Machine learning-based detection of missed inspiratory efforts using esophageal pressure during noisy pressure...

Missed inspiratory efforts represent one of the most frequent and clinically relevant forms of patient-ventilator asynchrony during assisted mechanical ventilation. Their detection typically requires expert inspection of ventilator waveforms or the use of esophageal pressure (PESO), which limits continuous bedside recognition. Machine-learning approaches may enable automated, real-time identification of these events.

In this pilot observational physiological study, mechanically ventilated patients receiving pressure support ventilation were recorded under three levels of ventilatory variability (0%, 45%, and 90%). Airway pressure (PAO), flow, and esophageal pressure (PESO) signals were acquired synchronously at 200 Hz. In a physiological subset, breaths were manually annotated as normal or missed efforts based on the presence of an inspiratory PESO deflection without ventilator triggering. Waveforms then underwent standardized preprocessing including denoising, breath segmentation, expiratory-phase normalization, and generation of processed Flow-PAO difference curves. From the automatically extracted Flow-PAO waveform dataset, a balanced subset of 408 breaths (204 normal breaths and 204 missed efforts) was manually annotated and used for supervised machine-learning training and internal validation. The trained classifiers were subsequently applied to the full 36,096-breath waveform dataset to assess robustness across ventilatory variability levels. Model performance was evaluated using hold-out, 10-fold, and leave-one-out cross-validation.

A total of 36,096 respiratory cycles were automatically extracted from ventilator waveforms, while a separate 1422-breath PESO subset was used to establish the physiological definition of missed inspiratory efforts. A balanced manually annotated subset of 408 breaths was used for machine-learning training and validation. The majority-vote ensemble consistently achieved the highest performance across validation strategies, with accuracy ranging from 93.1% to 94.6% and AUC-ROC up to 0.96.

A machine-learning ensemble trained on ventilator waveforms and anchored to PESO-validated ground-truth labeling enables accurate and reliable detection of missed inspiratory efforts during assisted ventilation. These findings support the feasibility of automated real-time monitoring of patient-ventilator interaction and provide a methodological framework for future integration into intelligent ventilator systems.