- A new method uses high-performance computing algorithms to improve virtual Fractional Flow Reserve simulations.
- The algorithm is based on machine learning and deep learning acceleration method called Gaussian Process regression.
- It can perfectly simulate what’s happening inside the heart within 1-2 minutes.
Most multi-scale model of cardiac mechanics look promising but when it comes to diagnosis and treatment, their capabilities are quite limited. Since they are not capable of efficiently handling clinical data, constraining unfixed variables, and handling computational complexity, they can’t provide effective assistance with clinical decisions and medical care.
The most common type of heart disease is Coronary Artery Disease (CAD), which affects nearly 16.5 million Americans adults. It is also the leading cause of death of both women and men in the US. According to the Cleveland Clinic, someone in the US has a heart attack every 40 seconds.
It’s a condition in which coronary arteries get blocked (or narrowed) due to the buildup of fatty and cholesterol (known as plaques) deposits on the inner walls of arteries. These plaques restrict the flow of blood which could lead to heart attack.
Simulating Blood Flow in Arteries
To enhance the diagnosis of such diseases, researchers are exploring new methods to examine blockage in arteries using a technique called virtual Fractional Flow Reserve (vFFR). It utilizes Computational Fluid Dynamics and X-ray angiograms to investigate the fluid movements and simulate the flow of blood in coronary arteries.
To observe the plaque inside arteries, a patient needs to undergo hyperemic agent injections. However, these types of simulations eliminate the requirement of a pressure wire catheter.
Existing vFFR based on computational fluid dynamic algorithms often takes more than a day to generate a complete simulation. To effectively utilize vFFR method(s), it’s necessary to improve the algorithms they are running on, without reducing the diagnostic accuracy. It should be able to compute a full simulation in minutes, providing a broader range-view of blocked arteries.
To meet these requirements, IBM researchers developed a new method that uses high-performance computing algorithms to improve vFFR simulations. The algorithm is based on machine learning and deep learning acceleration method called Gaussian Process regression. It can be used to assist optimization algorithms, even in tricky scenarios where objective functionals can’t be easily differentiated.
Reference: Frontiers Physiology | doi:10.3389/fphys.2018.01002 | IBM
The algorithm takes size, location and transmural depth of the infarct as input variables and model computed changes in simulations. It can execute 40 simulations of infarct within varying locations and shapes. After training on outcomes of finite element simulations, the algorithm provides a useful representation for investigating complex effects.
The hemodynamic simulations for vFFR-based diagnosis execute within 1-2 minutes on POWER9 systems with NVIDIA Tesla V100 GPUs. According to the researchers, this is the first simulation of its kind to be executed in almost real-time.
The quick model simulations can reduce the manual labor and help clinicians quickly examine heart conditions, easing the mental burden for patients waiting on test reports.
Read: Google Develops AI That Predicts Heart Disease By Scanning Your Eyes
This study is part of IBM”s work to build a complete and more precise picture of inner mechanisms of the heart with artificial intelligence and biophysical models. They have published new techniques to visualize what is going on inside the heart on a cellular and anatomical level.
