Keywords: machine learning, semiparametric, causal inference, finite sample, high-dimensional
Globally, trauma is a a leading cause of death and poses both clinical and statistical challenges. Using highly predictive measures to optimize treatment assignment is of great current interest. However, the impact of treatment rules assigned based on TEG (Thromboelastography) measures has not been closely examined. The goal of our study was to employ robust, semiparametric data-adaptive modeling procedures to estimate the potential impact of protocols for achieving hemostasis. Under the causal inference framework, we focused on estimating the impact of blood product ratios assigned based on TEG measures on patient’s hemostasis and mortality status. Since the covariates are high-dimensional, we deployed ensemble machine learning (SuperLearning) methods for modeling the prediction of outcomes versus both adjustment covariates and intervention variables. Given the common problem of high dimension and small sample, standard doubly-robust estimators (such as estimating equation and targeted learning approaches) can break down due to lack of experimentation in the data (positivity violations), caused by high correlation between covariates and exposures in trauma data. We develop estimators of our estimands of interest within the collaborative targeted minimum loss-based estimation (CTMLE) framework, that optimizes the variance-bias trade-off when modeling the so-called propensity score, not with respect to prediction intervention, but to the estimand of interest. This allows for more automated estimation using machine learning in situations with limited data. Under this estimation framework, the treatment protocol (using TEG values immediately after injury) showed significant improvement in trauma patient’s hemostasis status (control of bleeding), and a decrease in mortality rate at 6h compared to standard care. The estimation results did not show significant change in mortality rate at 24h after arrival.