Reproduce “Generating Multi-Table Time Series EHR from Latent Space with Minimal Preprocessing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Eunbyeol Cho, Jiyoun Kim, Minjae Lee, Sungjin Park, Edward Choi
RawMed is a framework that generates multi-table, time-series Electronic Health Records (EHRs) by preserving all columns and original values with minimal preprocessing. This approach yielded clinical prediction task AUROC scores within 0.02-0.03 of real data and achieved significantly lower time gap KS statistics (0.01-0.03) compared to baselines (0.41-0.76), demonstrating high data fidelity and robust privacy preservation.
Key insights
- Text-based representation combined with Residual Quantization enables high-fidelity compression of complex EHR events, crucial for preserving fine-grained details across diverse data types.
- Explicitly tokenizing timestamps and integrating them alongside compressed event representations within an autoregressive Transformer is critical for accurately modeling temporal dynamics and inter-event dependencies.
- A rigorous multi-stage postprocessing pipeline for both event-level and patient-level validation is essential to ensure the structural and semantic integrity of the generated raw synthetic EHRs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.