Rapid prediction of cerebral edema on CT scan after traumatic brain injury.
Authors
Affiliations (8)
Affiliations (8)
- Fischell Department of Bioengineering, University of Maryland College Park, College Park, Maryland, USA.
- Program in Trauma, University of Maryland Baltimore School of Medicine, Baltimore, Maryland, USA.
- Emergency Medicine, University of Maryland Baltimore School of Medicine, Baltimore, Maryland, USA.
- Charles McC. Mathias, Jr., National Study Center for Trauma & Emergency Medical Systems, University of Maryland Baltimore School of Medicine, Baltimore, Maryland, USA.
- Epidemiology and Public Health, University of Maryland Baltimore School of Medicine, Baltimore, Maryland, USA.
- Information Systems, University of Maryland Baltimore County, Baltimore, Maryland, USA.
- Department of Anesthesiology, University of Maryland Baltimore School of Medicine, Baltimore, Maryland, USA.
- Institute of Health Computing, University of Maryland Baltimore School of Medicine, Baltimore, Maryland, USA.
Abstract
Traumatic brain injury (TBI) affects 1.7 million individuals annually in the USA, with cerebral edema (CE) as a critical determinant of outcomes and neurosurgical interventions. Delays in TBI diagnosis and triage in prehospital or resource-limited settings contribute to suboptimal care. This study evaluated the predictive potential of machine learning models using clinical and physiological data to detect CE on initial head CT scans, addressing this gap. We conducted a mixed retrospective and prospective study of 1222 suspected TBI patients. Data included clinical characteristics, radiographic scores (Marshall and Rotterdam), and physiological features derived from ECG and photoplethysmography signals collected within 1 hour of admission. Statistical models employed logistic regression and gradient boosting, using Shapley additive explanations for feature importance. Model performance was assessed using the area under the receiver operating characteristic curve (AUROC). The models demonstrated high predictive accuracy for CE-related radiographic features, including midline shift and cisternal abnormalities (AUROC: 0.79-0.85). Clinical features such as Glasgow Coma Scale components and intubation status, combined with physiological variables like heart rate variability, contributed significantly to predictions. In contrast, predictions for other findings (eg, epidural hematoma) showed lower discriminatory power. Prehospital applicability was highlighted by the reliance on readily available physiological data. Machine learning models effectively predict CE prior to CT scans, offering a rapid decision support tool for triage and neurosurgical prioritization in austere and resource-limited settings. Early identification of CE could improve patient outcomes by optimizing transport and treatment strategies. Future research should focus on multicenter validation and streamlined data collection to enhance generalizability and clinical utility. Level III, Prognostic.