A55-15 Incorporating Respiratory Phases Improves Deep Learning Detection of Double-Triggered Breaths in Mechanically Ventilated Patients

Rationale Double-triggered breaths occur when a patient’s respiratory demand exceeds the ventilator’s inspiratory time, leading to the delivery of excessively large tidal volumes that are associated with harm. In intensive care units, the frequency of these asynchronies is rarely captured. Deep learning models using airway pressure and flow waveforms have shown promise for automated detection of such events. We hypothesized that explicitly providing inspiratory and expiratory phase information as an additional input feature would improve model performance in detecting double-triggered breaths. Methods Deidentified ventilator waveform recordings from 33 UC Davis Health patients, enriched for patient-ventilator asynchronies, were annotated by a critical care physician to mark inspiratory-to-expiratory transitions and double-triggered breaths. Each breath, represented by airway pressure and flow signals, was then preprocessed by resampling to 256 timesteps and normalizing signal values on a per-breath basis. A time channel was also included to preserve temporal information. Four deep learning architectures—a convolutional neural network combined with a long short-term memory network (CNN-LSTM), an inception-based model, a temporal convolutional network (TCN), and a transformer model—were trained using five-fold cross-validation, with folds grouped by patient to prevent data leakage and stratified by outcome to maintain balanced class distributions. Model performance was primarily evaluated using the mean area under the precision-recall curve (AUPRC), selected for its ability to more accurately characterize model performance on the minority class of double-triggered breaths. The models were then retrained with an additional binary channel encoding whether each timestep corresponded to inspiration or expiration, and the performance metrics were compared. Results The dataset included 9,719 breaths from patients with diverse causes of respiratory failure, 4% of which were double-triggered events. Incorporating the respiratory phase channel improved the performance across all models, with comparable AUPRC observed among architectures with the respiratory phase information (0.76-0.78; Table 1). This additional input did not significantly increase training time and led to higher F1 scores. Conclusions Incorporating respiratory phase information as a mechanistic feature improves deep learning-based detection of double-triggered breaths in real-world ventilator data. Though further validation in larger, external cohorts is needed, these findings suggest that integrating physiologic context can enhance model performance, particularly when training on smaller, manually annotated datasets.