Machine Learning for Military Trauma: Novel Massive Transfusion Predictive Models in Combat Zones.

BACKGROUND

Damage control resuscitation has become the standard of care in military and civilian trauma. Early identification of blood product requirements may aid in optimizing the clinical decision-making process while improving trauma related outcomes. This study aimed to assess and compare multiple machine learning models for predicting patients at highest risk for massive transfusion on the battlefield.

METHODS

Supervised machine learning approaches using logistic regression, support vector machine, neural network, and random forest techniques were used to create predictive models for massive transfusion using standard prehospital and arrival data points from the Department of Defense Trauma Registry, 2008-2016. Seventy percent of the population was used for model development and performance was validated using the remaining 30%. Models were tested for accuracy and compared by standard performance statistics.

RESULTS

A total of 22,158 patients (97% male, 58% penetrating injury, median age 25-29 y/o, average Injury Severity Score 9, with an overall mortality of 3%) were included in the analysis. Massive transfusion was required by 7.4% of patients. Overall accuracy was found to be above 90% in all models tested. Following cross validation and model training, the random forest model outperformed the alternatively tested models for precision, recall, and area under the curve.

CONCLUSION

Machine learning techniques may allow for more optimal and rapid identification of combat trauma patients at highest risk for massive transfusion. These powerful approaches may uncover novel correlations and help improve triage, activation of massive transfusion resources, and trauma-related outcomes. Further research seeking to optimize and apply these algorithms to trauma-centered research should be pursued.