Neural Network-Enabled Identification of Weak Inspiratory Efforts during Pressure Support Ventilation Using...

Area of Science:

• Critical Care Medicine
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Excessive pressure support ventilation (PSV) can lead to diaphragm atrophy and delayed weaning.
• Accurate identification of weak inspiratory efforts is crucial for optimizing PSV.
• Current methods for assessing inspiratory effort can be invasive or labor-intensive.

Purpose of the Study:

• To develop a neural network classifier for detecting weak inspiratory efforts during PSV.
• To utilize ventilator waveform data for automated identification of inspiratory effort.
• To provide a proof-of-concept for AI-driven personalized assisted ventilation.

Main Methods:

• A dataset of flow, airway, esophageal, and gastric pressures was created from 37 critically ill patients.
• A One-Dimensional Convolutional Neural Network (1D-CNN) was trained on data from 22 patients (45,650 breaths).
• The model classified breaths as having weak inspiratory effort based on a threshold of 50 cmH2O*s/min.

Main Results:

• The 1D-CNN model achieved 88% sensitivity and 96% negative predictive value in identifying weak inspiratory efforts.
• Specificity was 72%, and positive predictive value was 40% on data from 15 independent patients (31,343 breaths).
• The model demonstrated feasibility in distinguishing weak inspiratory efforts from ventilator waveforms.

Conclusions:

• A neural network model can effectively identify weak inspiratory efforts during PSV using ventilator data.
• This AI approach offers a non-invasive method for real-time assessment of patient effort.
• The findings support the potential for AI to enable personalized and optimized assisted ventilation strategies.