Code Like Humans: A Multi-Agent Solution for Medical Coding

2025-09-04

AI Summary

Medical coding requires mapping unstructured clinical text to over 70,000 ICD-10 diagnosis and procedure codes, with low accuracy and poor coverage for rare diseases representing a critical bottleneck. This paper proposes the first multi-agent generative coding framework designed for the full ICD-10 taxonomy: specialized agents emulate core expert decision steps—including terminology standardization, rule-based reasoning, and guideline verification—collaboratively adhering to official coding conventions. We introduce discriminative fine-tuning to enhance recognition of frequent codes and systematically identify model blind spots for rare codes. Experiments demonstrate state-of-the-art performance on rare diagnosis code prediction, the first end-to-end support for the complete ICD-10 code set, and the first empirical quantification of inherent systematic under-coding bias in generative models—revealing persistent gaps in rare-code recall despite high overall accuracy.

Technology Category

• Cognitive Modeling & Cognitive Systems: Agent Architectures
• Multiagent Systems: Modeling other Agents
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Economics, Online Markets and Human Computation: Trust and reliance of crowd workers and data experts on GenAI

Abstract

In medical coding, experts map unstructured clinical notes to alphanumeric codes for diagnoses and procedures. We introduce Code Like Humans: a new agentic framework for medical coding with large language models. It implements official coding guidelines for human experts, and it is the first solution that can support the full ICD-10 coding system (+70K labels). It achieves the best performance to date on rare diagnosis codes (fine-tuned discriminative classifiers retain an advantage for high-frequency codes, to which they are limited). Towards future work, we also contribute an analysis of system performance and identify its `blind spots' (codes that are systematically undercoded).

Problem

• Mapping unstructured clinical notes to medical codes
• Implementing official coding guidelines for experts
• Supporting full ICD-10 system with rare codes

Innovation

• Multi-agent framework with large language models
• Implements official human coding guidelines
• Supports full ICD-10 system with 70K+ labels