Air Medical Response to Traumatic Brain Injury: A Computer Learning Algorithm Analysis

Abstract

BACKGROUND: The role of air medicine in traumatic brain injury (TBI) has been studied extensively using trauma registries but remains unclear. Learning algorithms, such as artificial neural networks (ANN), support vector machines (SVM), and decision trees, can identify relationships between data set variables but are not empirically useful for hypothesis testing. OBJECTIVE: To use ANN, SVM, and decision trees to explore the role of air medicine in TBI. METHODS: Patients with Head Abbreviated Injury Score 3+ were identified from our county trauma registry. Predictive models were generated using ANN, SVM, and decision trees. The three best-performing ANN models were used to calculate differential survival values (actual and predicted outcome) for each patient. In addition, predicted survival values with transport mode artificially input as "air" or "ground" were calculated for each patient to identify those who benefit from air transport. For SVM analysis, chi was used to compare the ratio of unexpected survivors to unexpected deaths for air- and ground-transported patients. Finally, decision tree analysis was used to explore the indications for various transport modes in optimized survival algorithms. RESULTS: A total of 11,961 patients were included. All three learning algorithms predicted a survival benefit with air transport across all patients, especially those with higher Head Abbreviated Injury Score or Injury Severity Score values, lower Glasgow Coma Scale scores, or hypotension. CONCLUSION: Air medical response in TBI seems to confer a survival advantage, especially in more critically injured patients.