Generating Multi-Table Time Series EHR from Latent Space with Minimal Preprocessing

TL;DR

This work introduces RawMed, the first framework to synthesize multi-table, time-series EHR data that closely resembles raw EHRs, and proposes a new evaluation framework for multi-table time-series synthetic EHRs, assessing distributional similarity, inter-table relationships, temporal dynamics, and privacy.

Abstract

Electronic Health Records (EHR) are time-series relational databases that record patient interactions and medical events over time, serving as a critical resource for healthcare research and applications. However, privacy concerns and regulatory restrictions limit the sharing and utilization of such sensitive data, necessitating the generation of synthetic EHR datasets. Unlike previous EHR synthesis methods, which typically generate medical records consisting of expert-chosen features (e.g. a few vital signs or structured codes only), we introduce RawMed, the first framework to synthesize multi-table, time-series EHR data that closely resembles raw EHRs. Using text-based representation and compression techniques, RawMed captures complex structures and temporal dynamics with minimal preprocessing. We also propose a new evaluation framework for multi-table time-series synthetic EHRs, assessing distributional similarity, inter-table relationships, temporal dynamics, and privacy. Validated on two open-source EHR datasets, RawMed outperforms baseline models in fidelity and utility. The code is available at https://github.com/eunbyeol-cho/RawMed.

Figures (6)

• Figure 1: Conceptual overview of the RawMed pipeline. Left: Real-world EHR data. Center: Data generation process. Right: Resulting synthetic data. The bottom illustrates conventional approach focused on feature selection and domain-specific engineering, while the top illustrates RawMed, which minimizes domain-dependent preprocessing to generate raw EHR-like data, thereby enhancing user flexibility and utility.
• Figure 2: Comparison of VQ-VAE and RQ-VAE for the patientweight column in MIMIC-IV. Subfigures: (a) real data, (b) VQ reconstruction, (c) VQ with less compression, (d) VQ with doubled codebook, (e) RQ reconstruction, (f) RQ synthetic data. Shared x-axes with proportional y-axes.
• Figure 9: Table-specific visualizations of absolute differences in Pairwise Column Correlation (PCC) matrices between real and synthetic data for MIMIC-IV and eICU tables.
• Figure 15: Table-specific visualizations of absolute differences in Pairwise Column Correlation (PCC) matrices between real and synthetic data for MIMIC-IV and eICU tables.
• Figure 16: Cumulative Distribution Functions (CDFs) of real and synthetic data for MIMIC-IV and eICU datasets, showing Time Gap (with and without simultaneous events), Absolute Time from admission, and Event Counts per patient.