Automatic classification of patient status based on ventilation data

Abstract

Mechanical ventilation therapy is a vital treatment for a myriad of pulmonary complications. Optimally administering it demands continuous expert supervision from pulmonologists due to the patient's dynamic health status and requirements. The pulmonologist adjusts the mechanical ventilator's settings to meet the patient's ventilation needs. The needs are derived from performing skilled, invasive manoeuvres to determine the status of the respiratory system parameters, such as airway resistance and static compliance. However, practicality and insufficient expertise lead to error-prone, periodical assessments, leading to suboptimal mechanical ventilation and resource management. Therefore, a need persists for continuously monitoring and logging a patient's pulmonary parameters in a way that allows limited pulmonologists to still supervise patients practically at scale. An automated solution is proposed that assesses the mechanical ventilation time-series data to classify the ventilation mode, derive the ventilator settings, extract informative features, and predict the status of the airway resistance and static compliance. The ventilation mode classifier produced prediction accuracies over 99%. After a breath's ventilation mode is predicted, its corresponding ventilator settings determined, and its informative features extracted, the airway resistance and static compliance are predicted with accuracies exceeding the capabilities of human experts at over 98%. These findings not only show that autonomous monitoring of a patient's health status is possible, but also that it could be done continuously with high accuracies. Logging the patient's health status on a per-breath-basis at scale can reduce unnecessary complications and improve the efficiency of pulmonologist-to-patient ratios in demanding circumstances (such as the Covid-19 pandemic).