Work

Dynamic Prediction of 72-Hour Gastrointestinal Bleeding in Unselected ICU Patients Using MIMIC-IV and eICU-CRD: Deep-Learning Model Development with Dual-Database External Validation

Year: 2026
Type: preprint
Study design: Observational Study
Abstract: This is a retrospective cohort study based on two publicly available de-identified critical-care databases (MIMIC-IV v3.1 and eICU-CRD v2.0). The objective is to develop and externally validate a dynamic prediction model that, in unselected adult ICU patients, repeatedly estimates the risk of new-onset non-variceal gastrointestinal bleeding (GIB) within the next 72 hours after ICU admission. The design uses a two-tier architecture: a static baseline model (Model S) ensures clinical usability and comparability with existing literature, while a dynamic temporal model (Model D) is the core innovation. Model D uses a per-channel gated delta-rule linear attention family (KDA variants) combined with causal full attention, leveraging state continuity to perform incremental dynamic updates between landmarks. Validation follows a three-tier architecture: internal temporal extrapolation using the MIMIC-IV MetaVision period, quasi-external validation using the MIMIC-IV Epic period, and confirmatory external validation using eICU-CRD. The single confirmatory hypothesis H1: In the eICU external validation set, the frozen optimal dynamic model (Model D\*) has higher AUPRC than the frozen optimal static deep-learning model (Model S\*) for predicting new-onset GIB during the 72-hour window following the 24-hour landmark; ΔAUPRC is summarized by patient-clustered 5,000 paired percentile bootstrap (seed = 20260816), and confirmation requires the lower bound of the 95% CI to be >0. All other analyses are pre-registered exploratory or sensitivity analyses.
Author: Xuming Pan
Language: English
Field: Medicine
Open Access status: green
OpenAlex ID: W7204200952
Cites:
Cited by:
Related to: 10
Parts: 5