<![CDATA[Accurate hospital-wide mortality prediction is important for institutional clinical governance, quality improvement and resource planning. However, most machine learning studies have focused on patient-level data, intensive care populations, or disease-specific cohorts, with limited integration of hospital operational indicators. This retrospective single-center predictive modeling study used 36 monthly institutional observations from January 2022 to December 2024 at a regional referral hospital in Indonesia. Aggregated clinical severity indicators were combined with operational performance metrics, including length of stay, bed occupancy rate, and bed turnover rates. Random Forest, XGBoost, and feedforward neural network models were developed and compared with a linear regression baseline. The performance was internally evaluated using time-aware five-fold cross-validation with R², root mean squared error (RMSE), and mean absolute error (MAE). GDR and error metrics were expressed as deaths per 1,000 admissions. The stacked ensemble achieved the highest R² (0.841) and the lowest RMSE (4.49 deaths per 1,000 admissions), while the neural network achieved the lowest MAE (2.74 deaths per 1,000 admissions). In conclusion, operational indicators had modest direct effects but improved the model robustness through interaction effects. These findings support the use of explainable ensemble machine learning for institutional-level mortality prediction and hospital decision support.]]>
