Machine Learning for Pneumothorax in Trauma victims: cross-sectional validation study (PneumoDetect)

Abstract Background: Pneumothorax is a potentially fatal condition that requires early diagnosis and prompt management upon arrival at the Emergency Department(ED). The purpose of this study is to validate a Pneumothorax Machine learning (PneumoDetect) model designed on both an online and in-hospital dataset, and to compare its accuracy to that of radiologist and emergency physician Method: We conducted a cross-sectional study using an online available open access tool. We obtained a hospital dataset from January 1, 2010 to December 31, 2020, and extracted 4,788 DICOM X-ray images. A machine learning team manually labelled the images from hospital records. We performed internal validation using a supervised learning machine learning model with a Convolutional Neural Network architecture implemented in Python and Medcalc. We calculated Kappa statistics were calculated using STATA v14.2 to assess the model’s performance. Additionally, we generated AUROC curves using sensitivity, specificity, positive and negative predictive values, and accuracy metrics. Results: The initial training of the PneumoDetect model showed a validation accuracy of 96.4%, followed by pre-trained model with 98% accuracy & a fine-tuned model having 97.9% accuracy. The sensitivity was found to be 93.99%, specificity was 91.34, PPV was 92.88, NPV was 92.67, and the overall accuracy was 92.79%. PneumoDetect was highly accurate while there was only moderate agreement between the radiologist and emergency physician in presence of Pneumothorax. Conclusion: Our diagnostic investigation discovered that developing neural networks and advanced ML models may be used to diagnose pneumothorax using machine learning models. Integrating such AI systems into physician workflows for preliminary interpretations has the potential to provide physicians with early diagnostics and profound alerts that can help to better diagnose occult pneumothorax and reduce human errors, particularly in resource-constrained settings. This can improve overall accuracy and lower healthcare cost. Funding Source: Fogarty International Centre of the National Institutes of Health under Award Number D43TW007292