Interpretable Cancer Staging Using Knowledge Elicitation in Large Language Models

Cancer staging is critical for patient prognosis and treatment planning, yet extracting pathologic TNM staging from unstructured pathology reports is challenging. In this study, we introduce two LLM-based knowledge elicitation methods—one using a Long-Term Memory approach (KEwLTM) and another employing Retrieval-Augmented Generation (KEwRAG)—to enable large language models (LLMs) to induce domain-specific rules for cancer staging.