Page 60 - JSOM Spring 2026

AI-Assisted Lung Sliding Detection in
Point-of-Care Ultrasound by Marine Corps Corpsmen

A Multi-Reader Study

3
Melissa Côté, MSc *; Ross Prager, MD ; Khoa Tran, MSc ;
2
1
4
6
5
Nicolas Orozco, MD, MSc ; Delaney Smith, MSc ; Zoë Holliday ; Robert Arntfield, MD 7

ABSTRACT
Background: Artificial intelligence (AI) has the potential to Introduction
address training limitations and inter-operator variability that
constrain the use of lung ultrasound (LUS) in austere and pre- Lung ultrasound (LUS) is widely recognized as a rapid and
hospital settings. This pilot study evaluated whether AI-based reliable modality for ruling out pneumothorax in trauma and
decision support could improve the diagnostic accuracy and prehospital settings. Its advantages over chest radiograph,
confidence of United States Marine Corps Corpsmen in iden- particularly portability and diagnostic accuracy, has led to
tifying absent lung sliding, a key indicator of pneumothorax, the integration of point-of-care ultrasound (POCUS) proto-
during LUS interpretation. Methods: This pilot-prospective cols like the extended focused assessment with sonography
1,2
multi-reader, multi-case study involved five military med- in trauma (eFAST) exam, into military trauma algorithms.
ics, all novices in point-of-care ultrasound, each interpreting Despite these benefits, frontline adoption of LUS remains lim-
50 de-identified LUS video clips twice, once without AI as- ited among non-physician military providers due to the skill
3
sistance (control) and once with AI assistance (ATLAS, Deep required to obtain and interpret diagnostic-quality images.
Breathe Inc., London, Canada), in randomized order with at
least a 2-hour washout between sessions. Expert consensus Previous efforts to introduce ultrasound into combat casualty
served as a reference standard. Diagnostic performance was care have shown that brief, focused training can enable med-
assessed using area under the receiver operating characteristic ics to achieve reasonable accuracy in detecting absent lung
1
curve ( AUROC), sensitivity, specificity, and accuracy. Differ- sliding, the hallmark sonographic sign of pneumothorax.
ences were analyzed using the Random-Reader Random-Case However, performance remains variable across users, and skill
method. Per-clip reader confidence ratings were compared using degradation over time presents a barrier to reliable availability
4
the Stuart-Maxwell test. Results: AI assistance significantly im- in austere or operational environments. In response, artificial
proved diagnostic performance across all measured outcomes. intelligence (AI) has emerged as a potential tool to enhance
The mean AUROC increased from 0.72 (SD 0.16) without AI image interpretation and reduce user dependency. AI models
to 0.93 (SD 0.04) with AI (P=.03). Sensitivity rose from 0.63 capable of detecting key sonographic features such as lung
(SD 0.14) to 0.90 (SD 0.09), specificity from 0.70 (SD 0.15) sliding have demonstrated strong diagnostic performance in
5
6,7
to 0.86 (SD 0.10), and overall accuracy from 0.67 (SD 0.10) preclinical, retrospective and prospective studies, includ-
to 0.88 (0.06) (McNemar’s test, P<.001). Reader confidence ing those validating the generalizability of AI detection across
also improved, with high-confidence ratings nearly doubling multiple centers and diverse ultrasound vendors. However, the
from 20% to 37%, and low-confidence ratings decreasing impact of AI on real-time interpretation by novice military us-
8
from 38% to 33%. These distributional changes were statis- ers remains underexplored.
tically significant (Stuart-Maxwell χ², P<.001). Conclusion: AI
support markedly improved the diagnostic accuracy and con- The objective of this pilot study was to determine whether AI
fidence of novice LUS interpretation for detecting absent lung assistance could improve the diagnostic performance and in-
sliding. These findings suggest that real-time AI-based decision terpretive confidence of United States Marine Corps Corps-
support may help improve access to high-quality LUS in mili- men in identifying absent lung sliding with LUS. Employing a
tary and other resource-limited care settings. multi-reader, multi-case (MRMC) study, it was hypothesized
that AI support would significantly improve the area under
the receiver operating curve (AUROC) as the primary out-
Keywords: lung ultrasound; point-of-care ultrasound;
pneumothorax detection; artificial intelligence; eFAST; come. Additional analyses included evaluating potential im-
combat casualty care provements in sensitivity, specificity, accuracy, and diagnostic
confidence.
*Correspondence to mcote9@uwo.ca
1 Melissa Côté is a medical student and Canadian Armed Forces Biosciences Officer affiliated with Schulich School of Medicine, London, ON.
3
2 Dr. Ross Prager is an Intensivist affiliated with the London Health Sciences Centre, London, ON. Khoa Tran is a Software Engineer affiliated
with Deep Breathe Inc., London, ON. Nicolas Orozco is a Data Scientist associated with the Centro de Investigaciones Clínicas, Fundación Valle
4
5
6
del Lili, Cali, Colombia. Delaney Smith is a Data Scientist affiliated with Deep Breathe Inc., London, ON. Zoë Holliday is an undergraduate
student in Medical Biophysics at Western University, London, ON. Dr. Robert Arntfield is an Intensivist and Ultrasound Expert affiliated with
7
the London Health Sciences Centre and Deep Breathe Inc., London, ON.
58
58