Machine Learning for Mechanical Ventilation Control

Abstract

We consider the problem of controlling an invasive mechanical ventilator for pressure-controlled ventilation: a controller must let air in and out of a sedated patient's lungs according to a trajectory of airway pressures specified by a clinician. Hand-tuned PID controllers and similar variants have comprised the industry standard for decades, yet can behave poorly by over-or under-shooting their target or oscillating rapidly. We consider a data-driven machine learning approach: First, we train a simulator based on data we collect from an artificial lung. Then, we train deep neural network controllers on these simulators. We show that our controllers are able to track target pressure waveforms significantly better than PID controllers. We further show that a learned controller generalizes across lungs with varying characteristics much more readily than PID controllers do.

Deep Sequence Modeling for Pressure Controlled Mechanical Ventilation

“…A simplified respiratory circuit shows the airflow through the inspiratory pathway, entering and exiting the lung, and exiting through the expiratory pathway. (Suo et al, 2021)…”

“…The dataset was collected from a modified open-source ventilator, designed by the People's Ventilator Project (PVP) at Princeton University (LaChance et al, 2020), and connected to an artificial bellows test lung (IngMar, 2020) via a respiratory circuit (Suo et al, 2021). Figure 1 below illustrates the circuit, with the two control inputs highlighted in green and the state variable (airway pressure) to predict in blue.…”

“…A simplified respiratory circuit shows the airflow through the inspiratory pathway, entering and exiting the lung, and exiting through the expiratory pathway. (Suo et al, 2021) The dataset has eight features with approximately 75,000 training breaths and 50,000 test breaths (with over six million rows). It contains details regarding lung attributes, the input and output valves of the machinery, and the pressure inside the lungs connected to the machine.…”

“…Machine Learning for Mechanical Ventilation Control (Suo et al, 2021) is the most significant one. The paper presents two models: the real2sim creates a ventilator simulator based on deep feedforward networks and the sim2real where the data generated by this simulation is used to tune a controller.…”

“…A recent paper demonstrated the potential advantage of machine learning for ventilator control (Suo et al, 2021) on an open-source ventilator (LaChance et al, 2020) designed in response to the COVID-19 pandemic, paving the way for intelligent control methods that are robust and require less manual monitoring. We exploit this study and present a simulator based on a deep sequence model and leverage statistical learning techniques to improve the model's enhancement.…”