A researcher is studying patient-ventilator asynchrony (PVA) in critically ill children with machine learning to determine its impact on patient outcomes.

Critically ill children on ventilator support can experience a mismatch between their breathing efforts and) the rhythm delivered by the ventilator. This mismatch, called patient-ventilator asynchrony, is difficult to detect and can worsen patient outcomes. PVA is commonly associated with longer stays on a ventilator for adults and can raise the risks of infection, lung injury and brain damage. However, little is known about PVA in children, where it could be just as, if not more, common.

Robinder Khemani, MD, MsCI, Attending Physician in Pediatric Intensive Care at Children’s Hospital Los Angeles, is using machine learning to improve the outcomes of children put on ventilators.

A CHLA research team led by Khemani has received a $3.4 million grant from the National Institutes of Health to examine the frequency and risk factors for common types of PVA in critically ill children. Working with hospitals in Canada and the Netherlands, the researchers will investigate whether PVA is independently associated with poor clinical outcomes and determine the effects on the body when breathing doesn’t match the flow of air provided by the ventilator.

“It takes a very highly trained human to recognize PVA,” says Khemani. “But computers can do this very well. Our colleagues at the Virtual Pediatric Intensive Care Unit (vPICU) here at CHLA have been working with us on this project for a few years and have developed machine-learning algorithms that can identify different types of breathing asynchronies in children on ventilators.”

The study team will collect measurements from 200 children and combine this data with the analysis of 350 children in other studies, including a clinical trial that is testing a novel ventilator strategy.

“By the end of this project, we hope to have developed these algorithms and validate that they work in three different hospitals using data from many different children,” says Dr. Khemani. “Simultaneously we will build a tool to automatically detect PVA by analyzing ventilator data through machine-learning algorithms. We will test how well the tool helps providers to identify the minute-to-minute changes in patients and potentially alert the bedside team that an adjustment to the ventilator may be needed.”

To minimize the risks of ventilator support, medical teams want to keep patients participating in breathing for themselves as much as possible.

“So that’s where this study really comes into play, by constantly tracking the interaction between the child and the ventilator to ask if the ventilator is supplying just the right amount of help, precisely when needed,” says Khemani.

Children can need ventilator support for multiple reasons, including severe pneumonia or acute respiratory distress syndrome (ARDS), when infection or trauma causes swelling, inflammation and fluid buildup in the lungs. The body’s response to the initial injury can harm the lungs even more than the infection or trauma itself.

“Many of these very sick patients can develop unexpected complications from the very procedures that we use to help them,” says Khemani. Ventilator-induced lung injury can lead to heart and kidney damage, or can increase vulnerability to future lung disease, asthma or sleep-disordered breathing.

“Brain function can also be impaired by all the medications, anesthetics and sedation patients receive to help them to tolerate the ventilator. We weigh the risks and benefits to minimize potential harms and hopefully get them off the ventilator as soon as they are ready.”

Mismatches between patient breathing and the rhythm the ventilator provides can occur in different ways, as children’s breathing varies according to their weight, size and age. Respiration patterns can also change during the course of a child’s stay in the pediatric intensive care unit. But existing studies use different definitions for PVA subtypes and no study so far has been large enough to evaluate the relationship between different types of PVA and patient outcomes, or has yet focused on the highest-risk patients.

“There are many types of PVA, but we still don’t know which PVA subtypes are most harmful or are the most frequent,” says Khemani. “We need to develop a common set of definitions and measurements, especially for pediatric patients.”