Synthesizing mixed-type longitudinal EHR data with EHR-M-GAN

In this work, we propose a generative adversarial network (GAN) entitled EHR-M-GAN which simultaneously synthesizes mixed-type timeseries EHR data (e.g., continuous-valued timeseries and discrete-valued timeseries). EHR-M-GAN is capable of capturing the multidimensional, heterogeneous, and correlated temporal dynamics in patient trajectories.

This repository contains a Tensorflow implementation of EHR-M-GAN. For details, please see Generating Synthetic Mixed-type Longitudinal Electronic Health Records for Artificial Intelligent Applications. [Arxiv paper link]

Requirements

The code requires

• Python 3.6 or higher
• Tensorflow 1.14.0 or higher
• Numpy
• Sklearn
• Pickle
• Matplotlib
• Seaborn
• Pandas

Datasets

Download

All datasets are publicly available from PhysioNet, and can be downloaded from the following links:

1. MIMIC-III (v1.4)
2. eICU-CRD (v2.0)
3. HiRID (v1.1.1)

Pre-processing

In order to preprocess the datasets for running EHR-M-GAN, please refer to the following repository.

Training

Code explanation

• main_train.py : Use mixed-type timeseries EHR data as training set to generate synthetic data
• networks.py : Components (generators and discriminators in the sequentially coupled GANs, encoders and decoders in the dual-VAE) in the model
• m3gan.py : Pretrain the latent representations and optimize the adversarial learning networks
• Constrastivelosslayer.py : The contrastive loss function in learning the shared VAE representations
• Bilateral_lstm_cell.py : The proposed Bilateral LSTM cell (single-layer)
• Bilateral_lstm_class.py : The proposed Bilateral LSTM network with multiple layers
• init_state.py : Initial state function for recurrent neural networks
• utils.py : other utility fucntions for adversarial training

Running the code

To train the model(s) in the paper, simply run this command:

python main_train.py --dataset mimic --num_pre_epochs 500 --num_epochs 800 --epoch_ckpt_freq 100 

For training the conditional extension of EHR-M-GAN in the paper, run this command:

python main_train.py --dataset mimic --conditional True --num_labels 1 --num_pre_epochs 500 --num_epochs 800 --epoch_ckpt_freq 100