Scientists have proposed using artificial intelligence to develop new prosthetic heart valves. The combination of machine learning models together with mathematical optimization algorithms makes it possible to accelerate the selection of multiple parameters of the prosthesis flaps and obtain a heart valve of an "ideal" configuration several thousand times faster than in the case of classical development based on a multiple cycle of prototype manufacturing and its research. Optimization will help to reduce the stress that occurs in the wings of the prostheses during operation, which means that it will make them more durable. The results of the study, supported by the Russian Science Foundation (RNF), are published in the journal Frontiers in Bioengineering and Biotechnology.

More than 75 million people worldwide suffer from heart valve defects, that is, such malfunctions in which the valve either does not open or does not close completely. Because of this, the heart experiences excessive loads, which can lead to heart failure and other serious complications. The only effective way to treat heart valve defects is their prosthetics, that is, replacement with artificial analogues.

However, long-term observations of patients with artificial valves have shown that 10-15 years after installation, the prostheses begin to collapse due to the immune response of the body and constant mechanical loads. Therefore, scientists are trying to optimize the shape and design of prostheses so as to protect them from excessive stress and premature wear. To do this, using computer programs, an artificial heart valve is simulated and the loads acting on the resulting structure are evaluated. Since it is almost impossible to "guess" the optimal shape from the first time, its search turns out to be very long and resembles the "trial and error" method.

Scientists from the Research Institute of Complex Problems of Cardiovascular Diseases (Kemerovo) with colleagues from the Polytechnic University of Milan (Italy) and the University of Trent (Italy) proposed combining machine learning models and optimization algorithms to find the best geometry of heart valves.

The approach proposed by the authors is as follows: the optimizer program gives machine learning models some combinations of the most important geometric parameters of the valve: height, thickness, diameter, angle between its flaps, strength of the material for the prosthesis, places where the valve experiences maximum loads, and others. The machine learning model (regressor), in turn, based on the obtained parameters, predicts the stresses and the area of the valve clearance in the open state and returns the prediction results to the optimizer. Then the optimizer changes the parameters according to the chosen mathematical algorithm, and the process is repeated up to 2000 times until a better solution is found. If with manual design this process can take several months, then in this case — only a few hours.

At the first stage of the study, scientists generated more than 11 thousand variants of heart valves and performed computer modeling for each of them. This made it possible to obtain a set of data on the tension of the flaps and the area of the valve lumen in the open state for each of the configurations of the prosthesis. Then these data were used for machine learning of more than 340 models, the combination of the best of which allowed to achieve prediction accuracy of 96-98%. Thus, the result of the first stage was a program capable of predicting the performance of prostheses based on the entered characteristics. In the future, it will allow us to abandon classical numerical methods.

At the second stage of the study, the authors tested six optimization programs that differ in the principle of selecting geometric characteristics. For example, some of them created combinations of height, diameter and other valve parameters each time randomly, while others generated each subsequent variant based on the result of previous selections. It turned out that the Tree-structured Parzen Estimator and Nondominated Sorting Genetic Algorithm optimizers worked the fastest and with the least errors. The latter worked according to principles similar to natural selection in biology, and therefore followed the optimal path for finding new variants of geometries.

"The use of new methods in the design of medical devices will help to improve the quality of devices, speed up the development process and reduce the cost of their production. As a result, such an innovation can provide greater access to quality medical care for patients and stimulate innovation in other areas of medical science. In the future, we plan to produce a prototype based on the results of generative design in order to evaluate its hydrodynamic parameters in the real world and compare it with existing models," says Yevgeny Ovcharenko, Ph.D., Head of the Laboratory of New Biomaterials at the CPSSZ Research Institute, the head of the project supported by the RNF grant.

Information provided by the press service of the Russian Science Foundation

Photo source:

The information is taken from the portal "Scientific Russia" ( /)

Certificate of registration of mass media ЭЛ № ФС 77 - 78868 issued by Roskomnadzor on 07.08.2020