Clinical Text Summarization by Adapting LLMs | Nature Medicine

Official implementation from Stanford University

Title: Adapted Large Language Models Can Outperform Medical Experts in Clinical Text Summarization
Authors: Dave Van Veen, Cara Van Uden, Louis Blankemeier, Jean-Benoit Delbrouck, Asad Aali, Christian Bluethgen, Anuj Pareek, Malgorzata Polacin, Eduardo Pontes Reis, Anna Seehofnerova Nidhi Rohatgi, Poonam Hosamani, William Collins, Neera Ahuja, Curtis P. Langlotz, Jason Hom, Sergios Gatidis, John Pauly, Akshay S. Chaudhari
Contact: {vanveen} [at] stanford [dot] edu

Datasets

We use six pre-existing open-source datasets which are publicly accessible at the sources cited in our manuscript. Additionally, for datasets which do not require PhysioNet access, we provide our versions in data/:

opi: Open-i (radiology reports)
chq: MeQSum (patient/consumer health questions)
d2n: ACI-Bench (dialogue)

Models

In addition to proprietary models GPT-3.5 and GPT-4, we adapt the following open-source models available from HuggingFace:

Code

Set-up

Use these commands to set up a conda environment:

conda env create -f env.yml
conda activate clin-summ

In src/constants.py, create your own project directory DIR_PROJECT outside this repository which will contain input data, trained models, and generated output.
Move input data from this repo to DIR_PROJECT, i.e. mv data/ DIR_PROJECT
(optional) To add your own dataset, follow the format of example datasets opi, chq, and d2n in DIR_PROJECT/data/

Usage

Below is a description of relevant scripts:

./main.sh: Fine-tune open-source models, query, and compute metrics
python api/main.py: Query OpenAI models and compute metrics
- first enter information for your Azure deployment in src/constants.py via RESOURCE and API_KEY
python src/gen_faiss_idx.py: (new datasets only) Determine set of nearest neighbors training examples for each sample. Alternatively you can sample training examples at random.
src/UMLSScorer.py: Class definition for the MEDCON metric. To implement, follow these steps:
1. Acquire approval for a UMLS license
2. Follow the UMLS download instructions
3. Adapt the provided script, src/UMLSScorer.py
4. Call using the following two lines:
  - scorer = UMLSScorer()
  - medcon_score = scorer(string1, string2)

Citation

@article{vanveen2024clinical,
  title={Adapted Large Language Models Can Outperform Medical Experts in Clinical Text Summarization},
  author={Van Veen, Dave and Van Uden, Cara and Blankemeier, Louis and Delbrouck, Jean-Benoit and Aali, Asad and Bluethgen, Christian and Pareek, Anuj and Polacin, Malgorzata and Collins, William and Ahuja, Neera and Langlotz, Curtis P. and Hom, Jason and Gatidis, Sergios and Pauly, John and Chaudhari, Akshay S.},
  journal={Nature Medicine},
  year={2024},
  doi={10.1038/s41591-024-02855-5},
  url={https://doi.org/10.1038/s41591-024-02855-5},
  published={27 February 2024}
}

License

src/UMLSScorer.py requires a UMLS license.
The remainder of this project is licensed under the MIT License. See the LICENSE file for details.

StanfordMIMI/clin-summ