Knowledge Elicitation with Large Language Models for Interpretable Cancer Stage Identification from Pathology Reports

TL;DR

This work tackles the challenge of extracting AJCC TNM T and N stages from unstructured pathology reports using minimal labeled data. It introduces two Knowledge Elicitation approaches for LLMs: KEwLTM, a label-free, long-term memory–driven rule induction from a small set of unannotated reports, and KEwRAG, a retrieval-augmented approach that distills external guideline content into an explicit rule set. Evaluated on TCGA BRCA pathology reports with open-source LLMs, KEwLTM excels when zero-shot reasoning is strong, whereas KEwRAG performs better when retrieval-based guidance is advantageous, with both methods yielding interpretable rule interfaces. The results indicate that these memory- and rule-based elicitation strategies offer scalable, transparent, and data-efficient paths to clinical cancer staging, motivating future generalization to other cancers and clinical tasks.

Abstract

Cancer staging is critical for patient prognosis and treatment planning, yet extracting pathologic TNM staging from unstructured pathology reports poses a persistent challenge. Existing natural language processing (NLP) and machine learning (ML) strategies often depend on large annotated datasets, limiting their scalability and adaptability. In this study, we introduce two Knowledge Elicitation methods designed to overcome these limitations by enabling large language models (LLMs) to induce and apply domain-specific rules for cancer staging. The first, Knowledge Elicitation with Long-Term Memory (KEwLTM), uses an iterative prompting strategy to derive staging rules directly from unannotated pathology reports, without requiring ground-truth labels. The second, Knowledge Elicitation with Retrieval-Augmented Generation (KEwRAG), employs a variation of RAG where rules are pre-extracted from relevant guidelines in a single step and then applied, enhancing interpretability and avoiding repeated retrieval overhead. We leverage the ability of LLMs to apply broad knowledge learned during pre-training to new tasks. Using breast cancer pathology reports from the TCGA dataset, we evaluate their performance in identifying T and N stages, comparing them against various baseline approaches on two open-source LLMs. Our results indicate that KEwLTM outperforms KEwRAG when Zero-Shot Chain-of-Thought (ZSCOT) inference is effective, whereas KEwRAG achieves better performance when ZSCOT inference is less effective. Both methods offer transparent, interpretable interfaces by making the induced rules explicit. These findings highlight the promise of our Knowledge Elicitation methods as scalable, high-performing solutions for automated cancer staging with enhanced interpretability, particularly in clinical settings with limited annotated data.

Figures (6)

• Figure 1: Overview of the KEwLTM workflow
• Figure 2: Overview of the KEwRAG workflow, where relevant text chunks are retrieved before rules are elicited.
• Figure 3: Impact of the number of training reports on KEwLTM performance for T-category classification. The plots show average Precision, Recall, and F1-score over eight random test splits.
• Figure 4: Impact of the number of training reports on KEwLTM performance for N-category classification. The plots show average Precision, Recall, and F1-score over eight random test splits.
• Figure 5: Average long-term memory length evolution for T-category rules with different edit distance threshold conditions during KEwLTM memory induction.