2063 Novel Assessment of the Brain Injury Guidelines (BIG) Using Artificial Intelligence: The BIG 3A and 3B Study

INTRODUCTION: The Brain Injury Guidelines (BIG) were developed to risk stratify patients with traumatic brain injury (TBI). The BIG are highly sensitive but lack specificity for predicting hemorrhage expansion or need for neurosurgical intervention (NSI). METHODS: We abstracted data from a prospective quality database and retrospective medical records for adult patients categorized as BIG 3 (most severe) at a single Level I Trauma Center from 6/1/2020-9/30/2023. All patients were hospitalized, underwent index head CT, at least one repeat head CT (RHCT), and neurosurgical evaluation. Patients were divided into BIG 3a and BIG 3b cohorts based on clinical course. Recursive feature elimination and random forest search were used to select BIG variables most discriminant between 3a and 3b patients. Selected variables were input into 8 ML models for classification of patients as BIG 3a or 3b. Variable importance calculations identified the most discriminant features for each algorithm. RESULTS: 1115 patients were analyzed (757 BIG 3a; 358 BIG 3b). The best algorithms produced AUC = 0.77 (95% CI: 0.75-0.79). ML over-triage (falsely predicting BIG 3a as BIG 3b) and under-triage (falsely predicting BIG 3b as BIG 3a) rates were compared to over and under triage rates of the BIG at our institution. Variable importance plots identified abnormal neurologic examination, loss of consciousness, subdural hemorrhage size, and presence of skull fracture as the most discriminating features. CONCLUSIONS: ML models can reduce over-triage rates of TBIs compared to the BIG. Under-triage rates remain too high for safe clinical application. Future work balancing cohort sample size and exploration of additional clinical variables will be utilized to create a triage system that maximizes patient safety and resource allocation.