Abstract
PURPOSE: Pulmonary complications, including pneumonia and respiratory failure, continue to be major contributors to morbidity and mortality in patients with chest trauma. Although several artificial intelligence (AI) models have been developed to predict trauma mortality, there remains a lack of AI-based prediction models specifically targeting pulmonary complications in chest trauma. To address this gap, we developed and validated an explainable AI model for predicting pulmonary complications. METHODS: This retrospective analysis included 1,040 patients with blunt chest trauma who were treated at a single regional trauma center between January 2019 and March 2023. Pulmonary complications were defined as pneumonia, prolonged mechanical ventilation (>48 hours), or other major thoracic complications necessitating surgical intervention. Machine learning algorithms, including extreme gradient boosting (XGBoost), random forest, adaptive boosting (AdaBoost), light gradient boosting machine (LightGBM), and a deep neural network, were trained using hyperparameter tuning and threefold cross-validation. Model performance was evaluated by sensitivity, specificity, accuracy, balanced accuracy, F1 score, and the area under the receiver operating characteristic curve (AUC). Model interpretability was assessed using Shapley Additive Explanations (SHAP) values. RESULTS: Among the total cohort, 188 patients (18.1%) developed pulmonary complications. In the independent testing dataset (n=208), XGBoost achieved the highest AUC (0.856), while AdaBoost demonstrated the highest balanced accuracy (0.779). All machine learning models outperformed conventional scoring systems. SHAP analysis identified key predictors of pulmonary complications, including age, Injury Severity Score, Glasgow Coma Scale score, Abbreviated Injury Scale of the extremity or head, initial PaO2 to fraction of inspired oxygen ratio, location of the primary rib fracture, and presence of flail motion. CONCLUSIONS: The developed AI model accurately predicts pulmonary complications in patients with chest trauma and outperforms traditional prognostic tools. The model's explainability offers actionable clinical insights, supporting early risk stratification and evidence-based decision-making in trauma care.