Grounded ophthalmic AI beyond the electronic health record trap: a critical narrative review

Abstract

Abstract Electronic health records (EHRs) show that health information technology can digitise clinical work while increasing documentation burden, workflow rigidity and professional burnout. Artificial intelligence (AI) may enter a deeper layer of medicine than the EHR because it can draft summaries, interpret images, recommend follow-up and influence decisions. This critical narrative review uses ophthalmic vision-language models as a concrete case study. Diabetic retinopathy screening AI and lesion-level deep learning have achieved important successes in defined tasks, but clinical interpretation is not classification. A sentence such as ‘hard exudates have increased in the macula’ requires localisation of the fovea and macula, lesion segmentation, longitudinal registration, change quantification and faithful language generation. We argue that the unresolved informatics challenge is grounded in clinical interpretation: the auditable translation of images into finding-location-change triples that remain inspectable by clinicians. The article proposes a three-layer architecture: globally standardised objective observation, explicit local policy and workflow adaptation and clinician-authored AI operations governed by safety gates. Customisation should not be monetised as bespoke vendor work; it should be a safety function built into procurement, audit and lifecycle regulation. The three-layer architecture is a conceptual governance framework derived from published evidence rather than a prospectively validated implementation, and we specify the observable failure modes and metrics that would refute it. Ophthalmology illustrates the broader design principle for medical AI: automation should make evidence more visible, not convert care into another opaque information-processing layer.