Generating Multi-Table Time Series EHR from Latent Space

RawMed is the first framework to synthesize multi-table, time-series Electronic Health Records (EHR) data that closely resembles raw EHRs. Using text-based representation and compression techniques, RawMed captures complex structures and temporal dynamics with minimal preprocessing.

🎯 Key Features

• Multi-table time-series generation: Synthesizes raw EHR data across multiple tables
• Minimal preprocessing: Uses text-based representation to preserve complex structures
• Comprehensive evaluation: Assesses distributional similarity, inter-table relationships, temporal dynamics, and privacy

🚀 Quick Start

Installation

# Create conda environment
conda create -n rawmed python=3.9
conda activate rawmed

# Install dependencies
pip install sacred==0.8.5
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install tqdm pandas scikit-learn transformers wandb einops dython rapidfuzz xgboost h5py

Data Preprocessing

Preprocess your EHR dataset using the integrated pipeline (see integrated-ehr-pipeline/README.md), then convert to NumPy format:

python ehrsyn/datamodules/preprocess.py

Training

Train the model in two stages. Both scripts take the following arguments:

bash run_scripts/<script>.sh [dataset] [obs_window] [cuda_device] [data_root] [ckpt_root] [syn_data_root]

Argument	Description	Example
dataset	Dataset name (eicu or mimiciv)	eicu
obs_window	Observation window in hours (6, 12, or 24)	12
cuda_device	Comma-separated GPU IDs	0 or 0,1,2,3
data_root	Path to preprocessed NumPy data	/path/to/data
ckpt_root	Path to save/load model checkpoints	/path/to/ckpts
syn_data_root	Path to save generated synthetic data	/path/to/output

Stage 1: Train RQ-VAE for event compression

bash run_scripts/train_RQVAE.sh eicu 12 0 /path/to/data /path/to/ckpts /path/to/output

Stage 2: Train AR model and sample synthetic data

Training and sampling are separated within the script. Uncomment the training block to train, then run sampling:

bash run_scripts/train_AR.sh eicu 12 0,1,2,3 /path/to/data /path/to/ckpts /path/to/output

When multiple GPUs are specified (e.g., 0,1,2,3), sampling is automatically parallelized across GPUs via run_scripts/parallel_sample.py. Each GPU runs an independent process with a different random seed, and results are concatenated.

Evaluation

Evaluate the generated synthetic data using MTTSeval, a comprehensive evaluation framework for multi-table time-series EHRs. The evaluation pipeline consists of two steps:

1. Postprocess: Convert generated NumPy arrays to table format (CSV)
2. Evaluate: Run evaluation metrics on the postprocessed tables

cd MTTSeval
bash run_postprocess.sh [dataset] [obs_window] [cuda_device] [mode] [data_root] [syn_data_root]
bash run_eval.sh [dataset] [obs_window] [cuda_device] [data_root] [syn_data_root]

MTTSeval provides various evaluation metrics including:

• Statistical distribution analysis: Compares distributional similarity between real and synthetic data
• Correlation analysis: Evaluates inter-table relationships and correlations
• Temporal dynamics analysis: Assesses time-series patterns and temporal consistency
• Utility evaluation: Includes TSTR (Train on Synthetic, Test on Real) and prediction similarity metrics

See MTTSeval/README.md for detailed setup and usage instructions.

📖 Citation

If you use RawMed in your research, please cite:

@article{cho2025rawmed,
  title={Generating Multi-Table Time Series EHR from Latent Space with Minimal Preprocessing},
  author={Cho, Eunbyeol and Kim, Jiyoun and Lee, Minjae and Park, Sungjin and Choi, Edward},
  journal={arXiv preprint arXiv:2507.06996},
  year={2025}
}