Diagnostic accuracy and screening statistics — branching viva

Opening (candidate)

I would treat this as a structured appraisal rather than a result handover. First I would judge the study's validity with QUADAS-2 before trusting the headline sensitivity and specificity, then I would ask whether the quoted positive predictive value applies to the community population, and finally I would work out how the test's likelihood ratio moves the probability of disease in an individual child. [9] [4]

Branch A — Validity before results

Examiner: The study recruited from a tertiary referral service. What bias does that introduce, and why does it matter? [10]

Candidate: A tertiary referral population produces spectrum bias, because the cases are more severe and the controls more clearly well than the children in whom the test will actually be used. That inflates both sensitivity and specificity, so the 95 and 93 percent are likely overestimates. Lijmer and colleagues showed empirically that case-control and extreme-spectrum designs overestimate accuracy compared with consecutive-series designs, so I would treat these numbers as optimistic until a consecutive-series estimate confirms them. [10] [9]

Branch B — The predictive value in the community

Examiner: The abstract reports a positive predictive value of 80 percent. Will that hold in the community? [4]

Candidate: Almost certainly not. The positive predictive value is a function of prevalence, not a property of the test. In the tertiary service the pre-test probability was high, so 80 percent of positives were true; in the community, where the condition is rarer, the same test's positives will be mostly false, and the predictive value will fall sharply even though the sensitivity and specificity are unchanged. Adopting the test in the community would generate a flood of false positives, unnecessary follow-up, and family anxiety. [4] [10]

Branch C — Likelihood ratios and the individual child

Examiner: How would you use this test for a single child? [5]

Candidate: I would estimate the child's pre-test probability from the history, examination and setting, then apply the test because its likelihood ratio crosses my decision threshold. With a sensitivity of 0.95 and specificity of 0.93, the LR+ is 0.95 divided by 0.07, about 13.6, a large shift, and the LR− is 0.05 divided by 0.93, about 0.054, also large. I would convert the pre-test probability to odds, multiply by the relevant likelihood ratio, and read the post-test probability off the Fagan nomogram, then share that with the family. [5] [7]

Branch D — The screening analogy and its biases

Examiner: The team now wants to screen asymptomatic children with this test. What biases must you exclude before accepting any benefit? [13]

Candidate: I would hold the programme against the Wilson and Jungner principles first, then exclude the screening biases. Lead-time bias finds disease earlier without changing the moment of death, length-time bias finds the slower indolent cases, overdiagnosis detects disease that would never become clinically apparent, and volunteer bias ensures the children who accept screening are healthier than those who decline. I would demand a fall in cause-specific mortality, not just survival from diagnosis, ideally from a randomised design that controls for these biases, before endorsing the programme. [13]

Branch E — Failure mode

Examiner: A colleague tells the family the test is 'virtually diagnostic'. [7]

Candidate: I would not contradict my colleague harshly in front of the family, but I would not leave the overclaim uncorrected. I would acknowledge the reassurance, introduce the balanced picture — the likelihood ratio generates a large but not certain shift in probability, and the post-test probability still needs to be read against the treatment threshold — and align the team afterwards so the family hears one defensible message. [7] [4]

Close

Confirm understanding with teach-back, leave a written summary of the appraised accuracy, the likelihood ratio, and the post-test probability, name the next contact, and document the pre-test probability, the test, its likelihood ratio, the post-test probability, and the shared decision. [5] [9]