Incorporating Immunoproteins in the Development of Classification Models of Progression of Intracranial Hemorrhage After Traumatic Brain Injury. Issue 5 (September 2021)

• Record Type: Journal Article
• Authors: Hinson, H. E.; Li, Peter; Myers, Loren; Agarwal, Chinar; Pollock, Jeffrey; McWeeney, Shannon
• Is Part Of: Journal of head trauma rehabilitation. Volume 36:Issue 5(2021)
• Publication Date: 2021-09
• Subjects: predictive modeling -- progression of intracranial hemorrhage -- progressive hemorrhagic injury -- traumatic brain injury -- traumatic intracranial hemorrhage; Brain damage -- Patients -- Rehabilitation -- Periodicals; Brain damage -- Periodicals; 617.4810443
• DOI: 10.1097/HTR.0000000000000654
• Languages: English
• ISSNs: 0885-9701

Abstract

Abstract : Objective: To define clinical, radiographic, and blood-based biomarker features to be incorporated into a classification model of progression of intracranial hemorrhage (PICH), and to provide a pilot assessment of those models. Methods: Patients with hemorrhage on admission head computed tomography were identified from a prospectively enrolled cohort of subjects with traumatic brain injury. Initial and follow-up images were interpreted both by 2 independent readers, and disagreements adjudicated. Admission plasma samples were analyzed and principal components (PCs) composed of the immune proteins (IPs) significantly associated with the outcome of interest were selected for further evaluation. A series of logistic regression models were constructed based on (1) clinical variables (CV) and (2) clinical variables + immune proteins (CV+IP). Error rates of these models for correct classification of PICH were estimated; significance was set at P < .05. Results: We identified 106 patients, 36% had PICH. Dichotomized admission Glasgow Coma Scale ( P = .004), Marshall score ( P = .004), and 3 PCs were significantly associated with PICH. For the CV only model, sensitivity was 1.0 and specificity was 0.29 (95% CI, 0.07-0.67). The CV+IP model performed significantly better, with a sensitivity of 0.93 (95% CI, 0.64-0.99) and a specificity of 1.0 ( P = .008). Adjustments to refine the definition of PICH and better define radiographic predictors of PICH did not significantly improve the models' performance. Conclusions: In this pilot investigation, we observed that composites of IPs may improve PICH classification models when combined with CVs. However, overall model performance must be further optimized; results will inform feature inclusion included in follow-up models. Abstract : Supplemental Digital Content is Available in the Text.