Large-Scale Study on AI's Impact on Identifying Chest Radiographs with No Actionable Disease in Outpatient Imaging.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Diagnostic Accuracy

Background:

• Radiologists face high workloads due to the large volume of chest radiographs.
• A significant portion of outpatient chest radiographs show no actionable findings, impacting efficiency.
• Automated identification of normal or non-actionable cases can streamline reading workflows.

Purpose of the Study:

• To evaluate an AI tool's performance in identifying chest radiographs with no actionable disease (NAD) in an outpatient setting.
• To use objective and reproducible criteria for defining NAD in a large-scale validation study.
• To assess the AI's accuracy in distinguishing between NAD and actionable disease (AD).

Main Methods:

• An independent validation study of 14,057 chest radiographs from an outpatient imaging center.
• Ground truth established by reviewing CXR reports, classifying cases as NAD or AD.
• The AI NAD Analyzer, trained on millions of images, uses a tandem system for case-level output.

Main Results:

• The prevalence of NAD cases was 70.7% in the study population.
• The AI correctly identified NAD cases with a sensitivity of 29.1% and a yield of 20.6%.
• The AI achieved 98.9% specificity, with a miss rate of 0.3% for all cases and 0.06% for significant findings, missing no critical findings.

Conclusions:

• AI can identify approximately 20% of outpatient chest radiographs as having no actionable disease (NAD).
• The AI demonstrates a very low miss rate for significant and critical findings.
• Implementing AI for NAD identification could streamline reading protocols, improving radiologist efficiency and reducing workload.