Abstract
Individual patient data meta-analyses using the raw data from primary diagnostic accuracy studies are taking hold in systematic reviews evaluating tests. Conventional reviews and meta-analyses that summarise study-level data on test accuracy (sensitivity and specificity) have several disadvantages. The most fundamental limitation of this approach is that it estimates the rates of test result-given disease (sensitivity is probability of positive test result-given disease is present; and specificity is probability of negative test result-given disease is absent). This may be addressed by summarising predictive values, but estimating accuracy for individual tests without consideration of other tests in the test chains that make up everyday diagnostic work-ups remain a problem. To inform clinical practice it is essential that test evaluation generates information about probability of disease given test results, and that it does so in view of the preceding contribution to diagnosis of other tests, for example, symptoms and signs. A multivariable (logistic regression) framework generates disease probabilities taking into account the important factors that play a role in diagnosis. Most primary accuracy studies lack statistical power to do this, particularly because of the small absolute number of disease events per test included in the diagnostic work. Synthesis using their raw data can overcome this problem, but meta-analysts will have limited success if there are difficulties in obtaining the large majority of valid studies, without 'missing' data on the tests relevant in clinical decision-making. Successful individual patient data meta-analyses create the opportunity to calculate directly and reliably disease probabilities corresponding with realistic chains of tests, thereby making outputs of reviews of test accuracy clinically applicable.
Ter Riet, Gerben