Diagnostic Accuracy

Diagnostic accuracy refers to the ability of a test to correctly identify the presence or absence of a target condition in comparison to a reference standard (often called the gold standard). It is quantified through a set of interrelated metrics, including sensitivity (the proportion of people with the condition whom the test correctly identifies as positive), specificity (the proportion of people without the condition whom the test correctly identifies as negative), positive predictive value (the proportion of people with a positive result who truly have the condition), negative predictive value (the proportion of people with a negative result who truly do not have the condition), and likelihood ratios (which express how much a test result changes the probability of disease). Diagnostic accuracy studies follow reporting standards such as STARD (Standards for Reporting Diagnostic Accuracy Studies).

Accurate diagnosis is the foundation of appropriate clinical management. A test with poor sensitivity will miss patients who have the disease, delaying treatment and potentially causing harm. A test with poor specificity will produce false positives, leading to unnecessary further testing, treatment, and patient anxiety. Understanding diagnostic accuracy helps clinicians interpret test results in the context of patient characteristics and disease prevalence, choose between available tests, and communicate the meaning of results to patients. Guideline developers and health technology assessors rely on diagnostic accuracy data to determine whether a test should be recommended for clinical use and under what circumstances.

Diagnostic accuracy is assessed by comparing the results of an index test (the test being evaluated) against a reference standard (the best available method for determining whether the condition is truly present) in a sample of patients drawn from the population in which the test will be used clinically.

The core framework is the 2 by 2 table, which cross tabulates the test result (positive or negative) against the true disease status (present or absent according to the reference standard). This table yields four counts: true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). From these, the key metrics are calculated. Sensitivity equals TP divided by (TP plus FN). Specificity equals TN divided by (TN plus FP). Positive predictive value equals TP divided by (TP plus FP). Negative predictive value equals TN divided by (TN plus FN).

Sensitivity and specificity are each calculated within one group, the people with the condition or the people without it, so prevalence does not enter their formulas; they can still vary between settings with the spectrum of patients tested, such as how severe or advanced the condition is and which other conditions are present. Predictive values, in contrast, depend directly on the prevalence of the condition in the population being tested. A test with 95% sensitivity and 95% specificity will have a positive predictive value of 95% when the condition prevalence is 50% but only about 16% when the prevalence is 1%. This distinction is clinically critical and frequently misunderstood.

Likelihood ratios express how much a test result changes the pre test probability of disease. A positive likelihood ratio (sensitivity divided by 1 minus specificity) greater than 10 provides strong evidence of disease. A negative likelihood ratio (1 minus sensitivity divided by specificity) less than 0.1 provides strong evidence against disease. Likelihood ratios are useful because they can be applied regardless of prevalence using Bayes' theorem or Fagan's nomogram.

Receiver operating characteristic (ROC) curves plot sensitivity against (1 minus specificity) across all possible cut off values for continuous tests, and the area under the ROC curve (AUC) summarizes overall diagnostic performance. An AUC of 1.0 represents perfect discrimination, and an AUC of 0.5 represents performance no better than chance.

Diagnostic accuracy studies are appropriate when evaluating a new test, comparing the performance of two or more tests, or assessing whether a test performs differently in specific populations or clinical settings. Regulators often expect them: in the EU, for example, the In Vitro Diagnostic Medical Devices Regulation (IVDR) requires manufacturers to demonstrate the clinical performance of in vitro diagnostic devices, including diagnostic sensitivity and specificity where appropriate. Health technology assessments of diagnostic technologies also rely on diagnostic accuracy data. The STARD 2015 checklist of 30 items guides transparent reporting. Diagnostic accuracy studies are not appropriate for evaluating the clinical impact of testing (whether testing leads to better patient outcomes), which is best assessed in randomized trials of test and treat strategies.

A frequent error is evaluating a test in a population that differs from the one in which it will be used clinically, such as comparing clearly diseased patients with healthy volunteers rather than with patients presenting with similar symptoms who may or may not have the condition. This "case control" design inflates sensitivity and specificity because the distinction between disease and non disease is artificially clear. Another mistake is using a reference standard that is itself imperfect, which biases accuracy estimates (a problem called imperfect reference standard bias). Reporting only sensitivity and specificity without predictive values or likelihood ratios provides an incomplete picture that clinicians find hard to apply to individual patients. Finally, failing to report the STARD checklist items, such as the flow of patients through the study, the proportion who received both the index test and reference standard, and the handling of inconclusive results, makes it difficult to assess the study's internal validity.

A study evaluates a rapid antigen test for influenza in 1,200 adults presenting with influenza like illness during a single flu season. Using RT PCR as the reference standard, 420 of the adults have influenza and 780 do not. The test is positive in 302 of the 420 (118 false negatives) and negative in 764 of the 780 (16 false positives). This gives a sensitivity of 72% (95% CI: 67 to 76%), a specificity of 98% (95% CI: 97 to 99%), a positive likelihood ratio of 35, and a negative likelihood ratio of 0.29. At the observed prevalence of 35%, the positive predictive value is 95% (302 of 318 positive results) and the negative predictive value is 87% (764 of 882 negative results). The authors note that the sensitivity is lower than reported in studies that included only hospitalized patients with high viral loads, and recommend confirmatory testing for negative results in high risk patients during periods of high influenza activity.