HARMON-E: Hierarchical Agentic Reasoning for Multimodal Oncology Notes to Extract Structured Data

This paper introduces HARMON-E, a novel agentic framework leveraging LLMs for extracting structured oncology data from unstructured clinical notes. The approach addresses the limitations of existing methods by employing context-sensitive retrieval and iterative synthesis to handle variability, specialized terminology, and inconsistent document formats. The framework's ability to decompose complex extraction tasks into modular, adaptive steps is a key strength. The impressive F1-score of 0.93 on a large-scale dataset demonstrates the potential of HARMON-E to significantly improve the efficiency and accuracy of oncology data extraction, facilitating better treatment decisions and research. The focus on patient-level synthesis across multiple documents is particularly valuable.

• •HARMON-E uses LLMs as reasoning agents for oncology data extraction.
• •The framework achieves high accuracy (F1-score of 0.93) on a large dataset.
• •The approach focuses on patient-level synthesis across multiple clinical documents.