Real-time detection of respiratory circuit events in mechanical ventilation using deep learning

Respiratory circuit events, including fluid accumulation and circuit or cuff leakage during mechanical ventilation, increase ventilator-associated event risks but often go undetected. We developed a convolutional neural network analyzing 57,296 annotated breaths (26,768 training/internal validation; 30,528 external validation) from 48 patients. The algorithm detected fluid-accumulation-like patterns with an F1-score of 99.90% internally and 92.35% externally, while leakage detection exceeded 99% accuracy. Clinically, 91.7% of patients exhibited circuit events, with fluid-accumulation-like patterns observed in 77.1% of cases and associated with measurable airway pressure increases (median ΔPaw = 2 cmH₂O). The algorithm demonstrated high accuracy and generalizability in detecting respiratory circuit events from waveform data and may allow earlier intervention to reduce ventilator-associated complications through real-time detection.