Developing an Artificial Intelligence Tool for Personalized Breast Cancer Treatment Plans based on the NCCN Guidelines

TL;DR

This paper introduces two AI-driven methods, Agentic-RAG and Graph-RAG, to automate personalized breast cancer treatment planning according to NCCN guidelines, achieving high accuracy and adherence in recommendations.

Contribution

The study presents novel AI methodologies that automate NCCN guideline-based treatment recommendations for breast cancer, improving accuracy and efficiency over existing approaches.

Findings

Agentic-RAG achieved 100% adherence with no hallucinations.

Graph-RAG achieved 95.8% adherence with one incorrect treatment.

Both methods outperformed Chat GPT-4 in adherence and accuracy.

Abstract

Cancer treatments require personalized approaches based on a patient's clinical condition, medical history, and evidence-based guidelines. The National Comprehensive Cancer Network (NCCN) provides frequently updated, complex guidelines through visuals like flowcharts and diagrams, which can be time consuming for oncologists to stay current with treatment protocols. This study presents an AI (Artificial Intelligence)-driven methodology to accurately automate treatment regimens following NCCN guidelines for breast cancer patients. We proposed two AI-driven methods: Agentic-RAG (Retrieval-Augmented Generation) and Graph-RAG. Agentic-RAG used a three-step Large Language Model (LLM) process to select clinical titles from NCCN guidelines, retrieve matching JSON content, and iteratively refine recommendations based on insufficiency checks. Graph-RAG followed a Microsoft-developed framework with proprietary prompts, where JSON data was converted to text via an LLM, summarized, and mapped into graph structures representing key treatment relationships. Final recommendations were generated by querying relevant graph summaries. Both were evaluated using a set of patient descriptions, each with four associated questions. As shown in Table 1, Agentic RAG achieved a 100% adherence (24/24) with no hallucinations or incorrect treatments. Graph-RAG had 95.8% adherence (23/24) with one incorrect treatment and no hallucinations. Chat GPT-4 showed 91.6% adherence (22/24) with two wrong treatments and no hallucinations. Both Agentic RAG and Graph-RAG provided detailed treatment recommendations with accurate references to relevant NCCN document page numbers.