Augmenting Large Language Models With Automated, Bibliometrics-Powered Literature Search for Knowledge Distillation:...

Area of Science:

• Medical Informatics
• Bibliometrics
• Artificial Intelligence

Background:

• Medical literature is rapidly expanding, making it difficult to stay current with neurosurgical advancements.
• Large Language Models (LLMs) offer potential for text summarization but cannot autonomously conduct literature reviews and may hallucinate sources.
• A new strategy is needed to efficiently distill and synthesize vast amounts of medical research.

Purpose of the Study:

• To develop and demonstrate a novel strategy for automated summarization and citation of seminal neurosurgical articles.
• To combine Reference Publication Year Spectroscopy with LLMs for efficient literature review.
• To validate the approach for four common spinal conditions.

Main Methods:

• Reference Publication Year Spectroscopy identified foundational articles for cervical myelopathy, lumbar radiculopathy, lumbar stenosis, and adjacent segment disease.
• Article texts were processed, and relevant chunks were retrieved using chain-of-thought prompting in a vector database.
• LLMs generated summaries with cited facts and statistics, followed by manual verification and faculty surveys.

Main Results:

• The combined approach was cost-effective (<$1 per condition) and rapid (under 5 minutes).
• Generative Pre-trained Transformer-4 achieved 97.5% citation accuracy.
• AI-generated summaries provided high-fidelity, clinically relevant information, refined by expert feedback.

Conclusions:

• A generalizable workflow for rapid, automated summarization of seminal spinal pathology articles was demonstrated.
• This strategy fuses bibliometrics and AI to automate knowledge distillation, reducing the need for manual literature review.
• The method is implementable using standard hardware and offers a path toward fully automated scientific knowledge synthesis.