This Arkangel AI whitepaper reviews what supportive AI can and cannot do today in image-based ophthalmology, and what it would take to use it responsibly in Latin America. It is a narrative synthesis of 25 sources from targeted searches of PubMed/MEDLINE, Crossref, WHO and FDA with a June 14, 2026 cutoff, plus four focused Europe PMC searches for regional evidence. The evidence is organized by task (diabetic-retinopathy screening, retina and OCT, glaucoma, referral and governance) and by five maturity stages, from retrospective discrimination to routine implementation.
The most mature evidence is diabetic-retinopathy screening: one FDA De Novo-authorized system reported 87.2% sensitivity and 90.7% specificity for more-than-mild diabetic retinopathy in its pivotal trial, figures the authors stress do not transfer to other tools or to Latin American settings. For glaucoma, an SD-OCT meta-analysis reported a pooled sensitivity of 0.91 and specificity of 0.90 with substantial heterogeneity. Direct regional evidence was limited to two Mexican evaluations, a randomized reading experiment with 17 ophthalmologists and a tertiary-care study of 435 patients, and neither measured patient outcomes or routine implementation. The authors recommend AI for screening, reading support, prioritization and referral within governed care pathways, with local validation before any silent prospective evaluation. The document is a whitepaper, not peer reviewed and not a systematic review; all authors are affiliated with Arkangel AI, and the evidence it cites does not evaluate Arkangel products.
