Artificial intelligence as a second reader for pneumothorax detection: redirecting rather than reducing chest CT use.

Abstract

Background: Pneumothorax is a finding that is frequently missed in the emergency department, particularly in cases with subtle radiographic features, and can rapidly progress to tension pneumothorax when overlooked on the initial chest radiograph. Although artificial intelligence (AI)-assisted chest radiograph interpretation has been shown to improve diagnostic accuracy, its effect on the actual decision behaviour of emergency physicians - particularly on chest computed tomography (CT) ordering - remains insufficiently studied. This study evaluated the impact of AI-assisted chest radiograph interpretation on emergency physicians' pneumothorax diagnostic decisions and chest CT ordering behaviour, with a particular focus on the capacity of the AI alert to function as a second reader and capture diagnoses initially missed by the physician.

Methods: This was a nationwide, online, cross-sectional simulation study conducted between 1 and 15 April 2026. Emergency physicians practising in Türkiye were asked to interpret 20 standardised posteroanterior chest radiographs presented in a randomised order (10 pneumothorax-positive [5 overt and 5 subtle] and 10 pneumothorax-negative cases, all confirmed by chest CT). For each case, participants first indicated whether they suspected pneumothorax on the unannotated radiograph, and then - after being shown the same image annotated by a CE-certified, deep learning-based chest radiograph interpretation system - indicated whether they would order a chest CT. The primary endpoint was the paired pre-AI versus post-AI change in simulated CT-ordering intent, derived from a predefined survey coding scheme and analysed using the McNemar exact test. As the pre- and post-AI items do not capture CT-ordering intent identically, this endpoint reflects simulated imaging intent rather than observed CT-ordering behaviour.

Results: A total of 231 physicians (51.5% general practitioners, 25.5% emergency medicine residents, 22.9% emergency medicine specialists) completed the survey, yielding 4,620 physician-case response pairs. Simulated CT-ordering intent rose from 8.1% before AI exposure to 26.7% after AI exposure (absolute difference + 18.6% points; p < 0.001). The increase was concentrated in pneumothorax-positive cases (+ 31.6 points; p < 0.001) and remained limited in pneumothorax-negative cases (+ 5.7 points; p < 0.001). In subtle pneumothorax cases, the integrated diagnostic capture rate rose from 54.8% to 85.0% (+ 30.2 points); among the 339 cases in which physicians had initially answered "no pneumothorax," 211 (62.2%) were converted into a CT order after the AI alert. The effect was statistically significant across all professional titles and seniority strata (p < 0.001 for all). The mean perceived usability score of the system was 8.1 out of 10 (interquartile range 7-10).

Conclusions: In this simulation, AI-assisted chest radiograph interpretation did not reduce simulated CT-ordering intent; rather, the increase in intent was concentrated in true positive cases, with only a limited effect on negative cases. These findings are hypothesis-generating and require prospective confirmation with real CT-ordering data. The principal clinical value of the system lies in capturing pneumothorax diagnoses initially missed by emergency physicians, supporting its positioning as a second reader rather than as a CT-reduction tool.

Trial registration: Not applicable.