Theory 07/Radiology & Medical Imaging /Clinical AI/3 min
Triage Moves Errors, It Doesn't Remove Them
A triage tool that catches what you would have caught changes nothing but your attention.
AI triage changes the distribution of radiologist errors rather than reducing their number, so deployment should be judged on which errors move, and cases where the model and reader agree are where new misses concentrate.
01The evidence
- Evaluation of medical AI systems has shown that performance varies across clinical scenarios and that calibration, not only accuracy, determines whether a tool is trustworthy in deployment.Trustworthy medical MLLMs
- Out-of-distribution studies show that models degrade unevenly across acquisition conditions, which means their errors are correlated with particular kinds of case rather than spread at random.MEDI 2026
02The tension
AI triage is evaluated by sensitivity and specificity, which summarise the error rate. A tool can improve the rate and still make care worse if it moves attention away from the cases where human judgement was the safeguard.
03The theory
A reader with a triage tool allocates attention differently. Cases flagged as normal are processed faster and with less scrutiny, so errors migrate from the flagged category to the unflagged one. The total may fall while the distribution changes.
If errors migrate, then the relevant evaluation is not the error rate but the error map: which findings are now missed, and whether those are the ones with the highest consequence. A tool that reduces misses of common findings while increasing misses of rare ones can be net harmful.
The prediction is that deployment should show a characteristic redistribution, with more missed rare findings and fewer missed common ones, and that the net effect should depend on the case mix.
The practical consequence is that triage tools should be evaluated with stratified error analysis and deployed with safeguards on the unflagged stream, not just on the flagged one.
04What this is not
Aggregate metrics treat errors as interchangeable. This treats them as a distribution that shifts under deployment, which predicts harm from a tool that improves the average.
05Predictions
- P1Deployment should shift misses toward rare findings in unflagged cases.
- P2Net benefit should depend on case mix rather than on sensitivity alone.
- P3Stratified error analysis should reveal effects invisible in aggregate metrics.
06What would refute it
Errors falling uniformly across strata would leave aggregate evaluation adequate.
07What it would change
It would change AI evaluation standards and force stratified error reporting.
08Evidence base
- Trustworthy medical multimodal models — calibration and scenario variationhttps://zenodo.org/records/22258831/files/Trustworthy_Medical_Multimodal_Large_Language_Models_Arxiv.pdf
- Out-of-distribution generalisation under manufacturer domain shifts — MEDIhttps://dlnext.acm.org/doi/10.1007/978-3-032-19865-5_6