<< Back to the abstract archive
Development of A Novel Artificial Intelligence Large Language Model for Microsurgery: MicrosurgeryLlama2
Berk B Ozmen, MD; Ibrahim Berber, MSc; Graham S Schwarz, MD, MSE, FACS
Cleveland Clinic Department of Plastic Surgery
2025-01-10
Presenter: Berk B Ozmen, MD
Affidavit:
Yes
Director Name: Graham S Schwarz, MD, MSE, FACS
Author Category: Fellow Plastic Surgery
Presentation Category: Clinical
Abstract Category: General Reconstruction
Background: Large language models (LLMs) like ChatGPT have demonstrated remarkable capabilities in processing vast datasets and generating coherent text. However, their application in specialized medical fields remains limited due to a lack of domain-specific knowledge. Aim of this study is to develop a novel domain-specific LLM for microsurgery, addressing the need for AI-driven tools in surgical education and clinical decision support.
Methods: To develop a domain-specific LLM for microsurgery, we curated a corpus of 9,086 microsurgery research abstracts from PubMed, spanning from 2010 to 2024. The open-source Llama-2-13b model served as our foundation. We first trained the model on the curated research abstracts dataset, then fine-tuned it using PyTorch and HuggingFace frameworks. We used Parameter-Efficient Fine-Tuning (PEFT) with Low-Rank Adaptation (LoRA) to enhance model adaptability and reduce computational overhead.
Results: The resulting model, named MicrosurgeryLlama2 demonstrated superior performance compared to the base Llama-2 model across multiple metrics. We observed improvements in BLEU (0.0371 vs. 0.0209), METEOR (0.2249 vs. 0.1168), ROUGE-1 (0.2459 vs. 0.2280), and ROUGE-L (0.1806 vs. 0.1740) scores. These enhancements indicate an improved capability in generating domain-specific, coherent text relevant to microsurgery.
Conclusion: This study introduces MicrosurgeryLlama2 which represents a significant advancement in applying AI to microsurgery education and practice. By leveraging a specialized corpus and advanced fine-tuning techniques, our model outperforms the native LLM, Llama2 in generating relevant, accurate content for the field. This innovation paves the way for AI-assisted clinical decision support and enhanced research capabilities in microsurgery.
