SMART

Official implementation of SMART: Towards Pre-trained Missing-Aware Model for Patient Health Status Prediction

• Cardiology: https://physionet.org/content/challenge-2012/1.0.0/
• Sepsis: https://physionet.org/content/challenge-2019/1.0.0/
• MIMIC-III: https://physionet.org/content/mimiciii/1.4/

You need to follow the instructions on the PhysioNet website to access the data.

For Cardiology and Sepsis, please follow the jupyter notebook in the data folder to preprocess the data. Please move the files from zips in Cardiology to raw folder before running the notebook.

For MIMIC-III, please follow the instructions in the mimic3-benchmarks repository to extract data from the MIMIC-III database. Notably, we adopt different settings when generating decompensation and length-of-stay data. Please replace the original mimic3benchmark/scripts/create_decompensation.py and mimic3benchmark/scripts/create_length_of_stay.py with the scripts in the data/MIMIC-III folder. After that, you can run the scripts in the jupyter notebook in the data/MIMIC-III folder to generate the tasks.

To train the model, please run the following command:

bash run.sh

It is recommended to use a fixed seed when preprocessing datasets.

@inproceedings{yu2024smart,
  title={SMART: Towards Pre-trained Missing-Aware Model for Patient Health Status Prediction},
  author={Yu, Zhihao and Chu, Xu and Jin, Yujie and Wang, Yasha and Zhao, Junfeng},
  booktitle={NeurIPS 2024},
  year={2024}
}