Get started
Work

An Explainable Machine Learning Framework for In-Hospital Mortality Prediction in the ICU: A Comparative Study of Advanced Tree-Based Ensembles and Multi-Faceted Interpretability

Year: 2026
Type: article
Study design: Observational Study
Abstract: Objective: While machine learning models show promise for predicting ICU mortality, their clinical adoption remains limited by black-box behavior and insufficiently interpretable, patient-specific risk insights. Methods: We retrospectively analyzed Intensive Care Unit (ICU) demographic, vital, and laboratory data to predict in-hospital mortality. Tree-based ensemble models were evaluated using accuracy, sensitivity, and specificity, with interpretability assessed through permutation importance, Shapley Additive Explanations (SHAP), and Individual Conditional Expectation (ICE) to identify global and individual risk patterns and support a clinically interpretable prediction framework. We retrospectively analyzed clinical and laboratory data from 3,600 ICU patients to predict in-hospital mortality. LASSO feature selection identified the 25 most contributive variables, which were used to train and compare tree-based machine learning models. Interpretability was evaluated using permutation importance, SHAP, and ICE analyses. Results: Performance was broadly comparable across the ten algorithms and no model was superior on every metric; XGBoost was designated as the final model because it combined top-ranked accuracy (0.861 ± 0.010) and a high specificity (0.939 ± 0.009) with the lowest across-fold variability, a profile suited to rule-out oriented risk stratification. Explainable AI analyses consistently identified impaired consciousness, renal dysfunction, advanced age, and metabolic abnormalities as key mortality predictors at both global and patient-specific levels. Conclusion: The proposed explainable framework provides interpretable, patient-specific mortality predictions in this cohort. Validation was internal only; external validation and prospective evaluation are required before any clinical use.
DOI: 10.66251/pspnpj56
Language: English
Open Access status: hybrid
OpenAlex ID: W7171557073
FWCI:
Cites:
Cited by:
Related to: 10