AI-Assisted Lung Sliding Detection in Point-of-Care Ultrasound by Marine Corps Corpsmen: A Multi-Reader Study.
Authors
Affiliations (6)
Affiliations (6)
- Schulich School of Medicine, London, ON.
- London Health Sciences Centre, London, ON.
- Deep Breathe Inc., London, ON.
- Centro de Investigaciones ClÃnicas, Fundación Valle del Lili, Cali, Colombia.
- Western University, London, ON.
- London Health Sciences Centre and Deep Breathe Inc., London, ON.
Abstract
Artificial intelligence (AI) has the potential to address training limitations and inter-operator variability that constrain the use of lung ultrasound (LUS) in austere and prehospital settings. This pilot study evaluated whether AI-based decision support could improve the diagnostic accuracy and confidence of United States Marine Corps Corpsmen in identifying absent lung sliding, a key indicator of pneumothorax, during LUS interpretation. This pilot-prospective multi-reader, multi-case study involved five military medics, all novices in point-of-care ultrasound, each interpreting 50 de-identified LUS video clips twice, once without AI assistance (control) and once with AI assistance (ATLAS, Deep Breathe Inc., London, Canada), in randomized order with at least a 2-hour washout between sessions. Expert consensus served as a reference standard. Diagnostic performance was assessed using area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, and accuracy. Differences were analyzed using the Random-Reader Random-Case method. Per-clip reader confidence ratings were compared using the Stuart-Maxwell test. AI assistance significantly improved diagnostic performance across all measured outcomes. The mean AUROC increased from 0.72 (SD 0.16) without AI to 0.93 (SD 0.04) with AI (P=.03). Sensitivity rose from 0.63 (SD 0.14) to 0.90 (SD 0.09), specificity from 0.70 (SD 0.15) to 0.86 (SD 0.10), and overall accuracy from 0.67 (SD 0.10) to 0.88 (0.06) (McNemar's test, P<.001). Reader confidence also improved, with high-confidence ratings nearly doubling from 20% to 37%, and low-confidence ratings decreasing from 38% to 33%. These distributional changes were statistically significant (Stuart-Maxwell χ², P<.001). AI support markedly improved the diagnostic accuracy and confidence of novice LUS interpretation for detecting absent lung sliding. These findings suggest that real-time AI-based decision support may help improve access to high-quality LUS in military and other resource-limited care settings.