Information Extraction from Clinical Notes: Are We Ready for Adopting Large Language Models?

Overview

This repository contains the code and resources for Kiwi, a LLM-based clinical Information Extraction (IE) system designed to extract medical problem, treatment, test, and drug entities, and their modifiers from clinical note.

Data

For the MTSample dataset, we are providing both entity offsets and texts, as well as the instruction tuning format. The MIMIC-III and 2010 i2b2 datasets, we are only providing entity offsets and texts without the original notes.

Environment

git clone https://github.com/BIDS-Xu-Lab/Kiwi-LLaMA.git
cd Kiwi-LLaMA
pip install -r requirements.txt

Instruction Tuning

To instruction tune the model from scratch, you will have to specify your huggingface token in train.py, and run it by:

### set CUDA_VISIBLE_DEVICES to multiple GPU for multi-GPU training
CUDA_VISIBLE_DEVICES=0,1 accelerate launch train.py

Inference

For inference, you will have to acquire model weights from https://kiwi.clinicalnlp.org/. Then specify the model directory in inference.py, and run it by:

python inference.py

Citation

@article{hu2024information,
  title={Information Extraction from Clinical Notes: Are We Ready to Switch to Large Language Models?},
  author={Hu, Yan and Zuo, Xu and Zhou, Yujia and Peng, Xueqing and Huang, Jimin and Keloth, Vipina K and Zhang, Vincent J and Weng, Ruey-Ling and Chen, Qingyu and Jiang, Xiaoqian and others},
  journal={arXiv preprint arXiv:2411.10020},
  year={2024}
}