Personalized physician recommendation for critical care using the TreeSHAP method

Abstract
Physician performance plays a critical role in the care of patients admitted to intensive care units (ICUs), where individuals often face life-threatening conditions requiring complex medical interventions. Rigorous quality assessment is essential to maintaining high standards of care and promoting continuous professional improvement. Traditionally, physician performance is evaluated using 360-degree assessments, which measure competencies across seven domains, including communication, management, and professionalism. Although these evaluations provide valuable insights, they may not fully account for heterogeneity in physician characteristics or patient complexity. In this study, we apply tree ensemble methods combined with TreeSHAP to investigate the effects of ICU departments and physician 360-degree evaluation scores on patient outcomes. We further propose a data-driven physician ranking index to quantify overall performance. Tree ensemble models flexibly capture nonlinear relationships and high-order interactions among predictors and have demonstrated strong predictive performance across a wide range of applications. The TreeSHAP framework enables efficient and theoretically grounded computation of Shapley values, providing transparent model interpretability. To account for patient-level heterogeneity and enable statistical inference, we incorporate mixed-effects models for uncertainty quantification and efficient estimation of fixed effects. Our findings suggest that differences across ICU departments are not significantly associated with patient outcomes, whereas specific components of the 360-degree evaluations exhibit meaningful associations. Because incomplete observations are present in the dataset, our analyses assume a missing completely at random (MCAR) mechanism. Future work should examine the robustness of the findings under alternative missing data mechanisms, as violations of the MCAR assumption may influence the results.