Related Experiment Videos
Assessing the consistency of a prognostic transition to persistent critical illness across two U.S. ICU databases: a
Guoxin Wu1, Quankun Lv1, Yi Ye1
1Department of Emergency, School of Medicine, The Sixth Affiliated Hospital of South China University of Technology (Nanhai District People's Hospital of Foshan), Foshan 528200, Guangdong, China.
Objective:
To test whether an apparent prognostic transition to persistent critical illness persists after daily clinical-state updating.
Design:
Retrospective daily risk-set analysis using multicenter eICU discovery, single-health-system MIMIC-IV concept replication, cluster-resampled simultaneous inference, and post-primary explanatory updated-state analyses.
Setting:
Two deidentified U.S. critical care databases: eICU Collaborative Research Database v2.0 and MIMIC-IV v3.1.
Patients:
Adults alive and hospitalized within the prespecified continuous ICU episode at completed ICU days 3-14.
Interventions:
None.
Measurements And Main Results:
Predisposition (P) was compared with clinical state from episode hours 0-24, then fixed (I), or from the 24 h before each daily risk set (S). ΔP-X = L_X - L_P was the state-only minus P-only out-of-fold log-loss difference; negative values favor state. The outcome was in-hospital death within 14 days after risk-set time and before live discharge. Paired comparisons used hospital-cluster inference in eICU and patient-cluster inference in MIMIC-IV. Day-3 risk sets included 46,182 eICU episodes from 203 hospitals and 24,185 MIMIC-IV episodes. At day 10, ΔP-I was 0.0060 (simultaneous 95 % CI, -0.0010 to 0.0130) and 0.0126 (0.0033 to 0.0220), respectively; ΔP-S was -0.0360 (-0.0471 to -0.0249) and -0.0387 (-0.0543 to -0.0230). Every ΔP-S interval was below zero through day 14. eICU nominated no fixed-state transition; MIMIC-IV had an aggregate run from day 5, but care settings and predictor definitions disagreed. Post-result analyses retained negative ΔP-S after removing Glasgow Coma Scale and/or missingness indicators, using gradient boosting, and excluding documented entry treatment limitation.
Conclusions:
In these U.S. data, crossover occurred only against an aging admission snapshot. With the studied predictor blocks through ICU day 14, no consistent transition was identified; transitions in other health systems or later ICU periods remain possible.