Machine learning applied to a Cardiac Surgery Recovery Unit and to a Coronary Care Unit for mortality prediction.

Abstract

Most established severity-of-illness systems used for prediction of intensive care unit (ICU) mortality were developed targeted at the general ICU population, based on logistic regression (LR). To date, no dynamic predictive tool for ICU mortality has been developed targeted at the Cardiac Surgery Recovery Unit (CSRU) and Coronary Care Unit (CCU) using machine learning (ML). CSRU and CCU adult patients from the MIMIC-III critical care database were studied. The ML methods developed extract ICU data during a 5-h window and demographic features to produce mortality predictions and were compared to six established severity-of-illness systems and LR. In a secondary experiment, additional procedure/surgery and ICU features were added to the models. The ML models developed were the Tree Ensemble (TE), Random Forest, XGBoost Tree Ensemble (XGB), Naive Bayes (NB), and Bayesian network. The discrimination, calibration and accuracy statistics were assessed. The AUROC values were superior for the ML models reaching 0.926 and 0.924 for the XGB, and 0.904 and 0.908 for the TE for ICU mortality prediction in the primary and secondary experiments respectively. Among the conventional systems, the serial SOFA obtained the highest AUROC (0.8405). The Brier score was better for the ML models except the NB over the conventional systems. The accuracy statistics less sensitive to unbalanced cohorts were higher for all the ML models. In conclusion, the predictive power of XGB was excellent, substantially outperforming the conventional systems and LR. The ML models developed in this work offer promising results that could benefit CSRU and CCU.