Transcript Clinical Prediction Scores

Better Medical
Diagnostic Decisions
thru Science
Optimal Clinical Application of
Exercise ECG Testing
V. Froelicher, MD
Professor of Medicine
Stanford University
VA Palo Alto HCS
What are the Questions being
asked regarding Coronary
Disease and Exercise Testing
Does this patient have or not have
Coronary Disease?
Is this patient going to experience a
Cardiac Event?
Better decisions are made possible by
applying the following two methods to
clinical and exercise test data:
Scientific Decision
Methods
Statistical Prediction
Rules = Scores
Receiver Operator
Characteristic Curves
Statistical Prediction
Rules
Based on mathematical models
or equations that can be
simplified as scores
They increase accuracy by
enhancing the odds that any
decision will be correct (a
reliable second opinion)
Clinical Scores
1. Predicting Outcomes
Follow up required (time, complete)
Endpoint Limitations (Death, CABG)
No Natural History
2. Predicting Angiographic Findings
Instant Epidemiology
Limitations of Angiography
Sub-ischemic Lesions cause events
Making any of these Five Mistakes
Evaluating Diagnostic Tests can
invalidate Scores & Stats
Limiting the population Challenge
by choosing extremes
Failure to reduce Work up bias
Use of Heart rate targets
Inclusion of MI patients
Use of Surrogates
Making any of these Four Mistakes
Evaluating Prognostic Tests can
invalidate Scores & Stats
Limited Challenge and work
up bias
Incomplete Follow up
Failure to Censor
Using Misleading Endpoints
Clinical Scores
1. Survival Analysis
Based on Follow-up and Censoring
Cox Hazard Function; time to event rather than
proportion differences
Weighted Coefficients used to construct
Equations for Scores and Nomogram
2. Probability of Coronary Disease
Based on Angiography
Multiple Logistic Regression
Coded Variables x Coefficients added then
solved in Natural Log Equation to fit a Sigmoid
Curve
Paradigm for Matching the Clinical Management
Strategy to the Estimated Probability of CAD
Probability for clinically significant CAD
Low probability
Patient reassured symptoms most likely not due to
CAD
Intermediate probability
Require other tests, such a stress echo, nuclear, or
angiography to clarify diagnosis; anti-anginal
medications tried.
High probability
Anti-anginal treatment indicated; intervention if
clinically appropriate; angiography usually required
Meta Analysis of Prognosis in Stable CAD
Exercise Test and Cath (N=9)
Poor Exercise Capacity
CHF
ST Depression
Resting
Exercise
Exercise SBP
6/9
3/9
2/9
3/9
3/9
Meta Analysis of Prognosis in Stable CAD
Exercise Test and Cath (N=9)
Exercise induced ST depression not
consistently a predictor
Exercise Capacity usually a predictor
Two Studies have used Cox Hazards
Function to chose variables significantly
and independently associated with time
to CV event (hard events, not CABG)
Prognostic Scores in Stable CAD
DUKE SCORE
METs - 5 X [mm E-I ST Depression] 4 X [Treadmill Angina Index]
******see Nomogram*******
VA SCORE
5 X [CHF/Dig] + [mm E-I ST Depression]
+ change in SBP score - METs
E-I = Exercise Induced
Duke Treadmill Score (uneven lines)
Prediction of Prognosis
Censoring when lost to follow up or
when an intervention performed that
alters outcome
All-cause mortality, infarct-free survival or
cardiovascular death
Predictors of MI and death differ
Prediction of Prognosis
What to do with patients who have
Interventions that could alter
Outcomes?
Exclude from analysis
Ignore
Use as endpoints (after a time lag)
Censor (end time of follow up)
Partial censoring?
The HR Recovery Studies Hi-light
problems with Prediction of
Prognosis
Failure to censor results in prediction of
outcome after application of standard
therapies
Does not allow for prediction of who
should receive therapies or interventions
Failure to censor and use infarct-free
survival or cardiovascular death negates
development of strategies or scores for
treatment of CAD
Diagnostic Scores:
ACC/AHA guidelines state that multivariable equations should be used to
enhance the diagnostic
characteristics of the exercise
treadmill test.
The Equations are often not applied in
practice because of their complexity
Multi-Variable Logistic
Regression
Probability (0 to 1) =
(a
+
bx
+
cy
.
.
.
)
1 / (1 + e
)
where a = intercept, b and c are coefficients, x
and y are variable values.
For instance:
x = age, y= chest pain type, z = diabetes ….
Meta Analysis of 24 Studies
Predicting Angiographic CAD
Most consistent clinical variables
chosen were:
Gender, Chest pain type, Age and
Hypercholesterolemia
Most consistent exercise test
variables chosen were:
ST depression and slope, Maximal Heart
rate and exercise capacity (METs)
Problems with Scores
Censoring
Follow-up Confounded by Interventions
The major Mistakes for evaluating Tests
Differences between Studies as to Variables
and Their Coding
Skepticism that Scores Can Be Better
Than Physician Estimates
Require Nomograms or Computers to
Calculate the Prediction
Are they Portable?
Simplified Score:
Derivation of a simplified treadmill
score based on multi-variable
statistical techniques
Validation of this treadmill score in
another population and comparison to
the ST response alone and the Duke
treadmill score
Methods:
Clinical and exercise test variables were
coded: Continuous and dichotomous variables
all set from 0 to 5 (five cells for continuous,
yes = 5) for proportionality
Graded as 0 for good and 5 for bad
The coded variables were entered into a
standard logistic regression model to
discriminate between those with and without
angiographically significant CAD (equal or
greater than 50%)
Methods:
The derived equation was then Simplified
by dividing all Coefficients by the least
coefficient so that they all became
multiples of one
The Simplified score then was created by
adding the variables after scoring and
multiplication
It was compared to the logistic regression
equation results by ROC analysis and found
to be equivalent
Variable
Maximal Heart Rate
Circle response
Less than 100 bpm = 30
100 to 129 bpm = 24
130 to 159 bpm =18
160 to 189 bpm =12
190 to 220 bpm =6
Exercise ST Depression
1-2mm =15
> 2mm =25
Age
>55 yrs =20
Sum
Males
Choose
only one
per
group
40 to 55 yrs = 12
Angina History
Definite/Typical = 5
Probable/atypical =3
Non-cardiac pain =1
Hypercholesterolemia?
Yes=5
Diabetes?
Yes=5
Exercise test
induced Angina
Occurred =3
Reason for stopping =5
Total Score:
<40=low prob
40-60=
intermediate
probability
>60=high
probability
Scientific Decision
Methods
Statistical Prediction
Rules = Scores
Receiver Operator
Characteristic Curves
Receiver Operator
Characteristic Curves
Improve the utility of decision-making
approaches ensuring that the number
of true cases diagnosed does not
come at the cost of too many false
positives (“false alarms”)
Allows comparison of the diagnostic
ability of competing diagnostic
techniques and scores
Overlapping, not separate; the further apart, better the test
Other Mxmnts: EBCT Calcium, Echo WMA, Nuclear, ST depr
cutpoint
Specificity
Inverse relationship
Sensitivity
Chest Pain
AUC
Screening
AUC
But Can Physicians do as well
as the Scores?
954 patients - clinical/TMT reports
Sent to 44 expert cardiologists, 40
cardiologists and 30 internists
Scores did better than all three but
were most similar to the experts
Two ways to compare the
discriminatory/diagnostic
characteristics of a test/
measurement
1. Range of Characteristic curves –
unaffected by prevalence, can be used
to choose cut points, require continuous
variables
2. Predictive Accuracy – TP+TN/pop,
requires dichotomy, same prevalence to
compare
Methods of Test comparison:
ROC Plots


1 perfect discrimination, .50 none
Not dependent upon prevalence of
disease
Predictive Accuracy


Percent of total true calls (TP +TN)
Dependent upon prevalence of disease
Comparison of Tests
Grouping
Standard ET
ET Scores
Score Strategy
ThalliumScint
SPECT
Adenosine SPECT
Exercise ECHO
Dobutamine ECHO
Dobutamine Scint
Electron Beam
Tomography (EBCT)
# of
Total #
Studies Patients
147
24,047
24
11,788
2
>1000
59
6,038
16+14 5,272
10+4
2,137
58
5,000
5
<1000
20
1014
16
3,683
Sens Spec Predictive
Accuracy
68% 77%
73%
80%
85% 92%
88%
85% 85%
85%
88% 72%
80%
89% 80%
85%
84% 75%
80%
88% 84%
86%
88% 74%
81%
60% 70%
65%
Conclusions:
Scores can predict the presence of CAD
better than ST analysis alone.
Scores can predict Prognosis
Scores can provide a valuable second
opinion (as good as experts).
Scores reduce the effect of physician
bias
Scores provide a management Strategy
Conclusions:
Scores can optimize the clinical
application of the standard exercise
ECG Test.
The Duke Treadmill Test Score and
VA/WV Simple Score should be part
of every exercise test interpretation