Research - EHR misses 1 in 3 OCT-confirmed GA cases. AI screening catches them.

Tahm Spitznagel^1,2, Arin Sheikh^1,2,3, David Isztl^1,2, Rui Santos^1,2, Blaise Thomson⁴, Matthias Dieter Becker^1,2,4, Gábor Márk Somfai^1,2,5,6

Purpose

To quantify the accuracy of geographic atrophy (GA) documentation in electronic health records (EHR) by comparing it with an OCT-based manual grading and AI-assisted screening, and to identify discrepancies between documented and OCT-confirmed GA in clinical practice.

Methods

In this retrospective single-centre quality-control study, OCT examinations of 500 consecutive patients from the Retina Unit of Stadtspital Zürich were enrolled, beginning January 1, 2025. Examinations with insufficient image quality were excluded. A locally deployed deep-learning model from AltrisAI (Chicago, USA), executed on the Bitfount platform (Ltd., London, UK), automatically detected GA. Atrophic areas <0.1 mm² were not classified as GA. All examinations were manually reviewed to establish an OCT-based ground truth. The primary outcome was the accuracy of EHR documentation compared with ground truth; secondary outcomes included the diagnostic performance of the AI model against manual grading.

Results

A total of 929 OCT examinations were analysed. Manual grading identified GA in 179 eyes (19.3%). GA was documented in the EHR in 117 cases, yielding a sensitivity of 65.4%, whereas 62 confirmed GA cases were undocumented, leading to a documentation gap of 34.6%. Among 750 eyes without GA, 12 were labelled as GA in the EHR (all with subthreshold atrophic changes <0.1 mm²), resulting in a specificity of 98.4% and an overall documentation accuracy of 92.1%. The AI model classified 162 eyes as GA-positive and 767 as GA-negative. Compared with manual grading, this yielded 156 true positives, 6 false positives, 23 false negatives, and 744 true negatives, giving a sensitivity of 87.2%, specificity of 99.2%, and an overall accuracy of 96.9%. AI-derived false positives involved very small atrophic areas (median 0.254 mm²; IQR 0.109–0.584 mm²), whereas true positives showed substantially larger lesions (median 1.627 mm²; IQR 0.423–4.896 mm²).

Conclusions

AI-assisted OCT screening combined with manual validation revealed substantial underdocumentation of OCT-confirmed GA in the EHR. Although the AI model demonstrated high specificity, a relevant proportion of GA cases remained undetected by both automated screening and clinical documentation. These findings highlight the need for improved GA recognition and documentation workflows and support the incorporation of AI tools into clinical practice.

‍

^{1 Department of Ophthalmology, Stadtspital Zürich, Zurich, Switzerland}

^{2 Spross Research Institute, Zurich, Switzerland}

^{3 University of Zurich, Zurich, Switzerland}

^{4 Bitfount Ltd, London, UK}

^{5 Department of Ophthalmology, University of Heidelberg, Heidelberg, Germany}

‍^{6 Department of Ophthalmology, Semmelweis University, Budapest, Hungary}

‍