Answers are grounded in this paper and cite their sources.
Background/Objectives: This retrospective methodological case study evaluated outcome classification in the first 501 laboratory-confirmed COVID-19 cases in North Macedonia (37 deaths; 7.39%). Methods: Six machine learning classifiers were compared with ZeroR and ridge logistic regression using stratified 10-fold cross-validation, supplemented by five repetitions. The primary analysis used 19 demographic, symptom and comorbidity predictors without hospitalization status. Symptom timestamps were unavailable, so performance at initial assessment could not be established. Results: J48 matched the majority-class accuracy of 92.6%; no classifier exceeded it. Naïve Bayes achieved the highest AUROC (0.863), MCC (0.369) and Death-class sensitivity (0.486). Logistic regression achieved AUROC 0.800 (95% CI 0.709–0.891), numerically exceeding five of the six machine learning classifiers; the conventional benchmark was competitive. Its repeated cross-validation AUROC was 0.801 ± 0.003. Bagging had the lowest Brier score (0.058), whereas Naïve Bayes and Random Tree had negative scaled Brier scores despite their discrimination above chance. Including hospitalization increased the Naïve Bayes AUROC to 0.893 and sensitivity to 0.649. Decision curves suggested net benefit at selected thresholds, but uncertainty and miscalibration limit interpretation. Separate exploratory permutation and SHAP analyses of a scikit-learn Random Forest reconstruction identified difficulty breathing, age and cardiovascular disease as prominent contributors; these explanations cannot be attributed directly to the evaluated WEKA ensemble. Conclusions: The contribution is a cohort-specific demonstration of how accuracy, discrimination, calibration and net benefit can support different conclusions, rather than a new algorithm or deployable mortality model. The small event count, uncertain predictor timing and absence of external validation preclude clinical implementation.
No takes yet. Share an insight, caveat, or question.
Paneva et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: