Transcript The ACC/ESC Recommendation for 99th Percentile

The ACC/ESC Recommendation for
99th Percentile of the Reference
Normal Overestimates the Risk of an
Acute Myocardial Infarct
Jerard Kneifati-Hayek, Salman Haq, M.D,
Madeleine Schlefer, Larry Bernstein, M.D.
New York Methodist Hospital Cornell-Weill
Background
A joint committee of the American College of
Cardiology and European Society of Cardiology
(ACC/ESC) has established the criteria for acute
recent or evolving AMI predicated on
a typical increase in troponin
in the clinical setting of myocardial ischemia,
which includes the 99th percentile of a healthy
normal population.
Problems in Recommendation
1. A “Reference normal population doesn’t
present to the emergency room”
2. The cutoff used to make a decision has an
effect on overuse of hospital resources,
especially costly and limited “telemetry”
3. The cutoff “selection” is not scientifically
validated
4. The cutoff selection is informed by a
potential projected risk extending to
perhaps four months after the presenting
event
Recommendation Challenged
1. Lin JC, Apple FS, Murakami MM, Luepker
RV. Rates of positive cardiac troponin I and
creatine kinase MB mass among patients
hospitalized for suspected acute coronary
syndromes. Clin Chem 2004.
2. Zarich SW, Bradley K, Mayall ID, Bernstein,
LH. Minor Elevations in Troponin T Values
Enhance Risk Assessment in Emergency
Department Patients with Suspected
Myocardial Ischemia: Analysis of Novel
Troponin T Cut-off Values. Clin Chim Acta
2004.
Hypothesis
An AMI is better identified by a TnI value
exceeding the 99th percentile of patients
seen with acute coronary syndrome (ACS)
who are subsequently excluded
than using the ACC/ESC guideline.
Patient selection
1480 successive patients who presented to the
emergency room at New York Methodist Hospital in
two month periods in 2003 and 2004 and were
required to have troponin I (TnI) were identified.
No Randomization
Observational study
Exclusion Criteria: Pacemaker, ST elevation,
patients under 30 years old
Inclusion Criteria: ST depression, T wave inversion,
LBBB, anginal equivalent chest pain, short breath
Patient Demographics
Women 56%
Men 44%
46% white, 32% Negro, 16% Hispanic, 6% Asian.
ECG Findings: 24% were normal,
73.3% were nonspecific,
and 55.2% were other than NSSTT changes.
ECG findings of AMI were present 4.6
percent of the time.
Laboratory Methods
The CKMB and TnI were measured on the
Centaur (Bayer, Tarrytown, NY) at the time of
presentation and at least 4 hours later.
Data Collection
The information collected was –
serum cardiac marker concentration,
discharge diagnosis,
ECG finding,
chest pain characteristics.
all relevant cardiac diagnoses, including, CHF, atrial or
ventricular dysrhythmias, and other relevant diagnoses,
such as, hypothyroid disease, hypertension and type 2
diabetes were included.
Computerized ECGs were reviewed by two cardiologists
Reference Normal Population
A control population of 140 blood
donors was also measured for TnI.
The control population TnI values
were all less than 0.1 g/L.
Statistical evaluation
The data was organized for paired
comparisons of troponin I and CKMB.
SPSS 11.5 (Chicago, IL) software was
used for ROC curves and basic statistics.
SYSTAT 10.0 (Chicago, IL) was used for
logistic regression,
Ordinal Regression, GOLDminerTM 3.0
(Statistical Innovation, Inc., Belmont,
MA)
and Latent Class Cluster Analysis, Latent
Gold 2.0 (ibid).
Two Step Procedure
1. Trial set to determine best decision value
Trial set was itself partitioned by an initial
set for which half the charts were
unavailable because of transition to a
new information system
2. Second trial set had 619 patients
3. Validation set
ROC curve on 2nd Training Set
The area under the ROC curve is 99%. The
best cutoff selected is 0.60  0.77 µg/L at a
sensitivity of 98.9% and a false positive rate at
4%.
This is actually lower than the 99th percentile of
the non-MI population and significantly higher
than the 99th percentile of the healthy donor
population, approximately the 97.5th percentile of
the ruled-out MI population.
Receiver Operating Characteristic (ROC) Plot
1.0
Hit Rate
0.8
0.6
0.4
0.2
0.0
0.0
0.2
0.4
0.6
0.8
False Alarm Rate
1.0
TnI
.21
.22
.23
.24
.25
.26
.27
.29
.30
.31
.32
.33
.35
.36
.37
.38
.41
Coordinates of the ROC curve
Sens
TnI
Sens
1 - Spec
1.000
.101
.46
.989
1.000
.098
.49
.989
1.000
.094
.52
.989
1.000
.089
.56
.989
1.000
.085
.57
.989
1.000
.082
.59
.989
1.000
.081
.989
.64
.989
.079
.69
.989
.989
.075
.72
.989
.989
.073
.73
.989
.989
.072
.75
.989
.989
.069
.77
.989
.989
.068
.80
.977
.989
.066
.82
.966
.989
.062
.83
.966
.989
.060
.84
.966
.989
.058
1 - Spec
.056
.055
.052
.050
.049
.047
.046
.045
.043
.042
.040
.039
.039
.037
.036
.035
Condition
Troponin I
Comment
Not AMI (619)
0.00
Median
MI (70)
> 0.61 g/L
(49)
Neg
Pos
CHF
6
2
2 CKMBs > 6.25
tachycardia
6
6
3 TnI above 3.3
Cardiomyopathy
4
1
1 TnI above 3.3
Acute renal failure
0
1
Resp Failure
1
2
Respiratory failure
Sepsis
1
1
discordant
Hypothyroid
1
0
Hypothyroid
other
13
17
Latent Class Analysis
A latent class model partitions the data into separate,
nonoverlapping sets (the fit measured by chi square).
We assume that within each latent class, each variable
is statistically independent of every other variable.
This is referred to as “conditional independence”.
However, within a latent class that corresponds to a
distinct medical syndrome, while the presence/ absence
of one symptom is viewed as unrelated to
presence/absence of all others, this may NOT be the
case.
other
MI
NMI
Cluster1
Overall Probability
0.9162
Not MI
TNISCA
0
1.0000
1
0.9997
2
0.0001
3
0.0000
Cluster2
0.0838
MI
0.0000
0.0003
0.9999
1.0000
g/L
< 0.3
0.3 – 0.6
0.61 – 1.2
> 1.2
Eight percent of patients are in cluster 2, consistent with
MI. A cutoff value of less than 0.61 is consistent with
classification of not MI.
In the previous two models we fitted the data
to a binary classification using the diagnosis
for training (ROC), or using the information in
the data combinations (LCA) to cluster into
MI or not MI at a cutoff of 0.61.
We next explored a method that allows us
to look at the probability of MI using the
diagnosis as a training value, and scaled
values of troponin for the Y variable in an
ORDINAL RESPONSE.
Ordinal Regression
The situation we describe is an
ordered response with graded
categories.
Magidson has developed a unified maximum
likelihood methodology for simultaneously
assessing the statistical significance of
treatment effects and the model fit when the
response variable contains ordered categories.
He explores the fit for different logit model
extensions (log-odds) to data derived under
the assumption of bivariate normality and finds
that the “parallel log-odds” model based on
adjacent odds often gives a better fit than the
“proportional odds” model based on
cumulative odds.
Frequency cross-table, probabilities and
odds-ratios for scaled TnI versus
expected diagnosis
TnI
scaled
Range
Not MI
MI
N
Pct in
MI
Prob by
TnI
Odds
Ratio
0
< 0.45
655
2
657
2
0
1
1
0.46-0.6
7
0
7
0
0.03
13
2
0.610.75
4
0
4
0.
0.26
175
3
0.76-0.9
13
59
72
57.3
0.82
2307
4
> 0.9
0
42
42
40.8
0.98
30482
679
103
782
100
There are 13 TnI values between 0.76-0.9 that are
not MI, that would be falsely classified as MI at a
cutoff below 0.76,
17 at a cutoff below 0.61,
and 24 at a cutoff below 0.46.
The false positive rate at a cutoff of 0.61 is 2.5%.
We created a second data set including 322
patients from a population different than the
trial population to validate the troponin I
cutoff selected.
See the frequencies of TnI at a cutoff of 0.15
µg/L in rows versus the diagnosis of MI,
and the frequencies of TnI at 0.65 µg/L versus
the diagnosis of MI.
Cutoff
Test Result
Not MI
MI
0.15 g/L
Negative
239 100%
86.9%
0
0%
Positive
36
44.5%
13.1%
Total
275
Negative
0.65 g/L
Row total
239
100%
45
55.5%
100%
81
100%
45
100%
320
260 99.2%
94.5%
2
0.76%
4.4%
262
100%
Positive
15
25.8%
5.5%
43
74.2%
95.6%
58
100%
Total
275
45
320
100%
100%
0%
100%
Myocardial Risks








Previous MI
High Blood Pressure
High Cholesterol
Diabetes
Age>65
Stroke, peripheral vascular disease
Hypercholesterolemia
Hypertension
LDL Cholesterol
Level for Drug
Consideration
Goal of
Therapy
Without coronary heart disease
and with fewer than two risk
factors
190 mg/dL or
higher*
less than
160 mg/dL
Without coronary heart disease
and with two or more risk factors
160 mg/dL or
higher
less than
130 mg/dL
With coronary heart disease
130 mg/dL or
higher**
100 mg/dL
or less
Comparison of TnI Level with Number of
Active Medications in Groups of MI and Non
MI
5
4
TNI
3
2
Medscore
0= No Statins
1= 1 Statin Drug
1
0
0.0
DXMI
0.2
0.4
0.6
0.8
MEDSCORE2
1.0
1.2
0
1
0=Non MI
1=MI
The documentation of chest pain characteristics
in the medical record is inconsistent, sometimes
absent, and often unclear. This results in
reliance on TnI and emphasis on ECG evidence
for the diagnosis of MI, even though many
patients lack a definitive ECG pattern. This
places greater importance on the presence of
risk factors, such as, diabetes, hypertension,
hypercholesterolemia, previous MI, and anginal
equivalent findings in addition to TnI in
assessing the probability of MI. A similar model
can be constructed using TnI and risk factors.