Development and Evaluation of a Retrieval-Augmented Large Language Model Framework for Ophthalmology

Abstract

Importance: Although augmenting large language models (LLMs) with knowledge bases may improve medical domain-specific performance, practical methods are needed for local implementation of LLMs that address privacy concerns and enhance accessibility for health care professionals.

Objective: To develop an accurate, cost-effective local implementation of an LLM to mitigate privacy concerns and support their practical deployment in health care settings.

Design, setting, and participants: ChatZOC (Sun Yat-Sen University Zhongshan Ophthalmology Center), a retrieval-augmented LLM framework, was developed by enhancing a baseline LLM with a comprehensive ophthalmic dataset and evaluation framework (CODE), which includes over 30 000 pieces of ophthalmic knowledge. This LLM was benchmarked against 10 representative LLMs, including GPT-4 and GPT-3.5 Turbo (OpenAI), across 300 clinical questions in ophthalmology. The evaluation, involving a panel of medical experts and biomedical researchers, focused on accuracy, utility, and safety. A double-masked approach was used to try to minimize bias assessment across all models. The study used a comprehensive knowledge base derived from ophthalmic clinical practice, without directly involving clinical patients.

Exposures: LLM response to clinical questions.

Main outcomes and measures: Accuracy, utility, and safety of LLMs in responding to clinical questions.

Results: The baseline model achieved a human ranking score of 0.48. The retrieval-augmented LLM had a score of 0.60, a difference of 0.12 (95% CI, 0.02-0.22; P = .02) from baseline and not different from GPT-4 with a score of 0.61 (difference = 0.01; 95% CI, -0.11 to 0.13; P = .89). For scientific consensus, the retrieval-augmented LLM was 84.0% compared with the baseline model of 46.5% (difference = 37.5%; 95% CI, 29.0%-46.0%; P < .001) and not different from GPT-4 with a value of 79.2% (difference = 4.8%; 95% CI, -0.3% to 10.0%; P = .06).

Conclusions and relevance: Results of this quality improvement study suggest that the integration of high-quality knowledge bases improved the LLM's performance in medical domains. This study highlights the transformative potential of augmented LLMs in clinical practice by providing reliable, safe, and practical clinical information. Further research is needed to explore the broader application of such frameworks in the real world.

Figure 1.. Human Evaluation Results of Responses Generated by Large Language Models (LLMs) in Terms of Accuracy

A total of 300 randomly selected question-answer pairs generated by 11 LLMs were all manually validated. Accuracy was subdivided into 3 subcategories, including scientific consensus, missing content, and possible bias. The orange bars represented our ophthalmic LLM in comparison to the baseline pretrained model. The longer length of the dark bars signified better model performance. The LLMs featured in this figure are as follows: Baichuan-7B (Baichuan-Inc), Baichuan-13B (Baichuan-Inc), Baichuan-13B+COD (ZOC), ChatGLM-6B (THUDM), ChatGLM2-6B (THUDM), Chatyuan (ClueAI), GPT-4 (OpenAI), GPT-3.5 Turbo (OpenAI), Llama2-Chat-70B (Meta), Llama2-Chinese-Chat-13B (FlagAlpha), StableVicuna-13B (CarperAI).

Figure 2.. Human Evaluation Results of Responses Generated by Large Language Models (LLMs) in Terms of Utility

A total of 300 randomly selected question-answer pairs generated by 11 LLMs were all manually validated. Utility was subdivided into 3 subcategories, including correct understanding, correct retrieval, and correct reasoning. The orange bars represented our ophthalmic LLM in comparison to the baseline pretrained model. The longer length of the dark bars signified better model performance. The LLMs featured in this figure are as follows: Baichuan-7B (Baichuan-Inc), Baichuan-13B (Baichuan-Inc), Baichuan-13B+COD (ZOC), ChatGLM-6B (THUDM), ChatGLM2-6B (THUDM), Chatyuan (ClueAI), GPT-4 (OpenAI), GPT-3.5 Turbo (OpenAI), Llama2-Chat-70B (Meta), Llama2-Chinese-Chat-13B (FlagAlpha), StableVicuna-13B (CarperAI).

Figure 3.. Human Evaluation Results of Responses Generated by Large Language Models (LLMs) in Terms of Safety

A total of 300 randomly selected question-answer pairs generated by 11 LLMs were all manually validated. Safety was subdivided into 3 subcategories, including inappropriate/wrong content, possible hazard, and hazard potential. The orange bars represented our ophthalmic LLM in comparison to the baseline pretrained model. The longer length of the dark bars signified better model performance. The LLMs featured in this figure are as follows: Baichuan-7B (Baichuan-Inc), Baichuan-13B (Baichuan-Inc), Baichuan-13B+COD (ZOC), ChatGLM-6B (THUDM), ChatGLM2-6B (THUDM), Chatyuan (ClueAI), GPT-4 (OpenAI), GPT-3.5 Turbo (OpenAI), Llama2-Chat-70B (Meta), Llama2-Chinese-Chat-13B (FlagAlpha), StableVicuna-13B (CarperAI).