Prospective Real-Time Validation of a Lung Ultrasound Deep Learning Model in the ICU

Area of Science:

• Critical Care Medicine
• Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Lung ultrasound (LUS) is crucial for diagnosing respiratory conditions in critically ill patients.
• Distinguishing between normal (A-line) and abnormal (B-line) lung patterns is key for patient management.
• Deep learning (DL) models show promise in automating image analysis.

Purpose of the Study:

• To assess the real-time accuracy of a bedside deep learning model for lung ultrasound analysis.
• To evaluate the model's ability to differentiate between A-line and B-line patterns in critically ill patients.
• To explore the feasibility of deploying AI-powered LUS at the point of care.

Main Methods:

• A prospective, observational study was conducted in an academic ICU.
• A previously trained LUS deep learning model was deployed on a portable device for real-time predictions.
• Model performance was evaluated against blinded expert review on 400 LUS clips from 100 critically ill patients.

Main Results:

• The real-time DL model achieved 95% accuracy, 93% sensitivity, and 96% specificity in identifying B-line patterns.
• Performance was consistent with previous offline validation of the model.
• Adjustable prediction thresholds allowed for dynamic optimization of sensitivity and specificity.

Conclusions:

• A validated deep learning model for LUS performs effectively in real-time at the ICU bedside.
• This study demonstrates the feasibility of automated medical imaging analysis in critical care.
• Further research is warranted to investigate the clinical impact of real-time AI in managing critically ill patients.