Download Prognostic Value of a Novel Classification Scheme for Heart Failure

American Journal of Epidemiology
Copyright ª 2006 by the Johns Hopkins Bloomberg School of Public Health
All rights reserved; printed in U.S.A.
Vol. 164, No. 2
DOI: 10.1093/aje/kwj168
Advance Access publication May 17, 2006
Practice of Epidemiology
Prognostic Value of a Novel Classification Scheme for Heart Failure:
The Minnesota Heart Failure Criteria
Joseph Kim1,2, David R. Jacobs, Jr.2,3, Russell V. Luepker2, Eyal Shahar2, Karen L. Margolis4, and
Mark P. Becker5
1
Medical Statistics Unit, London School of Hygiene and Tropical Medicine, University of London, London, United Kingdom.
Division of Epidemiology and Community Health, School of Public Health, University of Minnesota, Minneapolis, MN.
3
Department of Nutrition, Faculty of Medicine, University of Oslo, Oslo, Norway.
4
Berman Center for Outcomes and Clinical Research, Minneapolis, MN.
5
Office of the Provost, University of South Carolina, Columbia, SC.
2
Received for publication August 11, 2004; accepted for publication January 25, 2006.
The authors present the Minnesota Heart Failure Criteria (MHFC), derived using latent class analysis from widely
available items in the Framingham Criteria. The authors used 1995 and 2000 data on hospitalized Minnesota Heart
Survey subjects discharged after myocardial infarction or heart failure (N ¼ 7,379). Selected Framingham Criteria
variables (dyspnea, pulmonary rales, cardiomegaly, interstitial or pulmonary edema on chest radiograph, S3 heart
sound, tachycardia) plus left ventricular ejection fraction were used. The discriminatory power of the MHFC was
evaluated using age- and sex-adjusted 2-year mortality. A five-class latent class analysis model was collapsed into
cases and noncases. Mortality estimates discriminated noncases (18%) from cases (43%) (p < 0.001). The MHFC
performed better than previous truncated criteria (Framingham Criteria: 26% noncases, 43% cases; Duke Criteria:
29%, 40%; Killip Score: 31%, 44%; Boston Score: 28%, 45%). In a subset of patients admitted for heart failure
(n ¼ 5,128), the MHFC identified all but 2% (116/4,746) of cases found with a nearly full version of the Framingham
Criteria. In terms of prognostic value, the MHFC are as precise as or more precise than several previous sets of
truncated criteria. They closely approximate a nearly full version of the Framingham Criteria but require many fewer
variables and can facilitate epidemiologic case-finding for heart failure.
cardiovascular diseases; classification; diagnosis; heart diseases; heart failure, congestive; validation studies
[publication type]
Abbreviations: ICD-9, International Classification of Diseases, Ninth Revision; LCA, latent class analysis; LVEF, left ventricular
ejection fraction; MHSAMI, Minnesota Heart Survey Acute Myocardial Infarction; MHSCHF, Minnesota Heart Survey
Community Surveillance of Heart Failure.
Heart failure is a major public health problem affecting
over five million persons in the United States and over 23
million worldwide (1). The epidemiologic study of heart
failure, however, has been partly impeded by the absence
of a universally agreed-upon case definition (2–12). Though
numerous case definitions have been introduced into the
literature—primarily in the form of epidemiologic classifi-
cation schemes (13–19)—most, if not all, of these criteria
are based solely on a composite of signs and symptoms and
ignore left ventricular ejection fraction (LVEF), a key index
of underlying cardiac physiology for which there was limited data availability when the criteria were being developed
(3, 5, 8, 20). In addition, existing criteria were evaluated
using standard measures of validity (e.g., sensitivity and
Correspondence to Dr. Joseph Kim, Medical Statistics Unit, Department of Epidemiology and Population Health, London School of Hygiene
and Tropical Medicine, Keppel Street, London WC1E 7HT, United Kingdom (e-mail: [email protected]).
184
Am J Epidemiol 2006;164:184–193
Prognostic Value of the Minnesota Heart Failure Criteria
specificity), all of which assume a perfect reference standard; given the absence of a diagnostic gold standard for
heart failure, there is considerable uncertainty in the estimates of validity obtained for these criteria.
In their original conception, the Framingham Criteria—
arguably the most widely used epidemiologic criteria for
heart failure—were defined by combining 17 ‘‘major’’ or
‘‘minor’’ heart-failure-related variables to form a diagnosis
of heart failure in dichotomous categories (heart failure/no
heart failure) (21). However, in practice, the Framingham
Criteria are often implemented using only a subset of these
variables, since data on many of the characteristics (e.g.,
circulation time) are no longer collected in routine clinical
practice. It would therefore be of interest to evaluate the
validity of a more abridged form of the criteria and compare
this result with that of the original Framingham Criteria.
Moreover, the incorporation of data on LVEF could enhance
the classification process.
We present a novel epidemiologic classification scheme
for heart failure—the Minnesota Heart Failure Criteria—
derived from widely available items used in the Framingham
Criteria, employing latent class analysis (LCA), a probabilistic approach to disease classification which allows for and
can potentially identify more precise categories of disease
conditions. Validation of the Minnesota Heart Failure Criteria
was performed under the assumption that persons with true
cases of heart failure will have a higher risk of mortality
than noncases in the 2 years following hospitalization.
MATERIALS AND METHODS
Study population
The study included a population-based sample of patients
whose records were abstracted as part of the Minnesota
Heart Survey for the hospitalization discharge years 1995
and 2000. Patients were selected from two independent
components of the Minnesota Heart Survey: the Acute Myocardial Infarction Study (22) and the Community Surveillance of Congestive Heart Failure Study (23).
Acute myocardial infarction study. Patients selected for
the Minnesota Heart Survey Acute Myocardial Infarction
(MHSAMI) Study for the 1995 hospital discharge year
were considered (22). Participants were residents of the
Minneapolis-St. Paul, Minnesota, metropolitan area during
1995 and were between the ages of 30 and 74 years. Patients
were hospitalized with an eligible International Classification of Diseases, Ninth Revision (ICD-9), discharge code for
acute myocardial infarction (i.e., ICD-9 codes 410 and 411)
in any of the participating regional acute-care hospitals.
Trained nurses abstracted data on a 40 percent sample of
men and an 80 percent sample of women. Relevant data on
signs and symptoms, medical history, cardiac enzyme levels, clinical complications, therapy, and autopsy results were
gathered.
Congestive heart failure study. All patients in the Minnesota Heart Survey Congestive Heart Failure (MHSCHF)
Study had an eligible ICD-9 discharge code for heart failure
(i.e., ICD-9 code 428 or another heart-failure-related disAm J Epidemiol 2006;164:184–193
185
charge code) during 1995 or 2000 (23). Consistent with
the MHSAMI, patients were residents of the MinneapolisSt. Paul metropolitan area at the time of the index hospitalization, but in MHSCHF they were between the ages of 35
and 84 years. Trained nurses abstracted data from 50 percent
of all hospital discharges and gathered data on patient demographic factors, medical history, physical examination,
electrocardiographic findings, chest radiography, and blood
laboratory results and, where available, basic information
on cardiac imaging (including current and historical LVEF)
and cardiac catheterization.
Combined data set. A combined data set was formed on
the basis of MHSCHF data for 1995 and 2000 and MHSAMI
data for 1995. The data were merged to broaden the spectrum
of disease (with regard to both disease severity and type of
heart failure) and to enhance the generalizability of the criteria. We hypothesized that the majority of noncases would
arise from MHSAMI while the majority of cases would arise
from MHSCHF.
Statistical methods
Development of the Minnesota Heart Failure Criteria. There
are numerous methods by which a classification scheme for
heart failure can be derived (13, 24). For this study, we chose
LCA, an established method of categorical data analysis
which has been used previously to validate diagnostic tests
in the absence of a perfect reference standard (25–29). LCA
is ideally suited for this purpose, because it makes no assumption about the independence of errors between the diagnostic test and the reference standard, which helps in
avoiding the known problem of inflated validity estimates
with the use of an imperfect reference standard (29, 30).
LCA uses a nonlinear statistical model to classify groups
of subjects who have shared characteristics into latent
classes (e.g., diagnostic subgroups) (30, 31). It achieves this
by maximizing a mathematical likelihood function based on
the patients’ distributional pattern of signs, symptoms, and
physiologic markers present at the time of evaluation. Given
that the distribution of criterion variables is biologically relevant to the disease classification and there are sufficiently
large numbers of cases and noncases to support these categorizations, LCA can identify the disease subcategories reliably. LCA is discussed further in other publications (28,
30–32) and in Web appendix 1, which is posted on the
Journal’s website (http://www.aje.oxfordjournals.org).
Variable selection. The variables defining the categories
of heart failure were chosen on the basis of clinical relevance using the short Framingham Criteria, plus LVEF
(table 1). They included dyspnea at rest or upon exertion,
pulmonary rales, cardiomegaly, interstitial or pulmonary
edema, the S3 heart sound, tachycardia (resting heart rate
>120 beats/minute), and LVEF (<40 percent, 40 percent,
or missing data).
LCA model selection. A challenge of LCA is to choose
the most appropriate number of latent classes using a parsimonious set of criterion variables. We set out to identify
a plausible classification structure for heart failure with
the fewest possible number of criterion variables. However,
given the limited availability of variables in MHSAMI, we
186 Kim et al.
TABLE 1. Sets of heart failure signs and symptoms compared with those of the Minnesota Heart Survey,
1995 and 2000
Existing criteria
Heart failure sign
or symptom
Framingham
Criteria
(46)
Duke
Criteria
(34)
Killip
Score
(47)
Dyspnea
Pulmonary rales
X
Cardiomegaly
X
Acute pulmonary edema
X
S3 heart sound
X
Tachycardia (120 beats/minute)
X
Minnesota Heart Survey data sets
Boston
Score
(19)
MHSCHF*
Study
(n ¼ 5,128)
Combined
MHSAMI*
data sety
Study
(n ¼ 2,251) (N ¼ 7,379)
X
Xz
Xz
Xz
X§
X
X
X
X{
X
X
X
X
X#
X
X**
X**
X
X
X
X
X
X
Xyy
X
X
X
X
X
X
Paroxysmal nocturnal dyspnea
X
X
X
Orthopnea
X
X
X
X
X
Left ventricular ejection fraction
percentage
Neck vein distension
X
X
Hepatojugular reflux
X
X
Weight loss
X
X
Ankle edema
X
X
Hepatomegaly
X
Pleural effusion
X
Hypotension (systolic blood
pressure <90 mmHg)
X
X
X
Night cough
X
Vital capacity (1/3 of maximum)
X
Increased venous pressure
(>16 cm of water)
X
Circulation time 25 seconds
X
Venous hypertension
X
Peripheral vasoconstriction
X
X
X
Elevated jugular venous pressure
X
Wheezing
X
Upper zone flow
X
* MHSCHF, Minnesota Heart Survey Community Surveillance of Heart Failure; MHSAMI, Minnesota Heart
Survey Acute Myocardial Infarction.
y Pooled data from the MHSCHF and MHSAMI studies.
z Dyspnea at rest or upon exertion.
§ Basilar or more than basilar rales.
{ Cardiomegaly on chest radiograph or cardiothoracic ratio 0.5.
# Alveolar or interstitial pulmonary edema.
** Interstitial or pulmonary edema.
yy Heart rate >110 beats/minute.
opted to include all seven variables in the initial LCA model.
We then compared the goodness of fit of this model with
varying numbers of diagnostic classes. Nested models were
evaluated through goodness-of-fit statistics: the Pearson chisquared (v2) test, the likelihood ratio chi-squared (L2) test,
Akaike’s information criterion, and the Bayesian information criterion. All LCA models were fitted using Latent
GOLD (version 2.0; Statistical Innovations, Belmont, Massachusetts) (33).
Mortality comparison: LCA model versus previous truncated
criteria. In the absence of a perfect gold standard, a diag-
nostic test can be validated by comparing its results against
a known consequence of the disease (34–37). Since a major
predictable outcome of heart failure is short-term mortality
(38–45), we hypothesized a priori that true heart failure
cases would have a greater risk of mortality than noncases
and that a positive deviation in mortality from noncases
would reflect the discriminatory power of the criterion.
Am J Epidemiol 2006;164:184–193
Prognostic Value of the Minnesota Heart Failure Criteria
To this end, we fitted age- and sex-adjusted linear models
to obtain absolute risk estimates of 2-year mortality for
each LCA category. All-cause mortality, rather than
heart-failure-specific mortality, was used to avoid possible
misclassification resulting from death due to multisystem
disease.
We compared these estimates against those obtained similarly with sets of commonly used heart failure criteria (i.e.,
the Framingham Criteria (46), the Killip Score (47), the
Duke Criteria (34), and the Boston Score (19)) which were
adapted for use with data available in the Minnesota Heart
Survey. The purpose here was to show the amount of additional information about the lethality of heart failure that is
contained in a set of existing criteria, as a base for a similar
examination of the LCA criteria. However, these results do
not reflect the full potential of each criterion, since only
those variables for which data are routinely collected in
hospitals and are readily available in the Minnesota Heart
Survey were used (referred to herein as ‘‘previous truncated
criteria’’).
In MHSCHF, it was possible to compute results for both a
‘‘short’’ set of Framingham Criteria (a six-item score based
on widely available items) and a ‘‘long’’ set of Framingham
Criteria (a 13-item score based on variables commonly
assessed when evaluating possible heart failure) (table 1).
It was only possible to compute results for the ‘‘short’’
Framingham Criteria in MHSAMI, since the collection of
data on heart-failure-related variables was not part of the
central aims of that study. Of the original 17 Framingham
Criteria items (table 1), the long Framingham Criteria excluded factors that are not currently commonly assessed in
patients being evaluated for possible heart failure: circulation
time (25 seconds), increased venous pressure (>16 cm of
water), vital capacity (1/3 of maximum), and night cough.
To allow comparability with previous truncated criteria,
LCA classes were collapsed into dichotomous disease categories (i.e., noncases and cases). We assessed possible interaction of the relation between diagnostic criteria and
mortality with LVEF. All statistical analysis (excluding
LCA) was performed using SAS for Windows (version
8.2; SAS Institute, Inc., Cary, North Carolina).
RESULTS
Development of the Minnesota Heart Failure Criteria
We began by examining the prevalence of each criterion
variable in all subjects, as well as in cases identified by the
short Framingham Criteria separately for MHSAMI and
MHSCHF (table 2). As expected, more subjects in MHSCHF
than in MHSAMI presented with each of the seven heart
failure criterion variables. Among cases identified by the
short Framingham Criteria, pulmonary rales were less
common in MHSAMI as compared with MHSCHF (40 percent vs. 95 percent), as were the S3 heart sound (9 percent vs.
22percent) and dyspnea(78 percentvs.93percent). MHSAMI
subjects who fulfilled the short Framingham Criteria for
heart failure more often had a high heart rate (30 percent vs.
16 percent) and pulmonary edema (77 percent vs. 55 percent).
Am J Epidemiol 2006;164:184–193
187
The prevalences of cardiomegaly and low ejection fraction
were similar between MHSCHF and MHSAMI.
Bearing these differences in mind, we found that a fiveclass LCA model had a good fit to the observed data (Web
appendix 1). The distributions of signs and symptoms by
diagnostic class are shown in table 3. The LCA model shows
that class 1 subjects had the lowest probability of presenting with the seven criterion variables: Few, if any, of these
subjects had the S3 heart sound, interstitial or pulmonary
edema, or tachycardia, though a relatively large proportion
(40 percent) of subjects had dyspnea. Therefore, on the basis
of the distributional pattern of conditional probabilities for
the seven criterion variables, we labeled class 1 subjects
‘‘noncases.’’
In contrast, subjects in class 5 had the largest probability
of presenting with a criterion variable: Nearly all of them
had dyspnea, pulmonary rales, and cardiomegaly, while they
also had the highest probability of presenting with depressed
LVEF. Few subjects had the S3 heart sound in any other
heart failure class, whereas over half of all class 5 subjects
had the S3 heart sound. Therefore, we labeled these subjects
as having ‘‘advanced heart failure.’’ Among the remaining
classes, the distinctions were less clear: A defining feature
of class 3 was that nearly all subjects had interstitial or
pulmonary edema, in contrast to class 2, where only 1 percent of subjects had the sign. Among class 3 subjects, the
presence of interstitial or pulmonary edema was consistent
with a high prevalence of dyspnea and pulmonary rales. We
hypothesized that subjects in classes 2, 3, and 4 had one
form or another of ‘‘heart failure.’’
The prevalence of each diagnostic class (i.e., unconditional probability) is also shown in table 3. The ‘‘noncases’’
formed a large group comprising 28 percent (2,057/7,379)
of the study population; the remaining 72 percent were classified as having some form of heart failure. The distribution
of conditional probabilities and the consequent formation of
the diagnostic classes were consistent with the population of
origin: The ‘‘noncases’’ had the largest proportion derived
from MHSAMI (84 percent), whereas nearly all subjects
with ‘‘advanced heart failure’’ originated from MHSCHF.
The majority of subjects with an LVEF under 40 percent
originated from the MHSCHF data set: Only 32 percent of
subjects in class 4 and 1 percent of subjects in class 5 were
from MHSAMI.
The posterior probabilities of diagnostic class membership for subjects with all signs and symptoms and a known
LVEF (27, or 128 patterns), as well as those with missing
data on LVEF (26, or 64 possible patterns), are shown in
Web appendix 2 (http://www.aje.oxfordjournals.org). One
hundred and fifty of the 192 possible response patterns were
observed in our study population; we observed at least 10
patients in 82 categories. The response patterns listed in
Web appendix 2, together with LCA class assignment based
on their respective posterior probabilities, form the Minnesota Heart Failure Criteria.
Application of the Minnesota Heart Failure Criteria in an
epidemiologic setting is illustrated through an example:
Suppose retrospectively abstracted medical records indicate
that a subject had pulmonary rales, the S3 heart sound, interstitial or pulmonary edema, cardiomegaly, and dyspnea
188 Kim et al.
TABLE 2. Distribution of heart failure signs and symptoms at the time of the index hospitalization,
by data set, Minnesota Heart Survey, 1995 and 2000
All patients
Heart failure sign/
symptom present?
MHSAMIy Study
(n ¼ 2,251)
Heart failure cases only*
MHSCHFy Study
(n ¼ 5,128)
No.
%
MHSAMI Study
(n ¼ 253)
MHSCHF Study
(n ¼ 3,983)
No.
No.
No.
%
%
%
No
1,209
54
461
9
55
22
264
7
Yes
1,042
46
4,667
91
198
78
3,719
93
2,115
94
762
15
151
60
190
5
136
6
4,366
85
102
40
3,793
95
1,577
70
1,423
28
42
17
540
14
674
30
3,705
72
211
83
3,443
86
1,993
89
2,911
57
57
23
1,806
45
258
11
2,217
43
196
77
2,177
55
2,219
99
4,242
83
231
91
3,116
78
32
1
886
17
22
9
867
22
2,122
94
4,473
87
177
70
3,356
84
129
6
655
13
76
30
627
16
No
906
40
1,297
25
64
25
1,023
26
Yes
323
14
1,735
34
104
41
1,435
36
1,022
45
2,096
41
85
34
1,525
38
Dyspnea
Pulmonary rales
No
Yes
Cardiomegaly on chest
radiograph
No
Yes
Interstitial or pulmonary
edema
No
Yes
S3 heart sound
No
Yes
Heart rate >120
beats/minute
No
Yes
Left ventricular ejection
fraction <40%
Missing data
* Based on the short Framingham Criteria (six variables for which data were available in both studies).
y MHSAMI, Minnesota Heart Survey Acute Myocardial Infarction; MHSCHF, Minnesota Heart Survey Community
Surveillance of Heart Failure.
but had missing data for LVEF. Web appendix 2 (observation number 187) would classify this person as belonging to
class 5 with 74 percent certainty. With regard to risk stratification, we can further state that the person has advanced
heart failure with a 53 percent risk of 2-year mortality. A
simple SAS program that provides an algorithm for classifying people into the corresponding LCA categories based
on the seven manifest variables for heart failure is downloadable from a University of Minnesota website (http://
www.epi.umn.edu/mhf).
Mortality comparison: previous truncated criteria
versus Minnesota Heart Failure Criteria
The results shown in tables 4 and 5, when interpreted together, establish the discriminatory power of the Minnesota
Heart Failure Criteria. Table 4 presents the 2-year mortality
difference between labeled cases and noncases according to
the previous truncated criteria (short Framingham Criteria:
26 percent noncases vs. 43 percent cases; Duke Criteria: 29
percent noncases vs. 40 percent cases; Killip Score: 31 percent noncases vs. 44 percent cases; Boston Score: 28 percent
noncases vs. 45 percent cases).
As compared with the previous truncated criteria, noncases
identified by the Minnesota Heart Failure Criteria had the
lowest mortality (18 percent), while cases had mortality equal
to that of the short Framingham Criteria (43 percent) (p for
difference < 0.001; table 5). Moreover, cases with advanced
heart failure had a significantly greater risk of mortality than
usual heart failure cases (47 percent vs. 42 percent; p ¼
0.003). The difference in mortality between heart failure
and advanced heart failure was greater in MHSAMI (33 percent heart failure vs. 53 percent advanced heart failure; p ¼
0.067) than in MHSCHF (45 percent heart failure vs. 50
percent advanced heart failure; p ¼ 0.014), though the
Am J Epidemiol 2006;164:184–193
Prognostic Value of the Minnesota Heart Failure Criteria
previous truncated criteria (table 4), persons with a normal
LVEF always had lower mortality than persons with a low
LVEF, while those with a low LVEF and missing data had
higher mortality. The lowest mortality was seen in noncases
with normal LVEF (table 4); it equaled that of Minnesota
Heart Failure Criteria noncases in the combined data set
(table 5). In both tables, persons with missing data on LVEF
always had an intermediate risk, and those noncases with
a low LVEF performed as poorly as the cases.
In addition, we found that the majority of persons in
MHSAMI had a favorable mortality estimate (9 percent),
while nearly all of the cases had an estimate comparable to
that of the noncases in MHSCHF (33 percent) (table 5).
Thus, the choice of using MHSAMI as the dominant source
of noncases was appropriate, given that most subjects appeared to be free of heart failure and the variables required
for performing LCA were easily accessible.
TABLE 3. Marginal percentage of persons with heart failure
symptoms, signs, and functional measures in each assigned
latent class analysis class, ordered by the probability of
occurrence in class 5, Minnesota Heart Survey, 1995 and 2000
Heart failure sign
or symptom
Dyspnea upon rest or
exertion
Pulmonary rales
Latent class analysis class
1
2
3
4
5
40
98
96
58
99
4
81
93
39
99
Cardiomegaly
18
62
77
65
91
Left ventricular ejection
fraction <40%
13
44
42
73
86
S3 heart sound
0
7
6
9
54
Interstitial or pulmonary
edema
2
1
95
30
52
Heart rate >120
beats/minute
2
12
19
7
14
2,057
2,088
1,773
735
726
28
28
24
10
10
189
Prevalence of latent class
No. of subjects
%
% of subjects from the
Minnesota Heart Survey
Acute Myocardial
Infarction Study
84
8
6
32
Cross-classification of the Framingham Criteria and the
Minnesota Heart Failure Criteria
We found that the classification achieved using the Minnesota Heart Failure Criteria was remarkably consistent
with the long Framingham Criteria, which approximate
the full Framingham Criteria (see upper half of table 6).
In MHSCHF, the two criteria agreed in 95 percent (4,848/
5,128) of participants, including 218 noncases and 4,630
cases. These two criteria disagreed only in 5 percent (280/
5,128) of all cases who had an ICD-9 heart-failure-related
discharge code. Among the 280 discordant pairs (116 identified by the long Framingham Criteria but missed by the
1
statistical significance was marginal because there were only
nine cases with advanced heart failure in MHSAMI.
We also found that the discriminatory power of each criterion was modified by LVEF. Among cases identified by
TABLE 4. Absolute risk of 2-year mortality* according to previous truncated criteria for heart failure, Minnesota Heart Survey,
1995 and 2000
Left ventricular ejection fraction
Classification
No. of
subjects
All patients
(N ¼ 7,379)
40%
(n ¼ 2,203)
<40%
(n ¼ 2,058)
Missing data
(n ¼ 3,118)
ARz
95% CIz
AR
95% CI
AR
95% CI
AR
95% CI
p for
interactiony
Short Framingham Criteria§
No heart failure
3,143
0.26
0.25, 0.28
0.17
0.14, 0.19
0.42
0.38, 0.46
0.27
0.25, 0.29
Heart failure
4,236
0.43
0.42, 0.44
0.36
0.33, 0.38
0.48
0.45, 0.50
0.44
0.42, 0.46
No heart failure
2,836
0.29
0.28, 0.31
0.20
0.18, 0.23
0.42
0.38, 0.46
0.31
0.29, 0.34
Heart failure
4,543
0.40
0.39, 0.41
0.31
0.29, 0.33
0.47
0.45, 0.50
0.39
0.37, 0.41
No heart failure
4,490
0.31
0.29, 0.32
0.21
0.19, 0.23
0.44
0.41, 0.47
0.31
0.29, 0.33
Heart failure
2,889
0.44
0.42, 0.46
0.36
0.33, 0.39
0.48
0.45, 0.51
0.46
0.43, 0.48
No heart failure
3,975
0.28
0.27, 0.30
0.18
0.16, .021
0.43
0.40, 0.47
0.29
0.26, 0.31
Heart failure
3,404
0.45
0.43, 0.46
0.38
0.35, 0.41
0.48
0.45, 0.51
0.47
0.44, 0.49
<0.0001
Duke Criteria
<0.0001
Killip Score
<0.0001
Boston Score
<0.0001
* The combined data set (pooled data from the Minnesota Heart Survey Community Surveillance of Heart Failure (MHSCHF) Study and the
Minnesota Heart Survey Acute Myocardial Infarction (MHSAMI) Study) was used in the calculation of 2-year absolute risks, adjusted for age and sex.
y Missing data on left ventricular ejection fraction were excluded from statistical testing.
z AR, absolute risk; CI, confidence interval.
§ The short Framingham Criteria were based on six variables for which data were available in both the MHSCHF Study and the MHSAMI Study.
Am J Epidemiol 2006;164:184–193
190 Kim et al.
TABLE 5. Absolute risk of 2-year mortality* according to the Minnesota Heart Failure Criteria, Minnesota Heart Survey, 1995 and 2000
Left ventricular ejection fraction
Classification
No. of
subjects
All patients
(N ¼ 7,379)
ARz
95% CIz
40%
(n ¼ 2,203)
AR
95% CI
<40%
(n ¼ 2,058)
AR
95% CI
Missing data
(n ¼ 3,118)
AR
p for
interactiony
95% CI
Combined data set §
No heart failure (LCAz class 1)
2,057
0.18
0.16, 0.20
0.14
0.11, 0.17
0.36
0.29, 0.44
0.18
0.15, 0.21
Heart failure (overall)
5,322
0.43
0.41, 0.44
0.35
0.33, 0.38
0.47
0.45, 0.49
0.43
0.42, 0.45
4,596
0.42
0.41, 0.43
0.35
0.33, 0.38
0.46
0.44, 0.49
0.43
0.41, 0.45
726
0.47
0.44, 0.50
0.37
0.30, 0.45
0.49
0.45, 0.54
0.51
0.45, 0.58
Heart failure
(LCA classes 2–4)
Advanced heart failure
(LCA class 5)
<0.0001
Minnesota Heart Survey Community Surveillance of Heart Failure Study
No heart failure (LCA class 1)
Heart failure (overall)
Heart failure
(LCA classes 2–4)
Advanced heart failure
(LCA class 5)
334
0.35
0.30, 0.40
0.29
0.21, 0.38
0.41
0.28, 0.53
0.38
0.30, 0.46
4,794
0.46
0.45, 0.47
0.40
0.38, 0.43
0.49
0.46, 0.51
0.47
0.44, 0.49
4,077
0.45
0.44, 0.47
0.40
0.37, 0.43
0.48
0.45, 0.51
0.46
0.44, 0.48
717
0.50
0.46, 0.54
0.41
0.32, 0.50
0.50
0.46, 0.55
0.54
0.47, 0.62
<0.0001
Minnesota Heart Survey Acute Myocardial Infarction Study
No heart failure (LCA class 1)
Heart failure (overall)
Heart failure
(LCA classes 2–4)
Advanced heart failure
(LCA class 5)
1,723
0.09
0.07, 0.11
0.06
0.04, 0.08
0.28
0.19, 0.37
0.10
0.07, 0.12
528
0.33
0.30, 0.36
0.24
0.19, 0.30
0.37
0.31, 0.44
0.33
0.29, 0.37
519
0.33
0.30, 0.35
0.23
0.18, 0.29
0.36
0.30, 0.43
0.33
0.29, 0.38
9
0.53
0.31, 0.75
—{
0.66
0.28, 1.04
—{
0.0596
* The combined data set (pooled data from the Minnesota Heart Survey Community Surveillance of Heart Failure (MHSCHF) Study and the
Minnesota Heart Survey Acute Myocardial Infarction (MHSAMI) Study) was used in the calculation of 2-year absolute risks, adjusted for age and sex.
y Noncases vs. cases. The distinction between heart failure and advanced heart failure was ignored. Missing data on left ventricular ejection
fraction were excluded from statistical testing.
z AR, absolute risk; CI, confidence interval; LCA, latent class analysis.
§ Pooled data from the MHSCHF Study and the MHSAMI Study.
{ Insufficient data for calculation of parameter estimates.
short Framingham Criteria; 164 identified only by the Minnesota Heart Failure Criteria), the proportions differently
classified by the two sets of criteria and their mortality
rates were similar; mortality estimates for the 280 subjects
were higher than in the 218 mutually classified noncases,
but not significantly so. The limitations of the short Framingham Criteria were apparent. No one was classified as
a noncase by the Minnesota Heart Failure Criteria but as
a case by the short Framingham Criteria. In contrast, 647
cases were missed by the Short Framingham Criteria; these
cases had mortality (44 percent) equivalent to that of the
3,983 persons classified as cases by both the Minnesota
Heart Failure Criteria and the long Framingham Criteria.
The lower half of table 6 compares the Minnesota Heart
Failure Criteria and the short Framingham Criteria in
MHSAMI, where results for the long Framingham Criteria
were not available. The two criteria agreed in 88 percent of
participants (1,976/2,251), including 1,723 noncases and 253
cases. The cases had a 2-year risk of mortality that was equivalent to that of the MHSCHF cases identified by both the
Minnesota Heart Failure Criteria and the long Framingham
Criteria, while the noncases had substantially lower mortal-
ity than the noncases in MHSCHF. A total of 275 cases misclassified as noncases by the short Framingham Criteria had
significantly greater mortality (32 percent) than noncases,
similar to the mortality risk in MHSCHF of noncases and
cases for whom the Minnesota Heart Failure Criteria
and the long Framingham Criteria disagreed.
DISCUSSION
In this study, we set out to derive a new classification scheme
for heart failure that was suitable for epidemiology—
specifically, one that could be useful for heart failure surveillance and mortality follow-up. We were cognizant in this
endeavor that heart failure surveillance and case-finding,
beyond potential cases already identified through physician
diagnosis, would be unlikely to include the depth of relevant
and decisive elements that the full Framingham Criteria call
for. Thus, a more pragmatic classification scheme for heart
failure is needed for epidemiologic research (especially if
retrospective), and the LCA approach was intended to fill
this gap.
Am J Epidemiol 2006;164:184–193
Prognostic Value of the Minnesota Heart Failure Criteria
TABLE 6. Absolute risk of 2-year mortalityy by heart failure
classification in two different Minnesota Heart Survey data sets,
1995 and 2000
Heart failure present?
Latent
class
analysisz
Long
Framingham
Criteria§
Short Framingham
Criteria{ (as a
subset of the long
Framingham Criteria)
No. of
subjects
Mortality
risk
Minnesota Heart Survey Community Surveillance of
Heart Failure Study (n ¼ 5,128 )
No
No
No
218
0.30
No
Yes
No
116
0.35
No
Yes
Yes
Yes
No
No
164
Yes
Yes
No
647
0.44*
Yes
Yes
Yes
3,983
0.44*
0
0.35
Minnesota Heart Survey Acute Myocardial Infarction
Study (n ¼ 2,251)#
No
No
No
Yes
1,723
0.15
Yes
No
275
0.32*
Yes
Yes
253
0.45*
0
* p < 0.0001 (compared with persons unanimously classified as
noncases).
y The 2-year mortality estimates shown were derived from ageand sex-adjusted linear models.
z The five-class latent class analysis model was dichotomized for
comparability (class 1 ¼ noncase; classes 2–5 ¼ heart failure case).
§ The long Framingham Criteria were based on 13 variables for
which data were available in the Minnesota Heart Survey Community
Surveillance of Heart Failure Study.
{ The short Framingham Criteria were based on six variables for
which data were available in both the Minnesota Heart Survey
Community Surveillance of Heart Failure Study and the Minnesota
Heart Survey Acute Myocardial Infarction Study.
# Results for the long Framingham Criteria could not be calculated
because of missing data on some variables.
The Minnesota Heart Failure Criteria were derived using
LCA with six widely available items from the Framingham
Criteria, plus LVEF, an important marker of underlying cardiac physiology that was not included in the Framingham
Criteria. The best-fitting model was a five-class model, which
was collapsed into three categories of risk strata (i.e., noncase, heart failure, and advanced heart failure). Using high
2-year mortality as a proxy for true cases of heart failure—
which was reasonable given the high case fatality for heart
failure—we showed that the LCA model provides better discrimination than the short Framingham Criteria and other
previous truncated criteria considered in both patients hospitalized for congestive heart failure and patients hospitalized
for acute myocardial infarction.
Thus, the Minnesota Heart Failure Criteria may be a suitable substitute for the long Framingham Criteria in situations that may hide heart failure, such as cases of acute
myocardial infarction and chronic obstructive pulmonary
disease. The fact that the long Framingham Criteria are unAm J Epidemiol 2006;164:184–193
191
likely to be implementable in such hospital records reduces
their value for general heart failure surveillance. The comparability of the Minnesota Heart Failure Criteria with the
long Framingham Criteria is notable, given the contrasting
mechanisms by which each set of criteria was derived: The
Minnesota Heart Failure Criteria are based on a statistical
model and maximum likelihood estimation, whereas the
Framingham Criteria are based on clinical plausibility.
We caution that we were only able to evaluate the value of
a scheme using limited variables originally selected for classifying myocardial infarction, because of the restricted set
of signs and symptoms available in MHSAMI. Therefore,
there was no ‘‘control’’ group in the MHSCHF data set;
rather, we used a comparison population in whom we
thought heart failure would be relatively rare. Our data cannot confirm or firmly establish how well the previous truncated criteria would have performed had we had complete
information on all of the required variables. Indeed, a study
would have to be specially designed to obtain such a control
set; this would probably be difficult, since very few patients
without suspected heart failure are tested for the full set of
17 Framingham Criteria, and several of the original signs
and symptoms are now rarely clinically assessed, even in
patients suspected of having heart failure. Seen in this
light, MHSAMI is a rich source for comparison with the
MHSCHF data set, given that several critical heart failure
signs and symptoms are included in MHSAMI and that there
are some bona fide heart failure cases in the MHSAMI data
set. We believe that most non-heart-failure data sets (e.g.,
data on chronic obstructive pulmonary disease patients) will
not include many of the long Framingham Criteria variables,
though they are more likely to include the six Minnesota
Heart Failure Criteria variables and possibly data on LVEF.
In this sense the LCA method, including LVEF, is capable of
doing the desired epidemiologic job—namely surveillance
of heart failure occurrence in a broad set of hospitalized patients, including many who were not specifically suspected
of having heart failure.
The generalizability of the Minnesota Heart Failure Criteria must be verified in other populations, both to check their
validity in other settings and to extend them to an outpatient
setting. Repeatability is of particular concern for response
patterns with a small number of subjects. Another limitation
of our criteria is that they did not identify a separate category
corresponding to clinical isolated diastolic dysfunction. Although LVEF is related to diastolic dysfunction, the other
signs and symptoms are apparently not sufficiently indicative of isolated diastolic dysfunction for the final LCA model
to identify a subgroup with isolated diastolic dysfunction.
Using LCA, we identified a plausible classification structure for heart failure with three categories of risk strata
based on six Framingham Criteria variables for which data
are readily available from medical records (i.e., dyspnea,
pulmonary rales, cardiomegaly, S3 heart sound, interstitial
or pulmonary edema, and elevated heart rate), plus LVEF.
We found that the LCA classification provides better discrimination between cases and noncases against several previous sets of truncated criteria and equal discrimination
against a truncated form of the Framingham Criteria that
closely approximates the full version but requires many
192 Kim et al.
fewer variables. The posterior probabilities from the LCA
model (Web appendix 2 (http://www.aje.oxfordjournals.org))
form the basis for the Minnesota Heart Failure Criteria,
which have the potential to facilitate the epidemiologic classification of heart failure.
ACKNOWLEDGMENTS
Funding was provided in part by grants RO1-HL-23727
and RO1-HL60959 from the National Heart, Lung, and
Blood Institute.
The authors thank the participating hospitals in MinneapolisSt. Paul for their long-lasting support of epidemiologic research. They are indebted to the study programmers and
dedicated nurse abstractors of the Minnesota Heart Survey.
Conflict of interest: none declared.
REFERENCES
1. Association AH. Heart and stroke statistical update. Dallas,
TX: American Heart Association, 2004.
2. Purcell IF, Poole-Wilson PA. Heart failure: why and how to
define it? Eur J Heart Fail 1999;1:7–10.
3. Adams KF Jr, Zannad F. Clinical definition and epidemiology
of advanced heart failure. Am Heart J 1998;135:S204–15.
4. Banerjee P, Banerjee T, Khand A, et al. Diastolic heart failure: neglected or misdiagnosed? J Am Coll Cardiol 2002;
39:138–41.
5. Cheng TO. Definition of heart failure. (Letter). Eur J Heart
Fail 1999;1:433–4.
6. Denolin H, Kuhn H, Krayenbuehl HP, et al. The definition of
heart failure. Eur Heart J 1983;4:445–8.
7. Harley A. The management of heart failure: a matter of definition? Cardiovasc Drugs Ther 1993;7:661–9.
8. Tavazzi L. Towards a more precise definition of heart failure
aetiology. Eur Heart J 2001;22:192–5.
9. Wagner S, Cohn K. Heart failure. A proposed definition and
classification. Arch Intern Med 1977;137:675–8.
10. Wilhelmsen L, Eriksson H, Svardsudd K, et al. Improving
the detection and diagnosis of congestive heart failure. Eur
Heart J 1989;10:13–18.
11. Yip G, Wang M, Zhang Y, et al. Left ventricular long axis
function in diastolic heart failure is reduced in both diastole and systole: time for a redefinition? Heart 2002;87:
121–5.
12. Young JB, Pratt CM. Hemodynamic and hormonal alterations
in patients with heart failure: toward a contemporary definition
of heart failure. Semin Nephrol 1994;14:427–40.
13. Mosterd A, Deckers JW, Hoes AW, et al. Classification of
heart failure in population based research: an assessment
of six heart failure scores. Eur J Epidemiol 1997;13:
491–502.
14. Eriksson H, Caidahl K, Larsson B, et al. Cardiac and pulmonary causes of dyspnoea—validation of a scoring test for
clinical-epidemiological use: The Study of Men Born in 1913.
Eur Heart J 1987;8:1007–14.
15. McKee PA, Castelli WP, McNamara PM, et al. The natural
history of congestive heart failure: The Framingham Study.
N Engl J Med 1971;285:1441–6.
16. Walma EP, Hoes AW, Prins A, et al. Withdrawing long-term
diuretic therapy in the elderly: a study in general practice in the
Netherlands. Fam Med 1993;25:661–4.
17. Schocken DD, Arrieta MI, Leaverton PE, et al. Prevalence
and mortality rate of congestive heart failure in the United
States. J Am Coll Cardiol 1992;20:301–6.
18. Gheorghiade M, Beller GA. Effects of discontinuing maintenance digoxin therapy in patients with ischemic heart disease
and congestive heart failure in sinus rhythm. Am J Cardiol
1983;51:1243–50.
19. Carlson KJ, Lee DC, Goroll AH, et al. An analysis of physicians’ reasons for prescribing long-term digitalis therapy in
outpatients. J Chronic Dis 1985;38:733–9.
20. Jaski BE. Basics of heart failure. Boston, MA: Kluwer Academic Publishers, 2000.
21. McKee PA, Castelli WP, McNamara PM, et al. The natural
history of congestive heart failure: The Framingham Study.
N Engl J Med 1971;285:1441–6.
22. McGovern PG, Jacobs DR Jr, Shahar E, et al. Trends in acute
coronary heart disease mortality, morbidity, and medical care
from 1985 through 1997: The Minnesota Heart Survey.
Circulation 2001;104:19–24.
23. Shahar E, Lee S, Kim J, et al. Hospitalized heart failure:
rates and long-term mortality. J Card Fail 2004;10:374–9.
24. Wagner S, Cohn K. Heart failure. A proposed definition and
classification. Arch Intern Med 1977;137:675–8.
25. Guggenmoos-Holzmann I, van Houwelingen HC. The
(in)validity of sensitivity and specificity. Stat Med 2000;19:
1783–92.
26. Torrance-Rynard VL, Walter SD. Effects of dependent errors
in the assessment of diagnostic test performance. Stat Med
1997;16:2157–75.
27. Walter SD, Irwig LM. Estimation of test error rates, disease
prevalence and relative risk from misclassified data: a review.
J Clin Epidemiol 1988;41:923–37.
28. Rindskopf D, Rindskopf W. The value of latent class analysis in medical diagnosis. Stat Med 1986;5:21–7.
29. Uebersax JS, Grove WM. Latent class analysis of diagnostic
agreement. Stat Med 1990;9:559–72.
30. Hagenaars J, McCutcheon A. Applied latent class analysis.
Cambridge, UK: Cambridge University Press, 2002.
31. McCutcheon A. Latent class analysis. Newbury Park, CA:
Sage Publications, 1987.
32. Kaldor J, Clayton D. Latent class analysis in chronic disease
epidemiology. Stat Med 1985;4:327–35.
33. Vermunt JK, Magidson J. Latent GOLD 2.0 user’s guide.
Belmont, MA: Statistical Innovations, Inc, 2000.
34. Harlan WR, Oberman A, Grimm R, et al. Chronic congestive
heart failure in coronary artery disease: clinical criteria. Ann
Intern Med 1977;86:133–8.
35. Feinstein AR. Clinimetrics. New Haven, CT: Yale University
Press, 1987.
36. Ascherio A, Rimm EB, Giovannucci EL, et al. A prospective study of nutritional factors and hypertension among US
men. Circulation 1992;86:1475–84.
37. Aldoori WH, Giovannucci EL, Rimm EB, et al. A prospective study of diet and the risk of symptomatic diverticular
disease in men. Am J Clin Nutr 1994;60:757–64.
38. Mosterd A, Cost B, Hoes AW, et al. The prognosis of heart
failure in the general population: The Rotterdam Study. Eur
Heart J 2001;22:1318–27.
39. Murdoch DR, Love MP, Robb SD, et al. Importance of
heart failure as a cause of death. Changing contribution to
overall mortality and coronary heart disease mortality in
Scotland 1979–1992. Eur Heart J 1998;19:1829–35.
Am J Epidemiol 2006;164:184–193
Prognostic Value of the Minnesota Heart Failure Criteria
40. Changes in mortality from heart failure—United States,
1980–1995. MMWR Morb Mortal Wkly Rep 1998;47:
633–7.
41. Pernenkil R, Vinson JM, Shah AS, et al. Course and prognosis in patients > or ¼ 70 years of age with congestive heart
failure and normal versus abnormal left ventricular ejection
fraction. Am J Cardiol 1997;79:216–19.
42. Hillege HL, Girbes AR, de Kam PJ, et al. Renal function,
neurohormonal activation, and survival in patients with
chronic heart failure. Circulation 2000;102:203–10.
43. Zannad F, Braincon S, Juilliere Y, et al. Incidence, clinical
and etiologic features, and outcomes of advanced chronic heart
failure: The EPICAL Study. Epidemiologie de l’Insuffisance
Am J Epidemiol 2006;164:184–193
44.
45.
46.
47.
193
Cardiaque Avancee en Lorraine. J Am Coll Cardiol 1999;33:
734–42.
Rector TS, Cohn JN. Prognosis in congestive heart failure.
Annu Rev Med 1994;45:341–50.
Cowie MR, Wood DA, Coats AJ, et al. Survival of patients
with a new diagnosis of heart failure: a population based study.
Heart 2000;83:505–10.
Ho KK, Pinsky JL, Kannel WB, et al. The epidemiology of
heart failure: The Framingham Study. J Am Coll Cardiol
1993;22:6A–13A.
Killip T 3rd, Kimball JT. Treatment of myocardial infarction in a coronary care unit. A two year experience with 250
patients. Am J Cardiol 1967;20:457–64.