246 Algorithmic CT Scan Analysis for Prognosticating Brain Injury: A Preliminary Effort

Abstract

INTRODUCTION: Predicting which trauma patients will progress to brain death is currently based on history, physical examination and radiographic findings. METHODS: 38 trauma subjects, 18 which progressed to brain death (BD), who suffered from high-velocity trauma with presumed diffuse axonal injury, cardiopulmonary/respiratory arrest, and/or found down were included. BLAST-CT was used to identify four lesion types: intraparenchymal hemorrhage, extra-axial hemorrhage, perilesional oedema, and intraventricular hemorrhage. Cerebrospinal fluid volumes were extracted using SimpleITK. These two sets of features were obtained from whole head CT scans and within 117 standardized brain atlas regions. Five ROC curves for classifying BD versus non-BD were generated using a support vector machine trained on (1) BLAST-CT (4 features), (2) Regional BLAST-CT (468 features), (3) Regional CSF volumes (117), (4) BLAST-CT + Regional CSF volumes (121 features), and (5) Regional BLAST-CT + Regional CSF volumes (585 features). RESULTS: For BLAST-CT alone, area under the curve (AUC), sensitivity, specificity, and accuracy in classifying BD were 0.7806, 61.11%, 95.00%, and 78.95% respectively. For regional BLAST CT, they were 0.7611, 55.56%, 100%, and 78.95% respectively. For regional CSF volume, they were 0.7056, 94.44%, 50.00%, and 71.05%. For BLAST-CT + regional CSF volumes, they were 0.8389, 94.44%, 65.00%, and 79.85%. For regional BLAST-CT + regional CSF volumes, they were 0.9083, 83.33%, 100%, 92.11%. CONCLUSIONS: Results indicate that the present model using regional BLAST-CT + regional CSF volumes is able to classify brain death with a large specificity, sensitivity, and accuracy above 90%. Given the clinical value of a tool to assist physicians in predicting progression to BD, additional prospective trials with larger samples and testing cohorts should be conducted.