A 13-Variable Hour-24 Model for 7-Day ICU Mortality in Neurocritical Care: Strong Cross-Database Discrimination with Unit-Level Calibration Differences

Abstract Background Short-horizon neurocritical risk estimates must remain accurate across care settings. We developed a 13-variable model at hour 24 for intensive care unit (ICU) mortality by day 7 and evaluated its transport to a second database. Methods In this retrospective cohort study, adults alive in the ICU at hour 24 were identified in the eICU Collaborative Research Database (eICU-CRD) version 2.0 for development and Medical Information Mart for Intensive Care IV (MIMIC-IV) version 3.1 for cross-database evaluation. Earlier analyses, including MIMIC-IV target data, informed the cohort definitions; the final fitted model was applied unchanged. We harmonized 13 predictors and compared six model families with hospital- and patient-grouped repeated nested validation. Log loss and a one-standard-error rule governed model selection. Results eICU-CRD contributed 20,810 admissions, 885 deaths, and 192 hospitals; MIMIC-IV contributed 6,821 admissions and 203 deaths. The selection rule chose cubic B-spline ridge logistic regression. The nested selection procedure yielded an internal area under the receiver operating characteristic curve (AUROC) was 0.8765 (95% confidence interval [CI], 0.8626 to 0.8872), with a Brier score of 0.03407 and calibration slope of 0.9861. In MIMIC-IV, AUROC was 0.9037 (95% CI, 0.8873 to 0.9204), Brier score 0.02456, and calibration slope 0.9750. The pooled observed-to-expected (O/E) ratio was 0.984; By initial care setting, O/E indicated underprediction in the neurologic surgical ICU (Neuro SICU; 1.588) and overprediction in intermediate/stepdown care (0.371). Conclusions A transparent 13-variable hour-24 model retained strong discrimination after frozen cross-database transport. Care-setting calibration exposed offsetting probability errors hidden by pooled estimates, supporting unit-specific prospective evaluation of neurocritical early-warning workflows.