Identifying Patient-Ventilator Asynchrony on a Small Dataset Using Image-Based Transfer Learning.

Area of Science:

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

Background:

• Mechanical ventilation is crucial for patients unable to breathe independently.
• Patient-ventilator asynchrony (PVA) is linked to adverse clinical outcomes.
• Deep learning for PVA detection requires extensive annotated data, limiting its use.

Purpose of the Study:

• To develop and evaluate a transfer learning approach for PVA detection using small datasets.
• To adapt pretrained CNNs for recognizing PVA by converting 1D signals to 2D images.
• To assess the efficacy of the proposed method compared to non-transfer learning techniques.

Main Methods:

• Utilized a transfer learning architecture with pretrained CNNs.
• Converted 1D respiratory signals into 2D images for feature extraction.
• Implemented a partial dropping cross-validation technique for performance evaluation on limited data.

Main Results:

• The transfer learning method achieved approximately 90% accuracy with only 1% of the data.
• Performance was comparable to non-transfer learning methods on large datasets.
• Non-transfer learning methods showed accuracy below 80% with 1% of the data.

Conclusions:

• The proposed transfer learning method provides satisfactory accuracy for PVA detection with small datasets.
• This approach facilitates the application of deep learning for identifying various PVA types across different ventilation modes.