Page 63 - JSOM Spring 2026

TABLE 1 Distribution of Reader Confidence Ratings AI-augmented ultrasound has practical implications for front-
No. (%) of readers line military medicine. Real-time decision support may enable
Corpsmen to more confidently diagnose or exclude pneumo-
Confidence scores AI-unassisted AI-assisted P value * thorax at the point of injury, allowing timely interventions such
Not at all confident 20 (8.2) 10 (4.1) as needle decompression when indicated. AI tools embedded
Slightly confident 73 (29.8) 71 (29.0) on portable ultrasound devices can offer immediate guidance
Moderately confident 102 (41.6) 73 (29.8) <.001 in the field, reducing reliance on extensive pre-deployment
Confident 40 (16.3) 70 (28.6) training or teleconsultation in situations where connectivity is
10
Very confident 10 (4.1) 21 (8.6) limited. This approach is especially valuable given the por-
2
*Paired Stuart–Maxwell χ test assessing marginal homogeneity of tability of handheld LUS systems and the higher sensitivity of
confidence score distributions between AI-unassisted and AI-assisted ultrasound compared to chest radiography for pneumothorax
reads. detection. Findings from this and prior studies suggest that
2
AI = artificial intelligence. AI support increases user confidence, which may encourage
3
providers to act more decisively when procedures such as de-
of 0.98. McNemar’s test indicated statistically significant dif- compression are warranted. By reducing the level of expertise
ferences in the distribution of correct versus incorrect classifi- required to use ultrasound effectively, AI may help expand the
cations when AI alone was compared with both the AI-assisted role of point-of-care imaging in operational environments such
(P<.001) and unassisted (P<.001) conditions. as rural clinics, humanitarian missions, and forward-deployed
military units.
Discussion
Limitations
AI assistance significantly improved the ability of United States This study has several limitations. First, the sample size was
Marine Corps Corpsmen with limited ultrasound experience small and consisted exclusively of Corpsmen, which may limit
to identify absent lung sliding on LUS, a key sonographic sign the generalizability of the findings to non-military providers or
of pneumothorax. In this pilot study, AI support increased sen- to operational contexts where different training backgrounds
sitivity, specificity, overall accuracy, and AUROC compared and clinical responsibilities exist. Second, the analysis focused
to unassisted interpretations. The addition of AI also boosted only on the identification of lung sliding, a single sonographic
user confidence, doubling the frequency of high-certainty as- feature, and therefore does not reflect the broader diagnostic
sessments. These findings support the hypothesis that real-time utility of LUS in trauma or critical care. Third, the AI predic-
AI guidance can enhance diagnostic performance and confi- tions were presented to participants as binary outputs with-
dence in prehospital military medical care, particularly in aus- out any interpretive overlay or confidence metric, which may
tere or resource-constrained environments. differ from future AI implementations. Although the AI tool
was developed by a collaborating institution, bias was mini-
These results align with prior research demonstrating that AI mized through independent expert consensus for ground truth
can narrow the performance gap between novice and expert labeling and predefined statistical methods applied uniformly
ultrasound users. For example, in a Vietnamese intensive care across conditions. Fourth, despite a 2-hour washout period,
unit, non-expert clinicians improved their real-time lung slid- participants reviewed the same set of 50 ultrasound clips in
ing interpretation accuracy from 68.1% to 93.4% with AI both sessions, while none reported explicitly recognizing clips,
assistance, closely matching the 95.0% accuracy achieved by a learning effect cannot be excluded. Finally, the study relied
expert users on the same task. That study also documented in- on clip-based interpretation of previously recorded ultrasound
3
creased user confidence and reduced interpretation time. Sim- videos rather than live scanning, so it does not fully capture
ilar effects were observed in this study, where AI support not the cognitive and technical demands of image acquisition,
only improved accuracy but also shifted users toward higher probe technique, and real-time clinical integration. Future
confidence in their assessments. In high-stakes settings such as studies should examine AI-assisted LUS in live operational set-
prehospital trauma care, increased diagnostic confidence may tings to assess its impact on procedural guidance and real-time
reduce hesitation and facilitate more timely clinical interven- clinical decision-making.
tions. Notably, these benefits were observed in participants
with minimal ultrasound training, suggesting that AI tools In summary, this study contributes to the growing evidence
may empower less-experienced providers to make accurate that AI can improve the accessibility and quality of POCUS
and confident decisions at the point of care. interpretation by novice users. With further development
and integration, AI-augmented LUS could enhance prehospi-
This study is among the first to evaluate AI-augmented LUS tal diagnostic capabilities and help enable more timely, con-
interpretation in a military population. Prior research has fident decision-making by frontline personnel operating in
shown that with focused training, medics can achieve reason- resource-limited environments. Overall, this work should be
able diagnostic accuracy using handheld ultrasound devices interpreted as an encouraging early step rather than a defin-
in simulated combat scenarios. In one study, medics detected itive conclusion. While the findings suggest that AI may sup-
pneumothoraces with 85%–91% sensitivity and 80% spec- port medics in ultrasound interpretation, validation in larger
ificity, substantially outperforming physical examination and more diverse populations, and in live operational settings,
alone, which yielded 72% sensitivity. However, maintaining will be necessary to establish broader applicability and impact.
1
ultrasound proficiency typically requires ongoing training and
practice. AI augmentation may offer a scalable solution to Conclusion
9
this expertise barrier. With minimal instruction, novice users
can outperform conventional techniques and approach the in- This pilot study demonstrates that an AI-based LUS interpreta-
terpretive accuracy of experienced practitioners. tion tool can significantly enhance the diagnostic performance

AI-Assisted Lung Sliding Detection | 61