C55-27 A Simple AI Model Predicts ICU Mortality From Routinely Collected Data in a Large Multi-Center Cohort

Abstract Rationale Simple, transparent risk models can help intensive care unit (ICU) teams triage care and communicate prognosis. We aimed to develop and evaluate a parsimonious mortality-prediction model using variables available in the first hours of ICU care. Methods We performed a retrospective cohort study using the publicly available eICU Collaborative Research Database (eICU-CRD). Adult ICU stays with an in-hospital mortality label were included. Candidate predictors were age, sex, Acute Physiology and Chronic Health Evaluation IVa (APACHE-IVa) score, lactate, hemoglobin, 6-hour urine output, and a binary indicator for missing urine output. Data were partitioned 80%/20% into training and evaluation sets using a deterministic hash of the patient/stay identifier to prevent patient leakage. A logistic-regression model (BigQuery ML) was trained and evaluated for discrimination (area under the receiver operating characteristic curve AUC) and log loss; operating-point performance across thresholds (precision, recall, F1, specificity, accuracy, negative predictive value NPV); and calibration (Brier score, Hosmer–Lemeshow χ², and a post-/hoc recalibration fit yielding intercept and slope). Results The cohort contained 349,391 ICU stays with 31,741 deaths (9.1%). On the 20% heldout evaluation set, overall performance was AUC 0.806 and log loss 0.243. Calibration was favorable by Brier score (0.0677) and recalibration parameters (intercept −0.0218, slope 1.0418), although the Hosmer–Lemeshow test was significant (χ² 714.8; p 0.001), which is common in very large samples. At the F1-optimal probability threshold of 0.20, precision was 0.409, recall 0.416, F1 0.412, specificity 0.940, accuracy 0.892, and NPV 0.941; at the default 0.50 threshold, precision was 0.672 and recall 0.155. Findings were robust despite high missingness in lactate (∼62%) and 6-hour urine output (∼52%), managed via first-available values and a missingness indicator. Conclusion A simple, explainable 7-variable logistic-regression model built from routinely collected ICU data achieved good discrimination and near-ideal average calibration in a large multicenter cohort. Such parsimonious models may support early triage and communication while remaining interpretable. Planned work includes temporal and external validation, decision-curve analysis, and prospective evaluation of clinical utility. References (optional—counts toward 400 words): Pollard TJ, Johnson AEW, Raffa JD, Celi LA, Mark RG, Badawi O. The eICU Collaborative Research Database, a freely available multi-center database for critical care research. Scientific Data. 2018;5:180178. doi:10.1038/sdata.2018.178. This abstract is funded by: None