Diagnostic Testing & Medical

Decision Making
C Diana Nicoll, MD, PhD, MPA, & Michael Pignone, MD, MPH
See www.cmdtlinks.com

The clinician's main task is to make reasoned deci- fecal occult blood test may incur significant cost, risk,
sions about patient care despite incomplete clinical in- and discomfort during follow-up colonoscopy.
formation and uncertainty about clinical outcomes. Furthermore, a false-positive test may lead to incorrect
Although data elicited from the history and physical diagnosis or further unnecessary testing. Classifying a
examination are often sufficient for making a diagno- healthy patient as diseased based on a falsely positive diag-
sis or for guiding therapy, more information may be nostic test can cause psychological distress and may lead
required. In these situations, clinicians often turn to to risks from unnecessary or inappropriate therapy. A di-
diagnostic tests for help. agnostic or screening test may identify cases of disease that
would not otherwise have been recognized and that
would not have affected the patient. For example, early-
stage low-grade prostate cancer detected by prostate-spe-
cific antigen screening in an 84-year-old man with known
BENEFITS; COSTS & RISKS severe congestive heart failure will probably not become
symptomatic or require treatment during his lifetime.
The costs of diagnostic testing must always be under-
When used appropriately, diagnostic tests can be of
stood and considered. Total costs may be high, or cost-ef-
great assistance to the clinician. Tests can be helpful for
fectiveness may be unfavorable. An individual test such as
screening, ie, to identify risk factors for disease and to
MRI of the head can cost more than $1400, and diagnos-
detect occult disease in asymptomatic persons.
tic tests as a whole account for approximately 20% of
Identification of risk factors may allow early interven-
health care expenditures in the United States. Even rela-
tion to prevent disease occurrence, and early detection of
tively inexpensive tests may have poor cost-effectiveness if
occult disease may reduce disease morbidity and
they produce very small health benefits.
mortality .through early treatment. Optimal screening
tests meet the criteria listed in Table 42-1. Deyo RA: Cascade effects of medical technology. Annu Rev Pub-
Tests can also be helpful for diagnosis, ie, to help lic Health 2002;23:23. [PMID: 11910053]
establish or exclude the presence of disease in sympto- Grimes DA et al: Uses and abuses of screening tests. Lancet
matic persons. Some tests assist in early diagnosis after 2002;359:881. [PMID: 118973041
onset of symptoms and signs; others assist in develop- Mushlin AI: Cost-effectiveness of diagnostic rests. Lancet
ing a differential diagnosis; others help determine the 2001;358:1353. [PMID: 11684235]
stage or activity of disease. Willinsky RA et al: Neurologic complications of cerebral angiogra-
Finally, rests can be helpful in patient management phy: prospective analysis of 2,899 procedures and review of
Tests can help (1) evaluate the severity of disease, (2) es- the literature. Radiology 2003;227:522. [PMID: 12637677]
timate prognosis, (3) monitor the course of disease (pro-
gression, stability, or resolution), (4) detect disease re-
currence, and (5) select drugs and adjust therapy.
When ordering diagnostic tests, clinicians should ■ PERFORMANCE OF
weigh the potential benefits against the potential costs
and disadvantages. Some tests carry a risk of morbid-
DIAGNOSTIC TESTS
ity or mortality—eg, cerebral angiogram leads to
stroke in 0.5% of cases. The potential discomfort asso-
ciated with tests such as colonoscopy may deter some
TEST PREPARATION
patients from completing a diagnostic workup. The Factors affecting both the patient and the specimen
result of a diagnostic test may mandate further testing are important. The most crucial element in a properly
or frequent follow-up—eg, a patient with a positive conducted laboratory test is an appropriate specimen.

1675
1676 1 CHAPTER 42 111 CMDT 2005

Table 42-1. Criteria for use of screening

procedures.
■ TEST CHARACTERISTICS
Characteristics of population
1. Sufficiently high prevalence of disease. Table 42-2 lists the general characteristics of useful cL.-
2. Likely to be compliant with subsequent tests and agnostic tests. Most of the principles detailed belci,
treatments. can be applied not only to laboratory and radiol*.:
Characteristics of disease tests but also to elements of the history and physical
1. Significant morbidity and mortality. examination.
2. Effective and acceptable treatment available.
3. Presymptomatic period detectable.
4. Improved outcome from early treatment. Accuracy
Characteristics of test The accuracy of a laboratory test is its correspondence
1. Good sensitivity and specificity. with the true value. An inaccurate test is one that diffees
2. Low cost and risk. from the true value even though the results may be re-
3. Confirmatory test available and practical. producible (Figure 42-1A). In the clinical laboratory,
accuracy of tests is maximized by calibrating laboratory'
equipment with reference material and by participation in
external quality control programs.
Patient Preparation
Preparation of the patient is important for certain Precision
tests eg, a fasting state is needed for optimal glucose Test precision is a measure of a test's reproducibility
and triglyceride measurements; posture and sodium when repeated on the same sample. An imprecise tax
intake must be strictly controlled when measuring is one that yields widely varying results on repeated
renin and aldosterone levels; and strenuous exercise measurements (Figure 42-1B). The precision of diag-
should be avoided before taking samples for creatine nostic tests, which is monitored in clinical laboratories
kinase determinations, since vigorous muscle activity by using control material, must be good enough TO
can lead to falsely abnormal results. distinguish clinically relevant changes in a patient's
status from the analytic variability of the test. For in-
Specimen Collection stance, the manual white blood cell differential count
is not precise enough to detect important changes in
Careful attention must be paid to patient identifica-
tion and specimen labeling. Knowing when the the distribution of cell types, because it is calculated by
subjective evaluation of a small sample (100 cells). Re-
specimen was collected may be important. For in-
stance, aminoglycoside levels cannot be interpreted peated measurements by different technicians on the
same sample result in widely different results. Auto-
appropriately without knowing whether the speci-
m ated differential counts are more precise because
men was drawn just before ("trough" level) or after
they are obtained from machines that use objective
("peak" level) drug administration. Drug levels can-
physical characteristics to classify a much larger sample
not be interpreted if they are drawn during the
(10,000 cells).
drug's distribution phase (eg, digoxin levels drawn
during the first 6 hours after an oral dose). Sub-
stances that have a circadian variation (eg, cortisol)
can be interpreted only in the context of the time of Table 42-2. Properties of useful diagnostic
day the sample was drawn. tests.
During specimen collection, other principles
should be remembered. Specimens should not be 1. Test methodology has been described in detail so that it
drawn above an intravenous line, as this may contanai- can be accurately and reliably reproduced.
nate the sample with intravenous fluid. Excessive tour- 2. Test accuracy and precision have been determined.
niquet time will lead to hemoconcentration and an in- 3. The reference range has been established appropriately.
creased concentration of protein-bound substances 4. Sensitivity and specificity have been reliably established by
such as calcium. Lysis of cells during collection of a comparison with a gold standard. The evaluation has used
blood specimen will result in spuriously increased a range of patients, including those who have different but
serum levels of substances concentrated in cells (eg, commonly confused disorders and those with a spectrum
lactate dehydrogenase and potassium). Certain test of mild and severe, treated and untreated disease. The pa-
tient selection process has been adequately described so
specimens may require special handling or storage (eg,
blood gas specimens). Delay in delivery of specimens that results will not be generalized inappropriately.
to the laboratory can result in ongoing cellular metab- 5. Independent contribution to overall performance of a test
olism and therefore spurious results for some studies panel has been confirmed if a test is advocated as part of
arkamko.f.te&ts.
 CMDT 2005 ■ DIAGNOSTIC TESTING & MEDICAL DECISION MAKING I 1677

A B C

Figure 42-1. Relationship between accuracy and precision in diagnostic tests. The
center of the target represents the true value of the substance being tested. A: A di-
agnostic test that is precise but inaccurate; on repeated measurement, the test yields
very similar results, but all results are far from the true value. B: A test that is imprecise
and inaccurate; repeated measurement yields widely different results, and the results
are far from the true value. C: An ideal test that is both precise and accurate.

Reference Range It is important to consider also whether published

reference ranges are appropriate for the patient being
Reference ranges are method and laboratory specific. evaluated, since some ranges depend on age, sex,
In practice, theroften represent test results found in
95% of a small population presumed to be healthy; by weight, diet, time of day, activity status, or posture.
definition, then, 5% of healthy patients will have an For instance, the reference ranges for hemoglobin con-
centration are age and sex dependent. Table 2 of the
abnormal test result (Figure 42-2). Slightly abnormal
results should be interpreted critically—they may be Appendix contains the reference ranges for commonly
used chemistry and hematology tests. Test perfor-
either truly abnormal or falsely abnormal. The practi-
tioner should also be aware that the more tests or- mance characteristics such as sensitivity and specificity
are needed to interpret results and are discussed below.
dered, the greater the chance of obtaining a falsely ab-
normal result. For a healthy person subjected to 20
independent tests, there is a 64% chance that one test
Interfering Factors
result will lie outside the reference range (Table 42-3). The results of diagnostic tests can be altered by exter-
Conversely, values within the reference range may not nal factors, such as ingestion of drugs; and internal
rule out the actual presence of disease since the refer- factors, such as abnormal physiologic states.
ence range does not establish the distribution of results External interferences can affect test results in vivo or
in patients with disease. in vitro. In vivo, alcohol increases g-glutamyl
transpeptidase, and diuretics can affect sodium and
potassium concentrations. Cigarette smoking can in-
duce hepatic enzymes and thus reduce levels of sub-
stances such as theophylline that are metabolized by the
liver. In vitro, cephalosporins may produce spuri-
ous serum creatinine levels due to interference with a
common laboratory method of analysis.

Table 42-3. Relationship between the number of

-2 -1 Mean '1 tests and the probability that a healthy person will
Abnormal Normal Abnormal have one or more abnormal results.
(2.5%) --""*---(95%) (2.5%) —'
Test results
Probability That One or More
(percent of population) Number of Tests Results Will Be Abnormal
Figure 42-2. The reference range is usually defined as
1 5%
within 2 SD of the mean test result (shown as -2 and 2) in
6 26%
a small population of healthy volunteers. Note that in this
example, test results are normally distributed; however, 12 46%
many biologic substances will have distribu-
20 64%
tions that are skewed.
1678 / CHAPTER 42 ■ CMDT 2005

Internal interferences result from abnormal physio-

logic states interfering with the test measurement. As groups who are young and well. Compared with the
an example, patients with gross lipemia may have spu- general population, this study group will have more
riously low serum sodium levels if the test methodol- results that are truly positive (because patients have
ogy used includes a step in which serum is diluted be- more advanced disease) and more results that are
fore sodium is measured. Because of the potential for truly negative (because the control group is healthy).
test interference, clinicians should be wary of unex- Thus, test sensitivity and specificity will be higher
pected test results and should investigate reasons other than would be expected in the general population,
than disease that may explain abnormal results, in- where more of a spectrum of health and disease is
cluding laboratory error. found. Clinicians should be aware of this spectrum
bias when generalizing published test results to their
own practice.
Sensitivity & Specificity
Test sensitivity and specificity depend on the ,
Clinicians should use measures of test performance threshold above which a test is interpreted to be ab-
such as sensitivity and specificity to judge the quality normal (Figure 42-3). If the threshold is lowered, sen-
of a diagnostic test for a particular disease. Test sen- sitivity is increased at the expense of decreased speci-
sitivity is the likelihood that a diseased patient has a ft-city. If the threshold is raised, sensitivity is decreased
positive test. If all patients with a given disease have a while specificity is increased.
positive test (ie, no diseased patients have negative Figure 42-4 shows how test sensitivity and speci-
tests), the test sensitivity is 100%. Generally, a test ficity can be calculated using test results from patients
with high sensitivity is useful to exclude a diagnosis previously classified by the gold standard as diseased or
because a highly sensitive test will render few results nondiseased.
that are falsely negative. To exclude infection with The performance of two different tests can be com-
the AIDS virus, for instance, a clinician might pared by plotting the sensitivity and (1 minus the
choose a highly sensitive test such as the HIV anti- specificity) of each test at various reference range cut-
body test. off values. The resulting receiver operator character-
A test's specificity is the likelihood that a healthy istic (ROC) curve will often show which test is better;
patient has a negative test. If all patients who do not a clearly superior test will have an ROC curve that al-
have a given disease have negative tests (ie, no ways lies above and to the left of the inferior test
healthy patients have positive tests), the test specific- curve, and, in general, the better test will have a larger
ity is 100%. A test with high specificity is useful to area under the ROC curve. For instance, Figure 42-5
confirm a diagnosis, because a highly specific test will shows the ROC curves for prostate-specific antigen
have few results that are falsely positive. For instance, (PSA) and prostatic acid phosphatase. (PAP) in the di-
to make the diagnosis of gouty arthritis, a clinician agnosis of prostate cancer. PSA is a superior test be-
might choose a highly specific test, such as the pres-
ence of negatively bireftingent needle-shaped crys-
tals within leukocytes on microscopic evaluation of
joint fluid.
To determine test sensitivity and specificity for a 5)
particular disease, the test must be compared against c6)
an independent "gold standard" test that defines the Cil cf,
E
true disease state of the patient. For instance, the z ;a
sensitivity and specificity of the ventilation-perfusion
scan for pulmonary emboli are obtained by compar-
ing the results of scans with the gold standard, pul- A BC
monary arteriography. Application of the gold stan- Test results
dard examination to patients with positive scans
establishes specificity. Failure to apply the gold stan- Figure 42-3. Hypothetical distribution of test results
dard examination following negative scans may result for healthy and diseased individuals. The position of
in an overestimation of sensitivity, since false nega- the "cutoff point" between "normal" and 'abnormal"
tives will not be identified. However, for many dis- (or "negative" and "positive") test results determines
ease states (eg, pancreatitis), an independent gold the test's sensitivity and specificity. If point A is the
standard test either does not exist or is very difficult cutoff point, the test would have 100% sensitivity but
low specificity. If point C is the cutoff point, the test
or expensive to apply—and in such cases reliable esti-
mates of test sensitivity and specificity are sometimes would have 100% specificity but low sensitivity. For
difficult to obtain. many tests, the cutoff point is determined by the refer-
Sensitivity and specificity can also be affected by ence range, ie, the range of test results that is within 2 SD
the population from which these values are derived. of the mean of test results for healthy individuals (point B).
For instance, many diagnostic tests are evaluated first In some situations, the cutoff is altered to en-
using patients who have severe disease and control hance either sensitivity or specificity.
 CMDT 2005 ■ DIAGNOSTIC TESTING & MEDICAL DECISION MAKING I 1679

Disease
Present Absent

Positive TP FP TP = (Sensitivity)(Pretest probability)

FP = (1- Specificity)(1- Pretest probability)
FN = (1- Sensitivity)(Pretest probability)
TN = (Specificity)(1- Pretest probability)
Negative EN TN

Number of diseased patients with positive test TP

Sensitivity =
Number of diseased patients TP FN

Number of nondiseased patients with negative test TN

Specificity =
Number of nondiseased patients TN ÷ FP

TP
Posttest probability
Probability of disease if test positive -
after positive test TP + FP

(Sensitivity)(Pretest probability)

(Sensitivity)(Pretest probability) + (1- Specificity)(1- Pretest probability)

Figure 42-4. Calculation of sensitivity, specificity, and probability of disease after a positive test (posttest
probability). (TP = true positive; FP -= false positive; FN = false negative; TN = true negative.)

cause it has higher sensitivity and specificity for all

cutoff values.
■ USE OF TESTS IN DIAGNOSIS
Bossuyt PM et al: Towards complete and accurate reporting of & MANAGEMENT
studies of diagnostic accuracy: The STARD Initiative. Ann
Intern Med 2003;138:40. [PMID: 12513403]
The value of a test in a particular clinical situation de-
Gilbert R et al: Assessing diagnostic and screening tests: part 1.
Concepts. West J Med 2001;174:405. [PMID: 11381009] - pends not only on the test's sensitivity and specificity
Loong TW: Understanding sensitivity and specificity with but also on the probability that the patient has the dis-
the right side of the brain. BMJ 2003;327:716. 1PMID: ease before the test result is known (pretest probabil-
14512479] ity). The results of a useful test will substantially
Miller TD er al: Effects of adjustment for referral bias on the sen- change the probability that the patient has the disease
sitivity and specificity of single photon emission computed (posttest probability). Figure 42-4 shows how post-
tomography for the diagnosis of coronary artery disease. Am J test probability can be calculated from the known sen-
Med 2002;112:290. [PMID: 11893368]
sitivity and specificity of the test and the estimated
Muiherin SA et al: Spectrum bias or spectrum effect? Subgroup pretest probability of disease (or disease prevalence).
variation in diagnostic test evaluation. Ann Intern Med
The pretest probability of disease has a profound
2002;137:598. [PMID: 12353947]
effect on the posttest probability of disease. As demon-
Chimney AJ et al: Bayesian regression methodology for estimating
strated in Table 42--4, when a test with 90% sensitiv-
a receiver operating characteristic curve with two radiologic
applications: prostate biopsy and spiral CT of ureteral
ity and specificity is used, the posttest probability can
stones. Acad Radio! 2001;8:713. [PMID: 11508750) vary from 8% to 99% depending on the pretest proba-
Punglia S et al: Effect of verification bias on screening for prostate
bility of disease. Furthermore, as the pretest probabil-
cancer by measurement of prostate-specific antigen. N Engl J ity of disease decreases, it becomes more likely that a
Med 2003;349:335. [PM1D: 12878740] positive test result represents a false positive.
1680 / CHAPTER 42 CMDT 2005

prostate cancer rises to 50% and the posttest proba-

bility using the same test is 69% (Figure 42-6B). Fi-
nally, if the clinician estimates the pretest probability
8 to be 98% based on a prostate nodule, hone pain, and
.7 lytic lesions on spine x-rays, the posttest probability
using PSA is 99% (Figure 42-6C), This example il-
lustrates that pretest probability has a profound effect
on posttest probability and that tests provide more in-
formation when the diagnosis is truly uncertain (pre-
test probability about 50%) than when the diagnosis
.4
is either unlikely or nearly certain.

Knotterus JA et al: Evaluation of diagnostic procedures. BM]

.2 2002;324:477. [PMID: 11859054]
.1
.2 .3 .4 .5 .6 .7
1 - Specificity
Figure 42-5. Receiver operator characteristic (ROC) ■ ODDS-LIKELIHOOD RATIOS
curves for prostate-specific antigen (PSA) and pros-
tatic acid phosphatase (PAP) in the diagnosis of pros-
Another way to calculate the posttest probability of
tate cancer. For all cutoff values, PSA has higher sen- disease is to use the odds-likelihood approach. Sensi-
sitivity and specificity; therefore, it is a better test tivity and specificity are combined into one entity
based on these performance characteristics. (Modi- called the likelihood ratio (LR).
fied and reproduced, with permission, from Nicoll D et
Probability of result in diseased persons
al: Routine acid phosphatase testing for screening and LR - Probability of result in nondiseased persons
monitoring prostate cancer no longer justified. Clin
Chem 1993;39:2540.)
When test results are dichotomized, every test has
two likelihood ratios, one corresponding to a positive
As an example, suppose the clinician wishes to cal- test (Lk') and one corresponding to a negative test
culate the posttest probability of prostate cancer using
the PSA test and a cutoff value of 4 jig/L. Using the robability that test is positive in diseased persons
LR' — P
data shown in Figure 42-5, sensitivity is 90% and Probability that test is positive in nondiseased persons
specificity is 60%. The clinician estimates the pretest Sensitivity
probability of disease given all the evidence and then 1 — Specificity
calculates the posttest probability using the approach
shown in Figure 42-4. The pretest probability that an LR — P
robability that test is negative in diseased persons

otherwise healthy 50-year-old man has prostate can- Probability that test is negative in nondiseased persons
cer is equal to the prevalence of prostate cancer in 1 — Sensitivity
that age group (probability = 10%) and the posttest Specificity
probability after a positive test is only 20—ie, even
though the test is positive, there is still an 80% For continuous measures, multiple likelihood ra-
tios can be defined to correspond to ranges of results.
chance that the patient does not have prostate cancer
(Figure 42-6A). If the clinician finds a prostate nod- (See Table-42-5 for an example.)
Lists of likelihood ratios can be found in some text-
ule on rectal examination, the pretest probability of
books, journal articles, and computer programs (see
Table 42-6 for sample values). Likelihood ratios can
be used to make quick estimates of the usefulness of
Table 42-4. Influence of pretest probability on
contemplated diagnostic tests in particular situations.
the posttest probability of disease when a test
The simplest method for calculating posttest probabil-
with 90% sensitivity and 90% specificity is used. ity from pretest probability and likelihood ratios is to
use a nomogram (Figure 42-7). The clinician places a
Pretest Probability Posttest Probability straightedge through the points that represent the pre-
test probability and the likelihood ratio and then reads
0.01 0.08
the posttest probability where the straightedge crosses
0.50 0.90 the posttest probability line.
A more formal way of calculating posttest probabil-
0.99 0.999
ities uses the likelihood ratio as follows:
 CMDT 2005 • DIAGNOSTIC TESTING & MEDICAL DECISION MAKING / 1681

A Pretest
probability

Posttest
probability
Positive
test

.2 .5 1
Pro5ability of disease

Pretest Posttest
B probability probability

Positive
test

.5 .69 1
Probability of disease

Pretest Posttest
probability probability

0 .5 .98-1
Probability of disease .99-
Figure 42-6. Effect of pretest probability and test sensitivity and specificity on the posttest
probability of disease. (See text for explanation.)

Pretest odds x Likelihood ratio = Posttest odds positive (LW = 24), then the posttest odds of having a
myocardial infarction are
To use this formulation, probabilities must be con-
verted to odds, where the odds of having a disease are 3 72
2 x 24 = — or 36:1 odds
expressed as the chance of having the disease divided 2
by the chance of not having the disease. For instance, 36/1
_
36 - 97% probability)
a probability of 0.75 is the same as 3:1 odds (Figure \(36/1)+ 1 37
42-8).
If the troponin I test is negative (LR = 0.01), then
To estimate the potential benefit of a diagnostic
test, the clinician first estimates the pretest odds of dis- the posttest odds of having a myocardial infarction are
ease given all available clinical information and then
multiplies the pretest odds by the positive and nega- 3 003
-x 0.01 = odds
tive likelihood ratios. The results are the posttest 2 2

(
odds, or the odds that the patient has the disease if the 0.03/2 _ 0.015 = 1.5% probability
test is positive or negative. To obtain the posttest (0.03/2) + 1 0.015 + 1
probability, the odds are converted to a probability
(Figure 42-8).
For example, if the clinician believes that the pa- Sequential Testing
tient has a 60% chance of having a myocardial infarc- To this point, the impact of only one test on the prob-
tion (pretest odds of 3:2) and the troponin I test is ability of disease has been discussed, whereas, during
1682 1 CHAPTER 42 ■ CMDT 2005

Table 42-5. Likelihood ratios of serum ferritin in the .1 99

diagnosis of iron deficiency anemia.
.2 -
Serum Ferritin (pgIL) LR for Iron Deficiency Anemia
100 0.08 -95
.5 -
45-99 0.54
35-44 1.83
1 - 1000 - 90
25-34 2.54
500
15-24 8.83
<15 51.85 200 - 80
100
From Guyatt G: Laboratory diagnosis of iron deficiency anemia. J
50 - 70
Gen Intern Med 1992;7:145.
5 -
20 - 60
10 - 50
10 -
5 - 40
most diagnostic workups, clinicians obtain clinical in-
formation in a sequential fashion. To calculate the 2 - 30
20 -
posttest odds after three tests, for example, the clini- cy.
- 20
cian might estimate the pretest odds and use the ap- 30
propriate likelihood ratio for each test: 40 -
- 10
50 -
Pretest odds x LR1 x LR2 x LR3 = Posttest odds 60 -
- 5
70 - - .02
When using this approach, however, the clinician - .01
should be aware of a major assumption: the chosen 80 - - .005 - 2
tests or findings must be conditionally indepen- - .002
dent. For instance, with liver cell damage, the aspar- 90 - - 1
-.001
tate aminotransferase (AST) and alanine aminotrans-
ferase (ALT) enzymes may be released by the same
95 - -.5
process and are thus not conditionally independent.
If conditionally dependent tests are used in this se-
quential approach, an inaccurate posttest probability - .2
will result.
99 .1
Blume JD: Likelihood methods for measuring statistical evidence.
Pretest Likelihood Posttest
Stat Med 2002;21:2563. [PMID: 12205699]
probability ratio probat
Sutton AJ et al: Bayesian, methods in meta-analysis and evi-
dence synthesis. Stat Methods Med Res 2001;10:277. [PMID: Figure 42-7. Nomogram for determining
11491414] probability from pretest probability and likelih&:
tios. To figure the posttest probability, place a -
edge between the pretest probability and the 1: -
hood ratio for the particular test. The posttest
Table 42-6. Examples of likelihood ratios.
probability will be where the straightedge cross -- -
posttest probability line. (Adapted and reprddi...-:
Target Disease Test LR'" LW with permission, from Fagan Ti: Nomogram for
Abscess Abdominal CT 9.5 1 0.06 theorem. [Letter.] N Engl J Med 1975;293:257.)

Coronary artery Exercise ECG (1 3.5 0.45

disease mm depression)
Threshold Approach to Decision Marc
Lung cancer Chest x-ray 15 0.42
A key aspect of medical decision making is the 5. _
Left ventricular Echocardiography 18.4 0.08
tion of a treatment threshold, ie, the prolx±.17
hypertrophy disease at which treatment is indicated. Figut
Myocardial infarction Troponin I 24 0.01 shows a possible way of identifying a treatment - -
old by considering the value (utility) of the foc7
Prostate cancer Digital rectal 213 037 ble outcomes of the treat/don't treat decision.
examination Use of a diagnostic test is warranted when i 7 _
From http://www.med.unc.eduirnedicine/edursrc/Irmain.htm could shift the probability of disease across thf-
 CMDT 2005 1111 DIAGNOSTIC TESTING & MEDICAL DECISION MAKING / 1683

below 25%. Using the same nomogram, the posttest

Probability probability after a negative test would be 33% (Figure
Odds =
- Probability 42-10C). Therefore, ordering the throat culture
would not be justified as it does not affect patient
Example: If probability = 0.75, then. management.
This approach to decision making is now being ap-
0.75 0.75
Odds = = plied in the clinical literature.
1 -0.75 0.25
Dorin RI et al: Diagnosis of adrenal insufficiency. Ann Intern
Med 2003;139:194. [PMID: 12899587]
Odds
Probability = Solomon DH et al: The rational clinical examination. Does this
Odds + 1 patient have a torn meniscus or ligament of the knee? Value
of the physical examination. JAMA 2.001;286:1610. [PMID:
Example: If odds = 3:1, then 11585485]

3/1 3 Decision Analysis

Probability = +1 - _ 0.75
(311)+1 3+1 Up to this point, the discussion of diagnostic testing
has focused on test characteristics and methods for
Figure 42-8. Formulas for converting between proba- using these characteristics to calculate the probability
bility and odds. of disease in different clinical situations. Although use-
ful, these methods are limited because they do not in-
corporate the many outcomes that may occur in clini-
ment threshold. For example, a clinician might decide cal medicine or the values that patients and clinicians
to treat with antibiotics if the probability of strepto- place on those outcomes. To incorporate outcomes
coccal pharyngitis in a patient with a sore throat is and values with characteristics of tests, decision analy-
greater than 25% (Figure 42-10A). If, after reviewing sis can be used.
evidence from the history and physical examination, The basic idea of decision analysis is to model the
the clinician estimates the pretest probability of strep options in a medical decision, assign probabilities to
throat to be 15%, then a diagnostic test such as throat the alternative actions, assign values (utilities) to the
culture (IR+ = 7) would be useful only if a positive test various outcomes, and then calculate which decision
would shift the posttest probability above 25%. Use of gives the greatest expected value (expected utility).
the nomogram shown in Figure 42-7 indicates that To complete a decision analysis, the clinician would
the posttest probability would be 55% (Figure 42— proceed as follows: (1) Draw a decision tree showing
10B); thus, ordering the test would be justified as it af- the elements of the medical decision. (2) Assign
fects patient management. On the other hand, if the probabilities to the various branches. (3) Assign val-
history and physical examination had suggested that ues (utilities) to the outcomes. (4) Determine the
the pretest probability of strep throat was 60%, the expected value (expected utility) (the product of
throat culture = 0.33) would be indicated only if probability and value [utility]) of each branch. And
a negative test would lower the posttest probability (5) select the decision with the highest expected
value (expected utility).
Figure 42-11 shows a decision tree where the deci-
sion to be made is whether to treat without testing,
perform a test and then treat based on the test result,
or perform' no tests and give no treatment. The clini-
1- eat cian begins the analysis by building a decision tree
e at showing the important elements of the decision. Once
the tree is built, the clinician assigns probabilities to all
B the branches. In this case, all the branch probabilities
D can be calculated from (1) the probability of disease
before the test (pretest probability), (2) the chance of a
Treatment
threshold positive test if .the disease is present (sensitivity), and
(3) the chance of a negative test if the disease is absent
100 (specificity). Next, the clinician assigns value (utility) to
0 Probability of disease (%) each of the outcomes.
Figure 42-9. The "treat/don't treat" threshold. A: Pa- After the expected value (expected utility) is calcu-
tient does not have disease and is not treated (highest lated for each branch of the decision tree, by multiply-
utility). B: Patient does not have disease and is treated ing the value (utility) of the outcome by the probabil-
(lower utility than A). C: Patient has disease and is ity of the outcome, the clinician can identify the
treated (lower,utility than A). D: Patient has disease and alternative with the highest expected value (expected
is not treated (lower utility than C). utility).

I•
1684 / CHAPTER 42 ■ CMDT 2005

A Treat/don't treat
threshold

Don't treat Treat

0 .5
Probability of disease

Pretest
B probability Posttest
probability

Positive
test

Don't treat Treat

.5
Probability of disease

Pretest
Posttest probability
C probability

Negative
test

y
Don't treat Treat

0 .5 1
Probability of disease

Figure 42-10. Threshold approach applied to test ordering. If the contemplated

test will not change patient management, the test should not be ordered. (See text
for explanation.)
f

Although time-consuming, decision analysis can Sculpher Met al: Assessing quality in decision analytic cost-effective-
help to structure complex clinical problems and to ness models. A suggested framework and example of applica-
tion. Pharmacoeconomics 2000;17:461. [PMID: 10977388]
make difficult clinical decisions.

Elwyn Get al: Decision analysis in patient care. Lancet 2001; Evidence-Based Medicine
358:571. [PMID: 11520546]
Evidence-based medicine stresses the use of evidence
Kuntz KM et al: Assessing the sensitivity of decision-analytic re-
sults to unobserved markers of risk: defining the effects of from clinical research—rather than intuition and
heterogeneity bias. Ivied Deets Making 2002;22:218. [PMID: pathophysiologic reasoning—as a basis for clinical de-
12058779] cision making. Evidence-based medicine relies on the
 CMDT 2005 ■ DIAGNOSTIC TESTING & MEDICAL DECISION MAKING / 1685

Outcomes
Disease
Treat, Disease +, No test
Treat

1- p No disease Treat, Disease -, No test

Test +
Treat, Disease +, Test done
Disease

Test -
Don't treat, Disease +, Test done

Test + Treat, Disease -, Test done

1- Spec
No disease

Spec Test -
Don't treat, Disease —, Test done

Disease Don't treat, Disease +, No test

Don't treat

1- p No disease
Don't treat, Disease -, No test
Figure 42-11. Generic tree for a clinical decision where the choices are (1) to treat the patient empirically,
(2) to do the test and then treat only if the test is positive, or (3) to withhold therapy. The square node is
called a decision node, and the circular nodes are called chance nodes. (p = pretest probability of disease;
Sens = sensitivity; Spec = specificity.)

identification of methodologically sound evidence, though some clinicians are concerned about the effect
critical appraisal of research studies, and the dissemi- of guidelines on professional autonomy and individual
nation of accurate and useful summaries of evidence to decision making, many organizations are trying to use
inform clinical decision making. Systematic reviews can compliance with practice guidelines as a measure of
be used to summarize evidence for dissemination, as can quality of care.
evidence-based synopses of current research. Systematic
reviews often use mera-analysis—statistical techniques to Guyatt G, Rennie D (editors): Users' Guides to the Medical Literature: A
combine evidence from different stud- Manualfir Evidence-Based Clinical Practice. AMA Press, 2002.
ies to produce a more precise estimate of the effect of an Hatala R et al: Evaluating the teaching of evidence-based medicine.
JAMA 2002;288:1110. [PMID: 12204080]
intervention or the accuracy of a test.
Clinical practice guidelines are systematically devel- Man-Son-Hing M et al: Determination of the clinical importance
of study results. J Gen Intern Med 2002;17:469. (PMID:
oped statements intended to assist practitioners and 12133163]
patients in making decisions about health care. Clini-
Newman DH et al: Evidence-based medicine, A primer for the
cal algorithms and practice guidelines are now ubiqui- emergency medicine resident. Ann Emerg Med 2002;39:77.
tous in Medicine. Their utility and validity depend on [PMID: 11782735]
the quality of the evidence that shaped the recommen- Shea B et al: A comparison of the quality of Cochrane reviews and
dations, on their being kept current, and on their ac- systematic reviews published in paper-based journals. Eval
ceptance and appropriate application by clinicians. Al- Health Prof 2002;25:116. [PMID: 11868441]