Clinicians should take the sensitivity and specificity of a diagnostic test into account when selecting a test. A sensitive test (i.e., one that is usually positive in the presence of disease) should be chosen when there is an important penalty for missing a disease. This would be so, for example, when there is reason to suspect a dangerous but treatable condition, such as tuberculosis, syphilis, or Hodgkin lymphoma, or in a patient suspected of having DVT. Sensitive tests are also helpful during the early stages of a diagnostic workup, when several diagnoses are being considered, to reduce the number of possibilities. Diagnostic tests are used in these situations to rule out diseases with a negative result of a highly sensitive test (as in the DVT example). As another example, one might choose the highly sensitive HIV antibody test early in the evaluation of lung infiltrates and weight loss to rule out an AIDS-related infection. In summary, a highly sensitive test is most helpful to the clinician when the test result is negative.

Specific tests are useful to confirm (or “rule in”) a diagnosis that has been suggested by other data. This is because a highly specific test is rarely positive in the absence of disease; it gives few false-positive results. (Note that in the DVT example, the D-dimer test was not specific enough [60%] to initiate treatment after a positive test. All patients with positive results underwent compression ultrasonography, a much more specific test.) Highly specific tests are particularly needed when false-positive results can harm the patient physically, emotionally, or financially. Thus, before patients are subjected to cancer chemotherapy, with all its attendant risks, emotional trauma, and financial costs, tissue diagnosis (a highly specific test) is generally required. In summary, a highly specific test is most helpful when the test result is positive.

It is obviously desirable to have a test that is both highly sensitive and highly specific. Unfortunately, this is often not possible. Instead, whenever clinical data take on a range of values, there is a trade-off between the sensitivity and specificity for a given diagnostic test. In those situations, the location of a cutoff point, the point on the continuum between normal and abnormal, is an arbitrary decision. As a consequence, for any given test result expressed on a continuous scale, one characteristic, such as sensitivity, can be increased only at the expense of the other (e.g., specificity). Table 8.1 demonstrates this interrelationship for the use of B-type natriuretic peptide (BNP) levels in the diagnosis of congestive heart failure among patients presenting to emergency departments with acute dyspnea (6). If the cutoff level of the test were set too low (≥50 pg/mL), sensitivity is high (97%), but the trade-off is low specificity (62%), which would require many patients without congestive heart failure to undergo further testing for it. On the other hand, if the cutoff level were set too high (≥150 pg/mL), more patients with congestive heart failure would be missed. The authors suggested that an acceptable compromise would be a cutoff level of 100 pg/mL, which has a sensitivity of 90% and a specificity of 76%. There is no way, using a BNP test alone, that one can improve both the sensitivity and specificity of the test at the same time.

Another way to express the relationship between sensitivity and specificity for a given test is to construct a curve, called a receiver operator characteristic (ROC) curve. An ROC curve for the BNP levels in Table 8.1 is illustrated in Figure 8.4. It is constructed by plotting the true-positive rate (sensitivity) against the false-positive rate (1 — specificity) over a range of cutoff values. The values on the axes run from a probability of 0 to 1 (0% to 100%). Figure 8.4 illustrates visually the trade-off between sensitivity and specificity.

Tests that discriminate well crowd toward the upper left corner of the ROC curve; as the sensitivity is progressively increased (the cutoff point is lowered), there is little or no loss in specificity until very high levels of sensitivity are achieved. Tests that perform less well have curves that fall closer to the diagonal running from lower left to upper right. The diagonal shows the relationship between true-positive and false-positive rates for a useless test—one that gives no additional information to what was known before the test was performed, equivalent to making a diagnosis by flipping a coin.

The ROC curve shows how severe the trade-off between sensitivity and specificity is for a test and can be used to help decide where the best cutoff point should be. Generally, the best cutoff point is at or near the “shoulder” of the ROC curve, unless there are clinical reasons for minimizing either false negatives or false positives.

ROC curves are particularly valuable ways of comparing different tests for the same diagnosis. The overall accuracy of a test can be described as the area under the ROC curve; the larger the area, the better the test. Figure 8.5 compares the ROC curve for BNP to that of an older test, ventricular ejection fraction determined by electrocardiography (7). BNP is both more sensitive and more specific, with a larger area
under the curve (0.89) than that for ejection fraction (0.78). It is also easier and faster to obtain in an emergency setting—test characteristics important in clinical situations when quick results are needed.

Obviously, tests that are both sensitive and specific are highly sought after and can be of enormous value. However, practitioners must frequently work with tests that are not both highly sensitive and specific. In these instances, they must use other means of circumventing the trade-off between sensitivity and specificity. The most common way is to use the results of several tests together, as discussed later in this chapter.

Difficulties may arise when the patients used to describe the test’s properties are different from those to whom the test will be applied in clinical practice. Early reports often assess the test’s characteristics among people who are clearly diseased compared with people who are clearly not diseased, such as medical student volunteers. The test may be able to distinguish between these extremes very well but perform less well when differences are subtler. Also, patients with disease often differ in severity, stage, or duration of the disease, and a test’s sensitivity will tend to be higher in more severely affected patients.

Some people in whom disease is suspected may have other conditions that cause a positive test, thereby increasing the false-positive rate and decreasing specificity. In the example of guidelines for ovarian cancer evaluation, specificity was low for all cancer stages (60%). One reason for this is that levels of the cancer marker, CA-125, recommended by guidelines, are elevated by many diseases and conditions other than ovarian cancer. These extraneous conditions decreased the specificity and increased the false-positive rate of the guidelines. Low specificity of diagnostic and screening tests is a major problem for ovarian cancer and can lead to surgery on many women without cancer.

Disease spectrum and prevalence of disease are especially important when a test is used for screening instead of diagnosis (see Chapter 10 for a more detailed discussion of screening). In theory, the sensitivity and specificity of a test are independent of the prevalence of diseased individuals in the sample in which the test is being evaluated. (Work with Fig. 8.2 to confirm this for yourself.) In practice, however, several characteristics of patients, such as stage and severity of disease, may be related to both the sensitivity and the specificity of a test and to the prevalence because different kinds of patients are found in high-and low-prevalence situations. Screening for disease illustrates this point; screening involves the use of the test in an asymptomatic population in which the prevalence of the disease is generally low and the spectrum of disease favors earlier and less severe cases. In such situations, sensitivity tends to be lower and specificity higher than when the same test is applied to patients suspected of having the disease, more of whom have advanced disease.

The sensitivity and specificity of a test should be established independently of the means by which the true diagnosis is established. Otherwise, there could be a biased assessment of the test’s properties. As already pointed out, if the test is evaluated using data obtained during the course of a clinical evaluation of patients suspected of having the disease in question, a positive test may prompt the clinician to continue pursuing
the diagnosis, increasing the likelihood that the disease will be found. On the other hand, a negative test may cause the clinician to abandon further testing, making it more likely that the disease, if present, will be missed.

Therefore, when the sensitivity and specificity of a test are being assessed, the test result should not be part of the information used to establish the diagnosis. In studying DVT diagnosis by D-dimer assay (discussed earlier), the investigators made sure that the physicians performing the gold standard tests (ultrasonography and follow-up assessment) were unaware of the results of the D-dimer assays so that the results of the D-dimer assays could not influence (bias) the interpretation of ultrasonography (10).