# JointCoder: A Multi-Agent Framework-Based Automatic ICD Coding System for Chinese Medical Records

> An open-source project for automatic ICD coding of real-world Chinese electronic medical records, using a multi-agent framework to jointly predict ICD-10 disease codes and ICD-9-CM-3 surgical procedure codes, including real hospital datasets and a complete demonstration system.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-05-08T01:45:47.000Z
- 最近活动: 2026-05-08T02:35:46.614Z
- 热度: 152.2
- 关键词: 医疗AI, ICD编码, 多智能体, 电子病历, 自然语言处理, 中文NLP, 机器学习, 临床决策支持, 医疗信息化
- 页面链接: https://www.zingnex.cn/en/forum/thread/jointcoder-icd
- Canonical: https://www.zingnex.cn/forum/thread/jointcoder-icd
- Markdown 来源: floors_fallback

---

## JointCoder Project Introduction: A Multi-Agent Framework-Driven Automatic ICD Coding System for Chinese Medical Records

JointCoder is an open-source project for automatic ICD coding of real-world Chinese electronic medical records. Its core uses a multi-agent framework to achieve joint prediction of ICD-10 disease codes and ICD-9-CM-3 surgical procedure codes. It also provides real hospital datasets and a complete web demonstration system, aiming to solve the industry pain points of low efficiency and high error rate in manual coding.

## Project Background: Automation Challenges of Medical ICD Coding

ICD coding is a core part of medical informatization. It requires professionals to convert clinical diagnoses and surgical procedures into standard codes, but manual coding is time-consuming, labor-intensive, and easily affected by subjective factors. JointCoder provides a solution to this pain point, supporting joint prediction of disease and surgical codes for Chinese electronic medical records.

## Core Innovation: Multi-Agent Collaborative Coding Framework

JointCoder adopts a multi-stage multi-agent architecture aligned with workflow, decomposing the coding task into multiple collaborative stages, each handled by a dedicated agent (different from traditional end-to-end single models). This design improves interpretability (intermediate results can be reviewed) and controllability, and can flexibly introduce domain knowledge. It also supports joint prediction of disease and surgical codes, capturing their correlations to improve overall accuracy.

## Dataset and Performance

The project provides a desensitized dataset from real tertiary hospitals (6747 inpatient records, including ICD-10/ICD-9-CM-3 codes), which has authenticity, completeness, and Chinese characteristics. Performance evaluation shows: JointCoder outperforms traditional machine learning, deep learning, and commercial APIs in disease and surgical coding tasks; joint coding achieves better results; ablation experiments verify the positive contributions of components like multi-agent collaboration and workflow alignment.

## System Functions and Demonstration

JointCoder provides a complete web demonstration system with four functional modules: 1. Demo case selection (4 preloaded samples, supporting OCR/desensitization/editing/parallel coding); 2. Single PDF upload (local PDF processing); 3. Batch PDF upload; 4. Excel batch submission (no OCR needed, higher efficiency). After coding, it supports feedback, marking as completed, and result export.

## Application Prospects and Limitations

Application Prospects: Hospitals can improve coding efficiency and accuracy; medical insurance departments can support reasonable payment and disease monitoring; medical research can help with epidemiological analysis and clinical decision support system development. Limitations: Data privacy (upload is disabled in public demonstrations), domain adaptability (needs fine-tuning to adapt to other hospitals/departments), coding rules need regular updates and maintenance.

## Open-Source Contributions and Community Value

JointCoder is released as open-source on GitHub, providing code, pre-trained models, detailed documentation, and examples. Open-source lowers the threshold for Chinese medical AI research, provides the community with real datasets, multi-agent frameworks, and demonstration systems, and promotes the development of the field.