Transfer Knowledge from Natural Language to Electrocardiography: Can We Detect Cardiovascular Disease Through Language Models?

Accepted by EACL 2023

Citation

If you feel our code or models helps in your research, kindly cite our papers:

@article{Qiu2023TransferKF,
  title={Transfer Knowledge from Natural Language to Electrocardiography: Can We Detect Cardiovascular Disease Through Language Models?},
  author={Jielin Qiu and William Han and Jiacheng Zhu and Mengdi Xu and Michael Rosenberg and Emerson Liu and Douglas Weber and Ding Zhao},
  journal={ArXiv},
  year={2023},
  volume={abs/2301.09017}
}

Dataset

Use the PTB-XL dataset: https://physionet.org/content/ptb-xl/1.0.2/

Preprocessing

Use the script ecg_preprocess.ipynb

Extract ECG features

Use the script ecg_extract_features.ipynb

(Note the current data is based on the PTB-XL database).

Set up Environment

Create a virtual environment and activate it.

python -m venv .env
source .env/bin/activate

Install basic requirements.

pip install -r requirements.txt

Translation

The original ECG reports from the PTB-XL dataset is in German. We use EasyNMT('opus-mt') model for translation. Please translate the data before training with the translate.py file. Please view the config.py file in tandem and customize as necessary.

Training

The main.py file is used for training selected models. Please view the config.py file in tandem and customize as necessary.

Finetuneing

The finetune.py file is used for finetuning the LLM model. This step is necessary to reproduce the paper results. We finetuned the LLM for 50 epochs. Please view the config.py file in tandem and customize as necessary.

Evaluation

During training in the main.py file, we do include an evaluation step but if you wanted to only run inference on the models, please use evaluate.py. Please view the config.py file in tandem and customize as necessary.

Contact

If you have any question, please contact wjhan@andrew.cmu.edu, jielinq@andrew.cmu.edu.