An Interpretable Machine Learning Based Model for Traumatic Severe Pneumothorax Evaluation

Traumatic pneumothorax is a complex condition that is challenging to diagnose, particularly in hospitals, underdeveloped areas, and during mass casualty events. This study aimed to evaluate the potential of machine learning (ML) for diagnosing and assessing traumatic pneumothorax. We extracted 33 vital signs and blood gas parameters from the MIMIC-IV database, selecting 12 clinically significant features as inputs to four ML algorithms: extreme gradient boosting (XGBoost), artificial neural network (ANN), support vector machine (SVM), and k-nearest neighbors (KNN). Five-fold cross-validation was used to train and test the models, with external validation performed on the EICU database. Model performance was evaluated using AUROC, recall, and accuracy, with SHAP interpretability employed to understand feature importance. In total, 3871 participants from the MIMIC-IV database and 22,022 participants from the EICU database were analyzed. Hemoglobin, Oxygenation Index, and pH were found to be key indicators of severe traumatic pneumothorax. XGBoost exhibited the best performance, achieving an AUROC of 0.979 (95% CI: [0.966, 0.989]) on the MIMIC-IV dataset and 0.806 (95% CI: [0.740, 0.864]) on the EICU dataset. The results suggest that ML, particularly XGBoost, is faster and more convenient than traditional imaging methods, making it well-suited for emergency or mass casualty situations. ML algorithms show promise for initial diagnosis of traumatic pneumothorax, with XGBoost demonstrating strong interpretability and robust external validation.