Deep learning based automated assessment of end-inspiratory pause maneuver reliability in invasive mechanical ventilation.

OBJECTIVE: Manual end-inspiratory pause maneuvers (EIPM) for plateau pressure (Pplat) measurement suffer from significant variability, yet no objective assessment tool exists. This study develops an automated framework to evaluate EIPM reliability during mechanical ventilation. APPROACH: We proposed an automated framework in which a one-dimensional convolutional neural network (1DCNN) was developed to extract discriminative features from pressure, flow, and volume waveforms. This model accurately classified candidate reliable EIPMs, unreliable EIPMs, and non-EIPM events. For candidate reliable EIPMs, the method subsequently calculated the hold time to definitively identify reliable EIPMs (hold time >= 2s) and hold-time subthreshold EIPMs (hold time < 2s). MAIN RESULTS: The 1DCNN model demonstrated robust classification performance, achieving F1 scores of 0.935 for candidate reliable EIPMs, 0.924 for unreliable EIPMs, and 0.992 for non-EIPM events. The subsequent hold-time calculation, applied to candidate reliable EIPMs, achieved 100% sensitivity in classifying reliable and hold-time subthreshold EIPMs. SIGNIFICANCE: The proposed method provides a reliable tool for the automated quality control of EIPM operations, holding significant potential to support a more standardized and protocolized implementation of lung-protective ventilation strategies.