Abstract
About this talk
Background: Community optometrists increasingly provide first-contact urgent eye care. Artificial intelligence (AI) decision support is applied mainly to fundus photography and macular optical coherence tomography (OCT), yet many time-critical conditions are diagnosed from symptoms, slit-lamp findings, intraocular pressure, peripheral retinal examination and systemic features that these images do not capture.
Aim: To examine whether image-based AI is fit for purpose in community triage of acute ocular presentations, focusing on diagnostic coverage and false-negative risk.
Methods: Critical narrative review. PubMed was searched on 28 October 2026 with structured queries on community urgent eye care, AI for specific sight-threatening conditions and large language model (LLM) triage, supplemented by UK guidance. UK urgent eyecare emergency conditions were mapped against the inputs needed for diagnosis and the modalities used by published algorithms.
Evidence synthesis: Image-based AI performs well on defined tasks such as referable diabetic retinopathy (AUC 0.99) and papilloedema (AUC 0.96), but performance is task- and modality-specific. Retinal tears, microbial keratitis, acute angle closure and giant cell arteritis (GCA) depend on history, slit-lamp examination, tonometry or peripheral fundus assessment; in a prospective cohort of community referrals, 15% of retinal tears were seen only on indentation ophthalmoscopy. Algorithms for these conditions use modalities rarely available in community practice and were mostly validated in case–control or enriched cohorts; case–control designs overestimate diagnostic odds ratios roughly threefold. No AI designed for first-contact recognition of endophthalmitis, orbital cellulitis, chemical injury or pre-visual-loss GCA was identified. LLM triage studies are vignette-based; in one, 32% of responses were potentially harmful. A safety-first framework is proposed: red flags mandate clinician-led escalation irrespective of AI output; AI is confined to its validated scope; and a normal result never reassures outside it.
Conclusion: AI should be a condition-specific adjunct, not an autonomous triage tool. Prospective evaluations should report missed emergencies, time to definitive care and visual outcomes.
Keywords: Artificial intelligence; triage; optometry; ophthalmic emergencies; spectrum bias; patient safety.
Speaker
About the speaker
