Welberry, HeidiAlanazi, Abdullah2026-01-152025https://hdl.handle.net/20.500.14154/77887This study investigated two critical questions in intensive care using the MIMIC-III database. First, we examined whether time-to-death differs across ICU types using survival analysis methods. Second, we developed a machine learning model to predict prolonged ICU stays (≥7 days) from early clinical features, addressing the class imbalance inherent in this outcome. Our survival analysis of 24,754 adult ICU admissions revealed significant mortality differences between ICU types, with SICU and TSICU patients showing approximately 50% lower hazard of death compared to MICU patients (adjusted HR 0.51, 95% CI: 0.44-0.60 and 0.51, 95% CI: 0.42-0.62, respectively). For prolonged stay prediction, our Random Forest model with balanced training achieved strong discrimination (AUC-ROC 0.84) and balanced accuracy (76.6%), outperforming traditional logistic regression. The most important predictive features were Glasgow Coma Scale measures, mechanical ventilation, and vasopressor use—all indicators of illness severity. These findings suggest that ICU type substantially influences mortality risk and that early prediction of prolonged stays is feasible using routinely collected clinical data.22enSurvival AnalysisCox Proportional HazardsICU Length of StayMIMIC-III DatasetMachine LearningRandom ForestHealth Data ScienceIntensive Care UnitThesis