Impact of Artificial Intelligence Assistance on Diagnosing Traumatic Pneumothorax: A Comparison of Specialist and Non-Specialist Emergency Physicians.

INTRODUCTION

Supine chest radiography is routinely used in trauma care; however, its sensitivity is limited in pneumothorax detection. Although artificial intelligence (AI) has recently been introduced as a diagnostic support tool for chest radiograph interpretation, its use in trauma radiographs and its interaction with physician experience remain unclear. We aimed to evaluate the diagnostic performance of AI-assisted image interpretation in traumatic pneumothorax detection among emergency medicine specialists and non-specialists.

METHODS

In this retrospective single-center study, 34 supine chest radiographs (17 pneumothorax and 17 non-pneumothorax cases confirmed with computed tomography) were interpreted by 20 emergency medicine physicians (10 specialists and 10 non-specialists). Each participant reviewed all radiographs twice: first without AI assistance and then with AI assistance after a 2-week washout period. Sensitivity, specificity, accuracy, and precision were calculated. To account for clustering of observations within readers and cases, mixed-effects logistic regression analysis was performed.

RESULTS

AI assistance significantly improved sensitivity and diagnostic accuracy across participants, whereas specificity and precision were not significantly affected. Specialists demonstrated higher baseline sensitivity and accuracy than non-specialists. In the mixed-effects logistic regression analysis, AI assistance was independently associated with improved diagnostic accuracy (odds ratio 2.37, p < 0.001). Additionally, the interaction between AI assistance and physician specialist status was significant (p = 0.010).

CONCLUSION

AI-assisted interpretation of supine chest radiographs is a potentially useful decision-support tool for detecting traumatic pneumothorax in emergency settings.