Engineering a Safety-First Agentic Retrieval-Augmented Generation Architecture for Pediatric Interventional Radiology

Abstract

Background / Purpose

As interventional radiology transitions toward a direct clinical care model, the demand for reliable, localized patient education tools has surged. However, generic large language models pose unacceptable hallucination risks in high-stakes pediatric settings. The purpose of this study is to detail the development of a bilingual artificial intelligence assistant, engineered with an agentic retrieval-augmented generation architecture, to safely deliver institution-specific pediatric guidance.

As interventional radiology transitions toward a direct clinical care model, the demand for reliable, localized patient education tools has surged. However, generic large language models pose unacceptable hallucination risks in high-stakes pediatric settings. The purpose of this study is to detail the development of a bilingual artificial intelligence assistant, engineered with an agentic retrieval-augmented generation architecture, to safely deliver institution-specific pediatric guidance.

Methods

The system was built using a graph-based multi-agent framework to move beyond standard black-box retrieval, prioritizing a five-phase safety-first pipeline. First, an emergency detection layer uses algorithmic screening to redirect urgent medical queries to emergency care. Second, an orchestration agent translates layman queries into clinical, conceptual, and layman search variants to maximize vector retrieval recall. Third, a retrieval and self-correction loop evaluates retrieved institutional documents, automatically rewriting queries if initial relevance is low. Fourth, clinical synthesis generates structured outputs capturing confidence scores, reasoning, and cited sources. Finally, a simulated senior review sub-agent systematically audits drafted answers for factual correctness and clinical safety before delivery.

The system was built using a graph-based multi-agent framework to move beyond standard black-box retrieval, prioritizing a five-phase safety-first pipeline. First, an emergency detection layer uses algorithmic screening to redirect urgent medical queries to emergency care. Second, an orchestration agent translates layman queries into clinical, conceptual, and layman search variants to maximize vector retrieval recall. Third, a retrieval and self-correction loop evaluates retrieved institutional documents, automatically rewriting queries if initial relevance is low. Fourth, clinical synthesis generates structured outputs capturing confidence scores, reasoning, and cited sources. Finally, a simulated senior review sub-agent systematically audits drafted answers for factual correctness and clinical safety before delivery.

Results

An automated expert-agent evaluation framework successfully assessed the pipeline across edge-case clinical queries. Furthermore, the system demonstrated hospital deployment feasibility by comparing a secure, locally-hosted open-weight model against a state-of-the-art cloud-based model, highlighting the trade-offs between local data privacy and advanced reasoning capabilities.

An automated expert-agent evaluation framework successfully assessed the pipeline across edge-case clinical queries. Furthermore, the system demonstrated hospital deployment feasibility by comparing a secure, locally-hosted open-weight model against a state-of-the-art cloud-based model, highlighting the trade-offs between local data privacy and advanced reasoning capabilities.

Conclusion(s)

By integrating a safety-first agentic retrieval-augmented generation pipeline with robust self-correction and clinical auditing loops, the developed system provides a highly transparent, scalable, and safe blueprint for deploying generative artificial intelligence in localized, high-risk clinical environments.

By integrating a safety-first agentic retrieval-augmented generation pipeline with robust self-correction and clinical auditing loops, the developed system provides a highly transparent, scalable, and safe blueprint for deploying generative artificial intelligence in localized, high-risk clinical environments.