Performance of machine learning algorithms in predicting the need for surgical fixation in pediatric craniomaxillofacial trauma

Comprehensive Summary

This study by Thornton et al. examines whether machine learning (ML) algorithms can reliably determine which pediatric patients with craniomaxillofacial (CMF) trauma are likely to need surgical care. The authors conducted a large retrospective analysis using data from over 121,000 children in the National Trauma Data Bank, building several ML models based on clinical, anatomic, and injury-severity data. 80% of the data was used to train 6 different models (Random Forest, RidgeClassifier, AdaBoost, XGBoost, Neural network, and Logistic regression) and because only 11.3% of the patients underwent surgical fixation, the data was resampled using Synthetic Minority Over-sampling Technique (SMOTE). Models were tested on the remaining 20% of data and assessed using ROC-AUC, F1 and F2 scores, precision, recall, and specificity. They found that the XGBoost model performed marginally better than the other models, with an ROC-AUC of 0.89, and identified key predictors of operative need (mandibular fractures, fracture multiplicity, moderate-to-severe facial injury, and neurologic sequelae); injuries like closed cranial base fractures and moderate head trauma tended to shift predictions toward conservative management. Shapley Additive Explanations (SHAP) values were used to explain which features influenced the model’s predictions most and highlighted predictable shortcomings. This demonstrated that fracture type, injury severity, and fracture multiplicity were key contributors in model prediction but other important features like neurologic and TMJ/dental injuries were underweighted. In the discussion, the authors explain that using different sensitivity and specificity thresholds depending on institutional resources may improve local performance and emphasize that although ML shows clear promise, real-world use will require models that incorporate imaging and undergo external validation.

Outcomes and Implications

This research is important because pediatric CMF trauma is highly variable, and clinicians do not currently have standardized tools to determine which injuries truly require early surgical intervention. By demonstrating how ML models can convert complex trauma data into actionable predictions, the study shows how ML can be used to make triage decisions that are more consistent and less dependent on institutional experience. In practice, such tools can help guide early referrals and improve surgical planning, especially in hospitals where pediatric CMF expertise is limited.