GlaucoRAG: A Retrieval-Augmented Large Language Model for Expert-Level Glaucoma Assessment

GlaucoRAG: A Retrieval-Augmented Large Language Model for Expert-Level Glaucoma Assessment | medRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-P4HH5NV'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search GlaucoRAG: A Retrieval-Augmented Large Language Model for Expert-Level Glaucoma Assessment View ORCID Profile Mohammad Aminan , View ORCID Profile S. Solomon Darnell , View ORCID Profile Mohammad Delsoz , View ORCID Profile Amin Nabavi , Claire Wright , View ORCID Profile Brian Jerkins , View ORCID Profile Siamak Yousefi doi: https://doi.org/10.1101/2025.07.03.25330805 Mohammad Aminan 1 Department of Ophthalmology, University of Tennessee Health Sciences Center Memphis , Tennessee, United States Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Mohammad Aminan S. Solomon Darnell 2 Department of Genetics, Genomics and Informatics, University of Tennessee Health Sciences Center Memphis , Tennessee, United States Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for S. Solomon Darnell Mohammad Delsoz 1 Department of Ophthalmology, University of Tennessee Health Sciences Center Memphis , Tennessee, United States Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Mohammad Delsoz Amin Nabavi 1 Department of Ophthalmology, University of Tennessee Health Sciences Center Memphis , Tennessee, United States Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Amin Nabavi Claire Wright 1 Department of Ophthalmology, University of Tennessee Health Sciences Center Memphis , Tennessee, United States Find this author on Google Scholar Find this author on PubMed Search for this author on this site Brian Jerkins 1 Department of Ophthalmology, University of Tennessee Health Sciences Center Memphis , Tennessee, United States Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Brian Jerkins Siamak Yousefi 1 Department of Ophthalmology, University of Tennessee Health Sciences Center Memphis , Tennessee, United States Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Siamak Yousefi For correspondence: siamak.yousefi{at}uthsc.edu Abstract Full Text Info/History Metrics Data/Code Preview PDF Abstract Purpose Purpose: Accurate glaucoma assessment is challenging because of the complexity and chronic nature of the disease; therefore, there is a critical need for models that provide evidence-based, accurate assessment. The purpose of this study was to evaluate the capabilities of a glaucoma specialized Retrieval-Augmented Generation (RAG) framework (GlaucoRAG) that leverages a large language model (LLM) for diagnosing glaucoma and answering to glaucoma specific questions. Design Evaluation of diagnostic capabilities and knowledge of emerging technologies in glaucoma assessment. Participants Detailed case reports from 11 patients and 250 multiple choice questions from the Basic and Clinical Science Course (BCSC) Self-Assessment were used to test the LLM based GlaucoRAG. No human participants were involved. Methods We developed GlaucoRAG, a RAG framework leveraging GPT-4.5-PREVIEW integrated with the R2R platform for automated question answering in glaucoma. We created a glaucoma knowledge base comprising more than 1,800 peer-reviewed glaucoma articles, 15 guidelines and three glaucoma textbooks. The diagnostic performance was tested on case reports and multiple-choice questions. Model outputs were compared with the independent answers of three glaucoma specialists, DeepSeek-R1, and GPT-4.5-PREVIEW (without RAG). Quantitative performance was further assessed with the RAG Assessment (RAGAS) framework, reporting faithfulness, context precision, context recall, and answer relevancy. Main Outcome Measures The primary outcome measure was GlaucoRAG’s diagnostic accuracy on patient case reports and percentage of correct responses to the BCSC Self-Assessment glaucoma items, compared with the performance of glaucoma specialists and two benchmark LLMs. Secondary outcomes included RAGAS sub scores. Results GlaucoRAG achieved an accuracy of 81.8% on glaucoma case reports, compared with 72.7% for GPT-4.5-PREVIEW and 63.7% for DeepSeek-R1. On glaucoma BCSC Self-Assessment questions, GlaucoRAG achieved 91.2% accuracy (228 / 250), whereas GPT-4.5-PREVIEW and DeepSeek-R1 attained 84.4% (211 / 250) and 76.0% (190 / 250), respectively. The RAGAS evaluation returned an answer relevancy of 91%, with 80% context recall, 70% faithfulness, and 59% context precision. Conclusions The glaucoma-specialized LLM, GlaucoRAG, showed encouraging performance in glaucoma assessment and may complement glaucoma research and clinical practice as well as question answering with glaucoma patients. Introduction Artificial intelligence (AI) models have demonstrated high accuracy in detecting several ophthalmic diseases, in-cluding diabetic retinopathy and glaucoma, from various imaging modalities over the past few years[ 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 ]. More recently, large language models (LLMs) have shown promise in answering clinical queries[ 10 , 11 , 12 , 13 , 14 ]. While LLMs like GPT-4.5-PREVIEW hold promise for automated question answering, there is a risk of producing hallucinated answers without grounding[ 15 , 16 ]. Retrieval-Augmented Generation (RAG), which combines external, domain-specific knowledge, has been identified as a valuable approach with the capacity to improve the factual accuracy of LLM outputs[ 16 ]. RAG combines dense document embeddings with an LLM to generate responses that are validated through the embedded knowledge and typically provide evidence and grounding that minimizes hallucinations[ 1 ]. Previous RAG systems have performed well in radiology and pathology questions[ 2 ]. However, RAG-based models in ophthalmology remain relatively understudied.16 A head-to-head comparison of ChatGPT versus ophthalmologists on exam questions revealed that a general purpose LLM still performs lower on knowledge relevant to ophthalmologists, highlighting the necessity of domain-specialized RAG systems[ 6 ]. In this study, we developed GlaucoRAG, an augmentedretrieval LLM-based system to assess glaucoma using a rigorous glaucoma-specific knowledge base to facilitate more evidence-based question answering in glaucoma. 1. Methods 1.1 Knowledge Base Sources and Pre-Processing The implementation of GlaucoRAG involves data curation, ancillary open-source tools, and a popular state-of-the-art LLM. The knowledge base for GlaucoRAG was compiled from three sources: three authoritative glau-coma textbooks, American Academy of Ophthalmology Basic and Clinical Science Course BCSC Section 10, Shields’ Textbook of Glaucoma (7th ed.), and Glaucoma (2nd ed.); 1,859 peer-reviewed articles published between 2000 and March 2020. Articles were retrieved from Ophthalmology, Progress in Retinal and Eye Research, Ophthalmology Glaucoma, American Journal of Ophthalmology, Survey of Ophthalmology, and Journal of Glaucoma using the query term “glaucoma” in the title or abstract; and 15 guideline documents on assessment and management of glaucoma published by American Academy of Ophthalmology, European Glaucoma Society, and Royal College of Ophthalmologists. A limited set of 250 multiple-choice questions, without answers or explanatory content, were input into LLMs solely to evaluate LLM models. This use aligns with the AAO’s permitted usage guidelines and is consistent with the principles of Fair Use under U.S. copyright law, given the non-commercial, transformative nature of the work. No AAO answers, explanations, or proprietary educational material were reproduced or distributed in the study. 1.2 GlaucoRAG Architecture GlaucoRAG’s inference components are R2R[ 17 ], an open source agentic RAG system and GPT-4.5-PREVIEW[ 18 ], a state-of-the-art LLM. Figure 1 shows the architecture of the GlaucoRAG. The system is composed of two coupled subsystems, an offline knowledge-ingestion pipeline and an online inference pipeline. Both system pipelines are orchestrated by Docker-Compose which exposes a single REST/WS endpoint via the open-source R2R frame-work (v3.6.5). Source code for every component is available in the public repository (commit paper-v1). Download figure Open in new tab Figure 1. Architecture of the glaucoma specialized retrieval augmented generation (GlaucoRAG) framework. 1.2.1 Offline knowledge ingestion pipeline During ingestion, r2r_ingest.py iteratively traversed the corpus, eliminating duplicate titles, and segmented each document using an adaptive, progressive-fallback strategy. Initial segmentation employed 300-token chunks with 15-token overlap; if this precise mode failed, chunk size was reduced to 150 tokens with 10-token overlap, and, if necessary, a final fast mode ingested ≈ 1000-token chunks without overlap. Every resulting chunk was embedded with the default R2R dense encoder (OpenAI text-embedding-3-large) and stored in a Postgres-backed HNSW vector index using cosine similarity. Provenance was maintained by recording SHA-256 checksums for every processed file, while r2r_log-ging.py generated color-coded, timestamped logs that were written to /app/logs/ingest.log. 1.2.2 Online inference pipeline At query time, the retriever component is invoked through either r2r_chat_v4.py or r2r_query.py. R2R conducts an approximate nearest-neighbor search that initially returns 20 chunks; an optional cross-encoder re-ranking module (r2r_chat_v3_crossencoder.py) rescales similarity scores without introducing external text. The eight highest-ranked passages (≈ 3200 tokens total) are concatenated with the user prompt and forwarded to GPT-4.5-preview with deterministic generation settings (temperature 0; max_tokens = 1024). GlaucoRAG’s post-processing layer resolves numeric citations into human-readable article or textbook titles with page numbers, integrates them into the JSON response, and renders them inline in the chat interface. Robustness safeguards include an automatic retry that reduces context to four passages if the language model returns an empty string or malformed JSON; persistent failure then triggers an error message to the client. 1.3 Evaluation Two evaluation tasks were used: (1) Case report task, a pilot set of 11 de-identified glaucoma case reports were drawn from EyeRounds, the open teaching repository of the University of Iowa’s Department of Ophthalmology and Visual Sciences[ 19 ]. Details of selection and redaction have been reported previously. Each case was presented to GlaucoRAG, two baseline LLMs (GPT-4.5-PREVIEW and DeepSeek-R1), and three fellowship-trained glaucoma specialists, who independently sub-mitted a single best diagnosis. Reference diagnoses were taken from the EyeRounds key[ 19 ]. BCSC multiple-choice question task. Two hundred fifty glaucoma-focused items were extracted from the BCSC Self-Assessment bank, each consisting of a stem and four or five answer options. The questions were fed to the same three LLMs; responses were scored automatically against the official answer key. Each item was sent programmatically to GPT-4.5-PREVIEW (OpenAI Chat Completions, temper-ature 0, max_tokens 256) and DeepSeek-R1 (DeepSeek API v1, temperature 0, max_tokens 256) as a single-turn prompt containing only the question stem and answer options; the raw JSON responses were parsed automatically to extract the top-ranked answer choice. For both tasks, accuracy, expressed as the percentage of cases or questions answered correctly (top-1 match for the case reports). Beyond conventional accuracy measurements, we utilized the RAG assessment (RAGAS) framework for a multi-faceted analysis. Four metrics were assessed by RAGAS: (1) Faithfulness, computed by extracting each claim in the answer and checking it against the cited passages; (2) Context Precision, the share of retrieved passages that contain ground-truth evidence; (3) Context Recall, the fraction of all ground-truth evidence found in the retrieved set; and (4) Answer Relevancy, a cross-encoder similarity score between the user’s question and the generated reply[ 20 , 21 ]. All four metrics are reported on a 0 to 1, with higher values indicating better performance. Accuracy was expressed as a binomial proportion (correct / total). Two-sided 95% confidence intervals (CI) were. Differences between independent proportions were assessed with a two-tailed two-proportion z-test. Statistical significance was defined as p < 0.05. All analyses were performed in Python 3.1. 2. Result 2.1 Case Report Tasks On the case report task, GlaucoRAG correctly identified the reference diagnosis in 9 of 11, accuracy of 81.8% (95% CI, 52.3–94.9%). This exceeded the scores of standalone GPT-4.5-PREVIEW (8 / 11, 72.7%, 95% CI 43.4–90.3%) and DeepSeek-R1 (7 / 11, 63.6%, 95% CI 35.4–84.8%). The mean accuracy of the three fellowship-trained glaucoma specialists was ≈ 7/11 cases (66.7%). ( Table 1 ) View this table: View inline View popup Download powerpoint Table 1. Accuracy of glaucoma experts, LLM Models, and GlaucoRAG at assessment of case reports, and BCSC self-assessment. LLM: Large language model; GlaucoRAG: Glaucoma specialized Retrieval-Augmented Generation; BCSC: Basic and Clinical Science Course; CI: Confidence interval. 2.2 BCSC multiple-choice question task Out of 250 glaucoma-focused items from the BCSC Self-Assessment bank, GlaucoRAG answered 228 correctly (91.2%, 95% CI 87.0–94.1%), versus 211 (84.4%, 95% CI 79.4–88.4%) for GPT-4.5-PREVIEW and 190 (76.0%, 95 %CI 70.3–80.9%) for DeepSeek-R1. A two-sample z-test confirmed that the 6.8-point margin over GPT-4.5-PREVIEW was statistically significant (p = 0.02), and the 15.2-point margin over DeepSeek-R1 was highly significant ( p < 0.001). ( Table 1 ) 2.3 RAGAS Quality Metrics Table 2 shows details of the RAGAS quality metrics. RA-GAS analysis on the 250 BCSC items yielded a faithfulness score of 0.70, context precision 0.59, context recall 0.80, and answer relevancy 0.91, indicating that most answers were well supported by retrieved evidence while maintaining high relevance despite a recall-oriented retrieval strategy. View this table: View inline View popup Download powerpoint Table 2. RAGAS evaluation results for GlaucoRAG framework. RAGAS: Retrieval-Augmented Generation Assessment; GlaucoRAG: Glaucoma Specialized Retrieval-Augmented Generation. 3. Discussion In this study, we introduced and evaluated a RAG system, GlaucoRAG, designed to integrate expert curated textual knowledge for high accuracy question answering in glaucoma. Our results demonstrate a statistically significant improvement over non-RAG baselines. The performance boost observed with GlaucoRAG, 81.8% on case-based queries and 91.2% on BCSC multiple choice questions, exceeds the documented performance of non-RAG ophthalmic models. Consistent gains across small scale expert curated questions (18.2 percentage point improvement) and large-scale standardized assessments (6.8 percentage point improvement) validate the robustness of our approach. The advent of LLM like GPT-4.5-PREVIEW has opened new horizons in ophthalmology, particularly in domains that require nuanced, evidence-based interpretation, such as glaucoma management. LLMs hold great promise for decision support, yet their utility in medicine is limited by well-known pitfalls, most notably the tendency to “hallucinate” clinically plausible but unverified statements, a problem rooted in generic training corpora, weak domain evaluation standards, and uneven data quality. Routing generation through a retrieval layer largely neutralizes this risk by obliging the model to anchor every statement in citable evidence, an approach that mirrors the way clinicians defend their conclusions. In addition, widespread adoption of LLMs in medicine still limited by privacy concerns. GlaucoRAG addresses these demands and demonstrating that a locally hosted, cost-effective system can deliver expert level accuracy while maintaining traceable, evidence-grounded responses. Similar to our model, Luo et. al. reported a retrieval augmented ophthalmic LLM aligning with expert consensus 84% of the time versus 46% without retrieval[ 16 ], and GPT4 alone achieved 48% accuracy on complex cases[ 22 ]. Comprehensive RAGAS evaluation revealed nuanced system performance characteristics beyond traditional accuracy metrics. The model achieved a faithfulness score of 70% indicates that seven out of ten generated responses matched retrieved source content factually, providing significant protection against hallucinations, a key requirement for clinical applications. Such a score places GlaucoRAG firmly within the clinical RAG space, where faithfulness values 65% are deemed acceptable. Notably, the relatively low context precision (59%) alongside high context recall (80%) shows that our retrieval policy prioritizes broad knowledge coverage over stringent precision filtering. This conservative approach minimizes information loss but increases contextual noise. Nevertheless, an 80% recall confirms that our knowledge base supplies the information needed to answer four of five clinical queries. In terms of relevancy, the score was 91% that closely supports the overall accuracy of 91.2% demonstrating robust question comprehension and appropriate response generation. Despite these promising capabilities, several limitations merit discussion. First, our current implementation does not fuse ophthalmic imaging with textual data, an essential step for truly robust AI support in glaucoma, where accurate image interpretation is pivotal. As ophthalmic care is inherently multi-modal, future work should develop multi-modal RAG frameworks capable of retrieving and reasoning over both textual and imaging data. However, current LLMs have limited native image understanding capability. Second, although RAG mitigates the static knowledge cutoff of LLMs by injecting external documents, performance remains sensitive to retrieval quality; critical evidence may be missed or ranked low, reducing answer fidelity[ 15 , 16 ]. Dynamic, continuously updated knowledge bases will be essential for near real-time clinical relevance. In conclusion, we have demonstrated that a RAG approach significantly enhances automated glaucoma question answering across expert curated and standardized evaluations. The GPT4.5 + R2R architecture offers a promising, scalable model for evidence grounded medical question and answer. Data Availability All data produced in the present study are available upon reasonable request to the authors, and will be made available online in a github repository. https://store.aao.org/basic-and-clinical-science-course-section-10-glaucoma.html https://webeye.ophth.uiowa.edu/eyeforum/glaucoma_cases.htm#gsc.tab=0 Acknowledgments This work was supported by NIH Grant R01EY033005 (SY). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. References [1]. ↵ Yeganeh Madadi , Mohammad Delsoz , Albert S. Khouri , Michael Boland , Andrzej Grzybowski , and Siamak Yousefi . Applications of artificial intelligence-enabled robots and chatbots in ophthalmology: recent advances and future trends . Current Opinion in Ophthalmology , 35 N2 - Purpose of review Recent advances in artificial intelligence (AI), robotics, and chatbots have brought these technologies to the forefront of medicine, particularly ophthalmology. These technologies have been applied in diagnosis, prognosis, surgical operations, and patient-specific care in ophthalmology. It is thus both timely and pertinent to assess the existing landscape, recent advances, and trajectory of trends of AI, AI-enabled robots, and chatbots in ophthalmology. Recent findings Some recent developments have integrated AI enabled robotics with diagnosis, and surgical procedures in ophthalmology. More recently, large language models (LLMs) like ChatGPT have shown promise in augmenting research capabilities and diagnosing ophthalmic diseases. These developments may portend a new era of doctor-patient-machine collaboration. Summary Ophthalmology is undergoing a revolutionary change in research, clinical practice, and surgical interventions. Ophthalmic AI-enabled robotics and chatbot technologies based on LLMs are converging to create a new era of digital ophthalmology. Collectively, these developments portend a future in which conventional ophthalmic knowledge will be seamlessly integrated with AI to improve the patient experience and enhance therapeutic outcomes.(3) , 2024 . [2]. ↵ Yeganeh Madadi , Hashem Abu-Serhan , and Siamak Yousefi . Domain Adaptation-Based deep learning model for forecasting and diagnosis of glaucoma disease . Biomedical Signal Processing and Control , 92 : 106061 , June 2024 . OpenUrl [3]. ↵ Siamak Yousefi , Louis R. Pasquale , Michael V. Boland , and Chris A. Johnson . Machine-Identified Patterns of Visual Field Loss and an Association with Rapid Progression in the Ocular Hypertension Treatment Study . Ophthalmology , 129 ( 12 ): 1402 – 1411 , December 2022 . OpenUrl CrossRef PubMed [4]. ↵ Siamak Yousefi , Tobias Elze , Louis R. Pasquale , Osamah Saeedi , Mengyu Wang , Lucy Q. Shen , Sarah R. Wellik , Carlos G. De Moraes , Jonathan S. Myers , and Michael V. Boland . Monitoring Glaucomatous Functional Loss Using an Artificial Intelligence–Enabled Dashboard . Ophthalmology , 127 ( 9 ): 1170 – 1178 , September 2020 . OpenUrl CrossRef PubMed [5]. ↵ Anshul Thakur , Michael Goldbaum , and Siamak Yousefi . Predicting Glaucoma before Onset Using Deep Learning . Ophthalmology Glaucoma , 3 ( 4 ): 262 – 268 , July 2020 . OpenUrl PubMed [6]. ↵ Varun Gulshan , Lily Peng , Marc Coram , Martin C. Stumpe , Derek Wu , Arunachalam Narayanaswamy , Subhashini Venugopalan , Kasumi Widner , Tom Madams , Jorge Cuadros , Ramasamy Kim , Rajiv Raman , Philip C. Nelson , Jessica L. Mega , and Dale R. Webster . Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs . JAMA , 316 ( 22 ): 2402 – 2410 , December 2016 . OpenUrl CrossRef PubMed [7]. ↵ Daniel Shu Wei Ting , Carol Yim-Lui Cheung , Gilbert Lim , Gavin Siew Wei Tan , Nguyen D. Quang , Alfred Gan , Haslina Hamzah , Renata Garcia-Franco , Ian Yew San Yeo , Shu Yen Lee , Edmund Yick Mun Wong , Charumathi Sabanayagam , Mani Baskaran , Farah Ibrahim , Ngiap Chuan Tan , Eric A. Finkelstein , Ecosse L. Lamoureux , Ian Y. Wong , Neil M. Bressler , Sobha Sivaprasad , Rohit Varma , Jost B. Jonas , Ming Guang He , Ching-Yu Cheng , Gemmy Chui Ming Cheung , Tin Aung , Wynne Hsu , Mong Li Lee , and Tien Yin Wong . Development and Validation of a Deep Learning System for Diabetic Retinopathy and Related Eye Diseases Using Retinal Images From Multiethnic Populations With Diabetes . JAMA , 318 ( 22 ): 2211 – 2223 , December 2017 . OpenUrl CrossRef PubMed [8]. ↵ Michael David Abràmoff , Yiyue Lou , Ali Erginay , Warren Clarida , Ryan Amelon , James C. Folk , and Meindert Niemeijer . Improved Automated Detection of Diabetic Retinopathy on a Publicly Available Dataset Through Integration of Deep Learning . Investigative Ophthalmology & Visual Science , 57 ( 13 ): 5200 – 5206 , October 2016 . OpenUrl CrossRef PubMed [9]. ↵ Yukun Zhou , Mark A. Chia , Siegfried K. Wagner , Murat S. Ayhan , Dominic J. Williamson , Robbert R. Struyven , Timing Liu , Moucheng Xu , Mateo G. Lozano , Peter Woodward-Court , Yuka Kihara , Andre Altmann , Aaron Y. Lee , Eric J. Topol , Alastair K. Denniston , Daniel C. Alexander , and Pearse A. Keane . A foundation model for generalizable disease detection from retinal images . Nature , 622 ( 7981 ): 156 – 163 , October 2023 . Publisher: Nature Publishing Group . OpenUrl CrossRef PubMed [10]. ↵ Mohammad Delsoz , Hina Raja , Yeganeh Madadi , Anthony A. Tang , Barbara M. Wirostko , Malik Y. Kahook , and Siamak Yousefi . The Use of ChatGPT to Assist in Diagnosing Glaucoma Based on Clinical Case Reports . Ophthalmology and Therapy , 12 ( 6 ): 3121 – 3132 , December 2023 . OpenUrl [11]. ↵ Hina Raja , Xiaoqin Huang , Mohammad Delsoz , Yeganeh Madadi , Asma Poursoroush , Asim Munawar , Malik Y. Kahook , and Siamak Yousefi . Diagnosing Glaucoma Based on the Ocular Hypertension Treatment Study Dataset Using Chat Generative Pre-Trained Transformer as a Large Language Model . Ophthalmology Science , 5 ( 1 ): 100599 , January 2025 . OpenUrl PubMed [12]. ↵ Yeganeh Madadi , Mohammad Delsoz , Priscilla A. Lao , Joseph W. Fong , T. J. Hollingsworth , Malik Y. Kahook , and Siamak Yousefi . ChatGPT Assisting Diagnosis of Neuro-Ophthalmology Diseases Based on Case Reports . Journal of Neuro-Ophthalmology , page doi: 10.1097/WNO.0000000000002274 , April 2022 . OpenUrl CrossRef [13]. ↵ Xiaoqin Huang , Hina Raja , Yeganeh Madadi , Mohammad Delsoz , Asma Poursoroush , Malik Y. Kahook , and Siamak Yousefi . Predicting Glaucoma Before Onset Using a Large Language Model Chatbot . American Journal of Ophthalmology , 266 : 289 – 299 , October 2024 . OpenUrl PubMed [14]. ↵ Fares Antaki , Samir Touma , Daniel Milad , Jonathan El-Khoury , and Renaud Duval . Evaluating the Performance of ChatGPT in Ophthalmology: An Analysis of Its Successes and Shortcomings . Ophthalmology Science , 3 ( 4 ): 100324 , December 2023 . OpenUrl PubMed [15]. ↵ Wan Zhang and Jing Zhang . Hallucination Mitigation for Retrieval-Augmented Large Language Models: A Review . Mathematics , 13 ( 5 ): 856 , January 2025 . Number: 5 Publisher: Multidisciplinary Digital Publishing Institute . OpenUrl [16]. ↵ Ming-Jie Luo , Jianyu Pang , Shaowei Bi , Yunxi Lai , Jiaman Zhao , Yuanrui Shang , Tingxin Cui , Yahan Yang , Zhenzhe Lin , Lanqin Zhao , Xiaohang Wu , Duoru Lin , Jingjing Chen , and Haotian Lin . Development and Evaluation of a Retrieval-Augmented Large Language Model Framework for Ophthalmology . JAMA Ophthalmology , 142 ( 9 ): 798 – 805 , September 2024 . OpenUrl PubMed [17]. ↵ Nolan Tremelling . R2R (Version 3.6.5) , June 2025 . [18]. ↵ Chat GPT-4.5-PREVIEW . [19]. ↵ Nayasha Madhan B , Heather A Stiff M , Ze Zhang M , Wallace L.M. Alward M , Geoffrey D. Lively M W. L. Alward M , John H. Fingert MP , Brittni A. Scruggs MP , Tyler S. Quist M , Nasreen A. Syed M, Rixen J , Young H. Kwon MP . 11 Cases. EyeRounds.org . [20]. ↵ Shahul Es , Jithin James , Luis Espinosa-Anke , and Steven Schockaert . Ragas: Automated Evaluation of Retrieval Augmented Generation , April 2025 . arXiv: 2309.15217 [cs]. [21]. ↵ S. Solomon Darnell , Rupert W. Overall , Andrea Guarracino , Vicenza Colonna , Flavia Villani , Erik Garrison , Arun Isaac , Priscilla Muli , Frederick Muriuki Muriithi , Alexander Kabui , Munyoki Kilyungi , Felix Lisso , Adrian Kibet , Brian Muhia , Harm Nijveen , Siamak Yousefi , David Ashbrook , Pengzhi Huang , G. Edward Suh , Muhammad Umar , Christopher Batten , Hao Chen , Śaunak Sen , Robert W. Williams , and Pjotr Prins . Creating a biomedical knowledge base by addressing GPT inaccurate responses and benchmarking context , October 2024 . Pages: 2024.10.16.618663 Section: New Results. [22]. ↵ Daniel Milad , Fares Antaki , Jason Milad , Andrew Farah , Thomas Khairy , David Mikhail , Charles-Édouard Giguère , Samir Touma , Allison Bernstein , Andrei-Alexandru Szigiato , Taylor Nayman , Guillaume A. Mullie , and Renaud Duval . Assessing the medical reasoning skills of GPT-4 in complex ophthalmology cases . The British Journal of Ophthalmology , 108 ( 10 ): 1398 – 1405 , September 2024 . OpenUrl Abstract / FREE Full Text View the discussion thread. Back to top Previous Next Posted July 07, 2025. Download PDF Data/Code Email Thank you for your interest in spreading the word about medRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following GlaucoRAG: A Retrieval-Augmented Large Language Model for Expert-Level Glaucoma Assessment Message Subject (Your Name) has forwarded a page to you from medRxiv Message Body (Your Name) thought you would like to see this page from the medRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share GlaucoRAG: A Retrieval-Augmented Large Language Model for Expert-Level Glaucoma Assessment Mohammad Aminan , S. Solomon Darnell , Mohammad Delsoz , Amin Nabavi , Claire Wright , Brian Jerkins , Siamak Yousefi medRxiv 2025.07.03.25330805; doi: https://doi.org/10.1101/2025.07.03.25330805 Share This Article: Copy Citation Tools GlaucoRAG: A Retrieval-Augmented Large Language Model for Expert-Level Glaucoma Assessment Mohammad Aminan , S. Solomon Darnell , Mohammad Delsoz , Amin Nabavi , Claire Wright , Brian Jerkins , Siamak Yousefi medRxiv 2025.07.03.25330805; doi: https://doi.org/10.1101/2025.07.03.25330805 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Ophthalmology Subject Areas All Articles Addiction Medicine (568) Allergy and Immunology (863) Anesthesia (299) Cardiovascular Medicine (4425) Dentistry and Oral Medicine (443) Dermatology (382) Emergency Medicine (607) Endocrinology (including Diabetes Mellitus and Metabolic Disease) (1507) Epidemiology (15221) Forensic Medicine (30) Gastroenterology (1123) Genetic and Genomic Medicine (6588) Geriatric Medicine (667) Health Economics (997) Health Informatics (4524) Health Policy (1368) Health Systems and Quality Improvement (1612) Hematology (540) HIV/AIDS (1264) Infectious Diseases (except HIV/AIDS) (15910) Intensive Care and Critical Care Medicine (1103) Medical Education (623) Medical Ethics (145) Nephrology (667) Neurology (6588) Nursing (346) Nutrition (998) Obstetrics and Gynecology (1143) Occupational and Environmental Health (956) Oncology (3331) Ophthalmology (970) Orthopedics (369) Otolaryngology (420) Pain Medicine (435) Palliative Medicine (129) Pathology (663) Pediatrics (1690) Pharmacology and Therapeutics (691) Primary Care Research (710) Psychiatry and Clinical Psychology (5440) Public and Global Health (9219) Radiology and Imaging (2195) Rehabilitation Medicine and Physical Therapy (1369) Respiratory Medicine (1196) Rheumatology (593) Sexual and Reproductive Health (710) Sports Medicine (529) Surgery (710) Toxicology (99) Transplantation (289) Urology (265) (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9ffbfa73bbf48e2e',t:'MTc3OTQ1NDU1OQ=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();