Explainable and evidence-linked recommendations for spine surgery via a retrieval-augmented LLM agent

Area of Science:

• Artificial Intelligence in Medicine
• Spine Surgery Decision Support
• Clinical Informatics

Background:

• Clinical decision-making for spinal diseases requires integrating imaging, neurological status, bone integrity, and patient goals.
• Large Language Models (LLMs) offer potential for intelligent decision support but need rigorous evaluation for reliability and interpretability.

Purpose of the Study:

• To propose and evaluate a novel retrieval augmented generation (RAG)-based framework for spinal disease decision support.
• To systematically compare the performance of four advanced LLM agents (Gemini 2.5 Flash, DeepSeek, GPT-4o, GPT-4o-mini) within this framework.

Main Methods:

• A RAG framework was developed, incorporating domain-specific prompting, structured responses, and evidence tracking.
• 200 real-world spinal cases were used to assess diagnostic, therapeutic, and follow-up tasks.
• Five spine surgeons evaluated LLM outputs using an 11-dimension rubric, with statistical analysis of inter-group differences.

Main Results:

• Gemini 2.5 Flash achieved the highest overall score, significantly outperforming other models.
• DeepSeek demonstrated superior performance in differential diagnosis completeness.
• Substantial inter-model variability was observed in spine-specific clinical reasoning.

Conclusions:

• Gemini 2.5 Flash and DeepSeek exhibited superior clinical accuracy, comprehensiveness, and usability for spine decision support.
• Domain-adapted RAG agents show promise for enhancing evidence-based spinal care.
• Future work should explore multimodal data integration and prospective clinical validation.