Skip to main content
. 2021 Sep 28;23(9):e30157. doi: 10.2196/30157

Figure 3.

Figure 3

Feature importance in the GRU-D recurrent neural network model as defined by the average drop in the area under the receiver operator characteristic curve (with 95% CI) when each feature is individually removed from the analysis. The top 5 features are seen to be age, Charlson comorbidity index, minimum oxygen saturation, fibrinogen levels, and serum iron levels. AUROC: area under the receiver operator characteristic.