Abstract
Abstract
Published artificial intelligence for keratoconus is unproven in fellow eyes that topography and tomography call normal, the eyes where surgical decisions are made, although it reports near-perfect accuracy for early disease. In a registered systematic review of 573 studies (PROSPERO CRD420261441197), none of 2,035 claim-by-stratum units from 212 adjudicated studies established discrimination in these eyes. Most units (90.0%) answered a different question, never constructed the target population or certified its normality; external validation did not test label generation, and five systematic reviews did not assess it. In a public cohort, a frozen score yielded an area under the receiver operating characteristic curve of 0.996 against a same-eye index label and 0.549 against a contralateral proxy; its inputs reconstructed the label-generating index (
r
= 0.991), so accuracy there measured agreement with the index. Anchor-gap evaluation measures accuracy in these eyes against the contralateral clinical diagnosis.