AI Model Based on Diaphragm Ultrasound to Improve the Predictive Performance of Invasive Mechanical Ventilation...

Area of Science:

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

Background:

• Point-of-care ultrasonography is crucial for assessing diaphragmatic function in mechanically ventilated patients.
• Conventional diaphragm ultrasound assessment is operator-dependent and subjective.
• Previous research utilized 2D speckle-tracking for automatic diaphragm excursion and velocity measurements.

Purpose of the Study:

• To develop an artificial intelligence-multimodal learning framework for predicting weaning failure.
• To enhance individualized weaning strategies using AI-driven predictions.
• To improve the accuracy and objectivity of diaphragm function assessment.

Main Methods:

• Prospective study of 88 critically ill patients undergoing mechanical ventilation for >48 hours.
• Collected diaphragm ultrasound videos and automated measurements of excursion and velocity.
• Developed an AI-multimodal learning framework integrating clinical, laboratory, and ultrasound video data.

Main Results:

• The multimodal co-learning model achieved the highest accuracy (0.8331) and AUC (0.894), outperforming single-modal approaches.
• Diaphragm ultrasound video data alone yielded an accuracy of 0.8095 and AUC of 0.852.
• The model demonstrated excellent calibration and high average precision (0.91).

Conclusions:

• Combining ultrasound and clinical data via co-learning significantly improves weaning outcome prediction accuracy.
• The AI-multimodal approach enhances practical operability and user-friendliness for clinical application.
• The proposed model shows significant potential for widespread use in intensive care settings.