Abstract
BACKGROUND: Post-traumatic brain edema is a critical factor influencing the prognosis of traumatic brain injury (TBI). However, current imaging scoring systems, such as the Marshall and Rotterdam scores, lack specific quantitative criteria for assessing brain edema. This study aims to develop a prognostic prediction model for TBI by integrating the novel A5(+ 1) CT brain edema grading system with advanced machine learning methodologies. METHODS: A total of 216 patients with traumatic brain injury (TBI) were retrospectively enrolled in this study. CT imaging characteristics and clinical parameters within 72 h post-injury were extracted. Variable selection was performed using least absolute shrinkage and selection operator (LASSO) regression analysis. Subsequently, nine machine learning models were developed and their performances were compared. The evaluation metrics included the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and decision curve analysis (DCA). RESULTS: Seven key predictors were identified: age, contusion type, midline shift, CT edema grade, Glasgow Coma Scale score, pulmonary infection, and perilesional CT value. The naive Bayes (NB) model demonstrated superior performance, achieving an AUC of 0.944 in the test set, with 84.6% sensitivity and 94.1% specificity. The CT edema grade was the most significant predictor. A grade of ≥ 3 (indicating bilateral or diffuse edema) was strongly associated with poor Glasgow Outcome Scale scores (p < 0.001), elevated intracranial pressure (p < 0.001), and reduced perilesional CT values (p < 0.001). DCA indicated substantial clinical net benefit when the intervention threshold probability exceeded 40%. CONCLUSION: The ML model incorporating the novel A5(+ 1) CT edema grading system enables precise prognostic prediction for TBI patients. This integrated approach provides a quantitative and dynamic assessment of edema severity, outperforming traditional scoring systems and offering a valuable tool for risk stratification and clinical decision-making.