Recent developments in machine learning (M L) have led to the creation of automated systems that help radiologists analyze chest X-rays. However, current ML models for diagnosing pneumothorax do not work well when tested on cases where the patient has not been treated yet. This happens because the models often rely on the presence of a chest tube to make predictions, instead of focusing on detecting the pneumothorax itself. Therefore, this study looks at whether adding the task of identifying chest tubes during training can help improve the model’s performance on untreated pneumothorax cases. The study tested different model designs using various chest X-ray datasets. We also used an extra model to estimate chest tube information when it was missing. To see if the models improved the study; we included a detailed analysis and used special tools to understand how the models were making decisions. The goal of this study is to make ML models better at detecting untreated pneumothorax which can help doctors make more accurate diagnoses.