A Distributed Local LLM and Centrally Governed RAG Architecture for Privacy-Preserving Cancer Registry Support

Open Access
Article
Conference Proceedings
Authors: Mitsuhiro Takasaki, Kazumi Sasaki, Narue Nakabayashi, Megumi Hori, Tomohiro Matsuda
Abstract

Hospital-based cancer registration requires registrars to interpret detailed and frequently updated coding rules, disease-specific manuals, staging classifications, and historical question-and-answer records. Retrieval-augmented generation (RAG) can provide natural-language access to these resources, but cloud-based implementations raise concerns about transmitting clinical information outside the hospital. Moreover, retrieval of the correct document does not guarantee correct application of a registration rule. This paper presents a prototype architecture for deploying a local large language model and vector database within each hospital, while centrally governing and distributing approved reference documents, metadata, retrieval policies, and selected deterministic validation rules. The prototype uses FastAPI, Ollama, ChromaDB, multilingual embeddings, and locally executable language models. Questions are routed to prioritized document groups including the standard registration form, ICD-O materials, staging and site-specific resources, and historical Q&A records. For closed, explicitly enumerated, or high-risk rules, generative processing is supplemented by deterministic safeguards. A formative evaluation involving prostate-specific antigen illustrates the limitation of conventional RAG: although PSA is clinically associated with prostate cancer, it is not one of the four combinations permitted as a site-specific tumor marker for diagnostic-basis coding in the Japanese cancer registry. The case illustrates that semantic plausibility and registry correctness are not necessarily equivalent. The proposed functional allocation assigns natural-language interpretation, retrieval, synthesis, and explanation to the LLM; exact membership checking to deterministic components; and final confirmation to the registrar. This hybrid design is intended to improve privacy, traceability, maintainability, and calibrated human reliance.

Keywords: Cancer Registry, Retrieval-Augmented Generation, Local Large Language Model, Privacy-Preserving AI, Human-in-the-Loop, Deterministic Guardrail

DOI: 10.54941/ahfe1008450

Cite this paper
Downloads
0
Visits
1
Download PDF

More from this volume

← Human and psychological factors and seasickness in transitions at seaPerceived Walking duration across Street-tree Spatial Patterns: Toward a Route Recommendation System →
View all articles in Human Factors in Design, Engineering, and Computing →