TarDiff: Target-Oriented Diffusion Guidance for Synthetic Electronic Health Record Time Series Generation

TL;DR

TarDiff tackles the challenge of generating synthetic EHR time series that are not only realistic but also maximally useful for downstream clinical tasks. It introduces an influence-guided diffusion framework that quantifies expected performance gains via influence functions and embeds this signal into the reverse diffusion process to produce task-optimized data. The method combines a pre-trained downstream model, gradient-based influence estimation, and diffusion sampling guided by this influence to yield high-utility synthetic sequences. Across six public datasets, TarDiff achieves state-of-the-art improvements, including up to $20.4\%$ in AUPRC and $18.4\%$ in AUROC, and demonstrates strong performance under class imbalance and varying data scarcity. This approach offers a practical pathway to augment healthcare data with privacy-preserving, task-tailored synthetic time series that boost clinical predictive performance.

Abstract

Synthetic Electronic Health Record (EHR) time-series generation is crucial for advancing clinical machine learning models, as it helps address data scarcity by providing more training data. However, most existing approaches focus primarily on replicating statistical distributions and temporal dependencies of real-world data. We argue that fidelity to observed data alone does not guarantee better model performance, as common patterns may dominate, limiting the representation of rare but important conditions. This highlights the need for generate synthetic samples to improve performance of specific clinical models to fulfill their target outcomes. To address this, we propose TarDiff, a novel target-oriented diffusion framework that integrates task-specific influence guidance into the synthetic data generation process. Unlike conventional approaches that mimic training data distributions, TarDiff optimizes synthetic samples by quantifying their expected contribution to improving downstream model performance through influence functions. Specifically, we measure the reduction in task-specific loss induced by synthetic samples and embed this influence gradient into the reverse diffusion process, thereby steering the generation towards utility-optimized data. Evaluated on six publicly available EHR datasets, TarDiff achieves state-of-the-art performance, outperforming existing methods by up to 20.4% in AUPRC and 18.4% in AUROC. Our results demonstrate that TarDiff not only preserves temporal fidelity but also enhances downstream model performance, offering a robust solution to data scarcity and class imbalance in healthcare analytics.

Figures (12)

• Figure 1: Overview of the Influence Guidance Diffusion framework. In Stage 1, we construct task-specific datasets from the original dataset $\mathcal{D}_{train}$ and train downstream models $f_{T_i}$ In Stage 2, we compute each sample's gradient-based influence for total influence $\mathbf{G}$ based on $\mathcal{D}_{T_i}$ and $f_{T_i}$. In Stage 3, we leverage influence signals guide the reverse diffusion process with computing $\Delta \mathcal{L}_{T}(\hat{z}) = \nabla_{\phi} \ell_T( \mathbf{x}_t,y_t; \phi)\cdot{G}$. All symbols are detailed in the legend on the right.
• Figure 2: Comparison of AUROC values for the Mortality and ICU Stay task on the MIMIC III and eICU dataset with synthetic-to-real data mix ratios from 0.2 to 1.0.
• Figure 3: Influence scale analysis conducted by generating samples from the Guidance-Val subset and assessing AUROC performance on the Guidance-Val subset for Mortality and ICU Stay tasks, with scales ranging from -1000 to 1000. The left panel illustrates sample influence value changes, while the right panel shows AUROC performance across different scales.
• Figure 4: Model performance with varying guidance set sizes on PTBrain dataset.
• Figure 5: Visualization of negative samples generated by different methods for ICU-Stay on eICU