A Bibliometric and Thematic Analysis of LLM ‐Based Agents for Clinical Decision Support: Comparative Topic Modelling and a Five‐Axis Taxonomy

ABSTRACT

Large language model (LLM)‐based agents combine reasoning, external knowledge access, tool use and multiagent coordination in emerging clinical decision support systems. This study maps both the knowledge structure of this literature and the expert‐system design choices represented within it. A PRISMA‐2020‐guided search of the Web of Science Core Collection yielded 40 publications from 2024 to 2025. Bibliometric networks were constructed with VOSviewer, while thematic structure was examined comparatively using NMF, LDA and BERTopic and assessed through reconstruction error, perplexity, log‐likelihood and c‐TF‐IDF–based consistency. The resulting evidence map was then interpreted through a five‐axis taxonomy covering agent organization, coordination and reasoning mechanism, knowledge grounding, clinical function and validation maturity. Annual output increased from 10 to 30 publications. Multiagent clinical decision support emerged as the largest thematic unit and was recovered by both NMF and BERTopic, whereas LDA underperformed at this sample size. BERTopic identified a Retrieval‐Augmented Generation (RAG) themed cluster whose specific composition, including semantically related but lexically dissimilar systems such as TimeCAP and the GPT‐4o cognitive framework, was not reproduced by bag‐of‐words methods at any tested . The taxonomy further shows that organizational complexity, grounding and validation maturity are distinct design dimensions: adding agents or debate does not itself provide external grounding or real‐world validation. The study contributes an integrated account of how the field's bibliometric structure, computational themes and expert‐system architectures relate, while identifying traceability, error control, grounding quality and progression beyond benchmark evaluation as priorities for clinically dependable systems.