Christopher Ball looks at how to apply the evidence to your patient, focusing on diagnosing conditions

Doctors spend their careers making guesses--they are constantly predicting the probability of certain diseases or outcomes in every patient they see. Most of the time this happens subconsciously, but if quizzed any doctor should be able to come up with rough figures. Doctors are often good at ruling in conditions--for example, angina or delirium--but much worse at ruling them out--for example, myocardial infarction or delirium. Tests provide information that can help alter these predictions, but by how much?

This article considers the types of numbers you will find in evidence relating to diagnosis and how you can adapt this information to your patient.

Pre-test probability of pulmonary embolism 9.5% (95% confidence interval 7.5% to 11.3%)

Diagnostic test Likelihood ratio of test Post-test probability Likelihood ratio of test Post-test probability

being positive (95% CI) being negative (95% CI)

Diagnosis other than 2.0 (1.9­2.2) 17% 0.023 (0.0032­0.16) 0.2%
low risk and d-dimer

Clinical scenario

A 36 year old woman arrives in the emergency department complaining of feeling suddenly breathless and having a sharp pain in her side. The house officer learns that she has no history of venous thromboembolism and is not taking the oral contraceptive pill. She is otherwise fit and well, and has had no episodes of haemoptysis. The examination is unremarkable. Using a clinical diagnosis rule, the house officer decides that she is at low risk for a pulmonary embolism.¹

The house officer recalls that a blood test called d-dimer can help in the diagnosis of deep vein thrombosis and wonders whether this test might help rule out a pulmonary embolism and avoid a ventilation-perfusion scan.

Should he order one, and what would a normal result mean?

Clinical question--In patients with a suspected pulmonary embolism (patient) can a normal d-dimer result (intervention) rule out a pulmonary embolism (outcome)?

Resource selected--Evidence-based on-call.

Search strategy--"d-dimer" and "embolism"

Result--The first structured abstract found gave the answer.

Answer--Patients with a normal d-dimer result and considered at low risk for a pulmonary embolism using a clinical diagnosis rule were unlikely to have a pulmonary embolism.²

Time to answer--12 seconds.

Reading through the structured abstract you find the information in the table.

A powerful way to think about diagnosis is to use probability to determine your patient's risk of a disease. Imagine your patient's probability of having a disease on a sliding scale from 0% to 100%. Tests can change this probability by pushing it up or down the scale. Your aim is to diagnose the disease (100% chance of having the disease) or rule it out (0% chance of having the disease). Even with multiple tests you are unlikely ever to exclude with confidence a disease and will only be able to confirm many if you have histology or autopsy results. Using probabilities can help you make decisions about when you might want to stop investigating further for that condition, or when you should start treatment.

These thresholds vary between diseases. For example, the threshold for giving antibiotics to somebody with suspected meningitis is low, the consequences of not treating patients are severe, and the potential risks from using antibiotics are low. You might accept only a 60% chance of meningitis before initiating treatment. Before starting chemotherapy for cancer, however, you should be confident (99% sure or more) that your patient really has cancer, since the treatment has severe adverse effects.

There are a number of ways to measure the effect tests can have on the diagnostic process. Most doctors are familiar with sensitivity and specificity as the way to measure the usefulness of tests, but these measurements have some limitations. Both provide information on how many patients with or without the disease end up having a specific test outcome, and hence are great for comparing different tests. This is at odds with clinical practice; doctors are not interested in determining test quality in patients who are known to have disease. They want to know how definitive results (abnormal or normal) help them make diagnoses in patients with suspected disease. So sensitivity and specificity present the information the wrong way around to clinicians--no wonder most people find them confusing.

The abstract reports the pre-test probability and the post-test probability. The pre-test probability refers to the risk of a pulmonary embolism in a patient attending an emergency department with chest pain and breathlessness (10%). The post-test probability reports the risk in these patients after a positive or negative d-dimer test (17% and 0.2% respectively). As long as you think your patient's risk is similar to that of the study patients, you can apply this information immediately.

Likelihood ratios are a more effective way to determine how useful a test is. Unlike sensitivity and specificity, likelihood ratios allow you to combine your clinical prediction of your patient's disease with a test result to work out your patient's risk afterwards. Likelihood ratios provide a measure of how much a test pushes your patient's probability of disease along the probability scale towards making the diagnosis (see studentbmj.com for more information). A positive likelihood ratio pushes the probability to the right (makes the disease more likely), and a negative likelihood ratio pushes probability to the left (makes it less likely). Likelihood ratios of 10 or more indicate a very useful test for diagnosing a condition, whereas likelihood ratios of 0.1 or less indicate a useful test for ruling out a condition. For example in a patient with chest pain, any new ST elevation on an electrocardiogram has a likelihood ratio of 11.2 making a myocardial infarction much more likely.³

If you are interested in learning some more about how to combine probabilities and likelihood ratios to customise diagnostic information to your patient, look at the resources in the further reading section. Using a simple diagram called a normogram, you can quickly learn how to combine test results to work out whether your patient's risk of a disease has crossed the "treat" or "stop investigating" thresholds.

Clinical diagnosis rules

Clinical diagnosis rules typically allow clinicians to combine multiple clinical features, blood tests, or imaging results using a point scoring system, and then rank patients for their probability of having a disease. A number make diagnosing difficult problems such as deep vein thrombosis much simpler and reduce the number of cases missed.⁴ Before using a clinical diagnosis rule though, you need to make sure that it has been tested out and validated in multiple settings and is not just an experimental tool. Keep a look out for them though--they can make your clinical work much faster and easier.

Resolution of clinical scenario

The house officer orders a d-dimer which comes back as normal. Using the table from the abstract he finds his patient's risk of having a pulmonary embolism to be 0.2%. In other words, the patient has a one in 500 chance of having a pulmonary embolism--low enough to discharge her if she agrees. After discussing this with your patient, she suddenly feels a lot better and decides to go home.