Towards Explainable Hospital Mortality Prediction Using Critical Care Data

Predicting in-hospital mortality in Intensive Care Unit (ICU) patients is critical for optimizing clinical outcomes and resource allocation. Conventional scoring systems such as, Acute Physiology Score (APS), Logistic Organ Dysfunction System (LODS), and Sequential Organ Failure Assessment (SOFA) struggle with high dimensional data due to their rigid design. This study presents an explainable Machine Learning (ML) framework that predicts ICU mortality using structured Electronic Health Record (EHR) data from the MIMIC-III database containing 58,976 ICU admission data. A subpopulation of 12,528 ICU patients from this dataset was systematically selected, and 37 features were extracted based on their importance to test eight different Artificial Intelligence (AI) prediction models. Light Gradient-Boosting Machine (LightGBM) achieved the highest performance in both accuracy and Area Under the Curve (AUC), while Extreme Gradient Boosting (XGBoost) and Random Forest (RF) also provided strong and competitive results. To interpret the model outcomes for the different features, SHapley Additive exPlanations (SHAP) analysis was applied to identify the key predictors. The proposed framework offers an accurate and transparent ICU mortality prediction model, supporting explainable clinical decision making.