New findings from a large retrospective study suggest an artificial intelligence (AI) algorithm provides robust detection of clinically actionable pneumothorax and an 86-minute reduction in median reporting time.
For the study, recently reported in Academic Radiology, researchers compared the use of AI-enabled pneumothorax detection (Critical Care Suite 1.0, GE HealthCare) with a PACS triage system for 12,728 chest X-rays (CXRs) versus unassisted pneumothorax detection in 14,669 CXR scans.
Overall, the study authors found that the AI platform had a 78 percent area under the curve (AUC), 97 percent sensitivity and 60 percent sensitivity. However, when examining the use of the AI algorithm for moderate and large pneumothorax cases, the study authors noted the specificity rate and significant increases in the AUC (93 percent) and sensitivity rate (89 percent).
“This represents reasonable accuracy within a real-world clinical practice dataset, and the sensitivity for moderate/large (pneumothorax) is even within range of reported sensitivities of radiologists in detecting pneumothorax, which has been reported at (85 percent),” wrote study co-author Amit Gupta, M.D., who is affiliated with the Department of Radiology at University Hospitals Cleveland Medical Center in Ohio, and colleagues.
Use of the AI-enabled platform also facilitated significant workflow efficiencies, according to the study authors.
For routine priority pneumothorax cases that occurred during on-call hours, researchers noted a 57.1 percent reduction in the median reporting time for the AI algorithm in comparison to the control group (148 minutes vs. 345 minutes).
Three Key Takeaways
1. Enhanced detection for clinically actionable pneumothorax. The AI algorithm demonstrated strong overall specificity (97 percent) for all positive pneumothorax cases, and a significantly increased area under the curve (AUC) and sensitivity rate for moderate and large pneumothorax cases.
2. Significant workflow efficiency. Using AI reduced the median reporting time for all positive pneumothorax cases by 46.2 percent, routine priority pneumothorax cases by 57.1 percent and STAT priority pneumothorax cases by 29.3 percent, showcasing considerable workflow improvements.
3. Broad applicability across imaging priority levels. The AI tool’s triage effectiveness extended across different priority levels, suggesting it can improve workflows even beyond existing STAT priority protocols, particularly in high-volume settings.
The study authors noted that the AI platform facilitated a 29.3 percent reduction in the median reporting time for STAT priority pneumothorax cases (65 minutes vs. 92 minutes), and a 46.2 percent reduction for all cases with positive findings for pneumothorax (100 minutes vs. 186 minutes),
“This substantial decrease in reporting time was present in PTx- positive CXRs ordered not only with routine priority, but also with STAT priority … . This finding suggests that use of a pneumothorax-detecting AI tool in a triaging role improves workflow beyond conventional prioritization practices like STAT priority ordering,” emphasized Gupta and colleagues.
(Editor’s note: For related content, see “Key Takeaways from Multiple Radiology Societies on AI Assessment and Integration,” “Study: AI Enhances Abnormality Detection on CXR Across Radiologist Experience Levels” and “FDA-Cleared AI Triage Software for Chest X-Rays Offers Enhanced Detection of Pleural Effusion and Pneumothorax.”)
In regard to study limitations, the authors acknowledged that an AI-enabled scanner was the primary scanner utilized in the ICU whereas non-AI scanners were used in in the ICU, non-ICU patient settings and the facility’s emergency department. These differences may have been a prevailing factor in the imbalance of confirmed pneumothorax cases between the control group and the AI cohort, according to the researchers.
