RRNorm: A Novel Framework for Chinese Disease Diagnoses Normalization via LLM-Driven Terminology Component Recognition and Reconstruction

Yongqi Fan, Yansha Zhu, Kui Xue, Jingping Liu, Tong RuanACL 2024 · DOI

Abstract

RRNorm is a novel framework designed to address the "multi-implication" challenge in Chinese clinical terminology normalization. By leveraging Large Language Models (LLMs) to decompose complex medical mentions into atomic components and employing a specialized recall-and-rank pipeline, RRNorm achieves state-of-the-art performance on standard benchmarks. The framework significantly improves the accuracy of mapping non-standard clinical notes to standardized medical terms by resolving semantic blurring in multi-diagnosis strings.

Introduction

Clinical Terminology Normalization (CTN) is a fundamental task in medical informatics, mapping colloquial clinical descriptions to standardized codes like ICD-10. In Chinese medical records, this task is frequently complicated by the multi-implication problem, where a single string contains multiple distinct diagnoses (e.g., "left knee degeneration with loose body"). Traditional models often fail by attempting to map these complex strings to a single term, resulting in semantic blurring and lost information.

Figure 1: Comparison between traditional normalization and the proposed RRNorm approach, highlighting the decomposition into atomic mentions.

Methodology

The RRNorm framework consists of three primary stages designed to handle complexity through decomposition and structured learning:

  1. Recognition and Reconstruction (RR): Utilizing an LLM (ChatGPT), the system identifies key components such as Disease Content (the pathology), Disease Scope (the anatomical site), and Modifiers. These components are then reconstructed into "atomic mentions"—simplified phrases each representing a single medical concept.
  2. Atomic-Sampling-based Contrastive Learning: To improve the "Recall" stage, the authors use an LLM for automatic annotation of atomic mentions in the training set. A Sentence-BERT (SBERT) model is then trained via contrastive learning to ensure semantically similar medical concepts are clustered together in the embedding space.
  3. Knowledge-Guided Term Selection: In the final "Ranking" stage, a BERT-based classifier compares atomic mentions against candidate standard terms. It utilizes component-specific masks (e.g., masking the anatomical site when identifying the pathology) to focus the model on the most discriminative textual features.
Figure 3: The overall architecture of the RRNorm framework, illustrating the flow from raw mention to normalized standard terms through recognition, reconstruction, and ranking.

Datasets and Evaluation

The framework was evaluated on the CHIP-CDN benchmark, the primary standard for Chinese clinical terminology normalization. Additionally, the authors introduced CHIP-CDN-RR, a reorganized version of the dataset where complex mentions are mapped to their constituent atomic mentions. This reconstruction improved the "uni-implication" ratio by 37%, providing a cleaner signal for model training.

Key Results

RRNorm achieved State-of-the-Art (SOTA) results on the CHIP-CDN dataset:

  • Micro-F1 Score: Reached 64.20, surpassing the previous benchmark of 63.10.
  • Recall Improvement: The "RR" strategy consistently enhanced various base models; for instance, it boosted the hit rate (HR@10) of the M3E model from 40.40% to 51.30%.
  • Multi-Implication Handling: The framework proved particularly robust in handling "inseparable" multi-implication cases where traditional word segmentation methods typically fail.

Conclusion and Future Work

RRNorm demonstrates that decomposing complex medical language into structured components is a highly effective strategy for normalization. By combining the structural parsing capabilities of LLMs with specialized contrastive learning and ranking techniques, the framework effectively mitigates the multi-implication problem. Future research will focus on generalizing these component types to broader clinical domains, such as surgical procedures and medication normalization, and reducing the dependency on high-cost LLM inference.

Summary updated

BibTeX
@inproceedings{fan2024rrnorm,
  author = {Yongqi Fan and Yansha Zhu and Kui Xue and Jingping Liu and Tong Ruan},
  title = {{RRNorm: A Novel Framework for Chinese Disease Diagnoses Normalization via LLM-Driven Terminology Component Recognition and Reconstruction}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
  doi = {10.18653/v1/2024.findings-acl.544},
}
Figure