Personalized physician recommendation for critical care using the TreeSHAP method

Abstract
In this study, we employ the tree ensemble methods, namely XGBoost, Random Forest, and Tree Boosting Mixed Model, with the Tree SHapley Additive exPlanations (TreeSHAP) to explore the effects of ICU departments and physician 360 evaluations on patient outcomes. Furthermore, we develop a data-driven physician-ranking index to assess the physician performance.

Tree ensemble models automatically incorporate non-linearities and interactions of the variables, and they outperform other machine learning models in many applications. The TreeSHAP method provides a fast and precise computation of Shapley values to explain the tree ensemble models. Using mixed effects models to account for heterogeneity among patients allows for the uncertainty quantification and efficient learning of the fixed effect. We find that the difference in ICU departments does not affect patient outcomes, but certain factors of 360 evaluations do affect them.

As incomplete observations are present in the data, our analyses here are conducted under the assumption of missing completely at random (MCAR) mechanism. It is interesting to further explore how the violation of MCAR mechanism may affect the analysis results. This research warrants a careful study.