1116 Radiographic Classification and Predictors of Outcome in Gunshot Wounds to the Head Using a Machine Learning Model

INTRODUCTION: Firearm-related fatalities from gunshot wounds to the head (GSWH) area a deadly and continually increasing public health crisis with over 30,000 firearm related TBI annually in the United States. METHODS: This was a single-center, retrospective cohort study of GSWH patients admitted between 2014-2022 to a large, metropolitan, level 1 trauma center. Patients with dural penetration who survived primary resuscitation were included. Good functional outcome was defined as GOS of 4-5. A M5P model was constructed in R. It was trained using 75% of the dataset and validated using the remaining 25%. Model performance was evaluated using the ROC-AUC and confusion matrix metrics. RESULTS: 332 patients (64.8%) met inclusion criteria. Mortality was 58.7%. Good outcome was 30.1%. GCS, bilaterally non-reactive pupils, bihemispheric injuries, and cerebrovascular injury were independent predictors of mortality. GCS, bilaterally non-reactive pupils, lobes traversed by the projectile, and cerebrovascular injury were independent predictors of worse outcome. Six distinct trajectory patterns captured 94.6% of injuries and were independent predictors of outcome. These patterns were: A) unihemispheric penetrating, B) unihemispheric perforating/ricocheting, C) bifrontal, D) bihemispheric perforating/ricocheting, E) bihemispheric penetrating, and F) transtentorial. The machine learning model incorporated clinical, laboratory, and radiographic admission variables. It is available as a web-based tool. This model performed better than existing models (Baylor, SPIN, and St. Louis) in predicting mortality and functional outcome in the full cohort and in patients with at least one reactive pupil. CONCLUSIONS: We propose a new radiographic classification of GSWH injuries and a web-based machine learning model for predicting mortality and functional outcome based on admission factors.