Figure 4. Feature importance ranking of clinical indices.

(A) The relative importance of the 80 clinical indices in separating the deceased patients from patients with severe COVID-19 (n = 30 each) was evaluated in a random forest analysis. In this random forest, an assembly of decision trees (n = 1000) was generated using randomly selected subsets of patients and features (clinical indices) to collectively arrive at the final model prediction (deceased vs. severe). The importance of a feature (i.e., clinical index) was evaluated by the decrease in prediction accuracy, when such a feature was excluded from the model, assessed on the basis of (B) Gini impurity following a node split (MDI) and (C) the permuted values of the feature (MDA). The feature importance was evaluated in 10 repeated random forest analyses. The top 30 features in B and C are shown (the color scheme is proportional to the importance score).