Detection of patient-ventilator asynchrony from mechanical ventilation waveforms using a two-layer long short-term...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Respiratory Care

Background:

• Patient-ventilator asynchrony (PVA) arises from mismatches in mechanical ventilation, leading to adverse clinical outcomes.
• Current automated PVA detection methods using ventilator waveforms lack sufficient efficiency.
• The potential of deep learning for PVA detection remains largely unexplored.

Purpose of the Study:

• To investigate the feasibility of deep learning for detecting patient-ventilator asynchrony.
• To develop and evaluate a long short-term memory (LSTM) network for identifying common PVA types.

Main Methods:

• A 2-layer LSTM network was proposed to detect double triggering (DT) and ineffective inspiratory effort during expiration (IEE).
• Performance was assessed using cross-validation on a combined dataset and cross-testing between two independent datasets.

Main Results:

• The LSTM network achieved superior performance compared to existing rule-based and machine learning models.
• High F1 scores of 0.983 for DT and 0.979 for IEE were obtained on the combined dataset.
• The LSTM network demonstrated strong performance in cross-testing scenarios.

Conclusions:

• Long short-term memory (LSTM) networks show significant promise for accurate PVA recognition in clinical settings.
• This deep learning approach can enhance patient-ventilator interaction by enabling timely detection and correction of PVA.