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

BACKGROUND Clinical decision-making in spinal and spinal cord diseases requires a comprehensive assessment of imaging findings, neurological status, bone integrity, and patient-centered goals. Recently, the emergence of Large Language Models (LLMs) has provided new tools for intelligent decision support; however, their reliability and clinical interpretability remain to be systematically evaluated.

METHODS We propose a novel retrieval augmented generation (RAG)-based framework specifically tailored for spinal disease decision support and systematically evaluated four advanced LLM agents (Gemini 2.5 Flash, DeepSeek, GPT-4o, GPT-4o-mini). The framework integrates domain-specific prompting, structured response formatting, and evidence citation tracking.We used 200 real-world spinal cases, each involving diagnostic, therapeutic, and follow-up tasks. Five spine surgeons independently evaluated model outputs using an 11-dimension rubric; each dimension rated on a 5-point Likert scale to evaluate both clinical and technical performance. Inter-group differences were analyzed using the Kruskal-Wallis and Dunn's tests (P < 0.05), with radar plots used for multidimensional visualization.

RESULTS The proposed expert-evaluated framework enables a comprehensive, real-case-based comparison of four LLM agents. Gemini 2.5 Flash achieved the highest overall score (49.25 ± 2.88), significantly outperforming DeepSeek (47.49 ± 3.34, P < 0.001), GPT-4o (45.59 ± 3.89, P < 0.001), and GPT-4o-mini (41.23 ± 5.96, P < 0.001). It demonstrated leading performance particularly in humanistic care, follow-up suggestion, and test recommendation. DeepSeek showed superior capability in differential completeness (mean = 4.76), significantly outperforming the other three models (P < 0.001). Although GPT-4o-mini demonstrated stable system performance (mean = 4.61), it underperformed in core clinical reasoning dimensions. These findings reveal substantial inter-model variability in spine-specific clinical reasoning, an aspect often overlooked in prior non-benchmark LLM evaluations.

CONCLUSION Among the four evaluated LLM agents, Gemini 2.5 Flash and DeepSeek demonstrated superior clinical accuracy, comprehensiveness, and usability in spine-related decision support. These findings support the potential of domain-adapted RAG agents to enhance evidence-based spinal care by providing accurate and comprehensive decision support. Future research should focus on multimodal integration data (e.g., imaging and clinical notes) and conducting prospective validation in real-world clinical environments.