Reliable Generation of Privacy-preserving Synthetic Electronic Health Record Time Series via Diffusion Models

TL;DR

This paper addresses privacy concerns that restrict access to Electronic Health Records by generating realistic, privacy-preserving synthetic EHR time series. It introduces TimeDiff, a mixed diffusion model that jointly handles continuous and discrete time-series variables via Gaussian diffusion for continuous data and multinomial diffusion for discrete data, trained with a time-conditioned BRNN denoiser. The authors show TimeDiff outperforms state-of-the-art baselines in data fidelity and privacy across multiple datasets, and requires less training time than GAN-based methods. The study demonstrates that synthetic data from TimeDiff supports downstream tasks such as in-hospital mortality prediction with data utility close to real data while reducing privacy leakage, making it a practical tool for safe data sharing and augmentation.

Abstract

Electronic Health Records (EHRs) are rich sources of patient-level data, offering valuable resources for medical data analysis. However, privacy concerns often restrict access to EHRs, hindering downstream analysis. Current EHR de-identification methods are flawed and can lead to potential privacy leakage. Additionally, existing publicly available EHR databases are limited, preventing the advancement of medical research using EHR. This study aims to overcome these challenges by generating realistic and privacy-preserving synthetic electronic health records (EHRs) time series efficiently. We introduce a new method for generating diverse and realistic synthetic EHR time series data using Denoising Diffusion Probabilistic Models (DDPM). We conducted experiments on six databases: Medical Information Mart for Intensive Care III and IV (MIMIC-III/IV), the eICU Collaborative Research Database (eICU), and non-EHR datasets on Stocks and Energy. We compared our proposed method with eight existing methods. Our results demonstrate that our approach significantly outperforms all existing methods in terms of data fidelity while requiring less training effort. Additionally, data generated by our method yields a lower discriminative accuracy compared to other baseline methods, indicating the proposed method can generate data with less privacy risk. The proposed diffusion-model-based method can reliably and efficiently generate synthetic EHR time series, which facilitates the downstream medical data analysis. Our numerical results show the superiority of the proposed method over all other existing methods.

Figures (46)

• Figure 1: Visualization of TimeDiff architecture.FC represents a fully connected layer, SiLU is sigmoid linear unit activation, SinuPos Embedding is a shorthand for sinusoidal positional embedding, and GeLU is Gaussian error linear unit activation.
• Figure 2: t-SNE visualization of the eICU ($1^{\text{st}}$ row) and the MIMIC-IV ($2^{\text{rd}}$ row) datasets. Synthetic samples in blue, real training samples in red, and real testing samples in orange.We observe that there is a significant overlap between synthetic samples from TimeDiff and real testing samples, suggesting TimeDiff produces realistic synthetic EHR data. DSPD-GP and HALO also yield noticeable overlap.
• Figure 3: UMAP visualization of the eICU and the MIMIC-IV datasets. Synthetic samples in blue, real training samples in red, and real testing samples in orange.We observe a similar result as the t-SNE visualizations, where there is an overlap between synthetic and real testing samples for TimeDiff. The overlap for other models is less significant.
• Figure 4: TSTR scores compared to TRTR scores (Top); TSRTR scores (Bottom).
• Figure 5: eICU, where the mean is the solid line and $\pm$ one standard deviation is the shaded area.

Theorems & Definitions (2)

• Definition 1
• Definition 2