Answers are grounded in this paper and cite their sources.
Artificial intelligence (AI) models for keratoconus frequently report near-perfect discrimination, with pooled sensitivity and specificity approaching 98% when manifest keratoconus is separated from normal corneas. These figures are increasingly invoked to answer a different, more consequential question: whether an apparently non-ectatic cornea can safely undergo refractive surgery. We argue this is a task mismatch rather than a performance shortfall. Three targets are routinely conflated-manifest keratoconus, subclinical phenotype and future post-refractive ectasia-with the evidence strongest for the target furthest from the clinical question and weakest for the one closest. Underlying this is a single methodological issue: the absence of a robust reference standard, or ground truth, appropriate to the surgical decision. These estimates also derive largely from enriched case-control datasets, with subclinical-keratoconus case proportions of roughly 25% against an estimated 1% clinical prevalence. Applied to 1000 candidates at 1% prevalence, the pooled estimates correspond to approximately one missed subclinical case (0.6-1.6) and 45 false-positive referrals (25-80), both ranges derived from the reported confidence bounds-modelled diagnostic outcomes, not postoperative ectasia risk. We propose that ectasia-risk AI be evaluated as decision support for surgical eligibility, along a decision-anchored evidence chain spanning clinical validity, calibration and threshold-specific error, transportability, decision utility and post-deployment surveillance-complementing, rather than replacing, TRIPOD+AI, PROBAST+AI, STARD-AI, DECIDE-AI and FUTURE-AI. For refractive surgery, trust should not be inferred from a high area under the curve; it should be earned through evidence that follows the decision from target definition to real-world surveillance.
No takes yet. Share an insight, caveat, or question.
Toprak et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: