Abstract
Intensive care unit (ICU) delirium affects 20–80% of critically ill patients and is associated with prolonged hospitalization, longterm cognitive impairment, and increased mortality; yet existing prediction models are largely single-cohort binary classiers that cannot represent an evolving patient state or support simulation-based decision making. We present a digital-twin-inspired framework for ICU delirium that represents each patient as a continuous latent state embedding from a Self-Attention Imputation for Time Series (SAITS) model—trained jointly for imputation and delirium prediction on hourly electronic health record (EHR) data—and unies prediction, risk-trajectory analysis, phenotyping, and similar-patient retrieval over that state. We benchmark SAITS on three large ICU cohorts—MIMIC-III ( n=4,047), MIMIC-IV ( n=25,697), and eICU-CRD ( n=6,631; mean missingness ≈83%, and broadly multi-center)—against seven baselines spanning classical machine learning (LR, RF, LightGBM, XGBoost), deep tabular models (TabPFN v2, FT-Transformer), and an EHR transformer (TransEHR). SAITS is not the strongest discriminator overall—TabPFN and gradientboosting trees lead on the well-observed MIMIC cohorts—but on the extreme-missingness eICU cohort it attains the best F1 (0.651; +0.082 over LightGBM) and best AUROC (0.765), indicating robustness when data are severely incomplete. Crucially, the 64-dimensional SAITS state embedding—which tree and in-context baselines do not provide—enables three digital-twin capabilities: phenotype clustering (four subgroups, delirium rates 6.3–57.1%, a 9-fold spread); temporal risk trajectories flagging high-risk patients within 1–2 hours; and similar-patient retrieval that recovers high-risk peers and surfaces a model false negative. The latent state further offers a natural substrate for future counterfactual intervention simulation—not yet implemented here—moving ICU delirium management beyond binary prediction toward clinical decision support.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DT4H_038.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=CVJgJMO3DS
BibTex
@InProceedings{HonSun_ADigital_MICCAISAT2026,
author = { Hong, Sung man AND Ahn, Sangjeong},
title = { { A Digital Twin Framework for ICU Delirium Using SAITS-Based Patient State Embeddings } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17275},
month = {pending},
page = {pending}
}