UC Berkeley

Trauma injury is one of the leading causes of death in the United States, accounting for over 120,000 deaths in 2010 according to the CDC. Understanding the underlying mechanisms and improving the treatment of trauma is of great clinical and public health interest. The systematic collection and study of critical care data originated in combat conflicts and wars and more recently to civilian centers. Improving patient outcomes, the quality of care received, and identifying high-risk patients are unmet needs in this field.

Clinicians rely on their intuition, training, and heuristic scoring systems to identify patients who are likely to die or experience other outcomes such as the need for a massive transfusion, which resuscitates the patient via the infusion of blood products such as plasma, platelets, and red blood cells. We assessed the ability of measured covariates to predict various clinical outcomes, demonstrate the utility of machine-learning prediction algorithms, and examined the predictive performance of a commonly-used score to predict massive transfusion. This highlights the need for a principled approach to predicting outcomes that does not rely only on ad hoc procedures.

In addition to the prediction of clinical outcomes, we defined a measure of variable importance for ranking predictors based on their relationship with the outcome of interest. This parameter was motivated by causal inference and requires a systematic approach to the question of interest that helps translate it into a parameter with a clinically meaningful interpretation rather and maintains transparency about the assumptions required to deem the parameter a causal effect. We apply this procedure to gene expression data from critically injured patients to illuminate how the coagulation and inflammation pathways react to trauma injury.

Finally, we compare the quality of care received at different trauma center types around the United States using another parameter motivated by causal inference. This allowed us to simulate what would have happened to a patient if they had been treated at a different trauma center and obtain an objective comparison that identified sites where severely injured patients would benefit most from being treated.

This research highlights the utility of causal inference for framing problems, motivating clinically meaningful statistical parameters, and interpreting the results. We also advocate for the use of semiparametric prediction algorithms to allow for greater flexibility in modeling assumptions and demonstrate their performance in practice.