Page 1

Impact of Reliance on CT Pulmonary Angiography on

Diagnosis of Pulmonary Embolism:

A Bayesian Analysis

Sumant R. Ranji, MD1

Kaveh G. Shojania, MD2

Robert L. Trowbridge, MD3

Andrew D. Auerbach, MD, MPH1

1Department of Medicine, University of California

San Francisco, San Francisco, California

2Department of Medicine, University of Ottawa,

and Ottawa Health Research Institute, Ottawa, On-

tario, Canada

3Department of Medicine, Maine Medical Center,

Portland, Maine

Dr Ranji had full access to all data in the study and

takes responsibility for the integrity of the data and

the accuracy of the data analysis. We thank C.

Seth Landefeld, MD, and Robert M. Wachter, MD,

for helpful editorial comments on earlier drafts of

the manuscript.

Dr Shojania is supported by a Canada Research

Chair in Patient Safety. Dr Auerbach is supported

by a K08 research and training grant from the

Agency for Healthcare Research and Quality.

BACKGROUND: Spiral computed tomographic pulmonary angiography (CTPA) has

become the primary test used to investigate suspected pulmonary embolism (PE)

at many institutions, despite uncertainty regarding its sensitivity and specificity.

Although CTPA-based diagnostic algorithms focus on minimizing the false-nega-

tive rate, we hypothesized that increasing use of CTPA also might lead to false-

positive diagnoses.

OBJECTIVE: Determine the frequency of possible false-positive diagnoses of PE

when CTPA is the primary diagnostic test.

DESIGN: Retrospective cohort study.

SETTING: Two academic teaching hospitals.

PARTICIPANTS: 322 patients with suspected PE evaluated with CTPA.

MEASUREMENTS: We used a validated prediction rule to determine the pretest

probability of PE in each patient. We combined these pretest probabilities with

published estimates of CTPA test characteristics to generate expected posttest

probabilities of PE. We compared these posttest probabilities to actual treatment

decisions to determine the rate of false-positive diagnoses of PE.

RESULTS: Among 322 patients investigated for PE, 37 (12%) had high pretest

probability, 101 (32%) moderate, and 184 (57%) low. CT scans were interpreted as

positive for PE in 57 patients (17.8%). Regardless of the pretest probability of PE,

96.5% of patients with a positive CTPA were treated with anticoagulants. Even

under an optimistic assumption of CTPA test characteristics, as many as 25.4% of

these patients may have been treated unnecessarily as a result of a false-positive

diagnosis. Most of these patients had a low pretest probability of PE.

CONCLUSIONS: FailuretoutilizeBayesianreasoningwheninterpretingCTPAmayleadto

false-positive diagnoses of pulmonary embolism in a substantial proportion of patients.

Journal of Hospital Medicine 2006;1:81–87. © 2006 Society of Hospital Medicine.

KEYWORDS: pulmonary embolism, CT pulmonary angiography, Bayes’ theorem,

diagnosis

S

pulmonary embolism (PE). At our institution CTPA became the

initial diagnostic study in 83% of patients with suspected PE

within 3 years of the introduction of CT,1and by 2001 CTPA had

become the most common diagnostic test performed nationwide

in patients diagnosed with PE.2Most scans are interpreted as

either positive or negative for pulmonary embolism, providing

clinicians with a greater sense of diagnostic certainty than with the

probabilistic results of lung scintigraphy. Initial studies of CTPA

supported this appearance of diagnostic certainty, reporting sen-

piral computed tomographic pulmonary angiography (CTPA)

is a common first-line test for the evaluation of suspected

ORIGINAL RESEARCH

© 2006 Society of Hospital Medicine

DOI 10.1002/jhm.71

Published online in Wiley InterScience (www.interscience.wiley.com).

81

Page 2

sitivity and specificity of greater than 90%,3,4but

several subsequent studies have failed to reproduce

these results.5–7Newer multidetector CT scans are

believed to be more accurate than earlier single-

detector CT,8but true estimates of CTPA test char-

acteristics will not be known until publication of

the forthcoming PIOPED II study.9

Even without these data, CT-based diagnostic

algorithms have already appeared.10–14These algo-

rithms generally focus on minimizing the false-neg-

ative rate through use of serial testing (involving

combinations of serum D-dimer, lower-extremity

ultrasound, and CTPA). A recent meta-analysis

demonstrated that negative CTPA is highly accurate

at ruling out PE, with test characteristics similar to

conventional pulmonary angiography.15Another

meta-analysis found that the 3-month rate of sub-

sequent venous thromboembolism after negative

CTPA was 1.4% (95% CI 1.1%-1.8%),16supporting

the strategy of withholding anticoagulants after

negative CTPA in combination with other tests.

However, use of serial testing to establish the diag-

nosis of PE and initiate anticoagulation has not

been systematically evaluated or recommended,

even for patients with a low pretest probability of

PE.17

To assess the potential impact of these algo-

rithms on the diagnosis of PE in clinical practice,

we analyzed the clinical presentation and treat-

ment of a cohort of patients at our institution who

underwent CTPA for suspected PE.1We calculated a

range of posttest probabilities for pulmonary em-

bolism for these patients, given the pretest proba-

bilities, test results, and estimates of CTPA test

characteristics. We then compared the treatment

decisions of clinicians to the posttest probabilities

of PE in order to establish the potential frequency

of false-positive and false-negative diagnoses and

to determine if patients were treated appropriately

based on these estimates.

METHODS

Sites and Subjects

Details of the sites, subjects, and methods used to

collect patient-level data in this analysis have been

previously published.1The study was performed at

Moffitt-Long Hospital and San Francisco General

Hospital, teaching hospitals affiliated with the Uni-

versity of California San Francisco School of Medi-

cine. At both sites, single-detector CT scans were

available 24 hours a day throughout the study pe-

riod and were read by attending radiologists who

specialized in thoracic imaging. We excluded pa-

tients whose CTPA was not completed as the initial

test in the evaluation of suspected PE, those who

underwent testing for any indication other than

suspected acute PE, and those with incomplete

medical records or technically inadequate CTPA.

We randomly selected 345 patients who under-

went CTPA between January 1, 1998, and December

31, 2000, from the Radiology Department data-

bases. One investigator (R.L.T.) then abstracted

charts of all patients. For each subject, we collected

data about history and clinical presentation, diag-

nostic impressions of the treating clinicians, treat-

ments administered both before and after diagnos-

tic testing, CTPA result, results of other diagnostic

tests for PE, and final clinical diagnosis. During the

study period, there were no institution- or depart-

ment-specific guidelines or decision aids available

for the diagnosis of PE. Ventilation-perfusion scan,

lower extremity ultrasound, and pulmonary angiog-

raphy were available, but highly sensitive D-dimer

assays were not in use. The study was approved by

the Institutional Review Boards of both sites, and

requirement for written informed consent from pa-

tients was waived.

Estimates of Pretest Probabilities of Pulmonary Embolism

and CTPA Test Characteristics

Several prediction rules18-20generate clinical pre-

test probabilities for patients with suspected PE. We

used the Wells score18to assign a pretest probabil-

ity of low, moderate, or high to each patient on the

basis of the following clinical variables: leg swelling,

hemoptysis, tachycardia, history of recent immobi-

lization, history of prior DVT or PE, active malig-

nancy, and lack of a more likely alternative diagno-

sis. We chose this rule as (unlike other prediction

rules such as the Geneva rule20) the Wells score has

been validated for hospitalized patients with sus-

pected PE and does not require arterial blood gas

measurements. The prevalence of PE reported in

the evaluation of the Wells score was 3.4%, 27.8%,

and 78.3% for low, moderate, and high pretest

probabilities, respectively.18

As in our previous study,1we assumed CTPA to

be 90% sensitive and 95% specific based on pub-

lished estimates.3,17These values correspond to a

positive likelihood ratio of 18 and a negative likeli-

hood ratio of 0.1.21We chose these values as a

best-case estimate of the test characteristics of

CTPA, although other studies have found less im-

pressive results.7Using these pretest probabilities

82Journal of Hospital MedicineVol 1 / No 2 / Mar/Apr 2006

Page 3

and likelihood ratios, we then used Bayes’ theorem

(Figure 1) to calculate the range of expected post-

test probabilities of pulmonary embolism.

Calculation of Posttest Probabilities and Comparison to

Treatment Outcomes

For each pretest probability category, we used the

posttest probabilities calculated above to deter-

mine the number of true-positive pulmonary em-

boli, as follows:

Number of true-positive pulmonary emboli

? ?posttest probability given positive result?

? ?number of patients with positive CTPA?

We then compared treatment decisions made by

clinicians at our hospital to the calculated posttest

probabilities and number of true-positive diag-

noses of PE. We considered the difference between

the number of patients treated for PE and the num-

ber of true-positive diagnoses of PE to represent

possible false-positive diagnoses. In a similar fash-

ion, we determined the number of likely true-neg-

ative diagnoses of PE and considered the difference

between the number of patients not treated for PE

and the number of true-negative diagnoses to rep-

resent possible false-negative diagnoses.

RESULTS

Patient Characteristics

After excluding 23 patients receiving anticoagulants

for other indications prior to CTPA, the study co-

hort included 322 patients (57.7% female), with an

average age of 58.6 years, of whom 20.5% had can-

cer and 4.5% had a prior history of thromboembolic

disease. Scans were primarily ordered by the med-

icine service (47.7% of cases) and emergency de-

partment (22.9%). CTPA was the initial test for 9%

of patients evaluated for suspected acute PE during

the first 6 months of the study period, increasing to

83% by the end of 2000.1The overall pretest prob-

ability distribution remained the same throughout

the entire study period.1

Test Results and Treatment Decisions

Most patients in our cohort had a low (n ? 184,

57.1%) or a moderate (n ? 101, 31.4%) pretest prob-

ability of PE (Table 1). The likelihood of a positive

CTPA increased as the pretest probability in-

creased, but even among patients with high clinical

risk, only 35.1% had positive CT scans. In total,

FIGURE 1. Bayes’ theorem.

TABLE 1

Study Results Stratified by Pretest Probability

Pretest probability of PE (number of CTPA performed) Low (N ? 184) Moderate (N ? 101) High (N ? 37)Total (N ? 322)

CTPA positive for PE (% of pretest probability group)

CTPA negative for PE (% of pretest probability group)

Patients with positive CT subsequently treated for PE

(% of pretest probability group)

Patients treated for PE despite negative CT

(% of pretest probability group)

Total patients treated for PE (% of pretest probability group)

22 (12.0%)

162 (88.0%)

22 (21.8%)

79 (78.2%)

13 (35.1%)

24 (64.9%)

57 (17.7%)

265 (82.3%)

21 (11.4%) 21 (20.8%)13 (35.1%) 55 (17.1%)

5 (2.7%)

26 (14.1%)

3 (3.0%)

24 (23.8%)

3 (8.1%)

16 (43.2%)

11 (3.4%)

66 (20.5%)

Low, moderate, and high pretest probabilities were determined using the Wells criteria.18The probability of PE in each category was 3.4%, 27.8%, and 78.3%, respectively.

Impact of CT on PE Diagnosis / Ranji et al.83

Page 4

scans were positive in 57 patients and negative in

265 patients. Clinicians treated 55 patients with a

positive CTPA (96.5%); none of these patients un-

derwent additional testing for DVT or PE after the

imaging study. Among patients with a negative

CTPA, 254 (95.8%) were not treated; none of the

patients in whom anticoagulation was withheld un-

derwent further testing, whereas the other 11 pa-

tients were treated on the basis of other tests (5

high-probability ventilation-perfusion scans, 3 pos-

itive leg ultrasounds, and 3 for unclear reasons).

Overall, 66 patients (20.5%) were treated for pulmo-

nary embolism.

Literature-Derived Estimates of Posttest Probabilities of

Pulmonary Embolism

Patients who have a low pretest probability of PE

and a positive CTPA have a posttest probability of

41.6% under our estimate of CTPA test characteris-

tics. Patients with moderate pretest probability

have a posttest probability of 87.4% and patients

with a high pretest probability will have a 98.5%

probability of embolism with a positive scan. The

traditional treatment threshold for PE is a posttest

probability of 90%.22

Observed Versus Expected PE Rates and Subsequent

Treatment

Only 9 of the 22 patients (41%) with a low pretest

probability and a positive CTPA likely represent

true-positive emboli. However, clinicians chose to

treat 21 of the 22 patients with this combination of

pretest probability and imaging findings. Thus, 12

emboli would be considered possible false-positive

diagnoses. Similarly, in the moderate pretest prob-

ability group, 2 of 21 patients with moderate pretest

probability and 0 of 13 patients with high pretest

probability treated for PE had a possibly false-pos-

itive diagnosis. Thus, in total, 25.4% (14 of 55) pa-

tients treated for PE had a possible false-positive

diagnosis of pulmonary embolism and may have

been unnecessarily administered anticoagulants

(Table 2). All patients who potentially had a false-

positive PE had either a low or moderate pretest

probability of PE; in fact, the majority (57.1%) of

patients with a low pretest probability of PE who

were subsequently treated for PE likely had a false-

positive diagnosis.

Clinicians were more likely to overtreat a pa-

tient with a possible false-positive CT scan than to

withhold treatment from a patient with a possible

false-negative diagnosis. Using the same estimates

of CTPA test characteristics, the incidence of pos-

sible false-negative diagnosis of PE was 1.6% (4

possible false-negative diagnoses among 254 pa-

tients with negative CTPA results who were not

treated for PE.) All these patients had a high pretest

probability of PE.

DISCUSSION

Physicians at our institution regarded CTPA results

as definitive, anticoagulating 96.5% of patients with

a positive CT and withholding treatment in 95.8% of

patients with a negative scan. This practice pattern

may result in unnecessary anticoagulation of many

patients with a low pretest probability of PE who

may have had false-positive CTPA findings. In con-

trast, the rate of possible false-negative diagnosis of

PE was low, consistent with the results of several

other studies.16

TABLE 2

Clinical Treatment Decisions Compared to Calculated Number of True-Positive Pulmonary Emboli in Patients Treated for PE

Pretest probability

Low

(n ? 184)

Moderate

(n ? 101)

High

(n ? 37)

Total

(n ? 322)

CTPA positive for PE (% of pretest probability group)

Patients with positive CTPA treated for pulmonary embolism (n, % treated in risk

group)

Calculated number and rate of probable true-positive evaluations

Number of true-positive PE (n, % treated in risk group)

Calculated number and rate of possible false-positive evaluations

Number of possible false-positive PE (n, % in risk group with unexpected PE)

22 (12.0%)22 (21.8%) 13 (35.1%) 57 (17.7%)

21 (95.4%) 21 (95.4%)13 (100%)55 (96.5%)

9 (42.9%) 19 (90.5%)13 (100%)41 (74.6%)

12 (58.1%) 2 (9.5%)0 14 (25.4%)

The number of false-positive pulmonary emboli in each group was determined by subtracting the calculated number of true-positive evaluations from the number of patients who were treated in each group. The

total number in each category was calculated as the sum of each pretest probability group.

84 Journal of Hospital MedicineVol 1 / No 2 / Mar/Apr 2006

Page 5

The use of CTPA is likely to increase because of

the publication of multiple algorithms advocating

that CTPA be the chief imaging study used in the

diagnosis of PE.10–14These algorithms recommend

serial testing on patients with a negative CTPA in

order to minimize the false-negative rate, but they

do not require systematic follow-up in patients with

a positive scan, even if the pretest probability was

low. In management trials, this approach resulted

in a low false-negative rate (1.0%-1.8% at 3-month

follow-up).11–14However, the rate of major bleeding

in patients treated for PE was 3.2%-6.0% at 3

months,12–14illustrating the potential risk of anti-

coagulating patients who may have false-positive

diagnoses. Furthermore, premature diagnostic clo-

sure after a CTPA “positive for PE” may result in

additional morbidity as a result of missing the true

diagnosis.

One potential explanation for the large number

of potential false-positive emboli seen in low-risk

patients is that it is difficult to accurately diagnose

distal pulmonary emboli with CTPA. The interrater

reliability of CTPA for diagnosis of subsegmental PE

is suboptimal,23and the clinical significance of

these emboli remains uncertain.24Thus, many em-

boli found in patients with low pretest probability

actually may have been subsegmental PE that

would not have been diagnosed by another radiol-

ogist. As CTPA is more accurate for diagnosing cen-

tral PE,25clinicians should consider reviewing “pos-

itive” scans with the interpreting radiologist,

especially when the pretest probability was low and

the filling defects identified are in distal vessels.

Our results may also illustrate that clinicians

have a lower treatment threshold when presented

with apparently definitive evidence of pulmonary

embolism. Previous proposals on the appropriate

treatment threshold for PE, which used Bayesian

decision-making methods similar to ours,22incor-

porated PIOPED26data on the pretest probability of

pulmonary embolism, the test characteristics of

ventilation-perfusion scans, and the clinical out-

comes of patients in each test result/pretest prob-

ability category. However, there is no correspond-

ing data for CTPA, as its test characteristics are still

uncertain, and long-term clinical outcomes have

not been documented for patients treated (or not

treated) on the basis of CT results.

Our study had several limitations. First, chart-

ing bias potentially was introduced by our using a

retrospective method of collecting data for calcu-

lating pretest probabilities. To address this poten-

tial bias, we collected data from the entire medical

record, including information available at and pre-

ceding the time of the CT scan. We believe this

method was effective, as the range of pretest prob-

abilities and the prevalence of PE in our study were

very similar to those seen in a number of prospec-

tive studies.18–20,26,27Although other risk indices

exist, the Wells score has been shown to have pre-

dictive powers equal to other algorithms and to

clinicians; implicit assessments.28,29In our cohort,

35.1% of patients with a high pretest probability

were diagnosed with PE; although this was lower

than that in the initial Wells cohort,18it was very

similar to a subsequent validation study using the

Wells algorithm, in which the prevalence of PE in

patients with high pretest probability was 37.5%.27

Plasma D-dimer testing is not routinely used at our

hospitals, but it is a component of some CTPA-

based diagnostic algorithms.11–14Although use of

D-dimer testing may have led to fewer scans in

patients with negative D-dimer test results and low

pretest probability,30the high false-positive rate for

D-dimer assays31makes it difficult to predict the

effect of widespread D-dimer use on the overall

pretest probability distribution. Using our assump-

tions about CT test characteristics, a pretest

probability of more than 30% is required to gener-

ate a posttest probability of PE of at least 90% (the

traditional treatment threshold for anticoagulant

therapy22) with a positive scan. Extensive D-dimer

use would be unlikely to cause such a shift in the

distribution of pretest probabilities.

Finally, CT technology has continued to ad-

vance, and many institutions now use 64-slice

scanners32in contrast to the single-slice scanners in

use at the time our data were collected. Our as-

sumptions were that CTPA has a positive likelihood

ratio of 18.0 and a negative likelihood ratio of 0.1

(corresponding to a sensitivity of 90% and a speci-

ficity of 95%), although many studies of single-

detector CTPA found less impressive values.5,7Mul-

tidetector CT is thought to be more accurate than

was earlier technology, but the true diagnostic per-

formance of multidetector CT is not yet known.

However, our findings pertain primarily to clini-

cians’ responses to test results, so even if newer

scanners are more accurate, Bayesian analysis will

still be required in order to appropriately treat pa-

tients. A recent meta-analysis of diagnostic strate-

gies for PE found CTPA to have a positive likelihood

ratio of 24.1, but even using this higher value, pa-

tients with a low pretest probability and positive

Impact of CT on PE Diagnosis / Ranji et al.85