In spring 2026, two landmark events occurred consecutively regarding AI screening for retinopathy of prematurity (ROP). On April 14, U.S. company Siloam Vision announced key results from its pivotal study: 16 practicing ophthalmologists interpreted 300 ROP examinations with AI assistance, and the diagnostic accuracy for the critical treatment-indicating sign—'plus disease'—improved from an average AUC of 0.93 to 0.95; inter-physician agreement also rose from κ 0.55 to 0.62. One month later, on May 14, the blindness-prevention organization Orbis announced that its AI assistant had entered the neonatal intensive care unit (NICU) at Mongolia’s National Center for Maternal and Child Health in Ulaanbaatar for the first time. Crucially, both announcements originated from press releases—not peer-reviewed papers—and the companies themselves disclosed that their systems have not yet received FDA clearance. News is outpacing evidence because this disease waits for no one: ROP is a leading cause of childhood blindness worldwide; untreated disease progresses relentlessly to retinal detachment and permanent blindness; and before irreversible damage occurs, ROP can only be detected through regular fundus screening. In low- and middle-income countries, neonatal survival rates are rising faster than the supply of pediatric ophthalmologists: the more babies saved, the more eyes awaiting screening.
## Two Established Approaches: Task Shifting and Clinical Prediction
Before AI arrived, viable solutions already existed. India’s KIDROP project reported in 2014 that across 81 neonatal units, trained technicians—not physicians—used portable cameras to capture fundus images, which remote experts then interpreted via smartphones. Across 6,339 imaging sessions involving 1,601 infants, sensitivity for treatment-requiring disease was 95.7%, specificity 93.2%, and the missed-diagnosis rate just 0.4%. 'Task shifting plus telemedicine' works. What AI aims to solve are the two remaining bottlenecks atop this framework: remote interpretation still depends on scarce experts, and expert interpretations themselves lack consistency.
The second established approach bypasses fundus images entirely, instead predicting who needs ophthalmologic examination using routine clinical data. A 2022 retrospective study of 815 infants achieved 100% sensitivity in internal validation—and maintained 100% sensitivity in external validation—but specificity plummeted to 7.5%: it flagged every infant as high-risk, rendering it clinically unhelpful. A November 2026 single-center study adopted an interpretable model using only conventional variables—gestational age, birth weight, etc.—and among 184 infants, correctly identified all 21 cases of Type 1 ROP while reducing the number requiring examination by 68.5% compared with local standards—but the authors themselves only cautiously called for prospective validation.
## Missing 20% Per Scan—Yet Missing Nothing Overall
The most aggressive route is fully autonomous AI screening—and here lies a set of figures demanding close scrutiny. A 2023 prospective study published in JAMA Ophthalmology reported 100% sensitivity for 'treatment-requiring ROP'—the most dangerous category, with zero misses—but only 80.0% sensitivity for the broader 'referral-requiring ROP' (which includes lesions not yet needing treatment but warranting referral for observation): among 156 infants in the test set, AI missed one-fifth of follow-up signals.
In 2024, multinational external validation published in the same journal pushed the paradox further: the i-ROP algorithm was tested across 11 U.S. and 48 Indian NICUs, totaling 11,880 examinations. At the single-examination level, sensitivity for 'moderate-to-severe ROP' ranged only from 80.8% to 83.5%; yet at the patient level, in both datasets, *every* infant who ultimately developed Type 1 ROP had been flagged positive by AI during *at least one* prior screening before diagnosis—100%.
Do these two numbers contradict? No—they share the same underlying mechanism: screening is repeated, and a single missed detection can be caught at the next visit—*provided the system ensures infants actually return for re-examination*. J. Peter Campbell, one of the study’s lead investigators, explained the design of the autonomous version in a 2025 interview: its output is not a diagnosis, but a binary decision—'refer, or repeat in one week.' When AI detects disease, images are routed for remote expert review; when it does not, the infant is scheduled for re-examination in one week. Thus, the safety net of autonomous screening lies not in the algorithm itself, but in the process. Conversely: this AI is permitted to miss—because 'not missing' is deliberately assigned to the repeated-screening process, not to the sensitivity of any single algorithm run.
## Assisting Doctors in Image Interpretation: Three Pitfalls That May Not Be Avoidable
Evidence for the assistive route is easier to read than for the autonomous route—but three pitfalls remain. First is device dependency: a 2023 Lancet Digital Health study trained a 'plus disease' classifier on fundus images from 1,370 newborns in London; external validation across U.S., Brazilian, and Egyptian sites yielded AUCs of 0.969–1.000 *on images from the same camera brand*, but dropped to 0.865–0.891 when another camera brand was used—the model may recognize not just disease, but the camera itself. Second is circular reference standards: an Indian team trained a screening model on 227,326 images accumulated over 11 years by KIDROP; sensitivity for Stage 1–3 ROP was 91.5%, specificity 91.2%, but the labels themselves came from human readers’ records within the tele-screening program—using human interpretation to train AI, then deploying AI to replace humans. The paper explicitly states that prospective clinical validation is 'under consideration' (i.e., not yet conducted), and many authors are employees or shareholders of the company. Third is evidence format: the earlier-cited jump from 0.93 to 0.95 currently exists only in a press release; peer review remains pending.
Yet the value of the assistive model lies elsewhere: it does not prove how accurate AI is alone, but rather demonstrates AI’s ability to narrow inter-physician interpretation variability. For a sign where even expert consensus reaches only κ 0.55, 'bringing 16 physicians closer together' may be more pragmatic than 'replacing physicians.'
## What Level of Evidence Has Been Reached—and Where Is Regulation?
Evidence in this field is often conflated, yet comprises three distinct tiers: retrospective benchmarking (re-running algorithms on existing images) ≠ prospective clinical utility (whether integrating into real-world screening improves referral and treatment decisions) ≠ patient outcomes (whether blindness is reduced). The strongest current evidence stops at 'Tier 1.5': multinational external validation uses real-world data, but endpoints remain 'detection,' not 'vision.' A July 2026 systematic review included 12 empirical studies—none compared AI screening versus conventional screening using blindness or vision loss as the primary endpoint.
Regulation mirrors this evidence gap: the FDA approved its first autonomous AI diagnostic device (for diabetic retinopathy) in 2018, yet *no* AI algorithm has been approved in the U.S. or Europe for ROP to date. The furthest along are assistive models—Siloam’s system received FDA Breakthrough Device designation and plans to pursue De Novo pathway approval. The systematic review’s conclusion is blunt: existing studies are limited by single-center designs, enriched samples, insufficient severe cases, and lack of fairness analyses—'routine autonomous implementation remains premature.' Economic analysis has just begun: a Chinese decision-tree model published in October 2026 showed community-based AI-assisted screening superior to both telemedicine and bedside expert screening in cost and quality-adjusted life-years (urban lifetime cost per infant: $1,507), but this is model-based projection—not real-world measurement—and ceases to hold if treatment-requiring ROP incidence falls below 0.7% or rural screening coverage drops to just 11%.
## Three Takeaways You Can Carry Away
For managers of neonatal networks: task-shifting plus telemedicine works *today*—you need not wait for AI; AI’s most reliable current use is improving interpretation consistency, not replacing interpretation. For payers and policymakers: cost-effectiveness evidence remains at the modeling stage; before procuring autonomous screening systems, demand local prospective validation and data on follow-up adherence. For families of preterm infants: what truly matters is not 'Is there AI?' but 'Does this hospital have standardized ROP screening and follow-up protocols?'—across all evidence, missing one follow-up visit is far more dangerous than an algorithm’s few percentage points of error.
Three developments merit watching next: whether Siloam’s De Novo application will be accepted and approved; whether the Mongolian deployment yields peer-reviewed prospective data; and whether the first true randomized trial emerges—directly comparing blindness outcomes between AI screening and conventional screening. The final question can be debated now: An autonomous AI screening system misses one-fifth of referral signals per scan, yet catches *all* treatment-requiring cases by relying on a 'repeat in one week' process—should regulators approve it? Should parents accept it? And whose responsibility is this 'safety net'—the algorithm’s, the process’s, or the person executing the process?
Source: Coyner AS, et al. Multinational External Validation of Autonomous Retinopathy of Prematurity Screening. JAMA Ophthalmology, 2024. https://doi.org/10.1001/jamaophthalmol.2024.0045