the logistic regression analyses. All of
the authors participated in the interpretation of the data and, along with
Dr Fonarow, had complete control over the contents of the manuscript and the
decision to submit the manuscript for publication.