Optimising The Follow-Up Time For The Dora Ai Tool Post Cataract Surgery In A Single Eye Unit

Artificial intelligence (AI) systems are increasingly deployed in post-operative care pathways, yet optimal follow-up timing remains undefined. Early follow-up may capture transient, self-limiting post-operative symptoms, inflating escalation rates without improving complication detection. This study aims to evaluate whether extending AI-led telephone follow-up from 2 weeks to 3-4 weeks post-cataract surgery reduces unnecessary clinical escalation while maintaining safety.

A single high-volume NHS district general hospital eye unit in the United Kingdom utilising the DORA AI telephone system for routine post-operative cataract surgery follow-up.

Prospective service evaluation comparing outcomes before and after implementation of delayed DORA follow-up (3-4 weeks post-operatively versus standard 2-week follow-up). All consecutive uncomplicated phacoemulsification cases during the study period will be included. Primary outcome: proportion of FAIL classifications. Secondary outcomes: proportion requiring in-person review, phone-managed concerns, unexpected management changes in PASS patients, and safety events. Call timing and clinical outcomes will be analysed using comparative statistics.

In the 2-week follow-up cohort (n=188), 52% of patients were classified as PASS and 48% as FAIL. Among PASS patients, 97% required no further follow-up and no serious harm occurred, demonstrating a strong safety profile. However, 77% of FAIL patients did not require in-person review (39% no concerns on nurse call; 38% managed by telephone), indicating substantial potentially avoidable escalation. These findings suggest opportunity to reduce false-positive FAIL classifications through temporal optimisation. Comparative data from the 3-4 week cohort are pending and will be presented at the conference.

AI-led 2-week post-operative follow-up after cataract surgery demonstrates a strong safety profile, with minimal missed pathology. However, the high proportion of FAIL classifications not requiring in-person review suggests over-escalation, likely driven by early transient symptoms. Follow-up timing may therefore represent a modifiable determinant of AI performance. Evaluating a 3-4 week model will determine whether temporal optimisation can reduce unnecessary escalation while preserving safety, informing evidence-based refinement of scalable AI-enabled cataract care pathways.