A model-based approach to generating annotated pressure support waveforms

Area of Science:

• Biomedical Engineering
• Respiratory Medicine
• Artificial Intelligence in Healthcare

Background:

• Patient-ventilator asynchronies during pressure support ventilation (PSV) are linked to increased patient discomfort, work of breathing, and mortality.
• Accurate detection of asynchronies is essential for implementing lung-protective ventilation strategies.
• Current machine learning (ML) approaches for asynchrony detection are hindered by the need for large, diverse, and high-quality clinical datasets.

Purpose of the Study:

• To propose and validate a novel method for generating large, realistic, and labeled synthetic datasets for training ML algorithms to detect diverse types of patient-ventilator asynchronies.
• To assess the feasibility of using synthetic data to train ML models for reliable asynchrony detection.

Main Methods:

• Developed a model-based approach utilizing a non-linear lung-airway model adapted for diverse patients and a first-order ventilator model.
• Generated labeled synthetic pressure, flow, and volume waveforms simulating PSV with various asynchronies.
• Evaluated the model's ability to reproduce basic lung mechanics and compared simulated waveforms against clinical data using expert clinician assessment (Fisher's exact test).

Main Results:

• The generated synthetic waveforms were indistinguishable from clinical data by experienced clinicians (P = 0.44).
• ML models trained on clinical data demonstrated comparable detection performance on both simulated (94.3% true positive rate, 93.5% positive predictive value) and clinical data (98% true positive rate, 98% positive predictive value).
• The model successfully generated labeled waveforms representing different types of asynchronies.

Conclusions:

• A model-based approach can generate realistic, labeled synthetic waveforms for PSV, effectively simulating various asynchronies.
• Synthetic datasets are suitable for training ML algorithms to detect patient-ventilator asynchronies, addressing data limitations in clinical settings.
• This method facilitates the development of robust ML-based decision support tools for lung-protective ventilation.