Anomaly Detection Outperforms Logistic Regression in Predicting Outcomes in Trauma Patients.

OBJECTIVE: Recent advancements in trauma resuscitation have shown a great benefit of early identification and control of hemorrhage, which is the most common cause of death in injured patients. We introduce a new analytical approach, anomaly detection (AD), as an alternative method to the traditional logistic regression (LR) method in predicting which injured patients receive transfusions, intensive care, and other interventions.

METHODS: We abstracted routinely collected prehospital vital sign data from patient records (adult patients who survived more than 15 minutes after being directly admitted to a level 1 trauma center). The vital signs of the study cohort were analyzed using both LR and AD methods. Predictions on blood transfusions generated by these approaches were compared with hospital records using the respective areas under the receiver operating characteristic curves (AUROC).

RESULTS: Of the patients seen at our trauma center between January 1, 2009, and December 31, 2010, 5,464 were included. AD significantly outperformed LR, identifying which patients would receive transfusions of uncrossmatched blood, transfusion of blood between the time of admission and 6 hours later, the need for intensive care, and in-hospital mortality (mean AUROC = 0.764 and 0.720, respectively). AD and LR provided similar predictions for the patients who would receive massive transfusion. Under the stratified 10 fold times 10 cross-validation test, AD also had significantly lower AUROC variance across subgroups than LR, suggesting AD is a more stable predictions model.

CONCLUSIONS: AD provides enhanced predictions for clinically relevant outcomes in the trauma patient cohort studied and may assist providers in caring for acutely injured patients in the prehospital arena.