MTGAN

Code for the paper: Multi-Label Clinical Time-Series Generation via Conditional GAN

Commands

Preprocess Data

1. Go to https://mimic.physionet.org/ for access. Once you have the authority for the dataset, download the dataset and extract the csv files to {data_path}/mimic3/raw/ and {data_path}/mimic4/raw/.
2. Check hyper-parameter settings:

python run_preprocess.py --help

3. Run preprocess program:

• For MIMIC-III

python run_preprocess.py --dataset mimic3 --train_num 6000

• For MIMIC-IV

python run_preprocess.py --dataset mimic4 --train_num 6000 --sample_num 10000

Add --from_saved if you run it again and want to load saved encoded data.

Training MTGAN

1. Check hyper-parameter setting:

python run_train.py --help

2. Run training program

• For MIMIC-III

python run_train.py --dataset mimic3

• For MIMIC-IV

python run_train.py --dataset mimic4

3. The parameters are saved in default at results/{dataset}/params/. The training plots are saved at results/{dataset}/records/.

Generating EHR data

1. Check hyper-parameter setting:

python run_generate.py --help

2. Run generating program

• For MIMIC-III

python run_generate.py --dataset mimic3

• For MIMIC-IV

python run_generate.py --dataset mimic4

3. The generated data are saved at results/synthetic_{dataset}.npz.

Requirements:

Environment

• Python >= 3.7
• Virtualenv (optional, recommended)
• CUDA (optional, recommended)
• RAM > 16GB

Installing packages:

• matplotlib
• numpy
• openpyxl
• pandas
• scipy
• pytorch