Development and temporal evaluation of an emergency severity index-informed Bayesian sequential model for hospital...

Triage tools in routine emergency care are largely static and may miss simple dynamic bedside cues available after presentation. We developed and temporally evaluated an emergency severity index (ESI)-informed Bayesian sequential prediction model for hospital admission using time-to-urination (TTU) in a prospective single-center cohort of ambulance-transported emergency department patients in Japan (February-August 2025; [Formula: see text]). The outcome was hospital admission at emergency department disposition. ESI was used as the initial pretest risk layer, TTU as a dynamic updating cue, and age and sex as refinement variables. Population-level fit to the cumulative admission curve was strong. In nested model comparison, ESI alone yielded an AUROC of 0.661 (95% CI 0.640-0.680), adding TTU improved discrimination to 0.677 (95% CI 0.658-0.698), and further adjustment for age and sex yielded the best performance (AUROC 0.741, 95% CI 0.722-0.760). Recalibration improved probability alignment without materially changing discrimination. Calibration deteriorated later in the post-arrival period, suggesting that the model is most informative in the early post-arrival window. This framework is designed to augment, rather than replace, existing triage systems.