Multi-Agent HCC Clinical Decision Support

Context

HCC clinical decision-making draws on hepatology, oncology, and radiology at once. A single LLM reasoning in isolation tends to miss cross-specialty nuance and lacks explicit grounding in current clinical guidelines — both of which matter for research intended to be evaluated by domain experts.

Problem

The goal was a system that reasons closer to a multidisciplinary tumor board — cross-checking specialty perspectives against each other — while staying traceable to guideline literature and auditable enough for clinical research use.

Architecture

A LangGraph-orchestrated multi-agent graph coordinates virtual specialist roles — hepatologist, oncologist, radiologist — each grounded via guideline-focused retrieval. A consensus step reconciles the specialist outputs into a single clinical reasoning trace.

Technical decisions

• Multi-agent consensus chosen over single large-context prompting: specialty-scoped retrieval and role framing reduced cross-domain reasoning errors compared to one model reasoning over everything at once.
• Guideline-grounded RAG used instead of relying on parametric model knowledge, keeping every recommendation traceable to source literature — critical for clinical research validity.
• Evaluated against a benchmark from the group’s public HCC LLM architecture benchmark work.

Evaluation & results

The architecture and its results were validated against domain-expert review criteria as part of the abstract submission process, leading to acceptance at ASCO 2026 — the American Society of Clinical Oncology’s annual meeting.