Better Care for Patients with Suspected Pulmonary Embolism

The diagnosis and treatment of pulmonary embolism engage the internist as few other diseases do, for many good reasons. Physicians often miss the diagnosis (1). The mortality rate for untreated pulmonary embolism is quite high (2, 3) and is much lower when the disorder is correctly treated (4). However, treatment-related bleeding carries a substantial fatality rate (5). Pulmonary angiography, still the definitive test, is inconvenient to do at night, and the sensitivity of noninvasive imaging tests is still too low to rule out disease when clinical suspicion is high (6). The medical literature contains 2 lines of inquiry regarding the diagnosis of pulmonary embolism. The first is the pragmatic management trial, which asks how often pulmonary embolism occurs when low-risk patients do not receive anticoagulant therapy. The second is concerned with estimating the pretest probability of pulmonary embolism and with using data related to test accuracy to calculate the post-test probability. The foundation is the Bayes theorem, which states that the post-test odds of a disease equal the pretest odds multiplied by the likelihood ratio of the test result. This relationship tells us that the interpretation of a test result depends on the pretest odds of disease. When should we withhold treatment? Some would say only when the probability of pulmonary embolism is zero. This policy will lead us to perform expensive tests in patients with a very low probability of pulmonary embolism. Decision theory states that we should withhold treatment when the probability of pulmonary embolism is lower than the treatment threshold probability for pulmonary embolism, which is the lowest probability at which it is reasonable to treat. This issue of Annals includes 2 articles about the management of pulmonary embolism. One reports that physicians often withhold anticoagulant treatment after a negative test result when decision guidelines say they should treat or do additional tests (7). The other article describes a clinical prediction rule for estimating the probability of pulmonary embolism (8). A third article in this issue deals with broader questions about clinical prediction rules in day-to-day practice (9). According to Roy and colleagues (7), performance needs improvement. They studied 1529 consecutive patients who presented with suspected pulmonary embolism to 117 emergency departments. On the basis of the literature, they established criteria for ruling out and confirming a diagnosis of pulmonary embolism for each level of probability (low, medium, and high) and for patients whose emergency department physician did not record a pretest probability of pulmonary embolism (28% of the patients). For example, in a patient with a high clinical probability, appropriate criteria for ruling out pulmonary embolism would include normal pulmonary arteriographic findings, negative results of a ventilationperfusion lung scan, and negative enzyme-linked immunosorbent assay d-dimer results. Two investigators independently reviewed each patient's data and the emergency department physician's final diagnosis. They classified the diagnostic criterion that the emergency department physician used as appropriate (that is, measures found on the list of the investigators' prespecified criteria) or inappropriate (that is, measures that were not on their list of criteria). Physicians followed established guidelines to confirm pulmonary embolism much more often than they followed guidelines for ruling out pulmonary embolism. Emergency department physicians diagnosed pulmonary embolism in 429 patients; in 393 (92%), they used appropriate criteria for confirming the diagnosis. They ruled out pulmonary embolism in 1100 patients. In 474 (43%), they used appropriate criteria for ruling out pulmonary embolism; in 626 (57%), they used inappropriate criteria. The incorrect use of tests to rule out pulmonary embolism had substantial costs. Among the 1100 patients in whom the emergency department physicians thought they had ruled out pulmonary embolism, 44 had a thromboembolic event during 3 months of follow-up. Most of these events (39 of 44) occurred in patients whose emergency department physicians used an inappropriate criterion for ruling out pulmonary embolism. The absolute risk for thromboembolism during follow-up increased from 1.2% to 7.7% when physicians used an inappropriate diagnostic criterion. If established guidelines represent correct reasoning, physicians used incorrect reasoning much more often when ruling out pulmonary embolism than when confirming the diagnosis. I have not seen previous reports of this phenomenon; if the researchers' findings are confirmed, it is an important observation. What does it mean? One possibility is that physicians are more likely to overestimate the effect of test results on probability when the results are negative, but it's also possible that their estimates of pretest probability are consistently too low. These hypotheses are testable. Some of these errors are easily avoidable. The best way to avoid errors in calculating post-test probability is to use the Bayes theorem, whereas errors in estimating pretest probability of pulmonary embolism can be prevented by using a clinical prediction rule. Computer-based decision support can be linked to the electronic health record, giving the physician the calculated post-test probability after a test and displaying evidence-based guidelines about the next steps in management. Such a system could also tell the physician how to estimate the pretest probability of pulmonary embolism. This proposal leads us to our second article on pulmonary embolism management. Le Gal and colleagues (8) have developed a clinical prediction rule for estimating the probability of pulmonary embolism (8). It is not the first. Wells and colleagues published a rule in 1998 (10), which was later refined in 2000 (11). The 1998 Wells rule classified patients into 3 categories of pretest probability: low (3%), intermediate (28%), and high (78%). The 1998 rule was awkward to use and required a subjective judgment about whether another diagnosis could be more likely than pulmonary embolism. The 2000 rule used a simple point-based scoring system but retained the judgment about other diagnoses. The Geneva rule avoided the subjective judgment but required arterial blood gas measurement (12). In the newest approach, Le Gal and colleagues developed a clinical prediction rule that would not require tests or subjective diagnostic judgments. The authors enrolled 965 consecutive patients who presented with suspected pulmonary embolism to the emergency department of 3 hospitals. Each underwent a standardized evaluation that was specific to the patient's clinical probability group as established by applying the Geneva rule. The overall prevalence of pulmonary embolism was 23%. The authors identified 8 independent predictors of pulmonary embolism by using multivariable statistical methods that assigned points to each finding. The sum of points assigned to these 8 findings determined the patient's total score. The authors defined 3 risk categories in which the prevalence of pulmonary embolism was 8%, 29%, and 74%. The authors did several things to evaluate the rule. First, the area under the receiver-operating characteristic curve was 0.74, which means that the probability is 0.74 that in a randomly selected pair of patients, one who has a pulmonary embolism and another who does not, the patient with a pulmonary embolism has a higher probability of pulmonary embolism. This probability is not high, but the rule's purpose is to estimate the pretest probability, not to make a final diagnosis. Second, they applied the rule to another sample of patients with suspected pulmonary embolism in the same clinical setting. The prevalence of pulmonary embolism in the 3 risk groups was similar, but this finding is a relatively weak test of external validity because the study took place in the same setting. The new prediction rule advances the field because it uses only history and physical examination to estimate pretest probability. Will it become an international standard? That verdict will depend on validation studies by other investigators who compare its predictions to another rule (13, 14). If 2 rules place every patient in the same probability category, they are both robust. Most people will use the simpler rule if both perform equally well. The ultimate validation study would be a management trial in which physicians use the rule to categorize the patient as having a high, intermediate, or low probability of pulmonary embolism and then consistently use a testing and treatment algorithm that is specific to each probability category. Ideally, each testing algorithm could lower the probability of pulmonary embolism to below the treatment threshold. Many management studies have documented the safety of managing low-risk patients without anticoagulant therapy after following a strategy to lower the probability of pulmonary embolism (15). A management trial for patients with an intermediate probability of pulmonary embolism might consist of a standard algorithm for using an imaging test to identify low-risk patients, for managing them without anticoagulant therapy, and for monitoring carefully for pulmonary embolism. Reilly and Evans (9) propose that we do an impact analysis on prediction rules. They propose a formal validation study of using the prediction rule to decide clinical management. An impact analysis is not a new concept. Management trials, such as those done in patients at low risk for pulmonary embolism, are impact analyses. Reilly and Evans have thought systematically about impact analyses, and they propose safety and efficiency as the outcome measures. They define safety as the proportion of people who have the target condition (for example, pulmonary embolism) and who receive the appropriate intervention (for exam