Ophthalmic Ai Triage: Why Patient Safety Depends On Fairness Auditing
Published 2026 - 44th Congress of the ESCRS
Reference: POREF209 | Type: Free paper | DOI: 10.82333/sr5v-dq21
Authors: Hamid Tavakoli* 1
1AI Fairness CIO,London,United Kingdom
Purpose
AI-assisted triage systems are increasingly used to prioritise ophthalmic patients before clinical assessment. While reported accuracy is often high, unequal performance across patient subgroups can remain undetected, leading to systematic under- or over-triage.
This work demonstrates why fairness auditing is a prerequisite for patient safety in ophthalmic AI triage and introduces a practical, clinician-accessible approach to detecting and mitigating unsafe decision disparities.
Setting
AI-assisted ophthalmic triage and referral prioritisation in outpatient and screening pathways, where algorithmic outputs are increasingly used to support clinical decision-making before specialist assessment.
Methods
We developed a no-code, outcome-based fairness auditing workflow designed for ophthalmic triage applications.
The Fairness Diagnostic Kit (FDK) enables independent auditing of AI decision outputs, including black-box systems, using clinically relevant fairness and safety metrics without requiring programming or access to model internals.
BiasClean and BiasClean+SVM were designed as complementary pre-deployment tools to mitigate bias at the data level before training or deployment. The workflow supports local execution, allowing clinicians to audit and monitor AI systems within their own clinical environments without external data transfer.
Results
Fairness auditing revealed clinically meaningful disparities in triage outcomes that were not apparent from aggregate accuracy alone.
Differences in error distribution, risk stratification, and worst-group performance indicated unequal safety profiles across patient subgroups.
Data-level bias mitigation measurably reduced these disparities, improving fairness metrics and stabilising triage behaviour across groups. Importantly, these improvements were achieved without modifying model architecture or relying on vendor intervention, demonstrating that safety auditing can be performed independently at the point of use.
Conclusions
In ophthalmic AI triage, accuracy alone is insufficient to assure patient safety. Un-audited bias constitutes a latent safety risk, analogous to an uncalibrated medical device.
Fairness auditing should therefore be treated as a routine safety control rather than an ethical afterthought. No-code, outcome-based tools allow eye practitioners to independently audit, mitigate, and continuously monitor AI triage systems in real-world clinical settings.
All tools described are non-commercial, open-source, and intended for independent clinical safety auditing rather than service provision. Embedding fairness auditing into ophthalmic AI workflows is essential for safe, trustworthy, and sustainable clinical deployment.