Home›Ophthalmology› AI keratoconus tools miss the normal fellow eyes that guide surgery
AI keratoconus tools miss the normal fellow eyes that guide surgeryArtificial intelligence shows mixed results for diagnosing keratoconus
medRxivPublished October 4, 2026Study authors: Jiang, H.; Gao, F.; Jie, Y.; Yang, L.; Li, Y.; Jiang, Y.DOI ↗Editorial oversight: Dr. Lars van Dijk, PhD · Surgical, Procedural & Diagnostic
AI-generated summary of the cited source, checked by automated accuracy review.
How we work
Share
Key Takeaway
AI keratoconus detection is unproven in fellow eyes with normal topography and tomography.
A systematic review examined whether artificial intelligence can identify keratoconus in fellow eyes that look normal on topography and tomography — the eyes where surgical decisions are actually made. The review screened 573 studies, adjudicated 212, and broke their claims into 2,035 claim-by-stratum units.
Not one of those 2,035 units established discrimination in fellow eyes with normal topography and tomography. Instead, 90.0% of units answered a different question, and the units did not construct the target population or certify its normality.
Reported performance looked strong but reflected the wrong task. Accuracy for early disease was near-perfect, and the AUC against a same-eye index label reached 0.996. Against a contralateral proxy, however, the AUC fell to 0.549. Inputs correlated with the label-generating index at r = 0.991, so accuracy there measured agreement with the index rather than true disease detection.
External validation did not test label generation, and five systematic reviews did not assess it either. Safety, funding, and conflicts were not reported.
For clinicians, the message is caution: AI for keratoconus remains unproven in fellow eyes with normal topography and tomography, and surgical decisions in those eyes should not rest on these tools.
How this fits prior evidence
This systematic review addresses a gap in the clinical utility of artificial intelligence for keratoconus. While prior coverage noted that iOCT-guided pneumodescemetopexy may achieve PDL/DM attachment in acute corneal hydrops, this review specifically addresses the diagnostic accuracy of AI tools. The finding that 0 out of 2,035 units successfully discriminated keratoconus in fellow eyes with normal topography and tomography suggests that AI currently lacks the specificity required for these critical surgical decision-making cases.
Living with keratoconus, a condition where the cornea thins and bulges, makes it hard for doctors to know exactly when to intervene. Because of this, many people look to artificial intelligence to help provide clear answers. A large review of over 500 studies looked at how well these AI tools actually perform when looking at the eyes.
The results were mixed. While the AI showed near-perfect accuracy when looking at eyes already showing signs of the disease, it failed to distinguish the condition in the other eye when that eye had normal measurements. This is a significant finding because doctors often rely on the healthy eye to make critical surgical decisions.
There are also some hurdles to clear before these tools can be used in a clinic. Many of the studies reviewed did not properly define the group of healthy eyes, and some did not test how the AI actually generates its labels. Because the technology is not yet proven for healthy eyes, it is still unclear how much it can help in everyday practice.
What this means for you:
AI shows high accuracy for existing keratoconus but cannot reliably identify the disease in healthy eyes.
Common questions
Is AI accurate at finding early signs of keratoconus?
The review found that artificial intelligence showed near-perfect accuracy when identifying the disease in cases of early disease. However, it is important to note that this accuracy measured how well the tool agreed with the specific index used in the study, rather than being a definitive clinical confirmation.
Can AI tell the difference between a healthy eye and one with keratoconus?
No, the study found that none of the 2,035 units tested were able to distinguish between eyes with keratoconus and fellow eyes with normal measurements. This is a key finding because doctors often use the healthy eye to make important decisions about treatment and surgery.
Is AI currently reliable for making surgical decisions?
The evidence shows that AI for keratoconus is currently unproven in eyes with normal measurements. Because these are the eyes where surgical decisions are often made, the technology is not yet proven to be a reliable tool for those specific cases.
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.
