P019 Automatic Detection of Patient-Ventilator Asynchrony during Non-Invasive Ventilation using Matrix Profiles

Automated Patient-Ventilator Asynchrony (PVA) detection and classification has been described previously, however most approaches target specific asynchrony types, are not applicable in non-invasive ventilation, and require manual pre-processing, removing the opportunity for real-time analysis. During our previous study of polysomnography-assisted NIV implementation (Hannan 2019), each participants’ sleep studies were manually, breath-by-breath scored for PVA. The data were collected over approximately 8 hours and included > 650,000 breath cycles. The source data for the machine learning consisted of the 21 recorded signal traces exported as text files. We employed an outlier detection approach to detect asynchrony with Artificial Intelligence (AI) models as the core engine. The fully automated approach starts with pre-processing to filter and de-noise the polysomnography data, with AI models, including Multidimensional Matrix Profiles, employed to find asynchronies from a signal subset. We initially investigated the best features for PVA and arousal detection in a subset of the 59 participant sleep studies with the highest number of labelled asynchronies and arousals. These six participants had a total of 6,942 events: 799 Arousal, 3,026 Ineffective Triggering, 1,983 Double Triggering and 1,134 Auto Triggering. All events were used in the training/testing stage. The best-balanced model employed a denoising technique with 300 milliseconds as the aggregating data points interval across: mask pressure, ventilator measured flow, abdominal, and thoracic band traces. This approach resulted in 0.80 for both sensitivity and specificity.