Automated assessment of early hypoxic brain edema in non-enhanced CT predicts outcome in patients after cardiac arrest

Introduction: Early prediction of potential neurological recovery in patients after cardiac arrest is challenging. Recent studies suggest that the densitrometic gray-white matter ratio (GWR) determined from cranial computed tomography (CT) scans may be a reliable predictor of poor outcome. We evaluated an automated, rater independent method to determine GWR in CT as an early objective imaging predictor of clinical outcome. Methods: We analyzed imaging data of 84 patients after cardiac arrest that underwent noncontrast CT within 24 h after arrest. To determine GWR in CT we applied two methods using a recently published automated probabilistic gray-white matter segmentation algorithm (GWR aut) and conventional manual measurements within gray-white regions of interest (GWR man). Neurological outcome was graded by the cerebral performance category (CPC). As part of standard routine CPC was assessed by the treating physician in the intensive care unit at admission and at discharge to normal ward. The performance of GWR measures (automated and manual) to predict the binary clinical endpoints of poor (CPC3-5) and good outcome (CPC1-2) was assessed by ROC analysis with increasing discrimination thresholds. Results of GWR aut were compared to GWR man of two raters. Results: Of 84 patients, 55 (65%) showed a poor outcome. ROC curve analysis revealed reliable outcome prediction of GWR aut (AUC 0.860) and GWR man (AUC 0.707 and 0.699, respectively). Predictive power of GWR aut was higher than GWR man by each rater (p = 0.019 and p = 0.021, respectively) at an optimal cut-off of 1.084 to predict poor outcome (optimal criterion with 92.7% sensitivity, 72.4% specificity). Interrater reliability of GWR man by intra-class correlation coefficient (ICC) was moderate (0.551). Conclusion: Automated quantification of GWR in CT may be used as an objective observer-independent imaging marker for outcome in patients after cardiac arrest. (C) 2016 Elsevier Ireland Ltd. All rights reserved.