Driven by rapid advancements in natural language processing technology powered by artificial intelligence (AI), large language models (LLMs), such as ChatGPT (OpenAI), are revolutionizing health care. LLMs, trained on diverse, high-quality datasets, encode extensive knowledge, and generate humanlike responses [1], showing promise in disease prevention, diagnosis, treatment, caregiving, and education [2].

Ophthalmic diseases significantly impact global health [3,4], yet a shortage of ophthalmologists, exacerbated by aging populations, has widened the gap between health care demand and supply, particularly in low- and middle-income countries [5]. LLMs are seen as a potential solution to this resource shortage [6], with studies demonstrating their capabilities in ophthalmology. For instance, Antaki et al [7] found that ChatGPT with GPT-4 [8] achieved an accuracy of more than 70% on simulated ophthalmology board-style exams, outperforming historical human performance. Similarly, Bernstein et al [9] demonstrated that LLMs could appropriately respond to patients’ eye health concerns, with response quality comparable to that of ophthalmologists.

However, LLMs are not always reliable [10]; they may generate convincing yet factually incorrect responses [9], posing risks to patient care. Thus, systematic evaluation of LLMs is essential before their clinical integration [6]. While numerous studies have assessed LLMs in ophthalmology [11-13], a comprehensive statistical analysis and synthesis of these evaluation practices is lacking. This gap hinders the standardization of future model evaluations, thereby postponing clinical deployment.

This study aims to systematically map the current landscape of LLM evaluations in ophthalmology. To achieve this, we first summarized existing LLM evaluation practices in ophthalmology across 6 dimensions: evaluated LLM, data modality, ophthalmic subspecialty, medical task, evaluation dimension, and clinical alignment. Moreover, we performed an additional meta-analysis of diagnostic performance evaluation, providing further evidence for the heterogeneity and fragmentation observed in current evaluation studies. By identifying gaps in current practices, we propose a standardized framework to guide future research and clinical implementation, ensuring the safe and effective integration of LLMs into ophthalmic care.

A systematic review was conducted in accordance with the relevant sections of the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) reporting guidelines [14], and the completed PRISMA checklist is provided in Multimedia Appendix 1.

Ethics approval was not required for this study because it did not involve the recruitment of patients.

Peer-reviewed studies and preprints were retrieved on November 17, 2024, from 4 databases, including PubMed, Web of Science, Embase, and IEEE Xplore. The key search strategies were based on a combination of 2 themes: LLM and ophthalmology, and synonyms or related terms of these 2 themes were also included in the search strategy to ensure a comprehensive coverage of related publications. The complete search strategies in 4 databases are listed in Multimedia Appendix 2 [15-33].

All records were screened in 3 sequential phases by 2 independent reviewers (ZZ and ZP). First, an automated title-matching script removed exact duplicates before any screening. Second, during title and abstract screening, reviewers applied the inclusion and exclusion criteria (Textbox 1) and manually discarded near-duplicate records (eg, preprint version of a published study). Third, full texts were assessed against the same criteria, and any additional duplicates were removed when the full-text content was found to be substantially identical despite different titles or abstracts. Disagreements were resolved through discussion, with a third reviewer (HZ) acting as arbiter when necessary.

In line with recent AI-focused reviews [35], we used domains that are more appropriate for AI model evaluations rather than the traditional patient-centered population, intervention, comparison, outcome, and study design headings.

We performed an exploratory meta-analysis limited to the diagnosis task, the only domain with relatively uniform methodology. Eligible studies were screened using 2 criteria: reporting an exact “correct or incorrect” proportion on open-ended diagnostic questions and avoiding multiple-choice or top-N scoring formats. The proportion of correct diagnoses was the primary effect size. Parallel model arms within a study were treated as independent evaluations. Data, including model, case count, modality, subspecialty, and accuracy, were extracted through dual-independent review with third-reviewer adjudication. Random-effects pooling (logit transformation; DerSimonian-Laird τ²) and inverse-variance weighting were applied in Python (version 3.8.19) using NumPy (version 1.24.3) and SciPy (version 1.10.1) libraries, and heterogeneity was quantified with I². Subgroup analyses by LLM, modality, and subspecialty were performed to explore sources of heterogeneity.

Bar plots and stacked bar plots were used to visualize the distribution of relevant studies across different categories, and forest plots were generated for the exploratory meta-analysis. A 2-sided Fisher exact test compared the prevalence of open-source LLMs in “evaluation-after-development” studies versus “pure-evaluation” studies. Data statistics and visualization were conducted in Python (version 3.8.19) using NumPy (version 1.24.3), Pandas (version 2.0.3, SciPy (version 1.10.1), and Matplotlib (version 3.7.3) libraries.

A total of 817 unique records were identified through our systematic search. After removing duplicates and unrelated papers based on title and abstract, 338 (41.4%) studies remained for full-text screening. Ultimately, a total of 187 (22.9%) studies met the study selection criteria and were included in the analysis (Figure 1). Key features of the included studies are described in Table 1.

Closed-source LLMs overwhelmingly dominated the included studies in this review. In total, 169 (90.4%) of the 187 studies evaluated 1 to 3 models, and more than half (n=100, 53.5%) compared multiple LLMs in parallel (Figure S1 in Multimedia Appendix 2). The 3 most frequently assessed LLM series were all closed-source ones—ChatGPT (n=170, 90.9%), Gemini (n=58, 31%; including Pathways Language Model, Bard, and Gemini), and Copilot (n=32, 17.1%; including Copilot [206] and Bing AI). In contrast, open-source LLMs were only evaluated in 25 (13.4%) studies, for example, LLaMA (n=15, 8%) and ChatGLM [207] (n=6, 3.2%; Table 2).

Notably, open-source models accounted for a substantially larger share of “evaluation-after-development” studies than “pure-evaluation” studies (17/22, 77.3% vs 8/165, 4.8%; P<1×10^-5, Fisher exact test; Figure S2 in Multimedia Appendix 2). Detailed definitions of the 2 study types (Table 1) demonstrate superior domain-specific versatility in ophthalmology.

Image-based evaluations were uncommon. Only 19 (10.2%) of the 187 studies assessed the ability of LLMs to process images and text in ophthalmology, whereas most (n=168, 89.8%) focused solely on text-only performance (Table 2). The ophthalmic images involved in these multimodality evaluations include, but are not limited to, slit lamp images, fundus photography of the posterior pole, optical coherence tomography, and ophthalmic ultrasonography. Other modalities such as voice, video, or documents were not included in any of the studies.

Evaluations clustered in comprehensive ophthalmology (72/187, 38.5%), leaving many subspecialties scarcely explored. Retina and vitreous (32/187, 17.1%) and glaucoma (20/187, 10.7%) followed at a distance, whereas ocular pathology and oncology, refractive surgery, and ophthalmic pharmacology appeared in no more than 3 studies each (Table 2).

Medical task coverage was unevenly distributed. Most studies concentrated on medical query (86/187, 46.0%), standardized examination (41/187, 21.9%), and diagnosis making (29/187, 15.5%) in ophthalmology, and other common tasks explored (>5%) included ophthalmic education, clinical decision-making, and medical text generation. In contrast, research assisting (5/187, 2.7%), patient triaging (3/187, 1.6%), and disease prediction (3/187, 1.6%) received comparatively less attention (Figure 2).

Evaluation dimensions were concentrated and unevenly distributed. Among 187 included studies, 118 (63.1%) evaluated 2 or more dimensions, yet only 15 (8%) assessed 4 or more dimensions, demonstrating the comprehensiveness of the existing evaluation (Figure S3 in Multimedia Appendix 2). Accuracy (n=176, 94.1%) was the most prevalent evaluation dimension, whereas calibration and uncertainty (n=5, 2.7%) were the least frequently examined. Notably, specific medical tasks showed distinct preferences for evaluation dimensions. For example, among 20 studies on ophthalmic education, 19 (95.0%) assessed readability and usability (Figure 2).

The use of real patient data was limited and often undocumented. As Figure 3A shows, of the 187 studies, 45 (24.1%) tested LLMs with real patient data, including clinical information, examination results, and imaging; 48 (25.7%) studies did not provide enough methodological detail to ascertain the data source; the remaining 94 (50.3%) studies used purely non–case-based ophthalmic problems or virtual patient data.

Comparisons between LLMs and humans were common but not universal. As Figure 3A shows, of the 187 studies, 63 (33.7%) compared LLM outputs with the performance of ophthalmologists at various training levels, providing a reference point for potential clinical integration.

Multilingual evaluations were scarce. As Figure 3A shows, only 24 (12.8%) of the 187 studies tested LLMs in non-English contexts. As Figure 3B shows, of these 24 studies, Chinese (n=14, 58.3%), Spanish (n=3, 12.5%), and Japanese (n=2, 8.3%) languages dominated these assessments; the remaining 7 languages (eg, French, German, and Finnish) appeared just once each.

Evaluations under real-world deployment were exceedingly rare. As Figure 3A shows, only 2.1% (4/187) of the studies evaluated LLMs in real ophthalmic clinical settings, while the remaining (183/187, 97.9%) studies were all conducted based on simulated clinical scenarios.

In total, 17 (58.6%) of the 29 diagnosis-making studies met our stricter inclusion criteria, providing 28 independent LLM evaluations of open-ended diagnostic accuracy (Figure S4 in Multimedia Appendix 2). Data extraction outcomes are detailed in Table S2 in Multimedia Appendix 2. Random-effects pooling yielded an overall correct-diagnosis proportion of 0.594 (95% CI 0.488-0.692) with extreme heterogeneity (I²=94.5%). Forest plots are provided in Figure S5 in Multimedia Appendix 2. Subgroup pooling did not resolve this variability: I² values remained greater than 80% for most models, modalities, and subspecialty strata (Table 3).

MOPH achieved the highest accuracy for diagnosis, though this estimate was based on a single evaluation. Counterintuitively, GPT-4 (pooled accuracy=0.559) underperformed compared to GPT-3.5 (pooled accuracy=0.649). Across modalities, the pooled accuracy for image-text tasks (0.407) was lower than for text-only tasks (0.619). Five subspecialties and “comprehensive ophthalmology” were included in the subspecialty analysis; pooled estimates ranged from 0.483 for retina and vitreous to 0.731 for external disease and cornea, all with wide confidence intervals.

Our synthesis of 187 studies shows a skewed evaluation landscape; certain aspects, such as question-and-answer tasks, comprehensive ophthalmology, and accuracy, dominate, whereas image-based testing, non-English settings, and low-incidence subspecialties (eg, ocular oncology [208]) are rarely explored. Commonly studied areas share 3 traits: broad task generality, easily accessible test data, and high disease prevalence [209-211]. Neglected topics often lack public datasets and have smaller specialist communities [212,213].

Few (19/187, 10.2%) studies included images despite imaging being central to eye care [214]; technical barriers (clinically usable multimodal LLMs only began to emerge in 2023 and are still maturing [8,199,215,216]), privacy concerns, and data-sharing problems all contribute to this low inclusion rate [217]. Open-source LLMs were underrepresented despite their several irreplaceable advantages over closed-source ones in clinical settings (Table 4), yet they appeared much more frequently after new model development, confirming their suitability for customization for specific ophthalmic tasks (eg, fundus image analysis [36-38]). Unlike pure evaluation studies that mainly relied on manual assessment with Likert scales [2,15,39], these postdevelopment evaluations used automated metrics, such as Bilingual Evaluation Understudy [218] and F₁-score, which are standard metrics in AI and natural language processing domains [40,41], enabling rapid testing after LLM customization.

Profound heterogeneity was observed in current evaluation practice and evidence, driven by 3 factors. The first factor was metric heterogeneity. Even within our unified 8D framework, measures under the same heading were not directly comparable; for example, “accuracy” could be reported as an exact-correct ratio, area under the curve, F₁-score, or a Likert scale score with varying anchors. The second factor was data heterogeneity. Assessment sets differed widely; in the “medical query” task, questions spanned multiple difficulty levels, originated from diverse sources, and covered numerous ophthalmic subspecialties. The third factor was model heterogeneity. LLMs are updated frequently (eg, GPT-4o 2024-05-13, 2024-08-06, 2024-11-20), yet most papers note only the major version, obscuring performance shifts between iterations.

To probe whether pooling was feasible despite this variability, we focused on the methodologically most consistent task, diagnosis making, as more than half of these studies (17/29, 59%) used open-ended questions with binary correct or incorrect scoring. However, the pooled accuracy of 0.594 still demonstrated extreme heterogeneity (I²=94.5%), indicating that this aggregate statistic failed to represent almost any specific clinical context. Subgroup pooling offered little relief. Some findings were merely self-evident (image-text tasks were harder than text-only tasks), others were contradictory (GPT-3.5 pooled accuracy was higher than GPT-4, probably because 1 study of 422 unusually difficult cases tested GPT-4 alone [16]), and most confidence intervals remained wide. Finally, continuous, undocumented model updates, such as minor ChatGPT and Gemini releases, occur every few months, making any version-based synthesis obsolete almost as soon as it is published. Taken together, these observations confirm that high methodological heterogeneity, small task-specific samples, and model drift currently preclude meta-analytic results from informing bedside use or system refinement. Therefore, rigorous, standardized, and multidimensional benchmarks must precede any performance aggregation if future syntheses are to provide actionable clinical guidance in ophthalmology.

Previous reviews focused on LLM applications in ophthalmology, but to the best of our knowledge, none systematically analyzed evaluation practice itself [6,222-224]. While cross-disciplinary surveys in medicine [204,205] flagged generic gaps in LLM evaluation, our study is the first to quantify them within ophthalmology.

We deepened the analysis by anchoring it in ophthalmology’s distinctive context. Firstly, we gave special weight to multimodal (image-text) evaluation because diagnosis hinges on imaging. Moreover, we paid special attention to non-English evaluation to assess global applicability because eye care resources are unevenly distributed across language regions [225]. Finally, we divided the field into highly granular subspecialties (eg, retina, glaucoma, and ocular oncology), which revealed differences in the degree of attention given to different subspecialties. To capture cutting-edge technical advances, we also expanded our search strategy to include the IEEE Xplore augmentation database, which yielded more than a 5-fold increase in included ophthalmology literature (from 36 to 187 studies) compared to previous reviews [204].

We propose 4 potential directions for future research. The first direction is the cocreation of a publicly available benchmark for LLMs in ophthalmology, covering all subspecialties, multiple languages, and image-text tasks. The second direction is the adoption of a standard reporting checklist that requires multidimensional metrics (eg, accuracy, comprehensiveness, and safety) and specifies the exact LLM version. The third direction is continuous evaluation of newly launched LLMs, such as the pioneering reasoning model o1, especially open-source models amenable to ophthalmic fine-tuning (eg, DeepSeek and Qwen [239]). The fourth direction includes clinically driven assessments targeting real-world needs, such as referral triage, postoperative counseling, and low-resource language support, progressing through the 3-phase road map to RCTs.

First, the search window closed on November 17, 2024, so publications released afterward, particularly work on the latest LLMs, were not captured. This was especially important given the rapid iteration of recent LLMs. We partially mitigated this by including preprints and 4 multidisciplinary databases, yet progress after the cutoff may have been underestimated.

Second, categorizing medical tasks and evaluation dimensions required subjective judgment; dual-independent extraction with third-reviewer adjudication reduced but did not eliminate misclassification risk.

Third, although we implemented stringent inclusion criteria, the absence of a formal risk-of-bias (RoB) assessment, reflecting the current lack of validated RoB tools specifically for medical AI studies, necessitated equal weighting of all included studies irrespective of methodological rigor. This limitation underscores the need for developing AI-specific RoB frameworks that enable future systematic reviews to appropriately weight evidence by study quality.

Fourth, extreme heterogeneity across studies precluded a definitive meta-analysis, highlighting the need for standardized evaluation protocols and datasets.

This systematic review mapped 187 ophthalmology-focused LLM evaluations and found a landscape that was extensive yet uneven. Research attention was concentrated on a few closed-source models, predominantly text-based tasks, and a limited set of subspecialties, medical tasks, and evaluation dimensions. Open-source LLMs, multimodal assessments, non-English testing, and real-world clinical studies were scarce. Fragmented methods yield heterogeneous evidence, impeding confident clinical adoption. These findings emphasize the necessity of standardized evaluation frameworks and highlight the critical gaps that must be closed before LLMs can be integrated into ophthalmic practice.