Evaluating Large Language Models in Ophthalmology: Systematic Review

Zitao Zhang, Haiyang Zhang, Z.p. Pan, Zhangqian Bi, Yao Wan, Xuefei Song, Xianqun FanJournal of Medical Internet Research 2025 · DOI

TL;DR

This systematic review reveals that while large language models (LLMs) are being extensively evaluated in ophthalmology, the research landscape is fragmented and heavily biased towards closed-source models and text-based tasks. The authors highlight a critical need for standardized, multimodal benchmarks and phased clinical validation to safely integrate LLMs into eye care.

What Problem Does This Paper Address?

The rapid advancement of Large Language Models (LLMs) like ChatGPT presents a significant opportunity to revolutionize healthcare, including ophthalmology, which faces a shortage of specialists. However, the evaluation of these LLMs in ophthalmology is currently scattered and lacks standardization. This fragmentation makes it difficult to synthesize findings, identify true clinical utility, and ensure patient safety. This paper aims to map the current evaluation landscape, identify gaps, and propose a path forward for robust LLM assessment in ophthalmology.

Methods

The researchers conducted a comprehensive systematic review of studies evaluating LLMs in ophthalmology. They searched major databases (PubMed, Web of Science, Embase, IEEE Xplore) up to November 17, 2024, including both peer-reviewed articles and preprints. Studies were included if they quantitatively assessed an LLM on ophthalmology-related tasks. Data was extracted across six dimensions: evaluated LLM, data modality, ophthalmic subspecialty, medical task, evaluation dimension, and clinical alignment. Descriptive statistics were used for analysis, and an exploratory meta-analysis was performed for diagnostic tasks.

Datasets

The paper does not introduce new datasets. Instead, it analyzes existing studies that utilized various datasets for LLM evaluation. These datasets varied widely and were often not explicitly detailed, contributing to the heterogeneity observed.

Key Insights and Ideas

  • Dominance of Closed-Source Models: The vast majority of studies evaluated closed-source LLMs like ChatGPT, with open-source models being significantly underrepresented.
  • Text-Centric Evaluation: Despite the visual nature of ophthalmology, most evaluations focused solely on text-based tasks, with image-text multimodal evaluations being rare.
  • Uneven Subspecialty Coverage: Research attention is heavily skewed towards comprehensive ophthalmology, retina, and glaucoma, leaving many other subspecialties largely unexplored.
  • Accuracy is King, But Other Dimensions Lag: While accuracy was the most frequently assessed evaluation dimension, crucial aspects like calibration and uncertainty were almost entirely neglected.
  • Real-World Application is Scarce: Very few studies involved real-world clinical deployment or comparisons with human performance, highlighting a significant gap between research and clinical integration.
  • Extreme Heterogeneity: The meta-analysis on diagnostic accuracy revealed extreme heterogeneity, making it impossible to draw reliable aggregate conclusions and underscoring the lack of standardized evaluation methods.

Key Results

Out of 817 identified records, 187 studies met the inclusion criteria.

  • LLM Usage: ChatGPT was the most frequently evaluated LLM (170 studies), followed by Gemini (58 studies) and Copilot (32 studies). Open-source LLMs were evaluated in only 25 studies.
  • Data Modality: 89.8% of studies used text-only data, while only 10.2% incorporated image-text tasks.
  • Subspecialty Focus: Comprehensive ophthalmology (38.5%), retina and vitreous (17.1%), and glaucoma (10.7%) were the most studied subspecialties.
  • Medical Tasks: Medical query (46.0%), standardized examination (21.9%), and diagnosis making (15.5%) were the predominant tasks.
  • Evaluation Dimensions: Accuracy was assessed in 94.1% of studies, while calibration and uncertainty were only evaluated in 2.7%.
  • Clinical Alignment: Only 24.1% of studies used real patient data, 33.7% compared LLMs to human performance, and a mere 2.1% evaluated LLMs in real-world clinical settings.
  • Diagnostic Accuracy Meta-Analysis: The pooled overall diagnostic accuracy was 0.594, but with extreme heterogeneity (I²=94.5%), indicating significant variability across studies.

Limitations and Future Work

The authors acknowledge several limitations:

  • Search Window: The search closed in November 2024, potentially missing the latest LLM developments.
  • Subjective Categorization: Categorizing medical tasks and evaluation dimensions involved subjective judgment.
  • Lack of Risk-of-Bias Assessment: The absence of validated risk-of-bias tools for AI studies meant all included studies were weighted equally.
  • Extreme Heterogeneity: This precluded a definitive meta-analysis and highlighted the need for standardization.

The authors propose future directions:

  • Publicly Available Benchmark: Creation of a comprehensive, publicly available benchmark for LLMs in ophthalmology, covering all subspecialties, languages, and image-text tasks.
  • Standardized Reporting Checklist: Development of a checklist requiring multidimensional metrics and specifying LLM versions.
  • Continuous Evaluation: Ongoing evaluation of newly launched LLMs, especially open-source models.
  • Clinically Driven Assessments: Assessments targeting real-world needs, progressing through a phased validation roadmap (technical benchmarking, retrospective validation, prospective trials) to ensure patient safety.

Summary updated

Abstract

Evidence on LLM evaluations in ophthalmology is extensive but heterogeneous. Most studies have tested a few closed-source LLMs on text-based questions, leaving open-source systems, multimodal tasks, non-English contexts, and real-world deployment underexamined. High methodological variability precludes meaningful performance aggregation, as illustrated by the heterogeneous meta-analysis. Standardized, multimodal benchmarks and phased clinical validation pipelines are urgently needed before LLMs can be safely integrated into eye care workflows.
BibTeX
@article{zhang2025evaluating,
  author = {ZITAO ZHANG and Haiyang Zhang and Z.P. Pan and Zhangqian Bi and Yao Wan and Xuefei Song and Xianqun Fan},
  title = {{Evaluating Large Language Models in Ophthalmology: Systematic Review}},
  journal = {Journal of Medical Internet Research},
  year = {2025},
  doi = {10.2196/76947},
}
Figure