Machine learning can customize oxygen levels for ventilated patients

Supplemental oxygen is one of the most widely prescribed treatments in the world, with an estimated 13 to 20 million patients worldwide needing oxygen provided by a ventilator each year. A ventilator is a type of life support machine that works like a bellows, moving breathable air in and out of the lungs. Ventilators have evolved far beyond the “iron lung” machines some may imagine, and have now evolved into sleek, small, digital machines that deliver oxygen through a tiny plastic tube that goes down the throat.

Despite technological advances, amount The amount of oxygen to administer to each patient remains a guessing game: Clinicians prescribe oxygen concentrations using equipment that records SpO.₂ Oxygen saturation measures the amount of oxygen in a patient's blood. However, previous studies have only measured SpO₂ The target is what's better for the patient.

The standard of care is to maintain oxygen saturation between 88 and 100. Within that range, physicians have had to choose a ventilator oxygen level without high-quality data on which to base their decision. Whether we like it or not, making that decision for each patient exposes the patient to the potential benefits or harms of the selected oxygen level.”

Kevin Buell, MBBS, Pulmonary and Critical Care Fellow, University of Chicago Medicine

To take the guesswork out of ventilators, Buell and a group of other researchers used machine learning models to study whether the effects of different oxygen levels depend on individual patient characteristics. JAMAsuggesting that individualized oxygenation targets may reduce mortality, which could have far-reaching implications for intensive care.

Previously, several research groups have conducted randomized trials to investigate whether higher or lower oxygen levels are better overall for patients, but most studies have not provided a clear answer. Buell and his colleagues say that neutral results do not indicate that oxygen levels have no effect on patient outcomes, but rather that the results of treatment with different oxygen levels vary from patient to patient, simply because the level of oxygen used is not related to the patient's overall condition. Average Randomized trials have shown zero effect.

As personalized medicine becomes more prevalent, there is growing interest in using machine learning to make predictions for individual patients. In the context of mechanical ventilation, these models might use specific patient characteristics to predict the ideal oxygen level for each patient. These characteristics include age, gender, heart rate, temperature, and reason for admission to the intensive care unit (ICU).

“We aimed to make evidence-based, individualized predictions of who would benefit from lower or higher oxygen targets when placed on a ventilator,” said Buell, co-lead author of the study.

The previous randomized trials were not a waste. Buell and his collaborators used data from those studies to design and train a machine learning model. After the model was developed using trial data collected in the United States, the collaborators applied it to data from patients around the world, from Australia and New Zealand. For patients who received oxygenation within the target range that the machine learning model predicted would benefit, mortality could have been reduced by 6.4% overall.

It's not possible to generalize predictions based on a single characteristic. For example, not all patients with brain injuries benefit from desaturation, although the data is biased in that direction. That's why clinicians need tools like the researchers' machine learning models to piece together a mosaic of each patient's needs. But while the algorithms themselves are complex, Buell noted that in the future, anyone could easily implement these kinds of tools because the variables the medical team inputs are all common clinical variables.

At UChicago Medicine, medical teams are already using algorithms integrated directly into the electronic health record (EHR) system to inform other areas of clinical decision-making. Buell hopes that ventilators will one day do the same. For hospitals that might not have the resources to integrate machine learning into their EHRs, he also envisions creating a web-based application where clinicians can input patient characteristics and get predictions — like an online calculator. There's a lot of validation, testing, and refinement to be done before clinical implementation becomes a reality, but given the end goal, future research is worth the investment.

In an editorial accompanying the article's publication, renowned critical care expert Derek Angus, MD, wrote: “If the results are true and generalizable, they will be astounding. If we could instantly assign every patient to the appropriate group of predicted benefit or harm, and assign oxygen targets accordingly, this intervention would theoretically produce the greatest single improvement in lives saved from critical illness in the history of the field.”