The below article has been reproduced from a BMJ series to help medical students make the leap to budding doctors. In this article, Richard Beasley and co-authors describe a systematic approach to the use and interpretation of diagnostic tests.
Medical students are often taught that diagnostic tests and investigative procedures give clear information on which a diagnosis can be made or refuted. In addition, the medical literature and lay media have created the impression that, with advances in technology, even the most obscure diagnosis can be made by the use of investigative tests. However, in reality, the situation is different. Most diagnoses in clinical medicine are made from the history or examination, or both, and the concomitant use of diagnostic tests and their interpretation is often suboptimal.
At the beginning of their careers in hospital-based practice, junior doctors often resort to a “fishing expedition” when admitting patients on the acute take, requesting multiple tests in an attempt to make a diagnosis. This can result in a cascade of unnecessary tests and inappropriate investigations, which may cause confusion and cause additional inconvenience to the patients. More specialised or invasive tests with which the junior doctor is unfamiliar may also lead to difficulties in interpretation.
To overcome these difficulties, junior doctors need to develop a logical clinical approach to the use and interpretation of diagnostic tests. This requires an understanding of:
- The main determinants of the properties of diagnostic tests
- The basic concepts of Bayesian theory
- Decision analysis
- Their integrated application to clinical medicine.1,2
Properties of diagnostic tests
To evaluate the usefulness of a clinical test, it is necessary to understand its properties. These are measures of the test’s ability to predict disease accurately. The main properties to be considered are the positive predictive value, negative predictive value, sensitivity, and specificity. The positive predictive value is the percentage of people with a positive test result that have the disease. Conversely, the negative predictive value is the percentage of people with a negative test that do not have the disease. These values depend on the clinical situation in which the test is being undertaken (see Example 1). The sensitivity of the test is the proportion of people with the disease who have a true positive test result. Specificity is the proportion of people who do not have the disease who have a true negative test result. The properties of a diagnostic test in terms of sensitivity and specificity are traditionally shown in figure 1.
Fig 1 The properties of plasma D-dimer measurements in the diagnosis of pulmonary embolism. The numbers in brackets represent different D-dimer level cut-off values4
An inherent trade-off exists between sensitivity and specificity; a cut-off point is usually chosen which is considered most appropriate to guide decision making (see Example 2). Different cut-off values (to define normal or abnormal) may be used depending on whether one attempts to confirm a disease through a positive result (high specificity) or to rule out the disease via a negative result (high sensitivity).
Bayesian theory
Bayes’s theory is commonly applied in medicine to predict the probability of a certain diagnosis. Simply put, Bayes’s theory proposes that to determine the probability of a specific diagnosis it is necessary to combine the information that is already known about the patient with the results of the test. In clinical medicine this means that the results of any test should not be considered in isolation but must be interpreted in conjunction with all other available information concerning the patient. In practice, to apply Bayes’s theorem to a diagnostic test we first need to determine the pre-test probability—that is, the probability of the patient having the disease in question before the test is done. The pre-test probability is determined from the pre-existing patient information and knowledge of the disorder under consideration. The test is then under- taken, from which a post-test probability is calculated—that is, the probability of a patient having the disease in question once the test result is known. The post-test probability is based on the pre-test probability and results of the test for which the properties are known. Figure 2 shows the use of Bayesian analysis in the interpretation of exercise electrocardiogram testing (see Example 3).
Fig 2 The probability of coronary artery disease is determined by the pre-test clinical probability and the results of the electrocardiogram exercise test, for which the properties are known. The same test result yields different probabilities in patients with different pre-test probabilities5
A number of observations can be made.
- Tests may have variable predictive values depending on the cut-off values used (for example, size of ST segment depression).
- The same test result may lead to an entirely different post-test probability of a diagnosis, depending on the pre-test probability—thus, the test results must be considered in conjunction with the clinical situation.
- In a situation of high or low pre-test probability, a test is unlikely to be informative—that is, if there is either no real clinical suspicion of the diagnosis or if the diagnosis has essentially already been confirmed, there is usually no justification for undertaking a further diagnostic test, and to do so is unlikely to be helpful.
- Junior doctors are not expected to assign a numerical pre-test probability to each patient. Rather, they are expected to categorise the pre- test probability as low, moderate, or high, based on a careful assessment of the clinical presentation and the underlying risk factors. This skill will improve with experience.
- It is not possible to interpret results of tests properly without knowing their properties. Even in situations in which it may be difficult to calculate a pre-test probability or know the precise properties of the tests undertaken, it is still worthwhile attempting to follow the Bayesian approach.
Decision analysis
Once the post-test probability has been determined, the next step is to decide whether the probability is sufficiently high to confirm the diagnosis, sufficiently low to exclude the diagnosis, or intermediate, in which case a further diagnostic test is required. This entails the process of decision analysis, in which the preferred course of action is calculated mathematically. The preferred course of action is the one which, when all possible outcomes are considered, yields the highest value calculated by the probability multiplied by the utility of each outcome.
In practical terms, this means determination of the test threshold (that is, the level above which the probability is not low enough to rule out the diagnosis and above which a further test is required) and the treatment threshold (that is, the level above which the probability is sufficiently high that treatment should be started and further tests are not required).
It is evident that:
- The test and treatment thresholds depend on a balance between the severity of the untreated disease, the efficacy of treatment, the risks associated with both the treatment and invasive tests, and the properties of the next test that would be performed. While these factors are intuitively considered when clinicians assess the requirement for further investigation or treatment, their integration is often difficult in complex medical conditions and in patients with comorbidity
- An understanding of how decision analysis applies to different clinical scenarios will enable junior doctors to exercise good judgment in the management of their patients.
As shown in Example 4/figure 3, the use of decision analysis enables the post-test probability to be used in a meaningful way as the basis for deciding the best approach for further investigation and treatment.
Fig 3 Relationship between pre-test and post-test probability for (a) lung scan (b) D-dimer (c) ultrasound scan through graphic representation of Bayes’s theorem. The shaded area in each diagram represents the range of post-test probabilities for which pulmonary angiography should be performed, if the test under consideration is the only available test, or the last in a sequence of tests. This range of post-test probabilities has been determined by decision analysis from which it has been calculated that the test threshold is 6% and the treatment threshold is 44%
Integration in algorithm
The properties of diagnostic tests, their determinants and application through Bayesian theory, and decision analysis can be brought together in the format of an algorithm. An algorithm presents a practical diagnostic and treatment pathway that can be followed to assist in patient management. Such algorithms should be tested before they are recommended for general use (figure 4).8
Fig 4 Validated algorithm for the outpatient investigation and management of patients with suspected deep vein thrombosis8
| Example 13 In a patient with acute coronary syndrome a raised troponin T suggests myocardial infarction (that is, it is likely to be a true positive result). In contrast, a raised troponin T in an asymptomatic patient with chronic renal failure does not suggest a myocardial infarction (that is, it is likely to be a false positive result). |
| Example 24 In a patient presenting with symptoms of pulmonary embolism, a D-dimer cut-off value of 500[micro ]g/l has high sensitivity (a value <500 [micro ]g/l effectively rules out pulmonary embolism) but low specificity (only half of patients with a value >500 [micro ]g/l have a pulmonary embolism). Using higher cut-off values improves the specificity but at the expense of sensitivity, to the extent that the D-dimer test loses its clinical utility. |
| Example 35 Results of an exercise test in a 50-year-old male smoker with exertional angina shows 2.3mm ST depression in contiguous leads, leading to an increase in the clinical probability of coronary artery disease from around 90% to over 99% and showing the requirement for coronary angiography. A 50-year-old male company director who is asymptomatic and has no major risk factors has an exercise test as part of a “well health” programme. An identical 2.3mm ST depression gives an increase in probability from around 5% to 53%. |
| Example 46,7 A patient with suspected pulmonary embolism (moderate clinical probability estimated at 50%) is investigated initially with a lung scan, which is low probability, yielding a post-test probability of 30%. This becomes the pre-test probability for the next test (D-dimer), which is positive, resulting in a post-test probability of 42%. This becomes the pre-test probability for the next test (ultrasound examination), which is negative, resulting in a post-test probability of 22%. Through decision analysis it is evident that the preferred approach is now to undertake pulmonary angiography as the probability is sufficiently low not to warrant treatment and sufficiently high not to be able to rule out a pulmonary embolism. |
Summary
Junior doctors need to understand the basic principles presented in this paper to ensure that they use and interpret investigations appropriately. These principles will become increasingly relevant to their practice as technology develops and with increasing reliance on investigations. In this burgeoning investigative culture a logical approach to the interpretation of tests needs to be mastered.
Original authors:
Richard Beasley, General physician and professor of medicine
Sarah Aldington, Senior research fellow.
Geoffrey Robinson, General physician and chief medical officer
Medical Research Institute of New Zealand and Capital Coast Health, Wellington, New Zealand.
We gratefully acknowledge the BMJ for permission to reproduce this article.
