This repository is the official implementation for UniFL

• 2023.09.16: Support MIMIC-IV 2.0 & Bug Fix

$ conda env create -n unifl -f env.yaml

• MIMIC-III, MIMIC-IV, eICU are public, but require some certificates

• Download the files from below links:

  • MIMIC-III
  • eICU
  • MIMIC-IV

• Then, execute the preprocessing codes

$ cd preprocess
$ bash preprocess_run.sh {MIMIC-III} {MIMIC-IV} {eICU} {save_path}

• Note that the preprocessing takes about 1 hours with AMD EPYC 7502 32-core processor, and it requires more than 60GB of RAM.

$ python src/main.py --device_num 0 --input_path <INPUT_PATH> --save_dir <SAVE_PATH> --train_type fedrated --algorithm fedpxn --pred_target mort --wandb_entity_name <ENTITY_NAME> --wandb_project_name <PROJECT_NAME>

or, you can execute multiple experiments simulatneously with scheduler.py

• Pause & Resume is only supported for fedrated learning (Kubernetis support)
• Pause & Resume is not verified with distributed environment
• We used one A100 80G gpu or two A6000 48G gpus for each run
• Distributed Data Parallel (DDP) with resume is not tested
• You can check hyperparameters on main.py

@article{kim2022universal,
  title={Universal EHR federated learning framework},
  author={Kim, Junu and Hur, Kyunghoon and Yang, Seongjun and Choi, Edward},
  journal={arXiv preprint arXiv:2211.07300},
  year={2022}
}