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

Technical Summary: Artificial Intelligence for ICU Mortality Prediction

Problem Statement
Intensive care units (ICUs) are high-acuity environments where accurate early prediction of in-hospital mortality is critical for triage, resource allocation, and clinical decision support. While traditional severity scores like APACHE II are widely used, their predictive accuracy may be limited in heterogeneous patient populations. Furthermore, the specific influence of surgical status on mortality modeling and the comparative performance of various machine learning (ML) architectures in this context require further investigation. This study addresses the need to compare multiple AI/ML models for ICU mortality prediction and to examine the role of surgical patients as a distinct determinant of model behavior.

Methodology
This retrospective, single-center observational study analyzed 769 adult ICU admissions from March 2023 to December 2025 at a tertiary education and research hospital in Mardin, Turkey. The study adhered to TRIPOD+AI reporting principles and was approved by the local ethics committee.

• Data and Preprocessing: The dataset included demographic data, diagnoses, comorbidities (Charlson Comorbidity Index), severity scores (APACHE II), and laboratory parameters (e.g., WBC, CRP, lactate, troponin, D-dimer, AST, ALT). Patients were categorized as surgical or medical admissions. Missing or erroneous records were excluded, resulting in a complete-case analytical cohort. To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) and class weighting were applied exclusively to the training partition to prevent data leakage.
• Modeling Approach: Five classifiers were evaluated: Decision Tree, Logistic Regression (L2 regularization), Random Forest, Tuned Random Forest (optimized via grid search), and Multilayer Perceptron (MLP). The dataset was split into 70% training and 30% testing.
• Outcome Definition: In-hospital mortality was designated as the positive class (1) and survival as the negative class (0) to align with clinical interpretability.
• Evaluation Metrics: Performance was assessed using accuracy, sensitivity, specificity, positive/negative predictive values (PPV/NPV), F1-score, balanced accuracy, and Area Under the Curve (AUC).
• Explainability: The study employed SHAP (SHapley Additive exPlanations) analysis, feature importance ranking, and correlation analysis to interpret model behavior and identify key predictors.

Key Results
In the independent test cohort (n = 231), ensemble methods demonstrated superior discrimination compared to linear and single-tree models:

• Random Forest achieved the highest AUC (0.91) and a high F1-score (0.902).
• Tuned Random Forest showed the highest sensitivity (0.961) but lower specificity (0.377).
• Logistic Regression performed strongly with an AUC of 0.88 and an F1-score of 0.902.
• MLP achieved an AUC of 0.85.
• Decision Tree showed the lowest performance with an AUC of 0.66.

Feature Importance and Clinical Insights:

• Top Predictors: The most influential variables were sepsis/infection, surgical operation status, trauma, APACHE II score, ICU length of stay, troponin, and D-dimer.
• Surgical Status: Surgical admission emerged as a major contributor to model behavior, ranking among the top influential variables. The authors note that surgical patients may generate more homogeneous mortality patterns due to the specific nature of trauma, operative interventions, and postoperative complications.
• Multimorbidity: The analysis highlighted that mortality risk is driven by the interaction between acute severity (e.g., sepsis, trauma) and background multimorbidity (e.g., heart failure, COPD, renal dysfunction, age).
• Correlations: Strong correlations were observed between liver enzymes (AST, ALT, LDH) and between admission type and sepsis/trauma. APACHE II scores correlated positively with comorbidity burden and age.

Significance and Claims
The paper claims that explainable AI approaches can effectively support ICU mortality prediction while highlighting clinically plausible patterns related to acute severity, multimorbidity, and the surgical pathway. The study suggests that:

1. Ensemble Superiority: Random forest models offer the best discrimination for this specific cohort, though logistic regression remains a strong, interpretable alternative.
2. Surgical Relevance: Surgical status is not merely a background characteristic but a significant determinant of mortality risk modeling, potentially offering distinct predictive signals compared to medical admissions.
3. Clinical Plausibility: The identified predictors (sepsis, trauma, APACHE II, troponin, D-dimer) align with existing literature regarding ICU mortality drivers.

Limitations and Cautions
The authors explicitly acknowledge several limitations:

• The study is retrospective, single-center, and lacks external or temporal validation.
• There is no reporting of calibration curves, decision-curve analysis, or subgroup-specific discrimination metrics.
• Data Leakage Concern: The inclusion of "ICU length of stay" as a predictor is acknowledged as a potential source of data leakage in an admission-time prediction framework, as it reflects post-admission evolution rather than baseline risk.
• Missing data were handled via complete-case analysis, which may introduce bias.

Conclusion
The study concludes that AI and ML models, particularly Random Forest, provide useful discrimination for in-hospital mortality prediction. However, the authors emphasize that future deployment should be objective-dependent (e.g., prioritizing sensitivity vs. specificity) and must be preceded by external validation, prospective testing, and local calibration before routine clinical implementation.