Improving COVID-19 Prognosis Predictions with Ensemble Learning Techniques Applied to Patient Health Data

Precise prognosis predictions of COVID-19 patients are critical to making clinical decisions, especially in limited resource settings. Existing models using patient data typically suffer from privacy issues, inconsistent utilization and reporting, and applicability to multiple health care settings. This study proposes an ensemble learning framework with a focus on facility level hospital utilization data to predict patient outcomes without relying on sensitive individual data. The proposed framework consists of staff, SVM, and XGBoost in a stack ensemble fashion to achieve better prediction performance and robustness. Important features like ICU bed occupancy, ventilator capacity, and total COVID-19 admissions, were engineered from time-series hospital administrative data to show resource stress in real time. Our experimental results compare the performance of facility-level ensemble model predictors with the previous models, achieving a model accuracy of 94.1%, F1-score of 0.89 and AUROC value of 0.96. Statistical tests confirm our findings were significant across multiple datasets, suggesting generalizability and clinical utility of our models. SHAP analysis provide explanations for prediction by determining the most influential predictors. Due to the framework and components being modularized and scale, it is a viable addition to Clinical Decision Support Systems (CDSS) to assist in early risk detection and proactive hospital resource planning (Rwe and Sewda) This study shows the distinction between predictive modelling vs practical hospital management predictions in real world emergent situations contributing to current pandemic response.