Fig. 3

Machine learning models for differentiating COVID-19 from CAP-sepsis. A Proteins above the 90th percentile of variable importance (dashed line) selected more frequently in the random forest models (RF-ML). B Lollipop plot showing the most frequently selected proteins in the logistic regression models with lasso regularization (LR-Lasso). C Accuracy radar plot comparing performance metrics of each model type calculated on testing datasets. The lines represent the mean value of the metric in position and the shadows represent the 95% CI (± 1.97 × SD) of the metric’s mean. D The LR-lasso model that had 100% accuracy in both training and testing data, which consisted of the smallest panel of proteins. Colors refer to the β coefficient as in B, and the ROC curve shows the model accuracy. The orange dashed line represents chance, the grey dotted lines represent AUCs for different values of lambda. E ROC curves demonstrating the diagnostic potential of existing clinical biomarkers in differentiating COVID-19 from CAP-sepsis. The dashed line represents chance. F ROC curves for the intersecting most frequently selected proteins in both RF and LR-Lasso models. Additional ROC curves of proteins, see Additional file 1: Fig. S4B, C