Purpose: To evaluate the prognostic performance of 4 CT scoring systems in pediatric patients with moderate-to-severe traumatic brain injury (TBI) and develop a simplified, interpretable predictive model based on machine learning.
Methods: This retrospective study included 103 pediatric patients with moderate-to-severe TBI admitted to a tertiary children's hospital in Southwest China from September 2020 to December 2023. CT images were assessed using the Marshall score, Rotterdam score, Helsinki score, and Stockholm score. Clinical outcomes were defined by the Glasgow outcome scale (GOS) and categorized as favorable results (GOS 4-5) or unfavorable results (GOS 1-3). The dataset was divided into a training set (n=83) and a test set (n=20). Class imbalance was corrected using the random over-sampling examples method. Eight classification models were compared through 5-fold cross-validation. The Naive Bayes model showed the best performance and was simplified to include 5 key predictors. An interactive online application was developed to provide individualized prognostic estimation and visualization.
Results: The Helsinki score demonstrated the highest predictive accuracy among the 4 CT scores (area under the curve (AUC)=0.906), followed by the Stockholm score (AUC=0.897), Rotterdam score (AUC=0.837), and Marshall score (AUC=0.764). The simplified Naive Bayes model achieved an AUC of 0.930, with 100% sensitivity and 65.9% specificity in the test set. The model enables real-time outcome prediction and visual interpretation of contributing factors.
Conclusions: The simplified Naive Bayes model outperforms traditional CT scoring systems in predicting outcomes of pediatric moderate-to-severe TBI. Its easy use, interpretability, and web-based implementation support its potential for clinical application. Further prospective and multicenter studies are needed to validate these findings.

