From Validation to Scalable Clinical Deployment: Imaging-Driven Failure Modes and Mitigation Strategies in Autonomous Diabetic Retinopathy Screening

Document Type

Conference Proceeding

Publication Date

6-2026

Publication Title

Investigative Ophthalmology and Visual Science

Abstract

Purpose : Autonomous AI for diabetic retinopathy (DR) screening is moving into primary care and handheld workflows. We tested the hypothesis that (1) imageability and insufficient-quality outputs are the main limiter of same-visit resolution and (2) device or workflow changes drive performance drift, requiring monitoring.

Methods : We performed a structured narrative review (PubMed/Google Scholar, 2010-2025) and extracted quantitative metrics from FDA decision summaries and pivotal trials for U.S.-cleared/authorized autonomous DR systems plus real-world deployment studies. Extracted items included on-label cameras, referral thresholds, sensitivity/specificity, imageability, or insufficient-quality rate, and reported completion/referral outcomes. Findings were mapped to a deployment pathway to identify failure modes and mitigation checkpoints.

Results : Three autonomous DR screening systems have U.S. FDA authorization or clearance for point-of-care outputs, each with specific on-label camera compatibility, referral thresholds, and insufficient-quality pathways. Pivotal evaluations report high diagnostic performance under prespecified conditions, while also documenting nontrivial insufficient-quality outputs, making imageability an upper bound on same-visit screening resolution. Across studies, recurrent implementation failure modes include camera-related domain shift, operator-dependent capture variability (particularly in handheld workflows), unclear escalation pathways for insufficient-quality results, and incomplete closed-loop referral tracking. Concurrently, regulatory expectations increasingly emphasize post-deployment monitoring, transparency, and defined human oversight alongside technical performance.

Conclusions : Scalable autonomous DR screening depends on imaging workflow performance and governance, not accuracy alone. Programs should standardize capture, track imageability/deferral rates, validate locally when devices or workflows change, and audit referral completion with prespecified remediation triggers.

Volume

67

Issue

7

First Page

PB0048

Comments

Association for Research in Vision and Ophthalmology (ARVO) Annual Meeting, May 3-7, 2026, Denver, CO

Last Page

PB0048

Link to Full Text

Share

COinS