I3PT develops an artificial intelligence model to diagnose flow asynchrony in critically ill patients

• The algorithm is able to recognize patterns in the patient's breathing signals from cases it has previously learned

• Its implementation will make it possible to detect and classify in time a type of asynchrony that is very damaging to the lung and is currently underdiagnosed

 When a patient cannot breathe effectively and must be connected to a mechanical ventilator, it is important that the rhythm of their breathing and that of the machine are synchronized. Often, however, the patient needs more air than the machine is giving him at that moment, either because of insufficient airflow or because of a high inspiratory effort. This lack of coordination between the patient and the ventilator is called flow asynchrony and if it is not detected in time it can be very harmful to the lungs.

Even today, the only way for ICU professionals to identify this type of asynchrony is by visual inspection of the mechanical ventilator waves, making its clinical diagnosis difficult and error-prone. "With the current method, flow asynchrony is very underdiagnosed. The ventilators are not prepared to automatically warn when one is occurring and the team cannot be waiting to watch the waves on the screen during the 24 hours of the day”, he explains Cande de Haro, intensivist doctor and researcher at the Parc Taulí Research and Innovation Institute (I3PT).

Intensivist Cande de Haro, principal researcher of the project, in an ICU cubicle

To solve this scenario, De Haro, together with a team made up of research and healthcare staff with multidisciplinary profiles, such as doctors and engineers, from the Institute, has developed a supervised artificial intelligence model to identify and classify flow asynchrony in time in the control volume mechanical ventilation mode – characterized to deliver a constant volume of air with each breath. This algorithm is able to recognize patterns in the patient's breathing signals from cases it has previously learned, and in this way detect when poor patient-ventilator interaction is occurring and the severity level of asynchrony.

“Our model uses advanced deep learning techniques (Deep Learning) what they mimic the workings of the human brain and can learn complex patterns based on data that we have previously given him", he explains Veronica Santos, I3PT engineer and one of the people in charge of setting up the model. Each of these data corresponds to one cycle of mechanical ventilation during a flow asynchrony and is labeled according to its severity. During training, the algorithm analyzed and memorized more than 6.500 different cycles.

The engineer Verónica Santos, one of the people in charge of developing and training the algorithm

In order to guarantee the reliability and accuracy of the model, a multicenter, observational study with adult critically ill patients connected to mechanical ventilation for more than 24 hours to obtain a large and representative data set from two different hospitals: the Parc Taulí University Hospital and the St. Michael's Hospital in Toronto. At the same time, five experts in intensive care have been counted on to classify and label each of the cycles, thus ensuring that the labels used to train the model are accurate and reliable. Analyzes of performance metrics to test the effectiveness of the model and a subanalysis of esophageal pressure have also been performed.

A model with route

One of the team's main challenges is the validation of this model in other modes of mechanical ventilation, like the pressure control – characterized to maintain a constant air pressure in the patient's lungs. "Before incorporating the algorithm into clinical practice, we want to make sure that it is able to adapt and function correctly in other very frequent modes of ventilation," says De Haro. Once this is achieved, he points out, "the intention will be to think about how to implement it on fans as an automatic alert and on platforms like BetterCare, a spin-off from Parc Taulí with which we collaborate and which already analyzes asynchronies with algorithms previously developed by us".