Heart-valve replacements can restore normal blood flow for patients with aortic valvular disease — but only when the prosthesis is a good fit.
A drawback of prosthetic heart valves is that they’re manufactured in fixed sizes.
In a presentation Monday at the GPU Technology Conference in San Jose, Iowa State professor Adarsh Krishnamurthy and graduate student Aditya Balu presented their research using deep learning to determine the exact valve shape to fit a patient.
Eventually, valve manufacturers could use the AI system to create prosthetic aortic valves tailored for each patient.
“The cardiac surgeon will choose whichever one has the closest size to match the patient, but you can’t get the best fit possible,” Krishnamurthy said.
An Open-and-Shut Case
The heart’s left ventricle supplies oxygenated blood to the entire body. Its outlet, the aortic valve, is a thin tissue structure consisting of three leaflets that fit together. These flaps need to open and close smoothly as the heart pumps blood, with enough overlap to prevent blood from leaking back into the heart’s left ventricle.
A valve replacement — 90,000 of which take place each year in the U.S. — may be needed if a patient’s aortic valve narrows or regurgitates blood, most often due to calcium buildup or congenital abnormalities.
By determining key measurements like a patient’s aortic diameter from an MRI or CT, alongside other parameters, AI can predict the valve geometry that optimizes the function of a prosthetic valve.
“Designing a heart valve can take thousands of simulations to test different input parameters. That’s three or four days of simulations.” Krishnamurthy said. “Instead of this time-consuming analysis, we can replace it with a deep learning model that can give you results within a second per simulation, or an hour or two total.”
This research was carried out in collaboration with Ming-Chen Hsu and Soumik Sarkar, assistant professors at Iowa State University. The team trained a deep learning model on NVIDIA P40 GPUs in under four hours. Running on an NVIDIA TITAN Xp GPU, the model takes less than 5 seconds to run a valve simulation.
Valves with a Longer Lifespan
Prosthetic valves come in two varieties: mechanical and tissue. Mechanical ones are durable, but can damage blood cells as the flaps quickly open and close, and require patients to take blood thinning medications. Tissue or bioprosthetic valves, developed in the last two decades, are made from the outer heart lining of pigs or cows.
Since they’re made from material similar to human heart valves, bioprosthetics don’t require patients to take additional medications — but are less durable than mechanical ones.
Durability matters, because with a bioprosthetic valve, a patient may need to undergo a second surgery to replace the prosthetic valve after 10 to 15 years. That could mean several open-heart surgeries over the lifetime of a patient who gets a bioprosthetic valve relatively early in life.
A valve well-fitted for a particular patient would be designed to minimize these stresses and increase durability.
“The heart valve opens and closes around 80 times a minute, or whatever the patient’s heart rate is,” Krishnamurthy said. “If the leaflets are not perfectly designed, the stresses and strains experienced by the leaflet are higher, leading to a shorter lifespan of the valve due to fatigue.”
Main image shows a simulated heart valve, created by the researchers’ deep learning model. Image courtesy Adarsh Krishnamurthy and Aditya Balu.