Table 1.
A methodological overview of the related works.
Fig 1.
The relative frequency of twelve diagnosed disease groups in the eICU-CRD v2.0.
Table 2.
Cross-validation results (AUROC) of the non-GBT baseline models using 10-fold cross-validation.
Table 3.
Cross-validation results (mean AUROC [standard deviation AUROC]) of GBT models using 10-fold cross-validation.
Table 4.
Cross-validation results (AUROC) of CatBoost and ICU illness severity scoring systems using 10-fold cross-validation.
Fig 2.
Feature importance plots based on Shapley values for burns/trauma, cardiovascular, neurologic, and oncology disease groups.
Table 5.
Top three most important features in mortality predictions across various disease categories based on Shapley values.
Fig 3.
Force plots of the most important features in mortality prediction of patients in endocrine and gastrointestinal disease groups.
Fig 4.
A detailed explanation of patient features with the highest mortality probability in endocrine and gastrointestinal disease groups using Shapley values (force plots).
Fig 5.
Feature importance of individual patients calculated using LIME in surgery, toxicology, burns-trauma, and cardiovascular disease groups.
Fig 6.
Partial dependence (PD) and individual conditional expectation (ICE) plots for the most important features in the prediction of ICU discharge status in surgery, toxicology, burns-trauma, and cardiovascular disease groups.
Fig 7.
Bar and KDE plots for the most important features in predicting ICU discharge status for surgery, toxicology, burns-trauma, and cardiovascular disease groups.
Table 6.
Cross-validation (mean AUROC [standard deviation AUROC]) results of E-CatBoost model using 10-fold cross-validation.