JMIR Publications

ABSTRACT

Background:

Medical AI agents are increasingly proposed as clinical decision support systems integrating multimodal data with iterative reasoning and tool use in clinical workflows.

Objective:

This review synthesizes the emerging literature on medical AI agents for clinical decision support and clarifies architectural design patterns, practices, and implementation challenges.

Methods:

We conducted a narrative literature search in PubMed, Google Scholar, and arXiv for articles published up to May 2025 and performed a thematic synthesis of peer-reviewed studies, reviews, conference proceedings, and selected high-impact preprints.

Results:

We organize agent-based clinical decision support systems using planning, acting, reflecting, and memory (PARM) as an analytical lens. Across diagnostic decision support, treatment planning, and monitoring, common patterns include target decomposition, tool-mediated information retrieval and computation, reflective safety controls, and longitudinal context retention. Major risks include inappropriate autonomy, workflow incompatibility, performance drift, and privacy concerns related to memory.

Conclusions:

Medical AI agents can improve clinical workflow support when designed with limited autonomy, transparent record-keeping, and evaluation beyond traditional accuracy criteria. For responsible deployment under continuous human supervision, governance and validation frameworks emphasizing troubleshooting behavior, cross-contextual reliability, and post-deployment monitoring are essential.