A machine-learning-based tool for stroke risk prediction in blunt cerebrovascular injury: development and preliminary evaluation

Clinical Snapshot

Key Findings

• P Value: Not reported in abstract

• Effect Size: Biffl grade strongest predictor: carotid aOR 2.02, vertebral aOR 1.44

• Primary Outcome: Stroke occurrence in BCVI patients

• Nnt Or Sensitivity: Sensitivity 66% (33%-100%), Specificity 74% (62%-82%), AUC 0.79 (0.57-0.95)

• Confidence Interval: Carotid 95% CI 1.62-2.53, vertebral 95% CI 1.18-1.77

Clinical Application

High - online calculator available for immediate use, requires only demographic and injury data Relevant to Australian trauma centers; would need validation in Australian population. Compatible with existing trauma protocols and could inform anti-thrombotic therapy decisions within TGA guidelines Trauma patients with blunt cerebrovascular injury requiring stroke risk stratification

Abstract

PURPOSE: Stroke risk correlates with the Biffl grading system in blunt cerebrovascular injury (BCVI). Although anti-thrombotic therapy is the mainstay of stroke prevention, no point-of-care clinical decision-support tool exists to guide timing for therapy. We sought to develop an interactive online calculator that incorporates patient-specific demographic and injury characteristics to estimate stroke risk and risk reduction with anti-thrombotic (AT) administration. METHODS: Data from BCVI patients (n = 1,197) at a Level I Trauma Center were retrospectively collected. Six machine learning methods were employed to predict stroke risk with and without AT therapy. Class imbalance was addressed using downsampling and/or class weighting. Model performance was assessed using 10-fold cross-validation. The model was implemented as an R-based Shiny online application. RESULTS: Stroke rate among the population was 4%, and the strongest predictors for stroke were the greatest Biffl grade of carotid (aOR [95%CI] = 2.02 [1.62-2.53]) and vertebral injuries (1.44 [1.18-1.77]). The least absolute shrinkage and selection operator (LASSO) model outperformed all others, achieving 66% [33%-100%] sensitivity and 74% [62%-82%] specificity for stroke prediction, with an area under the receiver operating characteristic curve of 0.79 [0.57-0.95]. This model was integrated into an interactive online tool ( https://grady-bcvi-calc.shinyapps.io/calculator/ ), where patient demographic and injury characteristics can be used to compute baseline stroke risk and estimate stroke risk with AT. CONCLUSION: We developed and evaluated a preliminary predictive model for personalized stroke risk assessment in patients with BCVI using key risk factors. The integration of patient-specific risk-benefit assessments into clinical decision-making could optimize and reduce variability in AT therapy. External validation is warranted to prepare this tool for broad clinical applicability.

References

1. 1.Wagner, V., Preston, J. D., De Leon Castro, A., Mueller, W. F., Nguyen, J., Garcia-Toca, M., Benjamin, E. R., Todd, S. R., & Sciarretta, J. D. (2026). A machine-learning-based tool for stroke risk prediction in blunt cerebrovascular injury: development and preliminary evaluation. European Journal of Trauma and Emergency Surgery. https://doi.org/10.1097/XCS.0000000000000771