Could machine learning algorithms help us predict massive bleeding at prehospital level?

OBJECTIVE: Comparison of the predictive ability of various machine learning algorithms (MLA) versus traditional prediction scales (TPS) for massive hemorrhage (MH) in patients with severe traumatic injury (STI).

DESIGN: On a database of a retrospective cohort with prehospital clinical variables and MH outcome, a treatment of the database was performed to be able to apply the different AML, obtaining a total set of 473 patients (80% training, 20% validation). For modeling, proportional imputation and cross validation were performed. The predictive power was evaluated with the ROC metric and the importance of the variables using the Shapley values.

SETTING: Out-of-hospital care of patients with STI.

PARTICIPANTS: Patients with STI treated out-of-hospital by a out-of-hospital medical service from January 2010 to December 2015 and transferred to a trauma center in Madrid.

INTERVENTIONS: None.

MAIN VARIABLES OF INTEREST: Obtaining and comparing the "Receiver Operating Characteristic curve" (ROC curve) metric of four MLAs: "random forest" (RF), "vector support machine" (SVM), "gradient boosting machine" (GBM) and "neural network" (NN) with the results obtained with TPS.

RESULTS: The different AML reached ROC values higher than 0.85, having medians close to 0.98. We found no significant differences between AMLs. Each AML offers a different set of more important variables with a predominance of hemodynamic, resuscitation variables and neurological impairment.

CONCLUSIONS: MLA may be helpful in patients with HM by outperforming TPS.