This study proposes a supervised machine learning (ML) framework to predict in-hospital mortality in patients in the Intensive Care Unit (ICU). This framework integrates a comprehensive and interpretable machine learning pipeline with intensive data preprocessing, such as K-nearest neighbors (KNN) imputation and feature selection using Recursive Feature Elimination with Cross-Validation (RFECV) guided by Random Forests. To address the inherent class imbalance in ICU mortality data, we employed SMOTETomek resampling technique. Exploratory data analysis showed higher mortality rates among patients aged over 70 years, females, those requiring respiratory support, and patients with comorbidities such as diabetes mellitus, immunosuppression, and solid tumors with metastasis. Logistic Regression, Random Forest, XGBoost, and a Stacking Ensemble model were evaluated using 10-fold cross-validation. Among the models, XG Boost performed the best on the balanced dataset with an AUROC of 87.69%, followed by the Ensemble model, which demonstrated competitive results across all metrics. SHAP (SHapley Additive exPlanations) was used to enhance model interpretability, which indicated that physiological and demographic parameters were the main predictors of mortality. This provided transparent and actionable insights on the model’s predictive patterns and clinical relevance. This study demonstrates that supervised ML models, when combined with rigorous preprocessing and interpretability techniques, can provide reliable, explainable, and clinically relevant tools for early prediction of ICU mortality.