Curriculum · Public Health and Communicable Diseases
Diagnostic test performance
What it is
Everything here comes off one 2 by 2 table, so hold the table first. Disease present and test positive is a, the true positive. Disease absent and test positive is b, the false positive. Disease present and test negative is c, the false negative. Disease absent and test negative is d, the true negative. Sensitivity is the proportion of people with the disease who have a positive test result: sensitivity = a/(a+c), that is true positives over true positives plus false negatives. Specificity is the proportion of people without the disease who have a negative test result: specificity = d/(b+d), that is true negatives over true negatives plus false positives. The short rule written beside them: specificity rules in the disease, sensitivity rules out the disease. The two that clinicians use run the other way round. Positive predictive value is the proportion of people with a positive test who have the disease, PPV = a/(a+b), and it says how likely it is that the patient has the disease. Negative predictive value is the proportion of people with a negative test who do not have the disease, NPV = d/(c+d). Both depend on the accuracy (likelihood ratio) of the test and on the prevalence, the pre-test probability, of the disease; prevalence is the patients who have the disease over the whole population, and predictive values are affected by it. Where the cut off is set decides both numbers: raise it and specificity improves, lower it and sensitivity improves. An ideal marker would separate the healthy population from the disease population completely, with no false positive cases, which means 100 percent specificity, and no false negative results, which means 100 % sensitivity. In the real world there is always an overlap between the disease population and the healthy population, so markers are not 100 % sensitive and not 100% specific, and the only lever left is choosing the cut off. Screening men above the age of 40 for prostate cancer shows the trade. A low cut off of 4 picks up all disease patients, 100 % sensitivity, but specificity will be 60 %, and the other 40 % who have levels above 4 are false positive because they are healthy. Moving the cutoff to 10 gives a specificity almost 80 % and a sensitivity near 80 %, close enough to each other to be used in screening; the healthy people who overlap above that cutoff are false positives and may have prostatitis, while those below it are false negatives from sampling error, small tumor sample or early disease. A cutoff of 20 gives specificity 100 % with zero false positives, but sensitivity around 50 %, so about half the patients with tumour are missed and the test cannot be used as screening. Higher cutoffs of tumour markers give better specificity; lower cutoffs give better sensitivity.
How it is diagnosed
Likelihood ratios are a direct link from pre-test probabilities to post-test probabilities, and are used in situations with more than 2 test outcomes. LR+ = sensitivity/(1-specificity), the likelihood of true positives over the likelihood of false positives. LR- = (1-sensitivity)/specificity, the likelihood of false negatives over the likelihood of true negatives. An LR of 1 has no diagnostic value. An LR >10 or <0.1 produces large changes in pre-test probability; an LR of 5 to 10 or 0.1 to 0.2 produces moderate changes; an LR of 1 to 2 or 0.5 to 1 produces small changes. Real tests show the trade. The Lachman test has a sensitivity of 68% for partial ruptures and 96% for complete ruptures, the anterior drawer test a sensitivity of 38% and a specificity of 81%. Light's criteria, which separate an exudate from a transudate, are three: 1. pleural protein over serum protein above 0.5, 2. pleural LDH over serum LDH above 0.6, 3. pleural LDH above 2/3 of the upper limit of normal for serum LDH; they are 98% sensitive and 83% specific to detect exudate. For Helicobacter pylori the whole accuracy table is worth reading down, given as sensitivity, specificity, PV+ and PV- as percentages, where the predictive values assume a prevalence of 33%:
- Urea breath test, carbon 13, noninvasive: 97, 100, 99, 1.5. Used for initial diagnosis and test of cure; expensive and inconvenient, and the patient must fast for six hours.
- Stool monoclonal antigen test by enzyme immunoassay: 92, 94, 89, 3.9. More expensive than immunochromatography.
- Stool monoclonal antigen test by immunochromatography: 69 to 87, 87 to 93, 72 to 85, 6.6 to 15. May be used in the office for rapid diagnosis; varying reliability.
- Antibody tests: 76 to 84, 79 to 80, 64 to 67, 9 to 13. Lower cost and easily available; PV+ depends on prevalence and it is not useful as a test of cure.
- Rapid urease test, invasive: 95, 100, 98.9, 2.4. Rapid and inexpensive; sensitivity is low in treated patients.
- Histology: 94, 99, 97, 3. Expensive.
- Culture: sensitivity not reported, specificity 100, predictive values not reported. Allows for susceptibility testing; not widely available and expensive.
- Polymerase chain reaction: sensitivity, specificity and both predictive values not reported. Allows for susceptibility testing; not standardized and not widely available.
A second page in our own material gives different numbers for the same two tests - culture 100% specific with a sensitivity of 70-92%, and the 13C-urea breath test 95% sensitive and 95-99% specific - so the two pages disagree on the numbers and are best read as two separate reports rather than one. The ROC curve plots sensitivity against specificity, and to be accurate the x axis should be (1 - specificity). The area under the graph is what rates the test: a perfect test = 1, an excellent test = 0.99, an acceptable test = 0.7-0.8, a poorly performing test = 0.5-0.6, and a useless test is below 0.5, because it means that 50 percent of the time you are right and 50 percent of the time you are wrong. A marker with a sensitivity of 50 and a specificity of 50 gives no discrimination. A worked example shows the steps. Cystic fibrosis with a prevalence of 40% in clinic: of 100 pts, 40 have CF and 60 are healthy, and sweat chloride has a specificity 90% and a sensitivity 80%. True positives = sensitivity x prevalence = (80 x 40)/100 = 32 %. The disease-free are 100 - 40 = 60 %, and true negatives = specificity x 60% = (90 x 60)/100 = 54 %, so 60 = 54 + FP and false positives = 6%. Positive predictive value = TP/(TP+FP) = 32/(32+6) = 84 %. Among the diseased, 40 = TP + FN = 32 + FN, so false negatives = 8, and NPV = TN/(TN+FN) = 54/(54+8) = 87%.
The cough stress test is worth following through to the end, because one test carries every part of how a test behaves in turn.
Its standing first: if stress incontinence is suspected, the cough stress test is the most reliable clinical assessment for confirming the diagnosis. When compared with more sophisticated multichannel urodynamic studies, it demonstrates good sensitivity and specificity for stress incontinence, although it requires further confirmatory urodynamic evaluation if the results are inconclusive.
Then how it is done, because the technique is what makes the result mean anything. With a full bladder, although not to the point of abrupt urination, the patient should be in the lithotomy position; women should separate the labia; the patient should relax the pelvic muscles and forcibly cough once. If the test is initially performed supine and no leakage is observed, the test should be repeated in the standing position, with a pad, or with the legs shoulder-width apart over a cloth or paper sheet on the floor to see the leakage. If urine leaks with the onset of the cough and terminates with its cessation, the test is positive for stress incontinence.
Then what a negative means. A negative test shows no leak or a delayed leak by five to 15 seconds, and rules out most cases of stress incontinence. But the delayed leak is not a plain negative result: a delayed leak may suggest a bladder spasm triggered by the cough, and not a weakness of the sphincter, which indicates possible urge incontinence. So the same test, read for timing rather than for yes or no, changes the diagnosis instead of excluding one.