Awesome

Warpformer

This is an offical implementation of Warpformer: A Multi-scale Modeling Approach for Irregular Clinical Time Series.

UPDATE (2024/06/09)

If you ran this script before 2024/06/09 and were unable to reproduce the results from the paper on the PhysioNet dataset, please follow these steps:

1. Remove the /path/to/datasets/physionet/ folder.
2. Set the --task in run.sh to physio.
3. Re-run the run.sh script.

Setup

Requirements

Your local system should have the following executables:

• conda
• Python 3.7 or later
• git
• PostreSQL (Opt.)

Create conda environment

All instructions below should be executed from a terminal.

1. clone this repository and run

cd Warpformer

2. create an environment warpformer and activate it.

conda create -n warpformer python=3.7
conda activate warpformer

3. install the required Python modules using file requirements.txt.

pip install -r requirement.txt

Tips: If you cannot install psycopg2 successfully, please try sudo apt-get install libpq-dev first and then run the command pip install psycopg2.

Download Clinical Data & Task Building

Our benchmark includes a diverse set of clinical tasks covering different clinical scenarios, with 61 common physiology signals and 42 widely used interventions in intensive care units. The following table summarizes the statistics of these tasks.

I. Access to MIMIC-III data

1. First you need to have an access to MIMIC-III Dataset, which can be requested here. The database version we used here is v1.4.
2. Download the MIMIC-III Clinical Database and place the MIMIC-III Clinical Database as either .csv or .csv.gz files somewhere on your local computer.

II. Generate datasets

1. Modify the mimic_data_dir variable to the path of MIMIC-III folder in the ./preprocess/preprocess_mimic_iii_large.py file, and run

   cd preprocess
   python preprocess_mimic_iii_large.py
   

2. (opt.) If you want to build the CIP task, it is necessary to obtain a file all_hourly_data.h5 first. Otherwise, you can skip this step.

   1. Start by referring to the MIT-LCP repository to create a database where you can host the MIMIC-III data.
   2. Next, you will need to generate the file named all_hourly_data.h5 using the MIMIC_Extract. Once generated, place the all_hourly_data.h5 file in the save_data_folder/tmp/ directory. You can locate the interv_outPath variable in the split_data_preprocessing_large.py file and ensure that it points to the correct destination, i.e., the save_data_folder/tmp/ folder.

3. Modify the data_root_folder and mimic_data_dir variables to the MIMIC-III folder path in the split_data_preprocessing_large.py, and run the following command for downstream tasks generation:

   python split_data_preprocessing_large.py
   

Run the Model

Please refer to the script run.sh for specific configurations used to obtain the main experimental results in the paper. To reproduce the results, please uncomment the corresponding commands and run:

bash run.sh

Example:

python Main_warp.py --task {task_name} \
               --data_path {path_to_data_folder} \
               --log {log_path} \
               --save_path {save_path}
               --epoch {epoch} \
               --seed {seed} \
               --lr 0.001 \
               --batch {batch_size} \
               --warp_num {warp_num} \
               --dp_flag

• task: the downstram task name, select from [mor, decom, cip, wbm, los, active, physio].
• seed: the seed for parameter initialization.
• warp_num: customize $L^{(n)}$ for each warp layer. The first layer is always 0, and each layer is split by '_', e.g., '0_12_6' (unnormalized version) or '0_0.2_1.2' (normalized version).
• dp_flag: use DataParallel for training or not.

For more details, please refer to run.sh and Main_warp.py.

License

The original MIMIC-III database is hosted and maintained on PhysioNet under PhysioNet Credentialed Health Data License 1.5.0, and is publicly accessible at https://physionet.org/content/mimiciii/1.4/.

Our code in this repository is licensed under the MIT license.