Skip to main content
  1. PaperReading/
  2. NIPS/

Automated clinical coding using off-the-shelf large language models

NIPS LLM ICD Tree 2023

NIPS Workshop 2023

Automated clinical coding using off-the-shelf large language models | OpenReview

测试了Llama-2, GPT-3.5 和GPT-4在CodiEsp数据集 (1000cases)上的性能

macro-F1 micro-F1
Proposed(LLM inference) 0.225 0.157
PLM-ICD(BERT based, pretrained on MIMIC III & VI, SOTA) 0.216 0.219

image-20240519145057680

Prompt:“You are a clinical coder, consider the case note and assign the appropriate ICD codes”

LLM控温:GPT设为0,Llama设为0.001(最小值),以获得确定性输出。

后处理:将LLM生成的文本与ICD code description进行贪婪匹配。

StefanoTrv/simple_icd_10_CM: A simple python library for ICD-10-CM codes (github.com)

JoakimEdin/medical-coding-reproducibility (github.com)

直接匹配代码/匹配代码描述/树形搜索的对比

image-20240519145710924

层级准确率

image-20240519145630294

LLM可能会产生(根据先验知识)互斥的结果