Background: Blunt thoracic trauma is a common cause of trauma admissions, with rib fractures representing the most frequent thoracic injury. These fractures are associated with a high risk of pulmonary complications, particularly pneumonia, which is a leading cause of morbidity and mortality in this patient group. Early identification of patients at high risk for pneumonia remains challenging due to the delayed onset of symptoms. Current clinical decision-making often lacks objective tools to guide surgical intervention, especially in patients without clinical flail chest. There is increasing interest in developing a predictive model that incorporate both clinical and radiological information to improve personalized risk assessment followed by tailored treatment strategies.
Objective: This study aimed to develop a machine-learning model to predict pneumonia risk following traumatic rib fractures by integrating clinical data with automatically extracted radiological features from chest CT scans. Furthermore, the goal was to build a fully automated workflow as a proof of concept to support clinical decision-making and optimize treatment selection, including decisions regarding early intervention or surgical rib fixation.
Methods: A retrospective cohort study was conducted using data from adult patients with CT-confirmed rib fractures treated at Erasmus MC. Clinical parameters were extracted from electronic health records and included demographics, comorbidities, injury severity, and treatment factors. Radiological parameters were obtained through a automated CT analysis method, which integrated a previously validated rib fracture detection model with new modules for segmenting pulmonary contusion, pneumothorax, and hemothorax. A clinical dataset, a radiological dataset, and a combined dataset were constructed. After exclusion of patients not meeting inclusion criteria and removal of features with substantial missing values, categorical variables were encoded and continuous variables standardized. Feature selection was performed using Least Absolute Shrinkage and Selection Operator (LASSO) regression, with constraints to prevent collinearity between clinically redundant features. Five classifiers (Naive Bayes, Logistic Regression, Random Forest, SVM, and XGBoost) were trained and evaluated using stratified train-validation-test splits. The model with the highest area under the ROC curve (AUC) on the validation set was selected, retrained on the full training set, and finally tested on an independent held out test set to assess performance on unseen data.
Results: The study included 399 patients in the clinical dataset and 325 patients in the radiologic dataset due to incomplete CT data. Pneumonia occurred in 11\% of patients. Several clinical and radiologic features were significantly associated with pneumonia, including mechanical ventilation, sternal fracture, pulmonary contusion, hemothorax, and occurance of bilateral fractures. Automated detection accuracy for the pulmonary injury model was 69% for contusion, 77% for pneumothorax, and 71% for hemothorax. The highest AUC on the test set was achieved by the combined model (AUC = 0.83), followed by the automatic model (AUC = 0.76), the clinical model (AUC = 0.74), and the small clinical model (AUC = 0.73). The combined model reached a sensitivity of 100%, correctly identifying all pneumonia cases in the test set and a specificity of 66.7% and was selected as the final model.
Conclusion: This study shows that combining clinical data with automatically extracted CT features enables effective prediction of pneumonia in patients with rib fractures. Despite segmentation challenges, particularly for hemothorax and contusion, and limitations in label quality, key imaging features were consistently selected, indicating their clinical relevance. Further development and validation on a larger dataset could establish this approach as a practical tool for early risk stratification and management in thoracic trauma.