Chatgpt As The Gatekeeper: Accuracy Of A Large Language Model In Extracting Rnfl Parameters And Detecting Glaucomatous Structural Damage From Oct Reports
Published 2026 - 44th Congress of the ESCRS
Reference: POGAU027 | Type: Free paper | DOI: 10.82333/v9w7-k107
Authors: Amit Toledano* 1 , Dr Yogev Giladi 1 , A/Prof. Adi Einan-Lifshitz 1 , A/Prof. Eran Pras 1 , Dr Biana Dubinsky-Pertzov 1
1Ophthalmology,Shamir (Assaf Harophe),Be'er Yaakov,Israel;Gray Faculty of Medical and Health Sciences,Tel Aviv University,Tel-Aviv,Israel
Purpose
To evaluate whether large language model–based interpretation of OCT RNFL reports can support glaucoma services, we assessed GPT’s accuracy for quantitative RNFL extraction and its OCT-only classification of glaucomatous structural damage compared with glaucoma specialist assessment.
Setting
Ophthalmology department (glaucoma service) of a tertiary medical center in Israel; retrospective analysis of de-identified Heidelberg Spectralis RNFL report printouts.
Methods
Spectral-domain OCT RNFL reports (Heidelberg Spectralis) were exported as de-identified RNFL printouts and uploaded via Python 3.12 to OpenAI GPT-5.2 through the OpenAI API (v1). GPT was prompted to (1) assess exam/data quality and extract numeric RNFL parameters, and (2) classify glaucomatous structural damage based solely on the RNFL image. Model outputs were compared with glaucoma specialist diagnosis (based on OCT plus clinical examination, IOP, and visual fields). Agreement was assessed using Cohen’s κ.
Results
GPT numeric RNFL values were available for 70/74 eyes (94.6%), and specialist diagnosis for 68/74 (91.9%). In eyes with paired measurements (n=70), true mean RNFL was 72.11±20.57 µm versus GPT 75.76±21.78 µm, with overestimation of 3.64±4.12 µm. Gatekeeper analyses (GPT “glaucoma” or “suspect” forwarded; n=64) showed GPT missed 5 specialist-defined glaucoma-spectrum eyes (5/53=9.4%, ≈1 in 11), with sensitivity 90.6%, specificity 27.3%, and κ=0.20. For severity (n=61), exact agreement was 49.2% with linear-weighted κ=0.577.
Conclusions
GPT showed high completeness for RNFL extraction and high sensitivity for glaucoma-spectrum triage, but a clinically meaningful miss rate and low specificity limit stand-alone gatekeeping. LLMs may be better suited for structured extraction and prioritization support, with safeguards if implemented clinically.