Machine Learning Techniques in Predicting Delayed Pneumothorax and Hemothorax Following Blunt Thoracic Trauma

Background: Delayed pneumothorax and hemothorax are among the possible fatal complications of blunt thoracic trauma. Objectives: Finding reliable criteria for timely diagnosis of high-risk patients has been an area of interest for researchers. Material sand Methods: We gathered a database including 616 patients among which, 17 patients experienced the delayed complications. Employing four classification techniques, namely, linear regression, logistics regression, artificial neural network, and naive Bayesian classifier, we tried to find a predictive pattern to recognize patients with positive results based on recorded clinical and radiological variables at the time of admission. Results: First, without using machine learning techniques, we tried to predict the complications based only on a single variable. We recognized chest wall tenderness as the best single criterion that enables to classify all high-risk patients with 100% sensitivity (95% CI, 82-100). This criterion potentially excludes 57% (95% CI, 53-61) of low-risk patients from further observation. Then we used the machine learning techniques to assess the effect of all admission time variables together. According to our results, amongst the utilized techniques, logistics regression model enables not only to exclude 81% (95% CI, 77-84) of patients without complications from unnecessary observation, but also to recognize all patients with true positive results for pneumothorax and hemothorax (95% CI, 82-100). Conclusions: Instead of serial chest X-ray, patients with blunt chest trauma could be initially evaluaed by a risk assessment model in order to avoid unnecessary work-up.