Artificial Intelligence for ICU Mortality Prediction:Comparative Machine Learning Models and the Role of Surgical Patients

Purpose: To compare artificial intelligence and machine-learning models for in-hospital mortality prediction in adult intensive care unit (ICU) patients and examine the contribution of surgical status to model behavior. Methods: This retrospective single-center study included 769 ICU admissions recorded between March 2023 and December 2025. Decision tree, logistic regression, random forest, tuned random forest, and multilayer perceptron models were evaluated. Mortality was treated as the positive class for clinically intuitive reporting. Results: In the 30% test cohort ( n = 231), random forest showed the highest discrimination (AUC 0.91), followed by tuned random forest (AUC 0.90), logistic regression (AUC 0.88), multilayer perceptron (AUC 0.85), and decision tree (AUC 0.66). Key clinically informative signals included sepsis or infection, surgical operation, trauma, APACHE II, troponin, and D- dimer. Surgical status emerged as an important contributor within the reported models, although subgroup-specific performance metrics were not available. Conclusion: Explainable artificial intelligence approaches may support ICU mortality prediction and clinical decision support while highlighting clinically plausible patterns related to acute severity, multimorbidity, and the surgical pathway. Trial registration: Not applicable; this study is a retrospective observational analysis and was not a clinical trial.