Patient-Ventilator Interaction using Autoencoder derived Magnitude of Asynchrony Breathing

Abstract

The occurrence of asynchronous breathing (AB) is prevalent during mechanical ventilation (MV) treatment. Despite studies being carried out to elucidate the impact of AB on MV patients, the asynchrony index (AI), a metric to describe the patient-ventilator interaction, may not be sufficient to quantify the severity of each AB fully in MV patients. This research investigates the feasibility of using a machine learning-derived metric, the ventilator interaction index (VI), to describe a patient's interaction with a mechanical ventilator. VI is derived using the magnitude of a breath's asynchrony to measure how well a patient is interacting with the ventilator. 1,188 hours of hourly AI and VI for 13 MV patients were computed using a convolution neural network and an autoencoder. Pearson's correlation analysis between patients’ AI and VI versus their levels of partial pressure oxygen (PaO2) and partial pressure of carbon dioxide (PaCO2) was carried out. In this patient cohort, the patients’ median AI is 38.4% [Interquartile range (IQR): 25.9-48.8], and the median VI is 86.0% [IQR: 76.5-91.7]. Results show that high AI does not necessarily predispose to low VI. This difference suggests that every AB poses a different magnitude of asynchrony that may affect patient's PaO2 and PaCO2. Quantifying hourly VI along with AI during MV could be beneficial in explicating the aetiology of AB.