PVADet: fast patient-ventilator asynchrony detection on waveforms

Patient-ventilator asynchrony (PVA) is a common and critically import clinical problem in patients receiving mechanical ventilation. However, PVAs are often underrecognized, underestimated and delayed, and there has been minimal success in automating their detection. In this study, we develop an efficient and fast end-to-end model to recognize PVAs on ventilator waveforms: running the model costs 106.5ms on CPUs and 7.8ms on GPUs. We propose label striping and stripe-mask encoding for efficient multi-class multi-target detecting. The model innovatively integrates causal convolutional, depth-wise separable convolutional, and recurrent neural networks to memorize long short-term causal features. With 60s waveform segments, our model performs a cross-validation mean average precision (mAP) of 88.1% and a testing mAP of 65.7% for comprehensive PVA detection. Our approach might be implemented as a monitoring tool to automatically identify PVAs for improving bedside and remote care and prioritizing patient comfort.