In a heartbeat: Machine learning speeds up heart-valve simulations

 

Simulation of fluid moving through a valve Simulation of heart movement Simulation of fluid moving through heart valves

Adarsh Krishnamurthy and Ming-Chen Hsu got to talking about their research interests – Krishnamurthy works in simulating cardiac mechanics and Hsu in simulating valve dynamics – when they had an idea.

“He has the heart without the valve, and I have the valve without the heart,” Hsu said. “We thought ‘why don’t we just put them together?’”

 

Modeling heart-valve replacements

Modeling heart valves and the mechanics of the fluid moving through the heart chambers can be used to predict deformations and diseases that affect the valves. Simulations can also add in a virtual prosthetic replacement valve to examine its effectiveness.

Replacing valves can be a very risky procedure for patients with valvular heart disease, so knowing ahead of time as much as possible about how a prosthetic valve will help a specific patient with specific circumstances will save money and lives.

But, one simulation of heart-valve movement (dynamic closure and opening) takes about four and half hours to complete, and one simulation of fluid movement (with fluid-structure interaction) through a valve takes about one week. In addition, development of a coupled fluid-structure interaction model with accurate movement of the ventricles is complicated, and this can take years.

 

From years to minutes

So Krishnamurthy, assistant professor of mechanical engineering, and Hsu, associate professor of mechanical engineering, knew who they needed to add to their team: Soumik Sarkar, associate professor of mechanical engineering, and an expert in machine learning. They also enlisted the help of Aditya Balu, graduate research assistant in mechanical engineering.

The team now hopes to teach a machine to predict heart valve simulations, and eventually fluid-structure interaction simulations of heart valves by providing enough data and simulations to learn from. Their initial tests of the machine learning have produced predictions that are 95% accurate according to the actual physics simulations that are run.

The potential with machine learning is simulations that normally take hours or days could be predicted by a machine within a matter of seconds. The long-term goal is to design custom-created heart valves unique to patients’ anatomy and health problems.

 

Combining expertise, solving problems

This combined research project falls under the umbrella of the Center for Multiphase Flow Research and Education (CoMFRE).

“CoMFRE is about the fluids,” Krishnamurthy said. “But with our expertise, we can also look at how the fluids interact with other physics. So, we are really working closely with the Multiphysics part of CoMFRE in this project.”

The team’s work is different from what people typically think of as mechanical engineering, but they are able to do this work because Iowa State encourages creative new lines of research that apply engineering across traditional discipline lines.

“A lot of people think, ‘Oh, you guys are doing biology.’” Hsu said. “No, it’s a very mechanical problem. At its core, a heart is a pump and it has four valves. They should open, and they should close – and that’s what we are doing here!”