Diagnostic information map
Every point is a sensitivity and specificity. Colour shows the share of diagnostic uncertainty a test there would resolve at a 20% starting probability. Move the probability and the entire map updates. White circles are the 273 published pooled estimates; switch on the programme's other evidence below. Dashed lines hold Youden's J fixed.
Expected to resolve
of the uncertainty about the diagnosis before testing
–
–
Sensitivity plus specificity is below 100%, so this test is oriented below chance: a positive result lowers the probability of disease. The information is the same as for the flipped test.
–
The diagnostic information library
The published pooled reference stays scientifically separate. Optional layers let you explore individual evidence without changing the manuscript's percentile scale. Search also covers 42 condition context profiles.
Clinical category comes from the clinical-feature source. Specialty and purpose come from clinical-calculator records; missing specialties are not guessed.
Loading records…
Companion to “The information in diagnostic tests”
At a stated starting probability, every diagnostic test has an expected information value: the share of the uncertainty about the diagnosis it is expected to resolve before its result is known. Describe any test here, a bedside rule, a symptom or sign, a laboratory threshold, an imaging study, a pair of likelihood ratios, or a result with several levels, and see that value, what each result would do, and where the test sits on the map of 273 published pooled estimates from 210 diagnostic reviews.
–
–
One row per result level at this starting probability.
| Result | Happens | Probability after | LR | Weight of evidence | Uncertainty after (change) | Share of information |
|---|
Share of starting uncertainty resolved from 1% to 99%. Grey bands summarise all 273 published pooled estimates; coloured lines are the tests you described.
–
–
| Test and condition | Source | Studies | Sens. | Spec. | J | Resolved |
|---|
| Quantity | Your test | Second test |
|---|
A yes-or-no question whose answer is a coin toss holds one bit of uncertainty. A question with a very likely or very unlikely answer holds less. At a 20% probability of disease the starting uncertainty is 0.722 bits; at 5% it is 0.286 bits.
H(p) = −p·log₂(p) − (1 − p)·log₂(1 − p)Each result leaves some uncertainty behind. Average that remaining uncertainty over how often each result occurs, subtract it from the starting uncertainty, and you have the expected diagnostic information in bits. Dividing by the starting uncertainty gives the share resolved (in statistics, the uncertainty coefficient). With several result levels the same average runs over every level.
I = H(p) − Σ P(result)·H(p after that result) · share = I ÷ H(p)A result moves the probability of disease (the likelihood ratio, expressed as a signed weight of evidence in bits, log₂ LR). Uncertainty after a result can be larger or smaller than before: a weak positive can move a rare diagnosis toward 50:50 and widen it. The size of the probability update itself is never negative; weighting it by how often each result occurs gives the expected information back.
273 pooled estimates from 210 diagnostic reviews (Cochrane, PubMed Central, OSF and Zenodo sources), each fitted with the same bivariate random-effects model with a continuity correction, equally weighted. Starting probabilities are analytical conditions, not prevalence estimates. The map describes published diagnostic evidence, not all of medicine.
Information is not clinical value. Choosing an action also needs the consequences of a missed or false-alarm result and the patient's preferences, which this page does not model.