Prediction-Time-Aware Multimodal Emergency Department Admission Prediction from Longitudinal EHR Data

Emergency department (ED) information accumulates asynchronously, making prediction landmarks, population eligibility, and input availability central to admission-prediction evaluation. In a retrospective, single-center internal evaluation of previously examined MIMIC-IV-ED v2.2 and MIMIC-IV v2.2 data, we evaluated adult ADMITTED-versus-HOME prediction with subject-level partitions and a primary 2-h landmark. Eligibility required remaining in the ED, without establishing that disposition decisions were pending. Longitudinal laboratory eligibility required both storetime and charttime to be no later than the cutoff. Among 56,161 test stays from 27,814 subjects, TextLabFusion achieved AUROC 0.855 (95% CI, 0.851–0.860) and AUPRC 0.814. Gradient-boosted trees achieved AUROCs of 0.859 using text and laboratories and 0.877 with additional clinical variables. For primary 2-h Fusion, the 365-day-minus-unrestricted-history AUROC difference was +0.0010 (95% CI, −0.0005 to +0.0023). Within the common-4h cohort, unrestricted-history models had higher AUROC than encounter-only models at 1–2 h and lower AUROC at 4 h and ED completion. Modality-dropout benefits depended on the missingness setting. Approximate-era analysis yielded Fusion AUROC 0.843 and showed overall overprediction. These findings support explicit reporting of prediction timing, eligibility, information scope, and missingness mechanisms; architectural superiority and clinical utility were not established.