Fig. 2

EHRFormer and XGBoost predict one-, three-, and five-year mortality with modest performance. (A). The number of features used in each of the 4 different models. A data dictionary of these features is available in Table S1. (B). The number of patients in each model split by class outcome of 1 year mortality. Teal indicates those who were excluded for modeling if they died within 90 days of their index stay. (C). Test set AUROCs for EHRFormer (red) and XGBoost (blue) prediction of 1 year mortality across the 4 models. Error bars indicate 95% CIs based on 50 bootstraps of the test set. (D). Heatmap of all test set metrics normalized within each metric (row-wise) including AUROC, accuracy, F1 score, precision, recall, and specificity for EHRFormer and XGBoost across the 4 models. E. Test set AUROCs for EHRFormer (red) and XGBoost (blue) prediction of 1-, 3-, and 5-year mortality. Error bars indicate 95% CIs based on 50 bootstraps of the test set. F. The number of patients split by class outcome of 1-, 3-, and 5-year mortality. Teal indicates those who were excluded for modeling if they died within 90 days of their index stay.