Predicting severe COVID-19 in elderly patients using routine laboratory indicators: Diagnostic accuracy of machine learning models

To enable early warning of severe 2019 coronavirus disease in elderly patients, this study collected routine laboratory indicators and clinical parameters to construct and validate clinical models to predict the risk of severe disease. A total of 123 elderly 2019 coronavirus disease patients (68 non-severe and 55 severe) were retrospectively enrolled and randomly split into training and test sets. Predictive variables were selected via univariate analysis and stepwise logistic regression (LR) with variance inflation factor testing. Four models - LR, Random Forest, Support Vector Machine (SVM), and eXtreme Gradient Boosting Tree - were built. Model stability was assessed with 100 rounds of Bootstrap resampling. Performance was evaluated using receiver operating characteristic curves, calibration curves, and decision curve analysis in the test set. Five variables were selected: lymphocyte percentage, red blood cell count, total iron-binding capacity, unsaturated iron-binding capacity, and red cell distribution width. Multivariate LR identified lymphocyte percentage, red blood cell count, and total iron-binding capacity as protective factors, while unsaturated iron-binding capacity and red cell distribution width as risk factors. Bootstrap showed SVM and LR had superior stability. In the test set, SVM achieved the highest area under the curve (best discrimination); Random Forest showed best calibration; LR and eXtreme Gradient Boosting yielded slightly higher net benefits. Based on routine laboratory indicators, 4 prediction models were constructed and compared. These models can quantify severe risk using routine data early after admission, demonstrating strong clinical application potential.