Artificial Intelligence-Guided Lung Ultrasound by Nonexperts.

Summary

In a multicenter diagnostic validation (n=176), trained nonexpert healthcare professionals using AI guidance achieved diagnostic-quality 8-zone lung ultrasound in 98.3% (95% CI 95.1%-99.4%) of cases, statistically indistinguishable from expert-acquired studies (difference 1.7%; 95% CI -1.6% to 5.0%). Masked expert readers validated image quality, and analyses followed intention to treat.

Key Findings

• AI-guided THCPs produced diagnostic-quality LUS in 98.3% (95% CI 95.1%-99.4%) of studies.
• No significant difference in image quality versus LUS experts (difference 1.7%; 95% CI -1.6% to 5.0%).
• Multicenter, masked expert panel validation; intention-to-treat analysis on 176 participants.

Clinical Implications

Enables task-shifting of LUS acquisition to trained nonexperts, supporting triage for dyspnea, heart failure, and pulmonary edema. Could underpin scalable point-of-care ultrasound programs and tele-expertise workflows.

Limitations

• Did not assess downstream clinical outcomes (e.g., management changes, patient outcomes)
• Generalizability beyond trained THCPs and participating sites remains to be tested

Future Directions

Evaluate impact on clinical decision-making, outcomes, and cost-effectiveness; assess performance in diverse settings, including prehospital and low-resource environments; integrate with tele-ultrasound workflows.