Automatic classification of patient-ventilator asynchronies with a 1D convolutional neural network: a pilot study

Critically ill patients often experience a mismatch between breathing efforts and mechanical ventilation, leading to adverse outcomes. However, patient-ventilator asynchronies (PVAs) go underdiagnosed because traditional bedside visual evaluation of respiratory waveforms is subjective and cannot be consistently performed. This study aims to develop a robust, automatic deep-learning method for continuous monitoring of respiratory waveforms and PVA classification. Retrospective analysis was conducted on 1-hour intervals of ventilation waveforms (pressure, flow, and volume curves) from 20 intensive care unit (ICU) patients at Humanitas Research Hospital (median age: 61.5 years; 7 women). Events (n=4402) were manually categorized as normal breaths (1008) or one of six PVA classes (flow starvation, delayed cycling, leak, auto-trigger, ineffective effort, double trigger). Classification models, including Random Forest, XGBoost, and a 1D deep convolutional neural network (CNN), were trained using Python. The 1D CNN emerged as the best-performing model in testing, with a weighted average precision, sensitivity, and F1 score of 0.885, 0.880, and 0.880, respectively. Class-specific performance varied, with the highest sensitivity in normal breath (99.0%) and auto trigger (91.3%) classification. Receiver operating characteristic (ROC) curves yielded an average area under the curve (AUC) of 0.98 (max 1.00, min 0.96). These preliminary results highlight the effectiveness of our model in classifying PVAs, showing significant potential for real-time and automated PVA detection in the ICU.