Research map / Time series / Foundation models for time series
Generating Multi-Table Time Series EHR from Latent Space with Minimal Preprocessing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces RawMed, which synthesizes raw multi-table time-series EHRs by serializing events as text, compressing them with residual quantization, and autoregressively modeling the latent sequence with a Transformer.
Problem
Existing EHR synthesis methods generate only expert-selected features and rely on lossy preprocessing, so they cannot produce raw, multi-table, time-series EHRs that retain all columns and original values.
Approach
Serialize each clinical event as text, compress its token embeddings into discrete residual-quantized latent codes, and model the interleaved timestamp and code sequence of each patient autoregressively, then decode back to relational tables with validation postprocessing.
Key results
- RawMed achieves CDE 0.04 and PCC 0.04 on MIMIC-IV, versus 0.06 and 0.08 for the next-best baseline ClavaDDPM in single-table evaluation.
- On clinical utility (MEDS-TAB, MIMIC-IV), RawMed reaches micro-AUROC 0.87 versus 0.90 for real data and 0.68 for ClavaDDPM.
- Time Gap KS statistic for RawMed is 0.01 (MIMIC-IV) and 0.03 (eICU), versus 0.41-0.76 for baselines; MIA accuracy stays near 0.50.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RawMed instead of ClavaDDPM | MIMIC-IV single-table synthesis · MIMIC-IV | CDE: 0.06 -> 0.04 | Helped |
| Use RawMed instead of ClavaDDPM | eICU single-table synthesis · eICU | CDE: 0.06 -> 0.05 | Helped |
| Use RawMed instead of ClavaDDPM | MIMIC-IV downstream clinical prediction · MIMIC-IV MEDS-TAB | AUROC: 0.68 ±0.19 -> 0.87 ±0.08 | Helped |
| Use RawMed instead of ClavaDDPM | eICU downstream clinical prediction · eICU MEDS-TAB | AUROC: 0.66 ±0.14 -> 0.83 ±0.10 | Helped |
| Use RawMed instead of ClavaDDPM | MIMIC-IV temporal synthesis · MIMIC-IV | Time Gap: 0.48 -> 0.01 | Helped |
| Use RawMed instead of RC-TGAN | eICU temporal synthesis · eICU | Event Count: 0.03 -> 0.05 | Hurt |
| Add Residual Quantization rather than use VQ (ablation) | RawMed · MIMIC-IV patientweight | KS statistic: 0.28 -> 0.09 | Helped |
| Add RQ (ablation) | RawMed · MIMIC-IV | CDE: 0.05 -> 0.04 | Helped |
| Add RQ (ablation) | RawMed · MIMIC-IV | Time Gap: 0.05 -> 0.01 | Helped |
| Add RQ (ablation) | RawMed · MIMIC-IV | Event Count: 0.13 -> 0.02 | Helped |
| Add Time Tokenization (ablation) | RawMed · MIMIC-IV | Time Gap: 0.04 -> 0.01 | Helped |
| Add Time Tokenization (ablation) | RawMed · MIMIC-IV | Event Count: 0.12 -> 0.02 | Helped |
Surprising findings
- RealTabFormer underperforms RawMed even though it models data without compression, with time gap KS of 0.17 vs 0.05 on MIMIC-IV under a 6-hour window.
- Vector quantization distorts independent numeric columns such as patientweight (KS 0.28 vs 0.09 for RQ), motivating residual quantization.
Limitations
- Focuses on only three primary tables, leaving scalability to dozens of tables untested.
- Only unconditional generation; static attributes such as gender or birth year are not integrated.
- Longer time series may require more advanced compression or sampling methods.
Experimental setup
Datasets: MIMIC-IV eICU MIMIC-IV-ED
Benchmarks: MIMIC-IV eICU
Metrics: CDE I-CDE PCC I-PCC ER SMAPE AUROC Next Event Prediction F1 Time Gap Event Count
Baselines: SDV RC-TGAN ClavaDDPM RealTabFormer
Builds on: VQ-VAE Residual Quantization GenHPF MEDS-TAB RealTabFormer
Closest papers by meaning
- Collaborative Synthesis of Patient Records through Multi-Visit Health State Inference (AAAI 2024 · similarity 0.88)
- IGAMT: Privacy-Preserving Electronic Health Record Synthesization with Heterogeneity and Irregularity (AAAI 2024 · similarity 0.84)
- Advancing Medical Multimodal Learning and Data Generation with Diffusion Model and LLM (AAAI 2025 · similarity 0.84)
- SM3-Text-to-Query: Synthetic Multi-Model Medical Text-to-Query Benchmark (NeurIPS 2024 · similarity 0.83)
- MOTOR: A Time-to-Event Foundation Model For Structured Medical Records (ICLR 2024 · similarity 0.83)
- EHRCon: Dataset for Checking Consistency between Unstructured Notes and Structured Tables in Electronic Health Records (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.