Deep Neural Networks for Ventilator Pressure Prediction

Abstract

Invasive mechanical ventilation is a common medical treatment required for applications spanning anesthesia, neonatal intensive care, and life support for the current COVID-19 pandemic. Despite its importance, the core technology of medical ventilation has remained largely unchanged for years. It is common for a clinician to continuously monitor and adjust the ventilator for a patient manually thus increasing workload. With machine learning (ML) taking center stage in healthcare recently, the question has been asked whether ML based control methods can be developed to replace manual intervention. The main challenge however remains the robustness, safety, and the high cost of development. In addition, the controller must be able to adapt reliably and quickly across varying clinical conditions and requirements not observable directly to the clinicians. Proportional-Integral-Derivative (PID) controllers have been to go to method of choice because of its limited parameters size, fewer samples for tuning, and its ability to generalize over the dynamic lung conditions. Current ventilator or patient simulators are trained by an ensemble of multiple models each simulating the parameters of a single lung. However, human lungs and respectively their parameters or attributes are dynamic and form a continuous space therefore, based on the physiological differences in patient lungs a parametric approach is better suited to improve generalization. This work centers around the possibility of developing an ML-based method which can improve performance simultaneously across a wide range of lung parameters based upon the ISO standard for performance of ventilatory support equipment (ISO 80601-2-80:2018). The results are compared against previously published data and closely match the expected outcome. This is a significant improvement towards a more robust alternative to PID tuning and more importantly cost-effective mechanical ventilators.