Skip to content

Machine learning models with SHAP explanations for predicting neurosurgical intervention and six month functional outcome after traumatic brain injury.

Oct 2026 · Injury · Vol 57 12, pp. 113825 · 0 citations · 32 references
Medicine

Abstract

Background

Admission information can help describe surgical treatment and subsequent functional outcomes following traumatic brain injury (TBI). We developed interpretable machine learning models to address these two tasks.

Methods

This retrospective study included 673 adults evaluated at a single center between December 2018 and December 2023. Outcomes recorded were neurosurgical intervention and six-month Glasgow Outcome Scale status, dichotomized as favorable (scores 4-5) or unfavorable (scores 1-3). Patients were randomly allocated to training and internal validation sets in a 7:3 ratio. Least absolute shrinkage and selection operator (LASSO) regression was used to select predictors for the XGBoost and random forest models. Receiver operating characteristic (ROC) analysis, along with classification and prediction error metrics, was employed for model evaluation. Shapley additive explanations (SHAP) were used to interpret model predictions.

Results

There were 75 unfavorable outcomes and 66 surgical interventions. The reported validation accuracy for XGBoost was 87.56% for predicting functional outcomes and 95.02% for surgical interventions. Validation Brier scores were 0.083 and 0.035, respectively. The Glasgow Coma Scale (GCS), pupillary light reflex, and blood glucose levels contributed significantly to functional outcome predictions, while GCS, fibrinogen, and white blood cell count were key predictors for surgical interventions. Interpretation of threshold-dependent measures is limited by class imbalance and incomplete specification of outcome coding.

Conclusion

The models combined routinely collected admission variables to characterize the treatment received and functional outcomes in a selected traumatic brain injury (TBI) cohort. These findings require confirmation through analysis replication and independent validation before they can be applied clinically.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.