A Second Opinion, Not a Replacement

In actual clinical use, a diagnosis-assist model sits alongside a radiologist or physician, not in place of one. It reviews the same scan or the same chart, surfaces cases it thinks deserve a closer look, and the clinician makes the actual call — the model is there to catch what a tired reviewer on their fortieth case of the day might miss, not to remove the reviewer.

That arrangement is not a temporary stepping stone toward full automation — it is the actual deployed shape almost everywhere one of these tools is in clinical use today. A model that is wrong in a way a busy clinician would catch is an acceptable cost of doing business; a model that is wrong with nobody checking is a different kind of product entirely, and regulators have so far authorised that second kind only for a handful of narrow, tightly scoped screening tasks.

Underneath the alert a doctor sees, the model itself never outputs “yes” or “no.” It outputs a number between 0 and 1 — its estimate of how likely this case is to be positive, based on everything similar it has seen before. Something has to turn that number into an actual flag, and that something is a threshold, chosen in advance by the people who built and deployed the system.

Two systems can share the exact same underlying number and behave completely differently once a hospital deploys them, because the number itself commits to nothing. A well-calibrated model that says 0.61 has not said “probably positive” or “probably negative” — it has said “more likely than not, by not very much,” and it is entirely up to the people deploying it to decide how much weight that deserves. (Many models' raw scores aren't even calibrated probabilities, which is one more reason the score alone commits to nothing.)

Move the slider below across forty simulated cases and watch what happens. Lower the threshold and more true cases get caught — but more healthy cases get flagged too, each one a false alarm that costs a clinician's time and a patient's anxiety on a follow-up that finds nothing. Raise it and false alarms drop — at the direct cost of missing real cases that scored just under the line.

Everything above describes properties of the test, not properties of your result. Sensitivity is the share of people who actually have the condition that the model correctly flags. Specificity is the share of people who don't have it that the model correctly clears. Both are honest, useful numbers — and neither one answers the question a patient is actually holding when a result comes back positive: given that I was flagged, how likely am I to actually have this?

That number has a name — positive predictive value, or PPV — and it depends on something sensitivity and specificity don't capture at all: how common the condition actually is in the population being tested. A test can be excellent by every measure a manufacturer would put on a spec sheet and still produce a flag that is wrong far more often than it is right, once the condition it is looking for is rare enough.

Work it out with real numbers rather than taking that on faith. Screen 10,000 people for a condition that actually affects 1% of them, using a test that is 90% sensitive and 90% specific — numbers a manufacturer would happily print on a box.

100 people actually have the condition. 9,900 don't. Sensitivity 90%              Specificity 90%  90 true positives            990 false positives  10 false negatives         8,910 true negatives Flagged positive: 90 + 990 = 1,080Of those, 90 actually have the condition. Positive predictive value: 90 / 1,080 ≈ 8%

A test that is right nine times out of ten in both directions still produces a positive flag that is wrong about eleven times out of twelve, because at 1% prevalence the healthy population is so much larger than the sick one that even a small false-positive rate applied to it swamps the true positives. That is not the test failing. That is arithmetic — sensitivity and specificity stay the same 90% and 90% no matter how rare the condition is; prevalence is the variable actually doing the damage.

This is why a screening programme aimed at an entire healthy population and a diagnostic test ordered for someone already showing symptoms are different products wearing the same underlying model, and why moving the threshold in the section above changes sensitivity and specificity without ever touching the deeper problem. A model tuned to catch more true cases at a low prevalence still returns mostly false alarms — it just returns a different mix of them.