The intersection of machine learning and medicine: new study from UW Medicine reveals top predictors of right heart failure after LVAD implantation News

If you’re a fan of “Grey’s Anatomy,” chances are the word “LVAD wire” probably means something to you. Unlike the show, left ventricular assistive devices (LVADs) have capabilities far beyond the realm of hospital romance.

An LVAD is a mechanical pump designed for patients with advanced heart failure. They are implanted in the top of the heart to help the lower left chamber (left ventricle) pump blood from the ventricle, through the aorta, and to the rest of the body. The pump is then attached to a cable that leads out of the body to a remote computer, which provides alarms and messages that help operate the system. LVAD devices extend the lives of thousands of heart failure patients each year.

However, these devices are not without the risk of serious complications. According to UW Medicine, more than 20% of LVAD recipients experience right heart failure (RHF) due to the right ventricle not being able to withstand the sudden revival of blood flow from the pump. This results in a reduced chance of survival, or even instant death, within days of implantation.

This outcome, often devastatingly unpredictable, piqued the curiosity of researchers at UW Medicine.

Using a machine learning (ML) system trained to search for 186 different factors, experts have identified the top 30 preimplantation patient factors strongly associated with right heart failure after LVAD implantation.

“Many patients, even if they survive, have a very poor quality of life and a major contributor to that is RHF,” Dr Song Lian assistant professor of cardiology at UW Medicine and one of its authors study, said. “It’s hard to predict in advance, so we were interested in trying a new method to improve these predictions.”

This new method references the groundbreaking logistics of explainable ML. The ability to analyze hundreds of variables simultaneously makes explainable ML much better equipped for the high-dimensional interactions between factors involved in this research.

“Many other AI machine learning models are really just black boxes, limiting their usefulness in medicine,” Li said. “We need an explainable ML technique to properly apply ML.”

Standard ML models are notoriously limited to proving correlations without explanation, often referred to as black boxes.

Based on a sample population of 20,000 LVAD patients, the study found that the top five predictors of RHF are the INTERMACS profile, model for terminal liver disease score, the number inotropic infusionshemoglobin and race.

Of the 186 preimplantation factors, narrowing the possible predictors of RHF down to five is an important discovery that will help physicians assess and manage a patient’s risk before surgery even takes place.

One factor in particular sparked further interest from Li and his team: race. African Americans had a higher risk of acute RHF after LVAD implantation compared to their white counterparts.

“It’s very puzzling to see why that is and we want to dig deeper to find out what might be driving that correlation,” Li noted. “It’s actually something we’re analyzing right now.”

The study also acknowledged the limitations of the data. The progression of RHF after LVAD implantation is not solely dependent on those pre-implant factors, and the conditions of operative and post-operative care must also be taken into account.

Going forward, Li explained how they plan to use the ML model to simulate various optimization strategies.

“Before we even think about testing it in actual patients, we can see how much of a difference it would really make,” Li said.

This unique intersection between machine learning and medicine is proving to be a successful endeavor – a collaboration that only a school like UW could discover and deliver.

Reach contributing writer Meha Singal at Twitter: @mehaha23

Do you like what you read? Support quality student journalism by donating here.

Leave a Comment