Advancing Question-Answering in Ophthalmology With Retrieval-Augmented Generation: Benchmarking Open-Source and Proprietary Large Language Models
Post-publication record
OpenAlex flags this work as retracted, but it carries no matching Retraction Watch record in this frame.
Bibliographic record
Abstract
Purpose: The purpose of this study was to evaluate the application of combining information retrieval with text generation using Retrieval-Augmented Generation (RAG) to benchmark the performance of open-source and proprietary generative large language models (LLMs) in question-answering in ophthalmology. Methods: Our dataset comprised 260 multiple-choice questions sourced from two question-answer banks designed to assess ophthalmic knowledge: the American Academy of Ophthalmology's (AAO) Basic and Clinical Science Course (BCSC) Self-Assessment program and OphthoQuestions. Our RAG pipeline retrieves documents in the BCSC companion textbook using ChromaDB, followed by reranking with Cohere to refine the context provided to the LLMs. Generative Pretrained Transformer (GPT)-4-turbo and 3 open-source models (Llama-3-70B, Gemma-2-27B, and Mixtral-8 × 7B) are benchmarked using zero-shot, zero-shot with Chain-of-Thought (zero-shot-CoT), and RAG. Model performance is evaluated using accuracy on the two datasets. Quantization is applied to improve the efficiency of the open-source models. Effects of quantization level are also measured. Results: Using RAG, GPT-4-turbo's accuracy increased by 11.54% on BCSC and by 10.96% on OphthoQuestions. Importantly, the RAG pipeline greatly enhances overall performance of Llama-3 by 23.85%, Gemma-2 by 17.11%, and Mixtral-8 × 7B by 22.11%. Zero-shot-CoT had overall no significant improvement on the models' performance. Quantization using 4 bit was shown to be as effective as using 8 bits while requiring half the resources. Conclusions: Our work demonstrates that integrating RAG significantly enhances LLM accuracy especially for smaller LLMs. Translation Relevance: Using our RAG, smaller privacy-preserving open-source LLMs can be run in sensitive and resource-constrained environments, such as within hospitals, offering a viable alternative to cloud-based LLMs like GPT-4-turbo.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.008 | 0.023 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.002 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.002 | 0.004 |
| Open science | 0.003 | 0.003 |
| Research integrity | 0.003 | 0.003 |
| Insufficient payload (model declined to judge) | 0.004 | 0.003 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".