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

Abstract

Machine learning offers a novel approach to improve surgical triage in pediatric craniomaxillofacial trauma, where decision-making often relies on clinician judgment without standardized criteria. This retrospective study used data from the National Trauma Data Bank (2017–2022) to evaluate the performance of machine learning algorithms in predicting the need for open surgical fixation among pediatric patients with facial fractures. Multivariable logistic regression identified fracture patterns, neurologic and dental injuries, and trauma center characteristics as significant predictors of open surgical fixation. XGBoost outperformed four other models (area under the receiver operating characteristic curve (ROC-AUC) 0.89) and demonstrated strong calibration. Shapley Additive Explanations (SHAP) confirmed that facial injury severity, mandible fractures, and fracture multiplicity were key drivers of operative predictions, aligning with clinical priorities to restore occlusion and stabilize the facial skeleton. Conversely, closed cranial base fractures and head injury severity reduced the predicted likelihood of surgery, consistent with conservative management. False positives and false negatives reflected gaps in radiographic detail and clinical context, including fracture displacement and surgical contraindications. These findings highlight the potential of machine learning to support the early identification of surgical candidates using structured data available at the time of presentation and underscore the need for multimodal approaches to further refine triage accuracy.