Explainable AI for Hospital Mortality Prediction Using the eICU Collaborative Research Database

An exploratory and leakage-aware machine-learning workflow for prediction of mortality of patients hospitalized in an ICU using the eICU Collaborative Research Database Demo version 2.0.1 is described in this study. From 2,520 ICU stays, features are extracted as combinations of demographic and admission parameters along with laboratory and vital signs data recorded during the first 24 hours. Repeated measurements are aggregated on the ICU stay level, highly incomplete features are removed and missing values are processed as part of the machine learning pipeline. As a result of preprocessing there are 2,092 stays and 170 hospital deaths in the final cohort. Logistic regression and random forest models are trained and tested using cross-validation with health-system stay-aware five folds and an unseen test set split to groups. Cross-validation ROC-AUC values for the models are 0.7644 ± 0.0215 for logistic regression and 0.7791 ± 0.0178 for random forest. For the test set, the values are 0.7554 and 0.7615 respectively, and the PR-AUC values are 0.3136 and 0.2097. Features with the largest SHAP attribution values in case of the random forest include respiration, age, heart rate, blood urea nitrogen and creatinine. In addition, issues with class imbalance, threshold selection, missingness and individual-level explanations are considered. Results are provided as an illustrative workflow rather than as a deployable prediction system.