An Agentic LLM Pipeline for Structured Adaptation of Clinical AI Literature: Construction and Validation of HOps

Bold Bayarbaatar

Keywords: LLM, Agentic AI, Unstructured data, Structured Data

Abstract:

Actionable clinical AI guidance is dispersed
across hundreds of heterogeneous, unstructured
PDF documents spanning reporting guidelines,
implementation science frameworks, and digital
health evaluation tools. We present a six-agent
sequential LLM pipeline that converts this unstructured
medical literature into a structured,
machine-readable tabular dataset at scale. Applied
to 55 clinical AI framework documents, the
pipeline extracted 2,386 classified items mapped
to eight predefined lifecycle stages, revealing systematic
coverage gaps across the literature. This
structured synthesis directly informed the development
of HOps, a 20-question guiding framework
for hospital digital product development, which
achieved 87% endorsement and 67% questionlevel
agreement across five domain experts. Our
work demonstrates that agentic structured adaptation
of clinical literature is feasible, produces reproducible
outputs, and yields structured datasets
with direct utility for clinical AI governance and
deployment research.

Submission Number: 136