Reliable and Calibrated Uncertainty Quantification for COVID-19 using Probabilistic Machine Learning Models

Abstract

Standard regression-based models have limited predictive capability because they often overlook nonlinear interactions among variables, including Peripheral Oxygen Saturation, metabolic markers, illness duration, and the need for Intensive Care Unit admission. The purpose of this study is to compare the generalization behavior and predictive performance of three supervised learning models, such as Random Forest (RF), Multinomial Logistic Regression(MLR), and Logistic Regression(LR), in the classification of Covid patient outcomes (Expired vs. Cured) and identify critical clinical determinants of mortality. In all the evaluations RF model has shown better performance in test AUC (≈0.95), accuracy (≈0.88), sensitivity (≈0.78), specificity (≈0.96), and Kappa (≈0.76). On the other hand, LR and MLR have a lower performance test AUC (≈0.65), accuracy (≈0.58), sensitivity (≈0.22), specificity (≈0.84), and Kappa (≈0.07). The 10-fold cross-validation endorsed the dominance of RF.   This research highlights the necessity for data-driven models to pinpoint high-risk COVID-19 patients. It uses clinical data from COVID-19 patients to compare the performance of RF, MLR, and LR classifiers for cured vs expired COVID-19 cases.