183 Predicting Vasospasm and Mortality in Early Severe TBI: A Model Using Serum Cytokines and Clinical Data

Abstract

INTRODUCTION: It is difficult to predict outcomes in patients with severe traumatic brain injury (sTBI) using current clinical tools. METHODS: Adult civilian patients were prospectively enrolled in the sTBI arm of the Surgical Critical Care Initiative (SC2i). Patient clinical and serum inflammatory and neuronal protein data were combined and evaluated using the machine learning methods of LASSO and CART to construct parsimonious models for predicting development of post-traumatic vasospasm and mortality. Cross validations were performed to assess the robustness of conclusions. Missing data were imputed with random forest techniques; only variables with less than 10% missingness were included. RESULTS: There were 53 patients, of whom 36 (67.9%) developed vasospasm and 10 (18.9%) died. The mean age was 39.2; 22.6% were female. There were an equal number of white (25) and black (25) patients. For vasospasm, LASSO identified Eotaxin and Marshall classification of traumatic brain injury as predictors (XV AUC = 0.77). CART identified S100B as a predictor (full data AUC = 0.75). For mortality, LASSO identified S100B and sRAGE (XV AUC = 0.97), while CART identified S100B (full data AUC = 0.912). CONCLUSION: Inflammatory and glial-specific protein levels following sTBI may have predictive value that exceeds conventional clinical variables for certain outcomes. Eotaxin, S100B and radiographic findings highly predict development of post-traumatic vasospasm. S100B and sRAGE highly predict mortality. These results warrant validation in a prospective cohort.