Mortality and Prolonged ICU Stay Analysis in the MIMIC-III Database: A Dual Analytical

Abstract

Abstract Background: Understanding mortality patterns across ICU types and identifying patients at risk for prolonged stays is crucial for resource allocation and clinical decision-making. Objectives: This study examined (1) differences in in-hospital mortality across ICU types after risk adjustment, and (2) developed a predictive model for prolonged ICU stays using early clinical data. Methods: Using MIMIC-III database, we analyzed 61,533 ICU admissions. Propensity score matching with logistic regression compared mortality across five ICU types against MICU. XGBoost classification predicted ICU stays ≥7 days using first 24-hour clinical features. Results: Overall mortality was 4.6%, with 16.1% prolonged stays. After propensity adjustment, CSRU demonstrated significantly lower mortality versus MICU (OR: 0.22, 95% CI: 0.16-0.29), while SICU, CCU, and TSICU showed no significant differences. The XGBoost model achieved excellent discrimination (AUC-ROC: 0.862, sensitivity: 0.264, specificity: 0.978). Glasgow Coma Scale mean score was the most important predictor, followed by vasopressor use and mechanical ventilation. Conclusions: Substantial mortality differences exist across ICU types after risk adjustment, with cardiac surgery patients showing superior outcomes. Early clinical data accurately identifies patients at risk for prolonged stays, enabling proactive resource planning.