Synthetic EHR Time-Series Generation Using Diffusion Models
Proceedings of the AAAI Symposium Series · 06.11.2026 · 10 (1) · 1231-1238
DOI: 10.1609/aaaiss.v10i1.43550 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the AAAI Symposium Series, online erschienen am 06.11.2026. Verfasst von Yagni Patel und 1 weiteren Autor:innen.
Abstract (Originalsprache)
Electronic Health Records (EHRs) provide rich longitudi- nal patient level data for clinical research, but privacy con- cerns and regulatory restrictions limit access to real pa- tient records, while de-identification does not eliminate re- identification risks. Synthetic EHR generation offers a po- tential solution, but realistic ICU time series generation re- mains challenging because clinical variables are observed at different frequencies and exhibit strong physiological de- pendencies. This study develops a diffusion based approach for generating realistic and privacy preserving synthetic ICU time series while modeling irregular observation patterns and cross feature relationships. Experiments were conducted on MIMIC III, MIMIC IV, and eICU using the first 24 hours of each ICU stay, represented by 35 hourly clinical fea- tures comprising 7 vital signs and 28 laboratory and res- piratory measurements, together with a binary observation mask. A Denoising Diffusion Probabilistic Model (DDPM) jointly generates clinical values and their observation mask using a shared bidirectional recurrent denoiser, with an ad- ditional cross feature correlation loss to preserve covariance relationships between clinical variables. Synthetic data are evaluated using discriminative score for realism, Train-on- Synthetic-Test-on-Real (TSTR) and Train-on-Real-Test-on- Real (TRTR) for predictive utility, Nearest-Neighbour Ad- versarial Accuracy (NNAA) for privacy, and Maximum Mean Discrepancy (MMD) for distributional fidelity. The proposed approach achieves strong privacy, with NNAA values rang- ing from 0.0022 to 0.0122, and near parity in predictive util- ity on the two MIMIC corpora, with TSTR/TRTR ratios of 1.11 to 1.14; the gap widens to 1.27 on the more heteroge- neous eICU corpus. However, discriminative scores of 0.23 to 0.36 indicate that synthetic sequences remain distinguishable from real data above chance, suggesting limitations in repro- ducing finer distributional structure. Overall, the framework demonstrates strong privacy and near-parity in predictive util- ity while preserving irregular observation patterns and cross- feature relationships, although the remaining discriminative gap indicates a need for improved modeling of higher-order dependencies and fine-grained clinical structure.
Zitieren
APA
Patel, Y., & Narayan, A. (2026). Synthetic EHR Time-Series Generation Using Diffusion Models. Proceedings of the AAAI Symposium Series, 10(1), 1231-1238. https://doi.org/10.1609/aaaiss.v10i1.43550
Vancouver
Patel Y, Narayan A. Synthetic EHR Time-Series Generation Using Diffusion Models. Proceedings of the AAAI Symposium Series. 2026;10(1):1231-1238. doi:10.1609/aaaiss.v10i1.43550
BibTeX
@article{Patel2026,
author = {Patel, Yagni and Narayan, Apurva},
title = {Synthetic EHR Time-Series Generation Using Diffusion Models},
journal = {Proceedings of the AAAI Symposium Series},
year = {2026},
volume = {10},
number = {1},
pages = {1231--1238},
doi = {10.1609/aaaiss.v10i1.43550}
}
EUVIMED